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 | */ | |
6fe6b2d6 | 887 | unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4; |
3a7053b3 | 888 | |
ac8e895b MG |
889 | static inline int task_faults_idx(int nid, int priv) |
890 | { | |
891 | return 2 * nid + priv; | |
892 | } | |
893 | ||
894 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
895 | { | |
896 | if (!p->numa_faults) | |
897 | return 0; | |
898 | ||
899 | return p->numa_faults[task_faults_idx(nid, 0)] + | |
900 | p->numa_faults[task_faults_idx(nid, 1)]; | |
901 | } | |
902 | ||
e6628d5b | 903 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
904 | static unsigned long source_load(int cpu, int type); |
905 | static unsigned long target_load(int cpu, int type); | |
906 | static unsigned long power_of(int cpu); | |
907 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
908 | ||
909 | struct numa_stats { | |
910 | unsigned long load; | |
911 | s64 eff_load; | |
912 | unsigned long faults; | |
913 | }; | |
e6628d5b | 914 | |
58d081b5 MG |
915 | struct task_numa_env { |
916 | struct task_struct *p; | |
e6628d5b | 917 | |
58d081b5 MG |
918 | int src_cpu, src_nid; |
919 | int dst_cpu, dst_nid; | |
e6628d5b | 920 | |
58d081b5 | 921 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 922 | |
58d081b5 MG |
923 | unsigned long best_load; |
924 | int best_cpu; | |
925 | }; | |
926 | ||
927 | static int task_numa_migrate(struct task_struct *p) | |
928 | { | |
929 | int node_cpu = cpumask_first(cpumask_of_node(p->numa_preferred_nid)); | |
930 | struct task_numa_env env = { | |
931 | .p = p, | |
932 | .src_cpu = task_cpu(p), | |
933 | .src_nid = cpu_to_node(task_cpu(p)), | |
934 | .dst_cpu = node_cpu, | |
935 | .dst_nid = p->numa_preferred_nid, | |
936 | .best_load = ULONG_MAX, | |
937 | .best_cpu = task_cpu(p), | |
938 | }; | |
939 | struct sched_domain *sd; | |
940 | int cpu; | |
941 | struct task_group *tg = task_group(p); | |
942 | unsigned long weight; | |
943 | bool balanced; | |
944 | int imbalance_pct, idx = -1; | |
e6628d5b | 945 | |
58d081b5 MG |
946 | /* |
947 | * Find the lowest common scheduling domain covering the nodes of both | |
948 | * the CPU the task is currently running on and the target NUMA node. | |
949 | */ | |
950 | rcu_read_lock(); | |
951 | for_each_domain(env.src_cpu, sd) { | |
952 | if (cpumask_test_cpu(node_cpu, sched_domain_span(sd))) { | |
953 | /* | |
954 | * busy_idx is used for the load decision as it is the | |
955 | * same index used by the regular load balancer for an | |
956 | * active cpu. | |
957 | */ | |
958 | idx = sd->busy_idx; | |
959 | imbalance_pct = sd->imbalance_pct; | |
960 | break; | |
e6628d5b MG |
961 | } |
962 | } | |
963 | rcu_read_unlock(); | |
964 | ||
58d081b5 MG |
965 | if (WARN_ON_ONCE(idx == -1)) |
966 | return 0; | |
967 | ||
968 | /* | |
969 | * XXX the below is mostly nicked from wake_affine(); we should | |
970 | * see about sharing a bit if at all possible; also it might want | |
971 | * some per entity weight love. | |
972 | */ | |
973 | weight = p->se.load.weight; | |
974 | env.src_stats.load = source_load(env.src_cpu, idx); | |
975 | env.src_stats.eff_load = 100 + (imbalance_pct - 100) / 2; | |
976 | env.src_stats.eff_load *= power_of(env.src_cpu); | |
977 | env.src_stats.eff_load *= env.src_stats.load + effective_load(tg, env.src_cpu, -weight, -weight); | |
978 | ||
979 | for_each_cpu(cpu, cpumask_of_node(env.dst_nid)) { | |
980 | env.dst_cpu = cpu; | |
981 | env.dst_stats.load = target_load(cpu, idx); | |
982 | ||
983 | /* If the CPU is idle, use it */ | |
984 | if (!env.dst_stats.load) { | |
985 | env.best_cpu = cpu; | |
986 | goto migrate; | |
987 | } | |
988 | ||
989 | /* Otherwise check the target CPU load */ | |
990 | env.dst_stats.eff_load = 100; | |
991 | env.dst_stats.eff_load *= power_of(cpu); | |
992 | env.dst_stats.eff_load *= env.dst_stats.load + effective_load(tg, cpu, weight, weight); | |
993 | ||
994 | /* | |
995 | * Destination is considered balanced if the destination CPU is | |
996 | * less loaded than the source CPU. Unfortunately there is a | |
997 | * risk that a task running on a lightly loaded CPU will not | |
998 | * migrate to its preferred node due to load imbalances. | |
999 | */ | |
1000 | balanced = (env.dst_stats.eff_load <= env.src_stats.eff_load); | |
1001 | if (!balanced) | |
1002 | continue; | |
1003 | ||
1004 | if (env.dst_stats.eff_load < env.best_load) { | |
1005 | env.best_load = env.dst_stats.eff_load; | |
1006 | env.best_cpu = cpu; | |
1007 | } | |
1008 | } | |
1009 | ||
1010 | migrate: | |
1011 | return migrate_task_to(p, env.best_cpu); | |
e6628d5b MG |
1012 | } |
1013 | ||
6b9a7460 MG |
1014 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1015 | static void numa_migrate_preferred(struct task_struct *p) | |
1016 | { | |
1017 | /* Success if task is already running on preferred CPU */ | |
1018 | p->numa_migrate_retry = 0; | |
06ea5e03 RR |
1019 | if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) { |
1020 | /* | |
1021 | * If migration is temporarily disabled due to a task migration | |
1022 | * then re-enable it now as the task is running on its | |
1023 | * preferred node and memory should migrate locally | |
1024 | */ | |
1025 | if (!p->numa_migrate_seq) | |
1026 | p->numa_migrate_seq++; | |
6b9a7460 | 1027 | return; |
06ea5e03 | 1028 | } |
6b9a7460 MG |
1029 | |
1030 | /* This task has no NUMA fault statistics yet */ | |
1031 | if (unlikely(p->numa_preferred_nid == -1)) | |
1032 | return; | |
1033 | ||
1034 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
1035 | if (task_numa_migrate(p) != 0) | |
1036 | p->numa_migrate_retry = jiffies + HZ*5; | |
1037 | } | |
1038 | ||
cbee9f88 PZ |
1039 | static void task_numa_placement(struct task_struct *p) |
1040 | { | |
688b7585 MG |
1041 | int seq, nid, max_nid = -1; |
1042 | unsigned long max_faults = 0; | |
cbee9f88 | 1043 | |
2832bc19 | 1044 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1045 | if (p->numa_scan_seq == seq) |
1046 | return; | |
1047 | p->numa_scan_seq = seq; | |
3a7053b3 | 1048 | p->numa_migrate_seq++; |
598f0ec0 | 1049 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1050 | |
688b7585 MG |
1051 | /* Find the node with the highest number of faults */ |
1052 | for_each_online_node(nid) { | |
745d6147 | 1053 | unsigned long faults; |
ac8e895b | 1054 | int priv, i; |
745d6147 | 1055 | |
ac8e895b MG |
1056 | for (priv = 0; priv < 2; priv++) { |
1057 | i = task_faults_idx(nid, priv); | |
745d6147 | 1058 | |
ac8e895b MG |
1059 | /* Decay existing window, copy faults since last scan */ |
1060 | p->numa_faults[i] >>= 1; | |
1061 | p->numa_faults[i] += p->numa_faults_buffer[i]; | |
1062 | p->numa_faults_buffer[i] = 0; | |
1063 | } | |
1064 | ||
1065 | /* Find maximum private faults */ | |
1066 | faults = p->numa_faults[task_faults_idx(nid, 1)]; | |
688b7585 MG |
1067 | if (faults > max_faults) { |
1068 | max_faults = faults; | |
1069 | max_nid = nid; | |
1070 | } | |
1071 | } | |
1072 | ||
6b9a7460 | 1073 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1074 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1075 | /* Update the preferred nid and migrate task if possible */ |
688b7585 | 1076 | p->numa_preferred_nid = max_nid; |
6fe6b2d6 | 1077 | p->numa_migrate_seq = 1; |
6b9a7460 | 1078 | numa_migrate_preferred(p); |
3a7053b3 | 1079 | } |
cbee9f88 PZ |
1080 | } |
1081 | ||
1082 | /* | |
1083 | * Got a PROT_NONE fault for a page on @node. | |
1084 | */ | |
b795854b | 1085 | void task_numa_fault(int last_nidpid, int node, int pages, bool migrated) |
cbee9f88 PZ |
1086 | { |
1087 | struct task_struct *p = current; | |
ac8e895b | 1088 | int priv; |
cbee9f88 | 1089 | |
10e84b97 | 1090 | if (!numabalancing_enabled) |
1a687c2e MG |
1091 | return; |
1092 | ||
9ff1d9ff MG |
1093 | /* for example, ksmd faulting in a user's mm */ |
1094 | if (!p->mm) | |
1095 | return; | |
1096 | ||
b795854b MG |
1097 | /* |
1098 | * First accesses are treated as private, otherwise consider accesses | |
1099 | * to be private if the accessing pid has not changed | |
1100 | */ | |
1101 | if (!nidpid_pid_unset(last_nidpid)) | |
1102 | priv = ((p->pid & LAST__PID_MASK) == nidpid_to_pid(last_nidpid)); | |
1103 | else | |
1104 | priv = 1; | |
ac8e895b | 1105 | |
f809ca9a MG |
1106 | /* Allocate buffer to track faults on a per-node basis */ |
1107 | if (unlikely(!p->numa_faults)) { | |
ac8e895b | 1108 | int size = sizeof(*p->numa_faults) * 2 * nr_node_ids; |
f809ca9a | 1109 | |
745d6147 MG |
1110 | /* numa_faults and numa_faults_buffer share the allocation */ |
1111 | p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN); | |
f809ca9a MG |
1112 | if (!p->numa_faults) |
1113 | return; | |
745d6147 MG |
1114 | |
1115 | BUG_ON(p->numa_faults_buffer); | |
ac8e895b | 1116 | p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids); |
f809ca9a | 1117 | } |
cbee9f88 | 1118 | |
fb003b80 | 1119 | /* |
b8593bfd MG |
1120 | * If pages are properly placed (did not migrate) then scan slower. |
1121 | * This is reset periodically in case of phase changes | |
fb003b80 | 1122 | */ |
598f0ec0 MG |
1123 | if (!migrated) { |
1124 | /* Initialise if necessary */ | |
1125 | if (!p->numa_scan_period_max) | |
1126 | p->numa_scan_period_max = task_scan_max(p); | |
1127 | ||
1128 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1129 | p->numa_scan_period + 10); | |
1130 | } | |
fb003b80 | 1131 | |
cbee9f88 | 1132 | task_numa_placement(p); |
f809ca9a | 1133 | |
6b9a7460 MG |
1134 | /* Retry task to preferred node migration if it previously failed */ |
1135 | if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry)) | |
1136 | numa_migrate_preferred(p); | |
1137 | ||
ac8e895b | 1138 | p->numa_faults_buffer[task_faults_idx(node, priv)] += pages; |
cbee9f88 PZ |
1139 | } |
1140 | ||
6e5fb223 PZ |
1141 | static void reset_ptenuma_scan(struct task_struct *p) |
1142 | { | |
1143 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1144 | p->mm->numa_scan_offset = 0; | |
1145 | } | |
1146 | ||
cbee9f88 PZ |
1147 | /* |
1148 | * The expensive part of numa migration is done from task_work context. | |
1149 | * Triggered from task_tick_numa(). | |
1150 | */ | |
1151 | void task_numa_work(struct callback_head *work) | |
1152 | { | |
1153 | unsigned long migrate, next_scan, now = jiffies; | |
1154 | struct task_struct *p = current; | |
1155 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1156 | struct vm_area_struct *vma; |
9f40604c | 1157 | unsigned long start, end; |
598f0ec0 | 1158 | unsigned long nr_pte_updates = 0; |
9f40604c | 1159 | long pages; |
cbee9f88 PZ |
1160 | |
1161 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1162 | ||
1163 | work->next = work; /* protect against double add */ | |
1164 | /* | |
1165 | * Who cares about NUMA placement when they're dying. | |
1166 | * | |
1167 | * NOTE: make sure not to dereference p->mm before this check, | |
1168 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1169 | * without p->mm even though we still had it when we enqueued this | |
1170 | * work. | |
1171 | */ | |
1172 | if (p->flags & PF_EXITING) | |
1173 | return; | |
1174 | ||
7e8d16b6 MG |
1175 | if (!mm->numa_next_reset || !mm->numa_next_scan) { |
1176 | mm->numa_next_scan = now + | |
1177 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
1178 | mm->numa_next_reset = now + | |
1179 | msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
1180 | } | |
1181 | ||
b8593bfd MG |
1182 | /* |
1183 | * Reset the scan period if enough time has gone by. Objective is that | |
1184 | * scanning will be reduced if pages are properly placed. As tasks | |
1185 | * can enter different phases this needs to be re-examined. Lacking | |
1186 | * proper tracking of reference behaviour, this blunt hammer is used. | |
1187 | */ | |
1188 | migrate = mm->numa_next_reset; | |
1189 | if (time_after(now, migrate)) { | |
598f0ec0 | 1190 | p->numa_scan_period = task_scan_min(p); |
b8593bfd MG |
1191 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); |
1192 | xchg(&mm->numa_next_reset, next_scan); | |
1193 | } | |
1194 | ||
cbee9f88 PZ |
1195 | /* |
1196 | * Enforce maximal scan/migration frequency.. | |
1197 | */ | |
1198 | migrate = mm->numa_next_scan; | |
1199 | if (time_before(now, migrate)) | |
1200 | return; | |
1201 | ||
598f0ec0 MG |
1202 | if (p->numa_scan_period == 0) { |
1203 | p->numa_scan_period_max = task_scan_max(p); | |
1204 | p->numa_scan_period = task_scan_min(p); | |
1205 | } | |
cbee9f88 | 1206 | |
fb003b80 | 1207 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1208 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1209 | return; | |
1210 | ||
19a78d11 PZ |
1211 | /* |
1212 | * Delay this task enough that another task of this mm will likely win | |
1213 | * the next time around. | |
1214 | */ | |
1215 | p->node_stamp += 2 * TICK_NSEC; | |
1216 | ||
9f40604c MG |
1217 | start = mm->numa_scan_offset; |
1218 | pages = sysctl_numa_balancing_scan_size; | |
1219 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1220 | if (!pages) | |
1221 | return; | |
cbee9f88 | 1222 | |
6e5fb223 | 1223 | down_read(&mm->mmap_sem); |
9f40604c | 1224 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1225 | if (!vma) { |
1226 | reset_ptenuma_scan(p); | |
9f40604c | 1227 | start = 0; |
6e5fb223 PZ |
1228 | vma = mm->mmap; |
1229 | } | |
9f40604c | 1230 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1231 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1232 | continue; |
1233 | ||
4591ce4f MG |
1234 | /* |
1235 | * Shared library pages mapped by multiple processes are not | |
1236 | * migrated as it is expected they are cache replicated. Avoid | |
1237 | * hinting faults in read-only file-backed mappings or the vdso | |
1238 | * as migrating the pages will be of marginal benefit. | |
1239 | */ | |
1240 | if (!vma->vm_mm || | |
1241 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1242 | continue; | |
1243 | ||
9f40604c MG |
1244 | do { |
1245 | start = max(start, vma->vm_start); | |
1246 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1247 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1248 | nr_pte_updates += change_prot_numa(vma, start, end); |
1249 | ||
1250 | /* | |
1251 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1252 | * at least one PTE is updated so that unused virtual | |
1253 | * address space is quickly skipped. | |
1254 | */ | |
1255 | if (nr_pte_updates) | |
1256 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1257 | |
9f40604c MG |
1258 | start = end; |
1259 | if (pages <= 0) | |
1260 | goto out; | |
1261 | } while (end != vma->vm_end); | |
cbee9f88 | 1262 | } |
6e5fb223 | 1263 | |
9f40604c | 1264 | out: |
f307cd1a MG |
1265 | /* |
1266 | * If the whole process was scanned without updates then no NUMA | |
1267 | * hinting faults are being recorded and scan rate should be lower. | |
1268 | */ | |
1269 | if (mm->numa_scan_offset == 0 && !nr_pte_updates) { | |
1270 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1271 | p->numa_scan_period << 1); | |
1272 | ||
1273 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); | |
1274 | mm->numa_next_scan = next_scan; | |
1275 | } | |
1276 | ||
6e5fb223 | 1277 | /* |
c69307d5 PZ |
1278 | * It is possible to reach the end of the VMA list but the last few |
1279 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1280 | * would find the !migratable VMA on the next scan but not reset the | |
1281 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1282 | */ |
1283 | if (vma) | |
9f40604c | 1284 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1285 | else |
1286 | reset_ptenuma_scan(p); | |
1287 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1288 | } |
1289 | ||
1290 | /* | |
1291 | * Drive the periodic memory faults.. | |
1292 | */ | |
1293 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1294 | { | |
1295 | struct callback_head *work = &curr->numa_work; | |
1296 | u64 period, now; | |
1297 | ||
1298 | /* | |
1299 | * We don't care about NUMA placement if we don't have memory. | |
1300 | */ | |
1301 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1302 | return; | |
1303 | ||
1304 | /* | |
1305 | * Using runtime rather than walltime has the dual advantage that | |
1306 | * we (mostly) drive the selection from busy threads and that the | |
1307 | * task needs to have done some actual work before we bother with | |
1308 | * NUMA placement. | |
1309 | */ | |
1310 | now = curr->se.sum_exec_runtime; | |
1311 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1312 | ||
1313 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1314 | if (!curr->node_stamp) |
598f0ec0 | 1315 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1316 | curr->node_stamp += period; |
cbee9f88 PZ |
1317 | |
1318 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1319 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1320 | task_work_add(curr, work, true); | |
1321 | } | |
1322 | } | |
1323 | } | |
1324 | #else | |
1325 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1326 | { | |
1327 | } | |
1328 | #endif /* CONFIG_NUMA_BALANCING */ | |
1329 | ||
30cfdcfc DA |
1330 | static void |
1331 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1332 | { | |
1333 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1334 | if (!parent_entity(se)) |
029632fb | 1335 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
1336 | #ifdef CONFIG_SMP |
1337 | if (entity_is_task(se)) | |
eb95308e | 1338 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 1339 | #endif |
30cfdcfc | 1340 | cfs_rq->nr_running++; |
30cfdcfc DA |
1341 | } |
1342 | ||
1343 | static void | |
1344 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1345 | { | |
1346 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1347 | if (!parent_entity(se)) |
029632fb | 1348 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1349 | if (entity_is_task(se)) |
b87f1724 | 1350 | list_del_init(&se->group_node); |
30cfdcfc | 1351 | cfs_rq->nr_running--; |
30cfdcfc DA |
1352 | } |
1353 | ||
3ff6dcac YZ |
1354 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1355 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1356 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1357 | { | |
1358 | long tg_weight; | |
1359 | ||
1360 | /* | |
1361 | * Use this CPU's actual weight instead of the last load_contribution | |
1362 | * to gain a more accurate current total weight. See | |
1363 | * update_cfs_rq_load_contribution(). | |
1364 | */ | |
bf5b986e | 1365 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 1366 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
1367 | tg_weight += cfs_rq->load.weight; |
1368 | ||
1369 | return tg_weight; | |
1370 | } | |
1371 | ||
6d5ab293 | 1372 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1373 | { |
cf5f0acf | 1374 | long tg_weight, load, shares; |
3ff6dcac | 1375 | |
cf5f0acf | 1376 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1377 | load = cfs_rq->load.weight; |
3ff6dcac | 1378 | |
3ff6dcac | 1379 | shares = (tg->shares * load); |
cf5f0acf PZ |
1380 | if (tg_weight) |
1381 | shares /= tg_weight; | |
3ff6dcac YZ |
1382 | |
1383 | if (shares < MIN_SHARES) | |
1384 | shares = MIN_SHARES; | |
1385 | if (shares > tg->shares) | |
1386 | shares = tg->shares; | |
1387 | ||
1388 | return shares; | |
1389 | } | |
3ff6dcac | 1390 | # else /* CONFIG_SMP */ |
6d5ab293 | 1391 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1392 | { |
1393 | return tg->shares; | |
1394 | } | |
3ff6dcac | 1395 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1396 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1397 | unsigned long weight) | |
1398 | { | |
19e5eebb PT |
1399 | if (se->on_rq) { |
1400 | /* commit outstanding execution time */ | |
1401 | if (cfs_rq->curr == se) | |
1402 | update_curr(cfs_rq); | |
2069dd75 | 1403 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1404 | } |
2069dd75 PZ |
1405 | |
1406 | update_load_set(&se->load, weight); | |
1407 | ||
1408 | if (se->on_rq) | |
1409 | account_entity_enqueue(cfs_rq, se); | |
1410 | } | |
1411 | ||
82958366 PT |
1412 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1413 | ||
6d5ab293 | 1414 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1415 | { |
1416 | struct task_group *tg; | |
1417 | struct sched_entity *se; | |
3ff6dcac | 1418 | long shares; |
2069dd75 | 1419 | |
2069dd75 PZ |
1420 | tg = cfs_rq->tg; |
1421 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1422 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1423 | return; |
3ff6dcac YZ |
1424 | #ifndef CONFIG_SMP |
1425 | if (likely(se->load.weight == tg->shares)) | |
1426 | return; | |
1427 | #endif | |
6d5ab293 | 1428 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1429 | |
1430 | reweight_entity(cfs_rq_of(se), se, shares); | |
1431 | } | |
1432 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1433 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1434 | { |
1435 | } | |
1436 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1437 | ||
141965c7 | 1438 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
1439 | /* |
1440 | * We choose a half-life close to 1 scheduling period. | |
1441 | * Note: The tables below are dependent on this value. | |
1442 | */ | |
1443 | #define LOAD_AVG_PERIOD 32 | |
1444 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1445 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1446 | ||
1447 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1448 | static const u32 runnable_avg_yN_inv[] = { | |
1449 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1450 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1451 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1452 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1453 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1454 | 0x85aac367, 0x82cd8698, | |
1455 | }; | |
1456 | ||
1457 | /* | |
1458 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1459 | * over-estimates when re-combining. | |
1460 | */ | |
1461 | static const u32 runnable_avg_yN_sum[] = { | |
1462 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1463 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1464 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1465 | }; | |
1466 | ||
9d85f21c PT |
1467 | /* |
1468 | * Approximate: | |
1469 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1470 | */ | |
1471 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1472 | { | |
5b51f2f8 PT |
1473 | unsigned int local_n; |
1474 | ||
1475 | if (!n) | |
1476 | return val; | |
1477 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1478 | return 0; | |
1479 | ||
1480 | /* after bounds checking we can collapse to 32-bit */ | |
1481 | local_n = n; | |
1482 | ||
1483 | /* | |
1484 | * As y^PERIOD = 1/2, we can combine | |
1485 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
1486 | * With a look-up table which covers k^n (n<PERIOD) | |
1487 | * | |
1488 | * To achieve constant time decay_load. | |
1489 | */ | |
1490 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
1491 | val >>= local_n / LOAD_AVG_PERIOD; | |
1492 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
1493 | } |
1494 | ||
5b51f2f8 PT |
1495 | val *= runnable_avg_yN_inv[local_n]; |
1496 | /* We don't use SRR here since we always want to round down. */ | |
1497 | return val >> 32; | |
1498 | } | |
1499 | ||
1500 | /* | |
1501 | * For updates fully spanning n periods, the contribution to runnable | |
1502 | * average will be: \Sum 1024*y^n | |
1503 | * | |
1504 | * We can compute this reasonably efficiently by combining: | |
1505 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
1506 | */ | |
1507 | static u32 __compute_runnable_contrib(u64 n) | |
1508 | { | |
1509 | u32 contrib = 0; | |
1510 | ||
1511 | if (likely(n <= LOAD_AVG_PERIOD)) | |
1512 | return runnable_avg_yN_sum[n]; | |
1513 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
1514 | return LOAD_AVG_MAX; | |
1515 | ||
1516 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
1517 | do { | |
1518 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
1519 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
1520 | ||
1521 | n -= LOAD_AVG_PERIOD; | |
1522 | } while (n > LOAD_AVG_PERIOD); | |
1523 | ||
1524 | contrib = decay_load(contrib, n); | |
1525 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
1526 | } |
1527 | ||
1528 | /* | |
1529 | * We can represent the historical contribution to runnable average as the | |
1530 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
1531 | * history into segments of approximately 1ms (1024us); label the segment that | |
1532 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
1533 | * | |
1534 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
1535 | * p0 p1 p2 | |
1536 | * (now) (~1ms ago) (~2ms ago) | |
1537 | * | |
1538 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
1539 | * | |
1540 | * We then designate the fractions u_i as our co-efficients, yielding the | |
1541 | * following representation of historical load: | |
1542 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
1543 | * | |
1544 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
1545 | * y^32 = 0.5 | |
1546 | * | |
1547 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
1548 | * approximately half as much as the contribution to load within the last ms | |
1549 | * (u_0). | |
1550 | * | |
1551 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
1552 | * sum again by y is sufficient to update: | |
1553 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1554 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1555 | */ | |
1556 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1557 | struct sched_avg *sa, | |
1558 | int runnable) | |
1559 | { | |
5b51f2f8 PT |
1560 | u64 delta, periods; |
1561 | u32 runnable_contrib; | |
9d85f21c PT |
1562 | int delta_w, decayed = 0; |
1563 | ||
1564 | delta = now - sa->last_runnable_update; | |
1565 | /* | |
1566 | * This should only happen when time goes backwards, which it | |
1567 | * unfortunately does during sched clock init when we swap over to TSC. | |
1568 | */ | |
1569 | if ((s64)delta < 0) { | |
1570 | sa->last_runnable_update = now; | |
1571 | return 0; | |
1572 | } | |
1573 | ||
1574 | /* | |
1575 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1576 | * approximation of 1us and fast to compute. | |
1577 | */ | |
1578 | delta >>= 10; | |
1579 | if (!delta) | |
1580 | return 0; | |
1581 | sa->last_runnable_update = now; | |
1582 | ||
1583 | /* delta_w is the amount already accumulated against our next period */ | |
1584 | delta_w = sa->runnable_avg_period % 1024; | |
1585 | if (delta + delta_w >= 1024) { | |
1586 | /* period roll-over */ | |
1587 | decayed = 1; | |
1588 | ||
1589 | /* | |
1590 | * Now that we know we're crossing a period boundary, figure | |
1591 | * out how much from delta we need to complete the current | |
1592 | * period and accrue it. | |
1593 | */ | |
1594 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
1595 | if (runnable) |
1596 | sa->runnable_avg_sum += delta_w; | |
1597 | sa->runnable_avg_period += delta_w; | |
1598 | ||
1599 | delta -= delta_w; | |
1600 | ||
1601 | /* Figure out how many additional periods this update spans */ | |
1602 | periods = delta / 1024; | |
1603 | delta %= 1024; | |
1604 | ||
1605 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
1606 | periods + 1); | |
1607 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
1608 | periods + 1); | |
1609 | ||
1610 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
1611 | runnable_contrib = __compute_runnable_contrib(periods); | |
1612 | if (runnable) | |
1613 | sa->runnable_avg_sum += runnable_contrib; | |
1614 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
1615 | } |
1616 | ||
1617 | /* Remainder of delta accrued against u_0` */ | |
1618 | if (runnable) | |
1619 | sa->runnable_avg_sum += delta; | |
1620 | sa->runnable_avg_period += delta; | |
1621 | ||
1622 | return decayed; | |
1623 | } | |
1624 | ||
9ee474f5 | 1625 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 1626 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
1627 | { |
1628 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1629 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1630 | ||
1631 | decays -= se->avg.decay_count; | |
1632 | if (!decays) | |
aff3e498 | 1633 | return 0; |
9ee474f5 PT |
1634 | |
1635 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1636 | se->avg.decay_count = 0; | |
aff3e498 PT |
1637 | |
1638 | return decays; | |
9ee474f5 PT |
1639 | } |
1640 | ||
c566e8e9 PT |
1641 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1642 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1643 | int force_update) | |
1644 | { | |
1645 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 1646 | long tg_contrib; |
c566e8e9 PT |
1647 | |
1648 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
1649 | tg_contrib -= cfs_rq->tg_load_contrib; | |
1650 | ||
bf5b986e AS |
1651 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
1652 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
1653 | cfs_rq->tg_load_contrib += tg_contrib; |
1654 | } | |
1655 | } | |
8165e145 | 1656 | |
bb17f655 PT |
1657 | /* |
1658 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
1659 | * representation for computing load contributions. | |
1660 | */ | |
1661 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
1662 | struct cfs_rq *cfs_rq) | |
1663 | { | |
1664 | struct task_group *tg = cfs_rq->tg; | |
1665 | long contrib; | |
1666 | ||
1667 | /* The fraction of a cpu used by this cfs_rq */ | |
1668 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
1669 | sa->runnable_avg_period + 1); | |
1670 | contrib -= cfs_rq->tg_runnable_contrib; | |
1671 | ||
1672 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
1673 | atomic_add(contrib, &tg->runnable_avg); | |
1674 | cfs_rq->tg_runnable_contrib += contrib; | |
1675 | } | |
1676 | } | |
1677 | ||
8165e145 PT |
1678 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
1679 | { | |
1680 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
1681 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
1682 | int runnable_avg; |
1683 | ||
8165e145 PT |
1684 | u64 contrib; |
1685 | ||
1686 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
1687 | se->avg.load_avg_contrib = div_u64(contrib, |
1688 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
1689 | |
1690 | /* | |
1691 | * For group entities we need to compute a correction term in the case | |
1692 | * that they are consuming <1 cpu so that we would contribute the same | |
1693 | * load as a task of equal weight. | |
1694 | * | |
1695 | * Explicitly co-ordinating this measurement would be expensive, but | |
1696 | * fortunately the sum of each cpus contribution forms a usable | |
1697 | * lower-bound on the true value. | |
1698 | * | |
1699 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
1700 | * (and the sum represents true value) or they are disjoint and we are | |
1701 | * understating by the aggregate of their overlap. | |
1702 | * | |
1703 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
1704 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
1705 | * cpus that overlap for this interval and w_i is the interval width. | |
1706 | * | |
1707 | * On a small machine; the first term is well-bounded which bounds the | |
1708 | * total error since w_i is a subset of the period. Whereas on a | |
1709 | * larger machine, while this first term can be larger, if w_i is the | |
1710 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
1711 | * our upper bound of 1-cpu. | |
1712 | */ | |
1713 | runnable_avg = atomic_read(&tg->runnable_avg); | |
1714 | if (runnable_avg < NICE_0_LOAD) { | |
1715 | se->avg.load_avg_contrib *= runnable_avg; | |
1716 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
1717 | } | |
8165e145 | 1718 | } |
c566e8e9 PT |
1719 | #else |
1720 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1721 | int force_update) {} | |
bb17f655 PT |
1722 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
1723 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 1724 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
1725 | #endif |
1726 | ||
8165e145 PT |
1727 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
1728 | { | |
1729 | u32 contrib; | |
1730 | ||
1731 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
1732 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
1733 | contrib /= (se->avg.runnable_avg_period + 1); | |
1734 | se->avg.load_avg_contrib = scale_load(contrib); | |
1735 | } | |
1736 | ||
2dac754e PT |
1737 | /* Compute the current contribution to load_avg by se, return any delta */ |
1738 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
1739 | { | |
1740 | long old_contrib = se->avg.load_avg_contrib; | |
1741 | ||
8165e145 PT |
1742 | if (entity_is_task(se)) { |
1743 | __update_task_entity_contrib(se); | |
1744 | } else { | |
bb17f655 | 1745 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
1746 | __update_group_entity_contrib(se); |
1747 | } | |
2dac754e PT |
1748 | |
1749 | return se->avg.load_avg_contrib - old_contrib; | |
1750 | } | |
1751 | ||
9ee474f5 PT |
1752 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
1753 | long load_contrib) | |
1754 | { | |
1755 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
1756 | cfs_rq->blocked_load_avg -= load_contrib; | |
1757 | else | |
1758 | cfs_rq->blocked_load_avg = 0; | |
1759 | } | |
1760 | ||
f1b17280 PT |
1761 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
1762 | ||
9d85f21c | 1763 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
1764 | static inline void update_entity_load_avg(struct sched_entity *se, |
1765 | int update_cfs_rq) | |
9d85f21c | 1766 | { |
2dac754e PT |
1767 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
1768 | long contrib_delta; | |
f1b17280 | 1769 | u64 now; |
2dac754e | 1770 | |
f1b17280 PT |
1771 | /* |
1772 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
1773 | * case they are the parent of a throttled hierarchy. | |
1774 | */ | |
1775 | if (entity_is_task(se)) | |
1776 | now = cfs_rq_clock_task(cfs_rq); | |
1777 | else | |
1778 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
1779 | ||
1780 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
1781 | return; |
1782 | ||
1783 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
1784 | |
1785 | if (!update_cfs_rq) | |
1786 | return; | |
1787 | ||
2dac754e PT |
1788 | if (se->on_rq) |
1789 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
1790 | else |
1791 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
1792 | } | |
1793 | ||
1794 | /* | |
1795 | * Decay the load contributed by all blocked children and account this so that | |
1796 | * their contribution may appropriately discounted when they wake up. | |
1797 | */ | |
aff3e498 | 1798 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 1799 | { |
f1b17280 | 1800 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
1801 | u64 decays; |
1802 | ||
1803 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 1804 | if (!decays && !force_update) |
9ee474f5 PT |
1805 | return; |
1806 | ||
2509940f AS |
1807 | if (atomic_long_read(&cfs_rq->removed_load)) { |
1808 | unsigned long removed_load; | |
1809 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
1810 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
1811 | } | |
9ee474f5 | 1812 | |
aff3e498 PT |
1813 | if (decays) { |
1814 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
1815 | decays); | |
1816 | atomic64_add(decays, &cfs_rq->decay_counter); | |
1817 | cfs_rq->last_decay = now; | |
1818 | } | |
c566e8e9 PT |
1819 | |
1820 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 1821 | } |
18bf2805 BS |
1822 | |
1823 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
1824 | { | |
78becc27 | 1825 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 1826 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 1827 | } |
2dac754e PT |
1828 | |
1829 | /* Add the load generated by se into cfs_rq's child load-average */ | |
1830 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
1831 | struct sched_entity *se, |
1832 | int wakeup) | |
2dac754e | 1833 | { |
aff3e498 PT |
1834 | /* |
1835 | * We track migrations using entity decay_count <= 0, on a wake-up | |
1836 | * migration we use a negative decay count to track the remote decays | |
1837 | * accumulated while sleeping. | |
a75cdaa9 AS |
1838 | * |
1839 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
1840 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
1841 | * constructed load_avg_contrib. | |
aff3e498 PT |
1842 | */ |
1843 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 1844 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
1845 | if (se->avg.decay_count) { |
1846 | /* | |
1847 | * In a wake-up migration we have to approximate the | |
1848 | * time sleeping. This is because we can't synchronize | |
1849 | * clock_task between the two cpus, and it is not | |
1850 | * guaranteed to be read-safe. Instead, we can | |
1851 | * approximate this using our carried decays, which are | |
1852 | * explicitly atomically readable. | |
1853 | */ | |
1854 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
1855 | << 20; | |
1856 | update_entity_load_avg(se, 0); | |
1857 | /* Indicate that we're now synchronized and on-rq */ | |
1858 | se->avg.decay_count = 0; | |
1859 | } | |
9ee474f5 PT |
1860 | wakeup = 0; |
1861 | } else { | |
282cf499 AS |
1862 | /* |
1863 | * Task re-woke on same cpu (or else migrate_task_rq_fair() | |
1864 | * would have made count negative); we must be careful to avoid | |
1865 | * double-accounting blocked time after synchronizing decays. | |
1866 | */ | |
1867 | se->avg.last_runnable_update += __synchronize_entity_decay(se) | |
1868 | << 20; | |
9ee474f5 PT |
1869 | } |
1870 | ||
aff3e498 PT |
1871 | /* migrated tasks did not contribute to our blocked load */ |
1872 | if (wakeup) { | |
9ee474f5 | 1873 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
1874 | update_entity_load_avg(se, 0); |
1875 | } | |
9ee474f5 | 1876 | |
2dac754e | 1877 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
1878 | /* we force update consideration on load-balancer moves */ |
1879 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
1880 | } |
1881 | ||
9ee474f5 PT |
1882 | /* |
1883 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
1884 | * transitioning to a blocked state we track its projected decay using | |
1885 | * blocked_load_avg. | |
1886 | */ | |
2dac754e | 1887 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1888 | struct sched_entity *se, |
1889 | int sleep) | |
2dac754e | 1890 | { |
9ee474f5 | 1891 | update_entity_load_avg(se, 1); |
aff3e498 PT |
1892 | /* we force update consideration on load-balancer moves */ |
1893 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 1894 | |
2dac754e | 1895 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
1896 | if (sleep) { |
1897 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
1898 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
1899 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 1900 | } |
642dbc39 VG |
1901 | |
1902 | /* | |
1903 | * Update the rq's load with the elapsed running time before entering | |
1904 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
1905 | * be the only way to update the runnable statistic. | |
1906 | */ | |
1907 | void idle_enter_fair(struct rq *this_rq) | |
1908 | { | |
1909 | update_rq_runnable_avg(this_rq, 1); | |
1910 | } | |
1911 | ||
1912 | /* | |
1913 | * Update the rq's load with the elapsed idle time before a task is | |
1914 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
1915 | * be the only way to update the runnable statistic. | |
1916 | */ | |
1917 | void idle_exit_fair(struct rq *this_rq) | |
1918 | { | |
1919 | update_rq_runnable_avg(this_rq, 0); | |
1920 | } | |
1921 | ||
9d85f21c | 1922 | #else |
9ee474f5 PT |
1923 | static inline void update_entity_load_avg(struct sched_entity *se, |
1924 | int update_cfs_rq) {} | |
18bf2805 | 1925 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 1926 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1927 | struct sched_entity *se, |
1928 | int wakeup) {} | |
2dac754e | 1929 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
1930 | struct sched_entity *se, |
1931 | int sleep) {} | |
aff3e498 PT |
1932 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
1933 | int force_update) {} | |
9d85f21c PT |
1934 | #endif |
1935 | ||
2396af69 | 1936 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1937 | { |
bf0f6f24 | 1938 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
1939 | struct task_struct *tsk = NULL; |
1940 | ||
1941 | if (entity_is_task(se)) | |
1942 | tsk = task_of(se); | |
1943 | ||
41acab88 | 1944 | if (se->statistics.sleep_start) { |
78becc27 | 1945 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
1946 | |
1947 | if ((s64)delta < 0) | |
1948 | delta = 0; | |
1949 | ||
41acab88 LDM |
1950 | if (unlikely(delta > se->statistics.sleep_max)) |
1951 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 1952 | |
8c79a045 | 1953 | se->statistics.sleep_start = 0; |
41acab88 | 1954 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 1955 | |
768d0c27 | 1956 | if (tsk) { |
e414314c | 1957 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
1958 | trace_sched_stat_sleep(tsk, delta); |
1959 | } | |
bf0f6f24 | 1960 | } |
41acab88 | 1961 | if (se->statistics.block_start) { |
78becc27 | 1962 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
1963 | |
1964 | if ((s64)delta < 0) | |
1965 | delta = 0; | |
1966 | ||
41acab88 LDM |
1967 | if (unlikely(delta > se->statistics.block_max)) |
1968 | se->statistics.block_max = delta; | |
bf0f6f24 | 1969 | |
8c79a045 | 1970 | se->statistics.block_start = 0; |
41acab88 | 1971 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 1972 | |
e414314c | 1973 | if (tsk) { |
8f0dfc34 | 1974 | if (tsk->in_iowait) { |
41acab88 LDM |
1975 | se->statistics.iowait_sum += delta; |
1976 | se->statistics.iowait_count++; | |
768d0c27 | 1977 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
1978 | } |
1979 | ||
b781a602 AV |
1980 | trace_sched_stat_blocked(tsk, delta); |
1981 | ||
e414314c PZ |
1982 | /* |
1983 | * Blocking time is in units of nanosecs, so shift by | |
1984 | * 20 to get a milliseconds-range estimation of the | |
1985 | * amount of time that the task spent sleeping: | |
1986 | */ | |
1987 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
1988 | profile_hits(SLEEP_PROFILING, | |
1989 | (void *)get_wchan(tsk), | |
1990 | delta >> 20); | |
1991 | } | |
1992 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 1993 | } |
bf0f6f24 IM |
1994 | } |
1995 | #endif | |
1996 | } | |
1997 | ||
ddc97297 PZ |
1998 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1999 | { | |
2000 | #ifdef CONFIG_SCHED_DEBUG | |
2001 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2002 | ||
2003 | if (d < 0) | |
2004 | d = -d; | |
2005 | ||
2006 | if (d > 3*sysctl_sched_latency) | |
2007 | schedstat_inc(cfs_rq, nr_spread_over); | |
2008 | #endif | |
2009 | } | |
2010 | ||
aeb73b04 PZ |
2011 | static void |
2012 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2013 | { | |
1af5f730 | 2014 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2015 | |
2cb8600e PZ |
2016 | /* |
2017 | * The 'current' period is already promised to the current tasks, | |
2018 | * however the extra weight of the new task will slow them down a | |
2019 | * little, place the new task so that it fits in the slot that | |
2020 | * stays open at the end. | |
2021 | */ | |
94dfb5e7 | 2022 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2023 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2024 | |
a2e7a7eb | 2025 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2026 | if (!initial) { |
a2e7a7eb | 2027 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2028 | |
a2e7a7eb MG |
2029 | /* |
2030 | * Halve their sleep time's effect, to allow | |
2031 | * for a gentler effect of sleepers: | |
2032 | */ | |
2033 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2034 | thresh >>= 1; | |
51e0304c | 2035 | |
a2e7a7eb | 2036 | vruntime -= thresh; |
aeb73b04 PZ |
2037 | } |
2038 | ||
b5d9d734 | 2039 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2040 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2041 | } |
2042 | ||
d3d9dc33 PT |
2043 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2044 | ||
bf0f6f24 | 2045 | static void |
88ec22d3 | 2046 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2047 | { |
88ec22d3 PZ |
2048 | /* |
2049 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2050 | * through calling update_curr(). |
88ec22d3 | 2051 | */ |
371fd7e7 | 2052 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2053 | se->vruntime += cfs_rq->min_vruntime; |
2054 | ||
bf0f6f24 | 2055 | /* |
a2a2d680 | 2056 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2057 | */ |
b7cc0896 | 2058 | update_curr(cfs_rq); |
f269ae04 | 2059 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2060 | account_entity_enqueue(cfs_rq, se); |
2061 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2062 | |
88ec22d3 | 2063 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2064 | place_entity(cfs_rq, se, 0); |
2396af69 | 2065 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2066 | } |
bf0f6f24 | 2067 | |
d2417e5a | 2068 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2069 | check_spread(cfs_rq, se); |
83b699ed SV |
2070 | if (se != cfs_rq->curr) |
2071 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2072 | se->on_rq = 1; |
3d4b47b4 | 2073 | |
d3d9dc33 | 2074 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2075 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2076 | check_enqueue_throttle(cfs_rq); |
2077 | } | |
bf0f6f24 IM |
2078 | } |
2079 | ||
2c13c919 | 2080 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2081 | { |
2c13c919 RR |
2082 | for_each_sched_entity(se) { |
2083 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2084 | if (cfs_rq->last == se) | |
2085 | cfs_rq->last = NULL; | |
2086 | else | |
2087 | break; | |
2088 | } | |
2089 | } | |
2002c695 | 2090 | |
2c13c919 RR |
2091 | static void __clear_buddies_next(struct sched_entity *se) |
2092 | { | |
2093 | for_each_sched_entity(se) { | |
2094 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2095 | if (cfs_rq->next == se) | |
2096 | cfs_rq->next = NULL; | |
2097 | else | |
2098 | break; | |
2099 | } | |
2002c695 PZ |
2100 | } |
2101 | ||
ac53db59 RR |
2102 | static void __clear_buddies_skip(struct sched_entity *se) |
2103 | { | |
2104 | for_each_sched_entity(se) { | |
2105 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2106 | if (cfs_rq->skip == se) | |
2107 | cfs_rq->skip = NULL; | |
2108 | else | |
2109 | break; | |
2110 | } | |
2111 | } | |
2112 | ||
a571bbea PZ |
2113 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2114 | { | |
2c13c919 RR |
2115 | if (cfs_rq->last == se) |
2116 | __clear_buddies_last(se); | |
2117 | ||
2118 | if (cfs_rq->next == se) | |
2119 | __clear_buddies_next(se); | |
ac53db59 RR |
2120 | |
2121 | if (cfs_rq->skip == se) | |
2122 | __clear_buddies_skip(se); | |
a571bbea PZ |
2123 | } |
2124 | ||
6c16a6dc | 2125 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2126 | |
bf0f6f24 | 2127 | static void |
371fd7e7 | 2128 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2129 | { |
a2a2d680 DA |
2130 | /* |
2131 | * Update run-time statistics of the 'current'. | |
2132 | */ | |
2133 | update_curr(cfs_rq); | |
17bc14b7 | 2134 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2135 | |
19b6a2e3 | 2136 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2137 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2138 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2139 | if (entity_is_task(se)) { |
2140 | struct task_struct *tsk = task_of(se); | |
2141 | ||
2142 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2143 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2144 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2145 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2146 | } |
db36cc7d | 2147 | #endif |
67e9fb2a PZ |
2148 | } |
2149 | ||
2002c695 | 2150 | clear_buddies(cfs_rq, se); |
4793241b | 2151 | |
83b699ed | 2152 | if (se != cfs_rq->curr) |
30cfdcfc | 2153 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2154 | se->on_rq = 0; |
30cfdcfc | 2155 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2156 | |
2157 | /* | |
2158 | * Normalize the entity after updating the min_vruntime because the | |
2159 | * update can refer to the ->curr item and we need to reflect this | |
2160 | * movement in our normalized position. | |
2161 | */ | |
371fd7e7 | 2162 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2163 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2164 | |
d8b4986d PT |
2165 | /* return excess runtime on last dequeue */ |
2166 | return_cfs_rq_runtime(cfs_rq); | |
2167 | ||
1e876231 | 2168 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2169 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2170 | } |
2171 | ||
2172 | /* | |
2173 | * Preempt the current task with a newly woken task if needed: | |
2174 | */ | |
7c92e54f | 2175 | static void |
2e09bf55 | 2176 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2177 | { |
11697830 | 2178 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2179 | struct sched_entity *se; |
2180 | s64 delta; | |
11697830 | 2181 | |
6d0f0ebd | 2182 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2183 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2184 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2185 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2186 | /* |
2187 | * The current task ran long enough, ensure it doesn't get | |
2188 | * re-elected due to buddy favours. | |
2189 | */ | |
2190 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2191 | return; |
2192 | } | |
2193 | ||
2194 | /* | |
2195 | * Ensure that a task that missed wakeup preemption by a | |
2196 | * narrow margin doesn't have to wait for a full slice. | |
2197 | * This also mitigates buddy induced latencies under load. | |
2198 | */ | |
f685ceac MG |
2199 | if (delta_exec < sysctl_sched_min_granularity) |
2200 | return; | |
2201 | ||
f4cfb33e WX |
2202 | se = __pick_first_entity(cfs_rq); |
2203 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2204 | |
f4cfb33e WX |
2205 | if (delta < 0) |
2206 | return; | |
d7d82944 | 2207 | |
f4cfb33e WX |
2208 | if (delta > ideal_runtime) |
2209 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2210 | } |
2211 | ||
83b699ed | 2212 | static void |
8494f412 | 2213 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2214 | { |
83b699ed SV |
2215 | /* 'current' is not kept within the tree. */ |
2216 | if (se->on_rq) { | |
2217 | /* | |
2218 | * Any task has to be enqueued before it get to execute on | |
2219 | * a CPU. So account for the time it spent waiting on the | |
2220 | * runqueue. | |
2221 | */ | |
2222 | update_stats_wait_end(cfs_rq, se); | |
2223 | __dequeue_entity(cfs_rq, se); | |
2224 | } | |
2225 | ||
79303e9e | 2226 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2227 | cfs_rq->curr = se; |
eba1ed4b IM |
2228 | #ifdef CONFIG_SCHEDSTATS |
2229 | /* | |
2230 | * Track our maximum slice length, if the CPU's load is at | |
2231 | * least twice that of our own weight (i.e. dont track it | |
2232 | * when there are only lesser-weight tasks around): | |
2233 | */ | |
495eca49 | 2234 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2235 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2236 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2237 | } | |
2238 | #endif | |
4a55b450 | 2239 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2240 | } |
2241 | ||
3f3a4904 PZ |
2242 | static int |
2243 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2244 | ||
ac53db59 RR |
2245 | /* |
2246 | * Pick the next process, keeping these things in mind, in this order: | |
2247 | * 1) keep things fair between processes/task groups | |
2248 | * 2) pick the "next" process, since someone really wants that to run | |
2249 | * 3) pick the "last" process, for cache locality | |
2250 | * 4) do not run the "skip" process, if something else is available | |
2251 | */ | |
f4b6755f | 2252 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 2253 | { |
ac53db59 | 2254 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 2255 | struct sched_entity *left = se; |
f4b6755f | 2256 | |
ac53db59 RR |
2257 | /* |
2258 | * Avoid running the skip buddy, if running something else can | |
2259 | * be done without getting too unfair. | |
2260 | */ | |
2261 | if (cfs_rq->skip == se) { | |
2262 | struct sched_entity *second = __pick_next_entity(se); | |
2263 | if (second && wakeup_preempt_entity(second, left) < 1) | |
2264 | se = second; | |
2265 | } | |
aa2ac252 | 2266 | |
f685ceac MG |
2267 | /* |
2268 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2269 | */ | |
2270 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2271 | se = cfs_rq->last; | |
2272 | ||
ac53db59 RR |
2273 | /* |
2274 | * Someone really wants this to run. If it's not unfair, run it. | |
2275 | */ | |
2276 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2277 | se = cfs_rq->next; | |
2278 | ||
f685ceac | 2279 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2280 | |
2281 | return se; | |
aa2ac252 PZ |
2282 | } |
2283 | ||
d3d9dc33 PT |
2284 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
2285 | ||
ab6cde26 | 2286 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2287 | { |
2288 | /* | |
2289 | * If still on the runqueue then deactivate_task() | |
2290 | * was not called and update_curr() has to be done: | |
2291 | */ | |
2292 | if (prev->on_rq) | |
b7cc0896 | 2293 | update_curr(cfs_rq); |
bf0f6f24 | 2294 | |
d3d9dc33 PT |
2295 | /* throttle cfs_rqs exceeding runtime */ |
2296 | check_cfs_rq_runtime(cfs_rq); | |
2297 | ||
ddc97297 | 2298 | check_spread(cfs_rq, prev); |
30cfdcfc | 2299 | if (prev->on_rq) { |
5870db5b | 2300 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2301 | /* Put 'current' back into the tree. */ |
2302 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2303 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2304 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2305 | } |
429d43bc | 2306 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2307 | } |
2308 | ||
8f4d37ec PZ |
2309 | static void |
2310 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2311 | { |
bf0f6f24 | 2312 | /* |
30cfdcfc | 2313 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2314 | */ |
30cfdcfc | 2315 | update_curr(cfs_rq); |
bf0f6f24 | 2316 | |
9d85f21c PT |
2317 | /* |
2318 | * Ensure that runnable average is periodically updated. | |
2319 | */ | |
9ee474f5 | 2320 | update_entity_load_avg(curr, 1); |
aff3e498 | 2321 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2322 | update_cfs_shares(cfs_rq); |
9d85f21c | 2323 | |
8f4d37ec PZ |
2324 | #ifdef CONFIG_SCHED_HRTICK |
2325 | /* | |
2326 | * queued ticks are scheduled to match the slice, so don't bother | |
2327 | * validating it and just reschedule. | |
2328 | */ | |
983ed7a6 HH |
2329 | if (queued) { |
2330 | resched_task(rq_of(cfs_rq)->curr); | |
2331 | return; | |
2332 | } | |
8f4d37ec PZ |
2333 | /* |
2334 | * don't let the period tick interfere with the hrtick preemption | |
2335 | */ | |
2336 | if (!sched_feat(DOUBLE_TICK) && | |
2337 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2338 | return; | |
2339 | #endif | |
2340 | ||
2c2efaed | 2341 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2342 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2343 | } |
2344 | ||
ab84d31e PT |
2345 | |
2346 | /************************************************** | |
2347 | * CFS bandwidth control machinery | |
2348 | */ | |
2349 | ||
2350 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2351 | |
2352 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2353 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2354 | |
2355 | static inline bool cfs_bandwidth_used(void) | |
2356 | { | |
c5905afb | 2357 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2358 | } |
2359 | ||
2360 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
2361 | { | |
2362 | /* only need to count groups transitioning between enabled/!enabled */ | |
2363 | if (enabled && !was_enabled) | |
c5905afb | 2364 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 2365 | else if (!enabled && was_enabled) |
c5905afb | 2366 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
2367 | } |
2368 | #else /* HAVE_JUMP_LABEL */ | |
2369 | static bool cfs_bandwidth_used(void) | |
2370 | { | |
2371 | return true; | |
2372 | } | |
2373 | ||
2374 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
2375 | #endif /* HAVE_JUMP_LABEL */ | |
2376 | ||
ab84d31e PT |
2377 | /* |
2378 | * default period for cfs group bandwidth. | |
2379 | * default: 0.1s, units: nanoseconds | |
2380 | */ | |
2381 | static inline u64 default_cfs_period(void) | |
2382 | { | |
2383 | return 100000000ULL; | |
2384 | } | |
ec12cb7f PT |
2385 | |
2386 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2387 | { | |
2388 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2389 | } | |
2390 | ||
a9cf55b2 PT |
2391 | /* |
2392 | * Replenish runtime according to assigned quota and update expiration time. | |
2393 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2394 | * additional synchronization around rq->lock. | |
2395 | * | |
2396 | * requires cfs_b->lock | |
2397 | */ | |
029632fb | 2398 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2399 | { |
2400 | u64 now; | |
2401 | ||
2402 | if (cfs_b->quota == RUNTIME_INF) | |
2403 | return; | |
2404 | ||
2405 | now = sched_clock_cpu(smp_processor_id()); | |
2406 | cfs_b->runtime = cfs_b->quota; | |
2407 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2408 | } | |
2409 | ||
029632fb PZ |
2410 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2411 | { | |
2412 | return &tg->cfs_bandwidth; | |
2413 | } | |
2414 | ||
f1b17280 PT |
2415 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2416 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2417 | { | |
2418 | if (unlikely(cfs_rq->throttle_count)) | |
2419 | return cfs_rq->throttled_clock_task; | |
2420 | ||
78becc27 | 2421 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2422 | } |
2423 | ||
85dac906 PT |
2424 | /* returns 0 on failure to allocate runtime */ |
2425 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2426 | { |
2427 | struct task_group *tg = cfs_rq->tg; | |
2428 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2429 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2430 | |
2431 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2432 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2433 | ||
2434 | raw_spin_lock(&cfs_b->lock); | |
2435 | if (cfs_b->quota == RUNTIME_INF) | |
2436 | amount = min_amount; | |
58088ad0 | 2437 | else { |
a9cf55b2 PT |
2438 | /* |
2439 | * If the bandwidth pool has become inactive, then at least one | |
2440 | * period must have elapsed since the last consumption. | |
2441 | * Refresh the global state and ensure bandwidth timer becomes | |
2442 | * active. | |
2443 | */ | |
2444 | if (!cfs_b->timer_active) { | |
2445 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2446 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2447 | } |
58088ad0 PT |
2448 | |
2449 | if (cfs_b->runtime > 0) { | |
2450 | amount = min(cfs_b->runtime, min_amount); | |
2451 | cfs_b->runtime -= amount; | |
2452 | cfs_b->idle = 0; | |
2453 | } | |
ec12cb7f | 2454 | } |
a9cf55b2 | 2455 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2456 | raw_spin_unlock(&cfs_b->lock); |
2457 | ||
2458 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2459 | /* |
2460 | * we may have advanced our local expiration to account for allowed | |
2461 | * spread between our sched_clock and the one on which runtime was | |
2462 | * issued. | |
2463 | */ | |
2464 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2465 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2466 | |
2467 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2468 | } |
2469 | ||
a9cf55b2 PT |
2470 | /* |
2471 | * Note: This depends on the synchronization provided by sched_clock and the | |
2472 | * fact that rq->clock snapshots this value. | |
2473 | */ | |
2474 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2475 | { |
a9cf55b2 | 2476 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2477 | |
2478 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2479 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2480 | return; |
2481 | ||
a9cf55b2 PT |
2482 | if (cfs_rq->runtime_remaining < 0) |
2483 | return; | |
2484 | ||
2485 | /* | |
2486 | * If the local deadline has passed we have to consider the | |
2487 | * possibility that our sched_clock is 'fast' and the global deadline | |
2488 | * has not truly expired. | |
2489 | * | |
2490 | * Fortunately we can check determine whether this the case by checking | |
2491 | * whether the global deadline has advanced. | |
2492 | */ | |
2493 | ||
2494 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
2495 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
2496 | cfs_rq->runtime_expires += TICK_NSEC; | |
2497 | } else { | |
2498 | /* global deadline is ahead, expiration has passed */ | |
2499 | cfs_rq->runtime_remaining = 0; | |
2500 | } | |
2501 | } | |
2502 | ||
2503 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2504 | unsigned long delta_exec) | |
2505 | { | |
2506 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 2507 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
2508 | expire_cfs_rq_runtime(cfs_rq); |
2509 | ||
2510 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
2511 | return; |
2512 | ||
85dac906 PT |
2513 | /* |
2514 | * if we're unable to extend our runtime we resched so that the active | |
2515 | * hierarchy can be throttled | |
2516 | */ | |
2517 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
2518 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
2519 | } |
2520 | ||
6c16a6dc PZ |
2521 | static __always_inline |
2522 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 2523 | { |
56f570e5 | 2524 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
2525 | return; |
2526 | ||
2527 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
2528 | } | |
2529 | ||
85dac906 PT |
2530 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
2531 | { | |
56f570e5 | 2532 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
2533 | } |
2534 | ||
64660c86 PT |
2535 | /* check whether cfs_rq, or any parent, is throttled */ |
2536 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2537 | { | |
56f570e5 | 2538 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
2539 | } |
2540 | ||
2541 | /* | |
2542 | * Ensure that neither of the group entities corresponding to src_cpu or | |
2543 | * dest_cpu are members of a throttled hierarchy when performing group | |
2544 | * load-balance operations. | |
2545 | */ | |
2546 | static inline int throttled_lb_pair(struct task_group *tg, | |
2547 | int src_cpu, int dest_cpu) | |
2548 | { | |
2549 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
2550 | ||
2551 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
2552 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
2553 | ||
2554 | return throttled_hierarchy(src_cfs_rq) || | |
2555 | throttled_hierarchy(dest_cfs_rq); | |
2556 | } | |
2557 | ||
2558 | /* updated child weight may affect parent so we have to do this bottom up */ | |
2559 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
2560 | { | |
2561 | struct rq *rq = data; | |
2562 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2563 | ||
2564 | cfs_rq->throttle_count--; | |
2565 | #ifdef CONFIG_SMP | |
2566 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 2567 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 2568 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 2569 | cfs_rq->throttled_clock_task; |
64660c86 PT |
2570 | } |
2571 | #endif | |
2572 | ||
2573 | return 0; | |
2574 | } | |
2575 | ||
2576 | static int tg_throttle_down(struct task_group *tg, void *data) | |
2577 | { | |
2578 | struct rq *rq = data; | |
2579 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2580 | ||
82958366 PT |
2581 | /* group is entering throttled state, stop time */ |
2582 | if (!cfs_rq->throttle_count) | |
78becc27 | 2583 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
2584 | cfs_rq->throttle_count++; |
2585 | ||
2586 | return 0; | |
2587 | } | |
2588 | ||
d3d9dc33 | 2589 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
2590 | { |
2591 | struct rq *rq = rq_of(cfs_rq); | |
2592 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2593 | struct sched_entity *se; | |
2594 | long task_delta, dequeue = 1; | |
2595 | ||
2596 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2597 | ||
f1b17280 | 2598 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
2599 | rcu_read_lock(); |
2600 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
2601 | rcu_read_unlock(); | |
85dac906 PT |
2602 | |
2603 | task_delta = cfs_rq->h_nr_running; | |
2604 | for_each_sched_entity(se) { | |
2605 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
2606 | /* throttled entity or throttle-on-deactivate */ | |
2607 | if (!se->on_rq) | |
2608 | break; | |
2609 | ||
2610 | if (dequeue) | |
2611 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
2612 | qcfs_rq->h_nr_running -= task_delta; | |
2613 | ||
2614 | if (qcfs_rq->load.weight) | |
2615 | dequeue = 0; | |
2616 | } | |
2617 | ||
2618 | if (!se) | |
2619 | rq->nr_running -= task_delta; | |
2620 | ||
2621 | cfs_rq->throttled = 1; | |
78becc27 | 2622 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
2623 | raw_spin_lock(&cfs_b->lock); |
2624 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
2625 | raw_spin_unlock(&cfs_b->lock); | |
2626 | } | |
2627 | ||
029632fb | 2628 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
2629 | { |
2630 | struct rq *rq = rq_of(cfs_rq); | |
2631 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2632 | struct sched_entity *se; | |
2633 | int enqueue = 1; | |
2634 | long task_delta; | |
2635 | ||
22b958d8 | 2636 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
2637 | |
2638 | cfs_rq->throttled = 0; | |
1a55af2e FW |
2639 | |
2640 | update_rq_clock(rq); | |
2641 | ||
671fd9da | 2642 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 2643 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
2644 | list_del_rcu(&cfs_rq->throttled_list); |
2645 | raw_spin_unlock(&cfs_b->lock); | |
2646 | ||
64660c86 PT |
2647 | /* update hierarchical throttle state */ |
2648 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
2649 | ||
671fd9da PT |
2650 | if (!cfs_rq->load.weight) |
2651 | return; | |
2652 | ||
2653 | task_delta = cfs_rq->h_nr_running; | |
2654 | for_each_sched_entity(se) { | |
2655 | if (se->on_rq) | |
2656 | enqueue = 0; | |
2657 | ||
2658 | cfs_rq = cfs_rq_of(se); | |
2659 | if (enqueue) | |
2660 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
2661 | cfs_rq->h_nr_running += task_delta; | |
2662 | ||
2663 | if (cfs_rq_throttled(cfs_rq)) | |
2664 | break; | |
2665 | } | |
2666 | ||
2667 | if (!se) | |
2668 | rq->nr_running += task_delta; | |
2669 | ||
2670 | /* determine whether we need to wake up potentially idle cpu */ | |
2671 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
2672 | resched_task(rq->curr); | |
2673 | } | |
2674 | ||
2675 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
2676 | u64 remaining, u64 expires) | |
2677 | { | |
2678 | struct cfs_rq *cfs_rq; | |
2679 | u64 runtime = remaining; | |
2680 | ||
2681 | rcu_read_lock(); | |
2682 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
2683 | throttled_list) { | |
2684 | struct rq *rq = rq_of(cfs_rq); | |
2685 | ||
2686 | raw_spin_lock(&rq->lock); | |
2687 | if (!cfs_rq_throttled(cfs_rq)) | |
2688 | goto next; | |
2689 | ||
2690 | runtime = -cfs_rq->runtime_remaining + 1; | |
2691 | if (runtime > remaining) | |
2692 | runtime = remaining; | |
2693 | remaining -= runtime; | |
2694 | ||
2695 | cfs_rq->runtime_remaining += runtime; | |
2696 | cfs_rq->runtime_expires = expires; | |
2697 | ||
2698 | /* we check whether we're throttled above */ | |
2699 | if (cfs_rq->runtime_remaining > 0) | |
2700 | unthrottle_cfs_rq(cfs_rq); | |
2701 | ||
2702 | next: | |
2703 | raw_spin_unlock(&rq->lock); | |
2704 | ||
2705 | if (!remaining) | |
2706 | break; | |
2707 | } | |
2708 | rcu_read_unlock(); | |
2709 | ||
2710 | return remaining; | |
2711 | } | |
2712 | ||
58088ad0 PT |
2713 | /* |
2714 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2715 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2716 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2717 | * used to track this state. | |
2718 | */ | |
2719 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
2720 | { | |
671fd9da PT |
2721 | u64 runtime, runtime_expires; |
2722 | int idle = 1, throttled; | |
58088ad0 PT |
2723 | |
2724 | raw_spin_lock(&cfs_b->lock); | |
2725 | /* no need to continue the timer with no bandwidth constraint */ | |
2726 | if (cfs_b->quota == RUNTIME_INF) | |
2727 | goto out_unlock; | |
2728 | ||
671fd9da PT |
2729 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
2730 | /* idle depends on !throttled (for the case of a large deficit) */ | |
2731 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 2732 | cfs_b->nr_periods += overrun; |
671fd9da | 2733 | |
a9cf55b2 PT |
2734 | /* if we're going inactive then everything else can be deferred */ |
2735 | if (idle) | |
2736 | goto out_unlock; | |
2737 | ||
2738 | __refill_cfs_bandwidth_runtime(cfs_b); | |
2739 | ||
671fd9da PT |
2740 | if (!throttled) { |
2741 | /* mark as potentially idle for the upcoming period */ | |
2742 | cfs_b->idle = 1; | |
2743 | goto out_unlock; | |
2744 | } | |
2745 | ||
e8da1b18 NR |
2746 | /* account preceding periods in which throttling occurred */ |
2747 | cfs_b->nr_throttled += overrun; | |
2748 | ||
671fd9da PT |
2749 | /* |
2750 | * There are throttled entities so we must first use the new bandwidth | |
2751 | * to unthrottle them before making it generally available. This | |
2752 | * ensures that all existing debts will be paid before a new cfs_rq is | |
2753 | * allowed to run. | |
2754 | */ | |
2755 | runtime = cfs_b->runtime; | |
2756 | runtime_expires = cfs_b->runtime_expires; | |
2757 | cfs_b->runtime = 0; | |
2758 | ||
2759 | /* | |
2760 | * This check is repeated as we are holding onto the new bandwidth | |
2761 | * while we unthrottle. This can potentially race with an unthrottled | |
2762 | * group trying to acquire new bandwidth from the global pool. | |
2763 | */ | |
2764 | while (throttled && runtime > 0) { | |
2765 | raw_spin_unlock(&cfs_b->lock); | |
2766 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
2767 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
2768 | runtime_expires); | |
2769 | raw_spin_lock(&cfs_b->lock); | |
2770 | ||
2771 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
2772 | } | |
58088ad0 | 2773 | |
671fd9da PT |
2774 | /* return (any) remaining runtime */ |
2775 | cfs_b->runtime = runtime; | |
2776 | /* | |
2777 | * While we are ensured activity in the period following an | |
2778 | * unthrottle, this also covers the case in which the new bandwidth is | |
2779 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
2780 | * timer to remain active while there are any throttled entities.) | |
2781 | */ | |
2782 | cfs_b->idle = 0; | |
58088ad0 PT |
2783 | out_unlock: |
2784 | if (idle) | |
2785 | cfs_b->timer_active = 0; | |
2786 | raw_spin_unlock(&cfs_b->lock); | |
2787 | ||
2788 | return idle; | |
2789 | } | |
d3d9dc33 | 2790 | |
d8b4986d PT |
2791 | /* a cfs_rq won't donate quota below this amount */ |
2792 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
2793 | /* minimum remaining period time to redistribute slack quota */ | |
2794 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
2795 | /* how long we wait to gather additional slack before distributing */ | |
2796 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
2797 | ||
2798 | /* are we near the end of the current quota period? */ | |
2799 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
2800 | { | |
2801 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
2802 | u64 remaining; | |
2803 | ||
2804 | /* if the call-back is running a quota refresh is already occurring */ | |
2805 | if (hrtimer_callback_running(refresh_timer)) | |
2806 | return 1; | |
2807 | ||
2808 | /* is a quota refresh about to occur? */ | |
2809 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
2810 | if (remaining < min_expire) | |
2811 | return 1; | |
2812 | ||
2813 | return 0; | |
2814 | } | |
2815 | ||
2816 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
2817 | { | |
2818 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
2819 | ||
2820 | /* if there's a quota refresh soon don't bother with slack */ | |
2821 | if (runtime_refresh_within(cfs_b, min_left)) | |
2822 | return; | |
2823 | ||
2824 | start_bandwidth_timer(&cfs_b->slack_timer, | |
2825 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
2826 | } | |
2827 | ||
2828 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
2829 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2830 | { | |
2831 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2832 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
2833 | ||
2834 | if (slack_runtime <= 0) | |
2835 | return; | |
2836 | ||
2837 | raw_spin_lock(&cfs_b->lock); | |
2838 | if (cfs_b->quota != RUNTIME_INF && | |
2839 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
2840 | cfs_b->runtime += slack_runtime; | |
2841 | ||
2842 | /* we are under rq->lock, defer unthrottling using a timer */ | |
2843 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
2844 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
2845 | start_cfs_slack_bandwidth(cfs_b); | |
2846 | } | |
2847 | raw_spin_unlock(&cfs_b->lock); | |
2848 | ||
2849 | /* even if it's not valid for return we don't want to try again */ | |
2850 | cfs_rq->runtime_remaining -= slack_runtime; | |
2851 | } | |
2852 | ||
2853 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2854 | { | |
56f570e5 PT |
2855 | if (!cfs_bandwidth_used()) |
2856 | return; | |
2857 | ||
fccfdc6f | 2858 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
2859 | return; |
2860 | ||
2861 | __return_cfs_rq_runtime(cfs_rq); | |
2862 | } | |
2863 | ||
2864 | /* | |
2865 | * This is done with a timer (instead of inline with bandwidth return) since | |
2866 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
2867 | */ | |
2868 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
2869 | { | |
2870 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
2871 | u64 expires; | |
2872 | ||
2873 | /* confirm we're still not at a refresh boundary */ | |
2874 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
2875 | return; | |
2876 | ||
2877 | raw_spin_lock(&cfs_b->lock); | |
2878 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
2879 | runtime = cfs_b->runtime; | |
2880 | cfs_b->runtime = 0; | |
2881 | } | |
2882 | expires = cfs_b->runtime_expires; | |
2883 | raw_spin_unlock(&cfs_b->lock); | |
2884 | ||
2885 | if (!runtime) | |
2886 | return; | |
2887 | ||
2888 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
2889 | ||
2890 | raw_spin_lock(&cfs_b->lock); | |
2891 | if (expires == cfs_b->runtime_expires) | |
2892 | cfs_b->runtime = runtime; | |
2893 | raw_spin_unlock(&cfs_b->lock); | |
2894 | } | |
2895 | ||
d3d9dc33 PT |
2896 | /* |
2897 | * When a group wakes up we want to make sure that its quota is not already | |
2898 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
2899 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
2900 | */ | |
2901 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
2902 | { | |
56f570e5 PT |
2903 | if (!cfs_bandwidth_used()) |
2904 | return; | |
2905 | ||
d3d9dc33 PT |
2906 | /* an active group must be handled by the update_curr()->put() path */ |
2907 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
2908 | return; | |
2909 | ||
2910 | /* ensure the group is not already throttled */ | |
2911 | if (cfs_rq_throttled(cfs_rq)) | |
2912 | return; | |
2913 | ||
2914 | /* update runtime allocation */ | |
2915 | account_cfs_rq_runtime(cfs_rq, 0); | |
2916 | if (cfs_rq->runtime_remaining <= 0) | |
2917 | throttle_cfs_rq(cfs_rq); | |
2918 | } | |
2919 | ||
2920 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
2921 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2922 | { | |
56f570e5 PT |
2923 | if (!cfs_bandwidth_used()) |
2924 | return; | |
2925 | ||
d3d9dc33 PT |
2926 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
2927 | return; | |
2928 | ||
2929 | /* | |
2930 | * it's possible for a throttled entity to be forced into a running | |
2931 | * state (e.g. set_curr_task), in this case we're finished. | |
2932 | */ | |
2933 | if (cfs_rq_throttled(cfs_rq)) | |
2934 | return; | |
2935 | ||
2936 | throttle_cfs_rq(cfs_rq); | |
2937 | } | |
029632fb | 2938 | |
029632fb PZ |
2939 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
2940 | { | |
2941 | struct cfs_bandwidth *cfs_b = | |
2942 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
2943 | do_sched_cfs_slack_timer(cfs_b); | |
2944 | ||
2945 | return HRTIMER_NORESTART; | |
2946 | } | |
2947 | ||
2948 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
2949 | { | |
2950 | struct cfs_bandwidth *cfs_b = | |
2951 | container_of(timer, struct cfs_bandwidth, period_timer); | |
2952 | ktime_t now; | |
2953 | int overrun; | |
2954 | int idle = 0; | |
2955 | ||
2956 | for (;;) { | |
2957 | now = hrtimer_cb_get_time(timer); | |
2958 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
2959 | ||
2960 | if (!overrun) | |
2961 | break; | |
2962 | ||
2963 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
2964 | } | |
2965 | ||
2966 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
2967 | } | |
2968 | ||
2969 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2970 | { | |
2971 | raw_spin_lock_init(&cfs_b->lock); | |
2972 | cfs_b->runtime = 0; | |
2973 | cfs_b->quota = RUNTIME_INF; | |
2974 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
2975 | ||
2976 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
2977 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2978 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
2979 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2980 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
2981 | } | |
2982 | ||
2983 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
2984 | { | |
2985 | cfs_rq->runtime_enabled = 0; | |
2986 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
2987 | } | |
2988 | ||
2989 | /* requires cfs_b->lock, may release to reprogram timer */ | |
2990 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
2991 | { | |
2992 | /* | |
2993 | * The timer may be active because we're trying to set a new bandwidth | |
2994 | * period or because we're racing with the tear-down path | |
2995 | * (timer_active==0 becomes visible before the hrtimer call-back | |
2996 | * terminates). In either case we ensure that it's re-programmed | |
2997 | */ | |
2998 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
2999 | raw_spin_unlock(&cfs_b->lock); | |
3000 | /* ensure cfs_b->lock is available while we wait */ | |
3001 | hrtimer_cancel(&cfs_b->period_timer); | |
3002 | ||
3003 | raw_spin_lock(&cfs_b->lock); | |
3004 | /* if someone else restarted the timer then we're done */ | |
3005 | if (cfs_b->timer_active) | |
3006 | return; | |
3007 | } | |
3008 | ||
3009 | cfs_b->timer_active = 1; | |
3010 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3011 | } | |
3012 | ||
3013 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3014 | { | |
3015 | hrtimer_cancel(&cfs_b->period_timer); | |
3016 | hrtimer_cancel(&cfs_b->slack_timer); | |
3017 | } | |
3018 | ||
38dc3348 | 3019 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3020 | { |
3021 | struct cfs_rq *cfs_rq; | |
3022 | ||
3023 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3024 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3025 | ||
3026 | if (!cfs_rq->runtime_enabled) | |
3027 | continue; | |
3028 | ||
3029 | /* | |
3030 | * clock_task is not advancing so we just need to make sure | |
3031 | * there's some valid quota amount | |
3032 | */ | |
3033 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3034 | if (cfs_rq_throttled(cfs_rq)) | |
3035 | unthrottle_cfs_rq(cfs_rq); | |
3036 | } | |
3037 | } | |
3038 | ||
3039 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3040 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3041 | { | |
78becc27 | 3042 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3043 | } |
3044 | ||
3045 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
3046 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
3047 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
3048 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 3049 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3050 | |
3051 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3052 | { | |
3053 | return 0; | |
3054 | } | |
64660c86 PT |
3055 | |
3056 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3057 | { | |
3058 | return 0; | |
3059 | } | |
3060 | ||
3061 | static inline int throttled_lb_pair(struct task_group *tg, | |
3062 | int src_cpu, int dest_cpu) | |
3063 | { | |
3064 | return 0; | |
3065 | } | |
029632fb PZ |
3066 | |
3067 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3068 | ||
3069 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3070 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3071 | #endif |
3072 | ||
029632fb PZ |
3073 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3074 | { | |
3075 | return NULL; | |
3076 | } | |
3077 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3078 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3079 | |
3080 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3081 | ||
bf0f6f24 IM |
3082 | /************************************************** |
3083 | * CFS operations on tasks: | |
3084 | */ | |
3085 | ||
8f4d37ec PZ |
3086 | #ifdef CONFIG_SCHED_HRTICK |
3087 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3088 | { | |
8f4d37ec PZ |
3089 | struct sched_entity *se = &p->se; |
3090 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3091 | ||
3092 | WARN_ON(task_rq(p) != rq); | |
3093 | ||
b39e66ea | 3094 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3095 | u64 slice = sched_slice(cfs_rq, se); |
3096 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3097 | s64 delta = slice - ran; | |
3098 | ||
3099 | if (delta < 0) { | |
3100 | if (rq->curr == p) | |
3101 | resched_task(p); | |
3102 | return; | |
3103 | } | |
3104 | ||
3105 | /* | |
3106 | * Don't schedule slices shorter than 10000ns, that just | |
3107 | * doesn't make sense. Rely on vruntime for fairness. | |
3108 | */ | |
31656519 | 3109 | if (rq->curr != p) |
157124c1 | 3110 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3111 | |
31656519 | 3112 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3113 | } |
3114 | } | |
a4c2f00f PZ |
3115 | |
3116 | /* | |
3117 | * called from enqueue/dequeue and updates the hrtick when the | |
3118 | * current task is from our class and nr_running is low enough | |
3119 | * to matter. | |
3120 | */ | |
3121 | static void hrtick_update(struct rq *rq) | |
3122 | { | |
3123 | struct task_struct *curr = rq->curr; | |
3124 | ||
b39e66ea | 3125 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3126 | return; |
3127 | ||
3128 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3129 | hrtick_start_fair(rq, curr); | |
3130 | } | |
55e12e5e | 3131 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3132 | static inline void |
3133 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3134 | { | |
3135 | } | |
a4c2f00f PZ |
3136 | |
3137 | static inline void hrtick_update(struct rq *rq) | |
3138 | { | |
3139 | } | |
8f4d37ec PZ |
3140 | #endif |
3141 | ||
bf0f6f24 IM |
3142 | /* |
3143 | * The enqueue_task method is called before nr_running is | |
3144 | * increased. Here we update the fair scheduling stats and | |
3145 | * then put the task into the rbtree: | |
3146 | */ | |
ea87bb78 | 3147 | static void |
371fd7e7 | 3148 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3149 | { |
3150 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3151 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3152 | |
3153 | for_each_sched_entity(se) { | |
62fb1851 | 3154 | if (se->on_rq) |
bf0f6f24 IM |
3155 | break; |
3156 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3157 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3158 | |
3159 | /* | |
3160 | * end evaluation on encountering a throttled cfs_rq | |
3161 | * | |
3162 | * note: in the case of encountering a throttled cfs_rq we will | |
3163 | * post the final h_nr_running increment below. | |
3164 | */ | |
3165 | if (cfs_rq_throttled(cfs_rq)) | |
3166 | break; | |
953bfcd1 | 3167 | cfs_rq->h_nr_running++; |
85dac906 | 3168 | |
88ec22d3 | 3169 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3170 | } |
8f4d37ec | 3171 | |
2069dd75 | 3172 | for_each_sched_entity(se) { |
0f317143 | 3173 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3174 | cfs_rq->h_nr_running++; |
2069dd75 | 3175 | |
85dac906 PT |
3176 | if (cfs_rq_throttled(cfs_rq)) |
3177 | break; | |
3178 | ||
17bc14b7 | 3179 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3180 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3181 | } |
3182 | ||
18bf2805 BS |
3183 | if (!se) { |
3184 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3185 | inc_nr_running(rq); |
18bf2805 | 3186 | } |
a4c2f00f | 3187 | hrtick_update(rq); |
bf0f6f24 IM |
3188 | } |
3189 | ||
2f36825b VP |
3190 | static void set_next_buddy(struct sched_entity *se); |
3191 | ||
bf0f6f24 IM |
3192 | /* |
3193 | * The dequeue_task method is called before nr_running is | |
3194 | * decreased. We remove the task from the rbtree and | |
3195 | * update the fair scheduling stats: | |
3196 | */ | |
371fd7e7 | 3197 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3198 | { |
3199 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3200 | struct sched_entity *se = &p->se; |
2f36825b | 3201 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3202 | |
3203 | for_each_sched_entity(se) { | |
3204 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3205 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3206 | |
3207 | /* | |
3208 | * end evaluation on encountering a throttled cfs_rq | |
3209 | * | |
3210 | * note: in the case of encountering a throttled cfs_rq we will | |
3211 | * post the final h_nr_running decrement below. | |
3212 | */ | |
3213 | if (cfs_rq_throttled(cfs_rq)) | |
3214 | break; | |
953bfcd1 | 3215 | cfs_rq->h_nr_running--; |
2069dd75 | 3216 | |
bf0f6f24 | 3217 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3218 | if (cfs_rq->load.weight) { |
3219 | /* | |
3220 | * Bias pick_next to pick a task from this cfs_rq, as | |
3221 | * p is sleeping when it is within its sched_slice. | |
3222 | */ | |
3223 | if (task_sleep && parent_entity(se)) | |
3224 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3225 | |
3226 | /* avoid re-evaluating load for this entity */ | |
3227 | se = parent_entity(se); | |
bf0f6f24 | 3228 | break; |
2f36825b | 3229 | } |
371fd7e7 | 3230 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3231 | } |
8f4d37ec | 3232 | |
2069dd75 | 3233 | for_each_sched_entity(se) { |
0f317143 | 3234 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3235 | cfs_rq->h_nr_running--; |
2069dd75 | 3236 | |
85dac906 PT |
3237 | if (cfs_rq_throttled(cfs_rq)) |
3238 | break; | |
3239 | ||
17bc14b7 | 3240 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3241 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3242 | } |
3243 | ||
18bf2805 | 3244 | if (!se) { |
85dac906 | 3245 | dec_nr_running(rq); |
18bf2805 BS |
3246 | update_rq_runnable_avg(rq, 1); |
3247 | } | |
a4c2f00f | 3248 | hrtick_update(rq); |
bf0f6f24 IM |
3249 | } |
3250 | ||
e7693a36 | 3251 | #ifdef CONFIG_SMP |
029632fb PZ |
3252 | /* Used instead of source_load when we know the type == 0 */ |
3253 | static unsigned long weighted_cpuload(const int cpu) | |
3254 | { | |
b92486cb | 3255 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3256 | } |
3257 | ||
3258 | /* | |
3259 | * Return a low guess at the load of a migration-source cpu weighted | |
3260 | * according to the scheduling class and "nice" value. | |
3261 | * | |
3262 | * We want to under-estimate the load of migration sources, to | |
3263 | * balance conservatively. | |
3264 | */ | |
3265 | static unsigned long source_load(int cpu, int type) | |
3266 | { | |
3267 | struct rq *rq = cpu_rq(cpu); | |
3268 | unsigned long total = weighted_cpuload(cpu); | |
3269 | ||
3270 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3271 | return total; | |
3272 | ||
3273 | return min(rq->cpu_load[type-1], total); | |
3274 | } | |
3275 | ||
3276 | /* | |
3277 | * Return a high guess at the load of a migration-target cpu weighted | |
3278 | * according to the scheduling class and "nice" value. | |
3279 | */ | |
3280 | static unsigned long target_load(int cpu, int type) | |
3281 | { | |
3282 | struct rq *rq = cpu_rq(cpu); | |
3283 | unsigned long total = weighted_cpuload(cpu); | |
3284 | ||
3285 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3286 | return total; | |
3287 | ||
3288 | return max(rq->cpu_load[type-1], total); | |
3289 | } | |
3290 | ||
3291 | static unsigned long power_of(int cpu) | |
3292 | { | |
3293 | return cpu_rq(cpu)->cpu_power; | |
3294 | } | |
3295 | ||
3296 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3297 | { | |
3298 | struct rq *rq = cpu_rq(cpu); | |
3299 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3300 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3301 | |
3302 | if (nr_running) | |
b92486cb | 3303 | return load_avg / nr_running; |
029632fb PZ |
3304 | |
3305 | return 0; | |
3306 | } | |
3307 | ||
62470419 MW |
3308 | static void record_wakee(struct task_struct *p) |
3309 | { | |
3310 | /* | |
3311 | * Rough decay (wiping) for cost saving, don't worry | |
3312 | * about the boundary, really active task won't care | |
3313 | * about the loss. | |
3314 | */ | |
3315 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
3316 | current->wakee_flips = 0; | |
3317 | current->wakee_flip_decay_ts = jiffies; | |
3318 | } | |
3319 | ||
3320 | if (current->last_wakee != p) { | |
3321 | current->last_wakee = p; | |
3322 | current->wakee_flips++; | |
3323 | } | |
3324 | } | |
098fb9db | 3325 | |
74f8e4b2 | 3326 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
3327 | { |
3328 | struct sched_entity *se = &p->se; | |
3329 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
3330 | u64 min_vruntime; |
3331 | ||
3332 | #ifndef CONFIG_64BIT | |
3333 | u64 min_vruntime_copy; | |
88ec22d3 | 3334 | |
3fe1698b PZ |
3335 | do { |
3336 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3337 | smp_rmb(); | |
3338 | min_vruntime = cfs_rq->min_vruntime; | |
3339 | } while (min_vruntime != min_vruntime_copy); | |
3340 | #else | |
3341 | min_vruntime = cfs_rq->min_vruntime; | |
3342 | #endif | |
88ec22d3 | 3343 | |
3fe1698b | 3344 | se->vruntime -= min_vruntime; |
62470419 | 3345 | record_wakee(p); |
88ec22d3 PZ |
3346 | } |
3347 | ||
bb3469ac | 3348 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3349 | /* |
3350 | * effective_load() calculates the load change as seen from the root_task_group | |
3351 | * | |
3352 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3353 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3354 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3355 | * |
3356 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3357 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3358 | * total group weight. | |
3359 | * | |
3360 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3361 | * distribution (s_i) using: | |
3362 | * | |
3363 | * s_i = rw_i / \Sum rw_j (1) | |
3364 | * | |
3365 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3366 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3367 | * shares distribution (s_i): | |
3368 | * | |
3369 | * rw_i = { 2, 4, 1, 0 } | |
3370 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3371 | * | |
3372 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3373 | * task used to run on and the CPU the waker is running on), we need to | |
3374 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3375 | * wakeup, compute the effect of the current task going to sleep. | |
3376 | * | |
3377 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3378 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3379 | * | |
3380 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3381 | * | |
3382 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3383 | * differences in waking a task to CPU 0. The additional task changes the | |
3384 | * weight and shares distributions like: | |
3385 | * | |
3386 | * rw'_i = { 3, 4, 1, 0 } | |
3387 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3388 | * | |
3389 | * We can then compute the difference in effective weight by using: | |
3390 | * | |
3391 | * dw_i = S * (s'_i - s_i) (3) | |
3392 | * | |
3393 | * Where 'S' is the group weight as seen by its parent. | |
3394 | * | |
3395 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3396 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3397 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3398 | */ |
2069dd75 | 3399 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3400 | { |
4be9daaa | 3401 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3402 | |
58d081b5 | 3403 | if (!tg->parent || !wl) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3404 | return wl; |
3405 | ||
4be9daaa | 3406 | for_each_sched_entity(se) { |
cf5f0acf | 3407 | long w, W; |
4be9daaa | 3408 | |
977dda7c | 3409 | tg = se->my_q->tg; |
bb3469ac | 3410 | |
cf5f0acf PZ |
3411 | /* |
3412 | * W = @wg + \Sum rw_j | |
3413 | */ | |
3414 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3415 | |
cf5f0acf PZ |
3416 | /* |
3417 | * w = rw_i + @wl | |
3418 | */ | |
3419 | w = se->my_q->load.weight + wl; | |
940959e9 | 3420 | |
cf5f0acf PZ |
3421 | /* |
3422 | * wl = S * s'_i; see (2) | |
3423 | */ | |
3424 | if (W > 0 && w < W) | |
3425 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3426 | else |
3427 | wl = tg->shares; | |
940959e9 | 3428 | |
cf5f0acf PZ |
3429 | /* |
3430 | * Per the above, wl is the new se->load.weight value; since | |
3431 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3432 | * calc_cfs_shares(). | |
3433 | */ | |
977dda7c PT |
3434 | if (wl < MIN_SHARES) |
3435 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3436 | |
3437 | /* | |
3438 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3439 | */ | |
977dda7c | 3440 | wl -= se->load.weight; |
cf5f0acf PZ |
3441 | |
3442 | /* | |
3443 | * Recursively apply this logic to all parent groups to compute | |
3444 | * the final effective load change on the root group. Since | |
3445 | * only the @tg group gets extra weight, all parent groups can | |
3446 | * only redistribute existing shares. @wl is the shift in shares | |
3447 | * resulting from this level per the above. | |
3448 | */ | |
4be9daaa | 3449 | wg = 0; |
4be9daaa | 3450 | } |
bb3469ac | 3451 | |
4be9daaa | 3452 | return wl; |
bb3469ac PZ |
3453 | } |
3454 | #else | |
4be9daaa | 3455 | |
58d081b5 | 3456 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 3457 | { |
83378269 | 3458 | return wl; |
bb3469ac | 3459 | } |
4be9daaa | 3460 | |
bb3469ac PZ |
3461 | #endif |
3462 | ||
62470419 MW |
3463 | static int wake_wide(struct task_struct *p) |
3464 | { | |
7d9ffa89 | 3465 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
3466 | |
3467 | /* | |
3468 | * Yeah, it's the switching-frequency, could means many wakee or | |
3469 | * rapidly switch, use factor here will just help to automatically | |
3470 | * adjust the loose-degree, so bigger node will lead to more pull. | |
3471 | */ | |
3472 | if (p->wakee_flips > factor) { | |
3473 | /* | |
3474 | * wakee is somewhat hot, it needs certain amount of cpu | |
3475 | * resource, so if waker is far more hot, prefer to leave | |
3476 | * it alone. | |
3477 | */ | |
3478 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
3479 | return 1; | |
3480 | } | |
3481 | ||
3482 | return 0; | |
3483 | } | |
3484 | ||
c88d5910 | 3485 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 3486 | { |
e37b6a7b | 3487 | s64 this_load, load; |
c88d5910 | 3488 | int idx, this_cpu, prev_cpu; |
098fb9db | 3489 | unsigned long tl_per_task; |
c88d5910 | 3490 | struct task_group *tg; |
83378269 | 3491 | unsigned long weight; |
b3137bc8 | 3492 | int balanced; |
098fb9db | 3493 | |
62470419 MW |
3494 | /* |
3495 | * If we wake multiple tasks be careful to not bounce | |
3496 | * ourselves around too much. | |
3497 | */ | |
3498 | if (wake_wide(p)) | |
3499 | return 0; | |
3500 | ||
c88d5910 PZ |
3501 | idx = sd->wake_idx; |
3502 | this_cpu = smp_processor_id(); | |
3503 | prev_cpu = task_cpu(p); | |
3504 | load = source_load(prev_cpu, idx); | |
3505 | this_load = target_load(this_cpu, idx); | |
098fb9db | 3506 | |
b3137bc8 MG |
3507 | /* |
3508 | * If sync wakeup then subtract the (maximum possible) | |
3509 | * effect of the currently running task from the load | |
3510 | * of the current CPU: | |
3511 | */ | |
83378269 PZ |
3512 | if (sync) { |
3513 | tg = task_group(current); | |
3514 | weight = current->se.load.weight; | |
3515 | ||
c88d5910 | 3516 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
3517 | load += effective_load(tg, prev_cpu, 0, -weight); |
3518 | } | |
b3137bc8 | 3519 | |
83378269 PZ |
3520 | tg = task_group(p); |
3521 | weight = p->se.load.weight; | |
b3137bc8 | 3522 | |
71a29aa7 PZ |
3523 | /* |
3524 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
3525 | * due to the sync cause above having dropped this_load to 0, we'll |
3526 | * always have an imbalance, but there's really nothing you can do | |
3527 | * about that, so that's good too. | |
71a29aa7 PZ |
3528 | * |
3529 | * Otherwise check if either cpus are near enough in load to allow this | |
3530 | * task to be woken on this_cpu. | |
3531 | */ | |
e37b6a7b PT |
3532 | if (this_load > 0) { |
3533 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
3534 | |
3535 | this_eff_load = 100; | |
3536 | this_eff_load *= power_of(prev_cpu); | |
3537 | this_eff_load *= this_load + | |
3538 | effective_load(tg, this_cpu, weight, weight); | |
3539 | ||
3540 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
3541 | prev_eff_load *= power_of(this_cpu); | |
3542 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
3543 | ||
3544 | balanced = this_eff_load <= prev_eff_load; | |
3545 | } else | |
3546 | balanced = true; | |
b3137bc8 | 3547 | |
098fb9db | 3548 | /* |
4ae7d5ce IM |
3549 | * If the currently running task will sleep within |
3550 | * a reasonable amount of time then attract this newly | |
3551 | * woken task: | |
098fb9db | 3552 | */ |
2fb7635c PZ |
3553 | if (sync && balanced) |
3554 | return 1; | |
098fb9db | 3555 | |
41acab88 | 3556 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
3557 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
3558 | ||
c88d5910 PZ |
3559 | if (balanced || |
3560 | (this_load <= load && | |
3561 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
3562 | /* |
3563 | * This domain has SD_WAKE_AFFINE and | |
3564 | * p is cache cold in this domain, and | |
3565 | * there is no bad imbalance. | |
3566 | */ | |
c88d5910 | 3567 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 3568 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
3569 | |
3570 | return 1; | |
3571 | } | |
3572 | return 0; | |
3573 | } | |
3574 | ||
aaee1203 PZ |
3575 | /* |
3576 | * find_idlest_group finds and returns the least busy CPU group within the | |
3577 | * domain. | |
3578 | */ | |
3579 | static struct sched_group * | |
78e7ed53 | 3580 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 3581 | int this_cpu, int load_idx) |
e7693a36 | 3582 | { |
b3bd3de6 | 3583 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 3584 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 3585 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 3586 | |
aaee1203 PZ |
3587 | do { |
3588 | unsigned long load, avg_load; | |
3589 | int local_group; | |
3590 | int i; | |
e7693a36 | 3591 | |
aaee1203 PZ |
3592 | /* Skip over this group if it has no CPUs allowed */ |
3593 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 3594 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
3595 | continue; |
3596 | ||
3597 | local_group = cpumask_test_cpu(this_cpu, | |
3598 | sched_group_cpus(group)); | |
3599 | ||
3600 | /* Tally up the load of all CPUs in the group */ | |
3601 | avg_load = 0; | |
3602 | ||
3603 | for_each_cpu(i, sched_group_cpus(group)) { | |
3604 | /* Bias balancing toward cpus of our domain */ | |
3605 | if (local_group) | |
3606 | load = source_load(i, load_idx); | |
3607 | else | |
3608 | load = target_load(i, load_idx); | |
3609 | ||
3610 | avg_load += load; | |
3611 | } | |
3612 | ||
3613 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3614 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
3615 | |
3616 | if (local_group) { | |
3617 | this_load = avg_load; | |
aaee1203 PZ |
3618 | } else if (avg_load < min_load) { |
3619 | min_load = avg_load; | |
3620 | idlest = group; | |
3621 | } | |
3622 | } while (group = group->next, group != sd->groups); | |
3623 | ||
3624 | if (!idlest || 100*this_load < imbalance*min_load) | |
3625 | return NULL; | |
3626 | return idlest; | |
3627 | } | |
3628 | ||
3629 | /* | |
3630 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
3631 | */ | |
3632 | static int | |
3633 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
3634 | { | |
3635 | unsigned long load, min_load = ULONG_MAX; | |
3636 | int idlest = -1; | |
3637 | int i; | |
3638 | ||
3639 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 3640 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
3641 | load = weighted_cpuload(i); |
3642 | ||
3643 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
3644 | min_load = load; | |
3645 | idlest = i; | |
e7693a36 GH |
3646 | } |
3647 | } | |
3648 | ||
aaee1203 PZ |
3649 | return idlest; |
3650 | } | |
e7693a36 | 3651 | |
a50bde51 PZ |
3652 | /* |
3653 | * Try and locate an idle CPU in the sched_domain. | |
3654 | */ | |
99bd5e2f | 3655 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 3656 | { |
99bd5e2f | 3657 | struct sched_domain *sd; |
37407ea7 | 3658 | struct sched_group *sg; |
e0a79f52 | 3659 | int i = task_cpu(p); |
a50bde51 | 3660 | |
e0a79f52 MG |
3661 | if (idle_cpu(target)) |
3662 | return target; | |
99bd5e2f SS |
3663 | |
3664 | /* | |
e0a79f52 | 3665 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 3666 | */ |
e0a79f52 MG |
3667 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
3668 | return i; | |
a50bde51 PZ |
3669 | |
3670 | /* | |
37407ea7 | 3671 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 3672 | */ |
518cd623 | 3673 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 3674 | for_each_lower_domain(sd) { |
37407ea7 LT |
3675 | sg = sd->groups; |
3676 | do { | |
3677 | if (!cpumask_intersects(sched_group_cpus(sg), | |
3678 | tsk_cpus_allowed(p))) | |
3679 | goto next; | |
3680 | ||
3681 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 3682 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
3683 | goto next; |
3684 | } | |
970e1789 | 3685 | |
37407ea7 LT |
3686 | target = cpumask_first_and(sched_group_cpus(sg), |
3687 | tsk_cpus_allowed(p)); | |
3688 | goto done; | |
3689 | next: | |
3690 | sg = sg->next; | |
3691 | } while (sg != sd->groups); | |
3692 | } | |
3693 | done: | |
a50bde51 PZ |
3694 | return target; |
3695 | } | |
3696 | ||
aaee1203 PZ |
3697 | /* |
3698 | * sched_balance_self: balance the current task (running on cpu) in domains | |
3699 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
3700 | * SD_BALANCE_EXEC. | |
3701 | * | |
3702 | * Balance, ie. select the least loaded group. | |
3703 | * | |
3704 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
3705 | * | |
3706 | * preempt must be disabled. | |
3707 | */ | |
0017d735 | 3708 | static int |
7608dec2 | 3709 | select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) |
aaee1203 | 3710 | { |
29cd8bae | 3711 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 PZ |
3712 | int cpu = smp_processor_id(); |
3713 | int prev_cpu = task_cpu(p); | |
3714 | int new_cpu = cpu; | |
99bd5e2f | 3715 | int want_affine = 0; |
5158f4e4 | 3716 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 3717 | |
29baa747 | 3718 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
3719 | return prev_cpu; |
3720 | ||
0763a660 | 3721 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 3722 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
3723 | want_affine = 1; |
3724 | new_cpu = prev_cpu; | |
3725 | } | |
aaee1203 | 3726 | |
dce840a0 | 3727 | rcu_read_lock(); |
aaee1203 | 3728 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
3729 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
3730 | continue; | |
3731 | ||
fe3bcfe1 | 3732 | /* |
99bd5e2f SS |
3733 | * If both cpu and prev_cpu are part of this domain, |
3734 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 3735 | */ |
99bd5e2f SS |
3736 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
3737 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
3738 | affine_sd = tmp; | |
29cd8bae | 3739 | break; |
f03542a7 | 3740 | } |
29cd8bae | 3741 | |
f03542a7 | 3742 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
3743 | sd = tmp; |
3744 | } | |
3745 | ||
8b911acd | 3746 | if (affine_sd) { |
f03542a7 | 3747 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
3748 | prev_cpu = cpu; |
3749 | ||
3750 | new_cpu = select_idle_sibling(p, prev_cpu); | |
3751 | goto unlock; | |
8b911acd | 3752 | } |
e7693a36 | 3753 | |
aaee1203 | 3754 | while (sd) { |
5158f4e4 | 3755 | int load_idx = sd->forkexec_idx; |
aaee1203 | 3756 | struct sched_group *group; |
c88d5910 | 3757 | int weight; |
098fb9db | 3758 | |
0763a660 | 3759 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
3760 | sd = sd->child; |
3761 | continue; | |
3762 | } | |
098fb9db | 3763 | |
5158f4e4 PZ |
3764 | if (sd_flag & SD_BALANCE_WAKE) |
3765 | load_idx = sd->wake_idx; | |
098fb9db | 3766 | |
5158f4e4 | 3767 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
3768 | if (!group) { |
3769 | sd = sd->child; | |
3770 | continue; | |
3771 | } | |
4ae7d5ce | 3772 | |
d7c33c49 | 3773 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
3774 | if (new_cpu == -1 || new_cpu == cpu) { |
3775 | /* Now try balancing at a lower domain level of cpu */ | |
3776 | sd = sd->child; | |
3777 | continue; | |
e7693a36 | 3778 | } |
aaee1203 PZ |
3779 | |
3780 | /* Now try balancing at a lower domain level of new_cpu */ | |
3781 | cpu = new_cpu; | |
669c55e9 | 3782 | weight = sd->span_weight; |
aaee1203 PZ |
3783 | sd = NULL; |
3784 | for_each_domain(cpu, tmp) { | |
669c55e9 | 3785 | if (weight <= tmp->span_weight) |
aaee1203 | 3786 | break; |
0763a660 | 3787 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
3788 | sd = tmp; |
3789 | } | |
3790 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 3791 | } |
dce840a0 PZ |
3792 | unlock: |
3793 | rcu_read_unlock(); | |
e7693a36 | 3794 | |
c88d5910 | 3795 | return new_cpu; |
e7693a36 | 3796 | } |
0a74bef8 PT |
3797 | |
3798 | /* | |
3799 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
3800 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
3801 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
3802 | * other assumptions, including the state of rq->lock, should be made. | |
3803 | */ | |
3804 | static void | |
3805 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
3806 | { | |
aff3e498 PT |
3807 | struct sched_entity *se = &p->se; |
3808 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3809 | ||
3810 | /* | |
3811 | * Load tracking: accumulate removed load so that it can be processed | |
3812 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
3813 | * to blocked load iff they have a positive decay-count. It can never | |
3814 | * be negative here since on-rq tasks have decay-count == 0. | |
3815 | */ | |
3816 | if (se->avg.decay_count) { | |
3817 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
3818 | atomic_long_add(se->avg.load_avg_contrib, |
3819 | &cfs_rq->removed_load); | |
aff3e498 | 3820 | } |
0a74bef8 | 3821 | } |
e7693a36 GH |
3822 | #endif /* CONFIG_SMP */ |
3823 | ||
e52fb7c0 PZ |
3824 | static unsigned long |
3825 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
3826 | { |
3827 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
3828 | ||
3829 | /* | |
e52fb7c0 PZ |
3830 | * Since its curr running now, convert the gran from real-time |
3831 | * to virtual-time in his units. | |
13814d42 MG |
3832 | * |
3833 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
3834 | * they get preempted easier. That is, if 'se' < 'curr' then | |
3835 | * the resulting gran will be larger, therefore penalizing the | |
3836 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
3837 | * be smaller, again penalizing the lighter task. | |
3838 | * | |
3839 | * This is especially important for buddies when the leftmost | |
3840 | * task is higher priority than the buddy. | |
0bbd3336 | 3841 | */ |
f4ad9bd2 | 3842 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
3843 | } |
3844 | ||
464b7527 PZ |
3845 | /* |
3846 | * Should 'se' preempt 'curr'. | |
3847 | * | |
3848 | * |s1 | |
3849 | * |s2 | |
3850 | * |s3 | |
3851 | * g | |
3852 | * |<--->|c | |
3853 | * | |
3854 | * w(c, s1) = -1 | |
3855 | * w(c, s2) = 0 | |
3856 | * w(c, s3) = 1 | |
3857 | * | |
3858 | */ | |
3859 | static int | |
3860 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
3861 | { | |
3862 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
3863 | ||
3864 | if (vdiff <= 0) | |
3865 | return -1; | |
3866 | ||
e52fb7c0 | 3867 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
3868 | if (vdiff > gran) |
3869 | return 1; | |
3870 | ||
3871 | return 0; | |
3872 | } | |
3873 | ||
02479099 PZ |
3874 | static void set_last_buddy(struct sched_entity *se) |
3875 | { | |
69c80f3e VP |
3876 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3877 | return; | |
3878 | ||
3879 | for_each_sched_entity(se) | |
3880 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
3881 | } |
3882 | ||
3883 | static void set_next_buddy(struct sched_entity *se) | |
3884 | { | |
69c80f3e VP |
3885 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
3886 | return; | |
3887 | ||
3888 | for_each_sched_entity(se) | |
3889 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
3890 | } |
3891 | ||
ac53db59 RR |
3892 | static void set_skip_buddy(struct sched_entity *se) |
3893 | { | |
69c80f3e VP |
3894 | for_each_sched_entity(se) |
3895 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
3896 | } |
3897 | ||
bf0f6f24 IM |
3898 | /* |
3899 | * Preempt the current task with a newly woken task if needed: | |
3900 | */ | |
5a9b86f6 | 3901 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
3902 | { |
3903 | struct task_struct *curr = rq->curr; | |
8651a86c | 3904 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 3905 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 3906 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 3907 | int next_buddy_marked = 0; |
bf0f6f24 | 3908 | |
4ae7d5ce IM |
3909 | if (unlikely(se == pse)) |
3910 | return; | |
3911 | ||
5238cdd3 | 3912 | /* |
ddcdf6e7 | 3913 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
3914 | * unconditionally check_prempt_curr() after an enqueue (which may have |
3915 | * lead to a throttle). This both saves work and prevents false | |
3916 | * next-buddy nomination below. | |
3917 | */ | |
3918 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
3919 | return; | |
3920 | ||
2f36825b | 3921 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 3922 | set_next_buddy(pse); |
2f36825b VP |
3923 | next_buddy_marked = 1; |
3924 | } | |
57fdc26d | 3925 | |
aec0a514 BR |
3926 | /* |
3927 | * We can come here with TIF_NEED_RESCHED already set from new task | |
3928 | * wake up path. | |
5238cdd3 PT |
3929 | * |
3930 | * Note: this also catches the edge-case of curr being in a throttled | |
3931 | * group (e.g. via set_curr_task), since update_curr() (in the | |
3932 | * enqueue of curr) will have resulted in resched being set. This | |
3933 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
3934 | * below. | |
aec0a514 BR |
3935 | */ |
3936 | if (test_tsk_need_resched(curr)) | |
3937 | return; | |
3938 | ||
a2f5c9ab DH |
3939 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
3940 | if (unlikely(curr->policy == SCHED_IDLE) && | |
3941 | likely(p->policy != SCHED_IDLE)) | |
3942 | goto preempt; | |
3943 | ||
91c234b4 | 3944 | /* |
a2f5c9ab DH |
3945 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
3946 | * is driven by the tick): | |
91c234b4 | 3947 | */ |
8ed92e51 | 3948 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 3949 | return; |
bf0f6f24 | 3950 | |
464b7527 | 3951 | find_matching_se(&se, &pse); |
9bbd7374 | 3952 | update_curr(cfs_rq_of(se)); |
002f128b | 3953 | BUG_ON(!pse); |
2f36825b VP |
3954 | if (wakeup_preempt_entity(se, pse) == 1) { |
3955 | /* | |
3956 | * Bias pick_next to pick the sched entity that is | |
3957 | * triggering this preemption. | |
3958 | */ | |
3959 | if (!next_buddy_marked) | |
3960 | set_next_buddy(pse); | |
3a7e73a2 | 3961 | goto preempt; |
2f36825b | 3962 | } |
464b7527 | 3963 | |
3a7e73a2 | 3964 | return; |
a65ac745 | 3965 | |
3a7e73a2 PZ |
3966 | preempt: |
3967 | resched_task(curr); | |
3968 | /* | |
3969 | * Only set the backward buddy when the current task is still | |
3970 | * on the rq. This can happen when a wakeup gets interleaved | |
3971 | * with schedule on the ->pre_schedule() or idle_balance() | |
3972 | * point, either of which can * drop the rq lock. | |
3973 | * | |
3974 | * Also, during early boot the idle thread is in the fair class, | |
3975 | * for obvious reasons its a bad idea to schedule back to it. | |
3976 | */ | |
3977 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
3978 | return; | |
3979 | ||
3980 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
3981 | set_last_buddy(se); | |
bf0f6f24 IM |
3982 | } |
3983 | ||
fb8d4724 | 3984 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 3985 | { |
8f4d37ec | 3986 | struct task_struct *p; |
bf0f6f24 IM |
3987 | struct cfs_rq *cfs_rq = &rq->cfs; |
3988 | struct sched_entity *se; | |
3989 | ||
36ace27e | 3990 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
3991 | return NULL; |
3992 | ||
3993 | do { | |
9948f4b2 | 3994 | se = pick_next_entity(cfs_rq); |
f4b6755f | 3995 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
3996 | cfs_rq = group_cfs_rq(se); |
3997 | } while (cfs_rq); | |
3998 | ||
8f4d37ec | 3999 | p = task_of(se); |
b39e66ea MG |
4000 | if (hrtick_enabled(rq)) |
4001 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4002 | |
4003 | return p; | |
bf0f6f24 IM |
4004 | } |
4005 | ||
4006 | /* | |
4007 | * Account for a descheduled task: | |
4008 | */ | |
31ee529c | 4009 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4010 | { |
4011 | struct sched_entity *se = &prev->se; | |
4012 | struct cfs_rq *cfs_rq; | |
4013 | ||
4014 | for_each_sched_entity(se) { | |
4015 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4016 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4017 | } |
4018 | } | |
4019 | ||
ac53db59 RR |
4020 | /* |
4021 | * sched_yield() is very simple | |
4022 | * | |
4023 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4024 | */ | |
4025 | static void yield_task_fair(struct rq *rq) | |
4026 | { | |
4027 | struct task_struct *curr = rq->curr; | |
4028 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4029 | struct sched_entity *se = &curr->se; | |
4030 | ||
4031 | /* | |
4032 | * Are we the only task in the tree? | |
4033 | */ | |
4034 | if (unlikely(rq->nr_running == 1)) | |
4035 | return; | |
4036 | ||
4037 | clear_buddies(cfs_rq, se); | |
4038 | ||
4039 | if (curr->policy != SCHED_BATCH) { | |
4040 | update_rq_clock(rq); | |
4041 | /* | |
4042 | * Update run-time statistics of the 'current'. | |
4043 | */ | |
4044 | update_curr(cfs_rq); | |
916671c0 MG |
4045 | /* |
4046 | * Tell update_rq_clock() that we've just updated, | |
4047 | * so we don't do microscopic update in schedule() | |
4048 | * and double the fastpath cost. | |
4049 | */ | |
4050 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4051 | } |
4052 | ||
4053 | set_skip_buddy(se); | |
4054 | } | |
4055 | ||
d95f4122 MG |
4056 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4057 | { | |
4058 | struct sched_entity *se = &p->se; | |
4059 | ||
5238cdd3 PT |
4060 | /* throttled hierarchies are not runnable */ |
4061 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4062 | return false; |
4063 | ||
4064 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4065 | set_next_buddy(se); | |
4066 | ||
d95f4122 MG |
4067 | yield_task_fair(rq); |
4068 | ||
4069 | return true; | |
4070 | } | |
4071 | ||
681f3e68 | 4072 | #ifdef CONFIG_SMP |
bf0f6f24 | 4073 | /************************************************** |
e9c84cb8 PZ |
4074 | * Fair scheduling class load-balancing methods. |
4075 | * | |
4076 | * BASICS | |
4077 | * | |
4078 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4079 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4080 | * time to each task. This is expressed in the following equation: | |
4081 | * | |
4082 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4083 | * | |
4084 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4085 | * W_i,0 is defined as: | |
4086 | * | |
4087 | * W_i,0 = \Sum_j w_i,j (2) | |
4088 | * | |
4089 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4090 | * is derived from the nice value as per prio_to_weight[]. | |
4091 | * | |
4092 | * The weight average is an exponential decay average of the instantaneous | |
4093 | * weight: | |
4094 | * | |
4095 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4096 | * | |
4097 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4098 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4099 | * can also include other factors [XXX]. | |
4100 | * | |
4101 | * To achieve this balance we define a measure of imbalance which follows | |
4102 | * directly from (1): | |
4103 | * | |
4104 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4105 | * | |
4106 | * We them move tasks around to minimize the imbalance. In the continuous | |
4107 | * function space it is obvious this converges, in the discrete case we get | |
4108 | * a few fun cases generally called infeasible weight scenarios. | |
4109 | * | |
4110 | * [XXX expand on: | |
4111 | * - infeasible weights; | |
4112 | * - local vs global optima in the discrete case. ] | |
4113 | * | |
4114 | * | |
4115 | * SCHED DOMAINS | |
4116 | * | |
4117 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4118 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4119 | * topology where each level pairs two lower groups (or better). This results | |
4120 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4121 | * tree to only the first of the previous level and we decrease the frequency | |
4122 | * of load-balance at each level inv. proportional to the number of cpus in | |
4123 | * the groups. | |
4124 | * | |
4125 | * This yields: | |
4126 | * | |
4127 | * log_2 n 1 n | |
4128 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4129 | * i = 0 2^i 2^i | |
4130 | * `- size of each group | |
4131 | * | | `- number of cpus doing load-balance | |
4132 | * | `- freq | |
4133 | * `- sum over all levels | |
4134 | * | |
4135 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4136 | * this makes (5) the runtime complexity of the balancer. | |
4137 | * | |
4138 | * An important property here is that each CPU is still (indirectly) connected | |
4139 | * to every other cpu in at most O(log n) steps: | |
4140 | * | |
4141 | * The adjacency matrix of the resulting graph is given by: | |
4142 | * | |
4143 | * log_2 n | |
4144 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4145 | * k = 0 | |
4146 | * | |
4147 | * And you'll find that: | |
4148 | * | |
4149 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
4150 | * | |
4151 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
4152 | * The task movement gives a factor of O(m), giving a convergence complexity | |
4153 | * of: | |
4154 | * | |
4155 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
4156 | * | |
4157 | * | |
4158 | * WORK CONSERVING | |
4159 | * | |
4160 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
4161 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
4162 | * tree itself instead of relying on other CPUs to bring it work. | |
4163 | * | |
4164 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
4165 | * time. | |
4166 | * | |
4167 | * [XXX more?] | |
4168 | * | |
4169 | * | |
4170 | * CGROUPS | |
4171 | * | |
4172 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
4173 | * | |
4174 | * s_k,i | |
4175 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
4176 | * S_k | |
4177 | * | |
4178 | * Where | |
4179 | * | |
4180 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4181 | * | |
4182 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4183 | * | |
4184 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4185 | * property. | |
4186 | * | |
4187 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4188 | * rewrite all of this once again.] | |
4189 | */ | |
bf0f6f24 | 4190 | |
ed387b78 HS |
4191 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4192 | ||
ddcdf6e7 | 4193 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4194 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4195 | #define LBF_DST_PINNED 0x04 |
4196 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4197 | |
4198 | struct lb_env { | |
4199 | struct sched_domain *sd; | |
4200 | ||
ddcdf6e7 | 4201 | struct rq *src_rq; |
85c1e7da | 4202 | int src_cpu; |
ddcdf6e7 PZ |
4203 | |
4204 | int dst_cpu; | |
4205 | struct rq *dst_rq; | |
4206 | ||
88b8dac0 SV |
4207 | struct cpumask *dst_grpmask; |
4208 | int new_dst_cpu; | |
ddcdf6e7 | 4209 | enum cpu_idle_type idle; |
bd939f45 | 4210 | long imbalance; |
b9403130 MW |
4211 | /* The set of CPUs under consideration for load-balancing */ |
4212 | struct cpumask *cpus; | |
4213 | ||
ddcdf6e7 | 4214 | unsigned int flags; |
367456c7 PZ |
4215 | |
4216 | unsigned int loop; | |
4217 | unsigned int loop_break; | |
4218 | unsigned int loop_max; | |
ddcdf6e7 PZ |
4219 | }; |
4220 | ||
1e3c88bd | 4221 | /* |
ddcdf6e7 | 4222 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
4223 | * Both runqueues must be locked. |
4224 | */ | |
ddcdf6e7 | 4225 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 4226 | { |
ddcdf6e7 PZ |
4227 | deactivate_task(env->src_rq, p, 0); |
4228 | set_task_cpu(p, env->dst_cpu); | |
4229 | activate_task(env->dst_rq, p, 0); | |
4230 | check_preempt_curr(env->dst_rq, p, 0); | |
6fe6b2d6 RR |
4231 | #ifdef CONFIG_NUMA_BALANCING |
4232 | if (p->numa_preferred_nid != -1) { | |
4233 | int src_nid = cpu_to_node(env->src_cpu); | |
4234 | int dst_nid = cpu_to_node(env->dst_cpu); | |
4235 | ||
4236 | /* | |
4237 | * If the load balancer has moved the task then limit | |
4238 | * migrations from taking place in the short term in | |
4239 | * case this is a short-lived migration. | |
4240 | */ | |
4241 | if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid) | |
4242 | p->numa_migrate_seq = 0; | |
4243 | } | |
4244 | #endif | |
1e3c88bd PZ |
4245 | } |
4246 | ||
029632fb PZ |
4247 | /* |
4248 | * Is this task likely cache-hot: | |
4249 | */ | |
4250 | static int | |
4251 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
4252 | { | |
4253 | s64 delta; | |
4254 | ||
4255 | if (p->sched_class != &fair_sched_class) | |
4256 | return 0; | |
4257 | ||
4258 | if (unlikely(p->policy == SCHED_IDLE)) | |
4259 | return 0; | |
4260 | ||
4261 | /* | |
4262 | * Buddy candidates are cache hot: | |
4263 | */ | |
4264 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
4265 | (&p->se == cfs_rq_of(&p->se)->next || | |
4266 | &p->se == cfs_rq_of(&p->se)->last)) | |
4267 | return 1; | |
4268 | ||
4269 | if (sysctl_sched_migration_cost == -1) | |
4270 | return 1; | |
4271 | if (sysctl_sched_migration_cost == 0) | |
4272 | return 0; | |
4273 | ||
4274 | delta = now - p->se.exec_start; | |
4275 | ||
4276 | return delta < (s64)sysctl_sched_migration_cost; | |
4277 | } | |
4278 | ||
3a7053b3 MG |
4279 | #ifdef CONFIG_NUMA_BALANCING |
4280 | /* Returns true if the destination node has incurred more faults */ | |
4281 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
4282 | { | |
4283 | int src_nid, dst_nid; | |
4284 | ||
4285 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || | |
4286 | !(env->sd->flags & SD_NUMA)) { | |
4287 | return false; | |
4288 | } | |
4289 | ||
4290 | src_nid = cpu_to_node(env->src_cpu); | |
4291 | dst_nid = cpu_to_node(env->dst_cpu); | |
4292 | ||
4293 | if (src_nid == dst_nid || | |
4294 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4295 | return false; | |
4296 | ||
4297 | if (dst_nid == p->numa_preferred_nid || | |
ac8e895b | 4298 | task_faults(p, dst_nid) > task_faults(p, src_nid)) |
3a7053b3 MG |
4299 | return true; |
4300 | ||
4301 | return false; | |
4302 | } | |
7a0f3083 MG |
4303 | |
4304 | ||
4305 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
4306 | { | |
4307 | int src_nid, dst_nid; | |
4308 | ||
4309 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
4310 | return false; | |
4311 | ||
4312 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) | |
4313 | return false; | |
4314 | ||
4315 | src_nid = cpu_to_node(env->src_cpu); | |
4316 | dst_nid = cpu_to_node(env->dst_cpu); | |
4317 | ||
4318 | if (src_nid == dst_nid || | |
4319 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4320 | return false; | |
4321 | ||
ac8e895b | 4322 | if (task_faults(p, dst_nid) < task_faults(p, src_nid)) |
7a0f3083 MG |
4323 | return true; |
4324 | ||
4325 | return false; | |
4326 | } | |
4327 | ||
3a7053b3 MG |
4328 | #else |
4329 | static inline bool migrate_improves_locality(struct task_struct *p, | |
4330 | struct lb_env *env) | |
4331 | { | |
4332 | return false; | |
4333 | } | |
7a0f3083 MG |
4334 | |
4335 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
4336 | struct lb_env *env) | |
4337 | { | |
4338 | return false; | |
4339 | } | |
3a7053b3 MG |
4340 | #endif |
4341 | ||
1e3c88bd PZ |
4342 | /* |
4343 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
4344 | */ | |
4345 | static | |
8e45cb54 | 4346 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
4347 | { |
4348 | int tsk_cache_hot = 0; | |
4349 | /* | |
4350 | * We do not migrate tasks that are: | |
d3198084 | 4351 | * 1) throttled_lb_pair, or |
1e3c88bd | 4352 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
4353 | * 3) running (obviously), or |
4354 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 4355 | */ |
d3198084 JK |
4356 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
4357 | return 0; | |
4358 | ||
ddcdf6e7 | 4359 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 4360 | int cpu; |
88b8dac0 | 4361 | |
41acab88 | 4362 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 4363 | |
6263322c PZ |
4364 | env->flags |= LBF_SOME_PINNED; |
4365 | ||
88b8dac0 SV |
4366 | /* |
4367 | * Remember if this task can be migrated to any other cpu in | |
4368 | * our sched_group. We may want to revisit it if we couldn't | |
4369 | * meet load balance goals by pulling other tasks on src_cpu. | |
4370 | * | |
4371 | * Also avoid computing new_dst_cpu if we have already computed | |
4372 | * one in current iteration. | |
4373 | */ | |
6263322c | 4374 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
4375 | return 0; |
4376 | ||
e02e60c1 JK |
4377 | /* Prevent to re-select dst_cpu via env's cpus */ |
4378 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
4379 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 4380 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
4381 | env->new_dst_cpu = cpu; |
4382 | break; | |
4383 | } | |
88b8dac0 | 4384 | } |
e02e60c1 | 4385 | |
1e3c88bd PZ |
4386 | return 0; |
4387 | } | |
88b8dac0 SV |
4388 | |
4389 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 4390 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 4391 | |
ddcdf6e7 | 4392 | if (task_running(env->src_rq, p)) { |
41acab88 | 4393 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
4394 | return 0; |
4395 | } | |
4396 | ||
4397 | /* | |
4398 | * Aggressive migration if: | |
3a7053b3 MG |
4399 | * 1) destination numa is preferred |
4400 | * 2) task is cache cold, or | |
4401 | * 3) too many balance attempts have failed. | |
1e3c88bd | 4402 | */ |
78becc27 | 4403 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
7a0f3083 MG |
4404 | if (!tsk_cache_hot) |
4405 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
4406 | |
4407 | if (migrate_improves_locality(p, env)) { | |
4408 | #ifdef CONFIG_SCHEDSTATS | |
4409 | if (tsk_cache_hot) { | |
4410 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
4411 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
4412 | } | |
4413 | #endif | |
4414 | return 1; | |
4415 | } | |
4416 | ||
1e3c88bd | 4417 | if (!tsk_cache_hot || |
8e45cb54 | 4418 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 4419 | |
1e3c88bd | 4420 | if (tsk_cache_hot) { |
8e45cb54 | 4421 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 4422 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 4423 | } |
4e2dcb73 | 4424 | |
1e3c88bd PZ |
4425 | return 1; |
4426 | } | |
4427 | ||
4e2dcb73 ZH |
4428 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
4429 | return 0; | |
1e3c88bd PZ |
4430 | } |
4431 | ||
897c395f PZ |
4432 | /* |
4433 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
4434 | * part of active balancing operations within "domain". | |
4435 | * Returns 1 if successful and 0 otherwise. | |
4436 | * | |
4437 | * Called with both runqueues locked. | |
4438 | */ | |
8e45cb54 | 4439 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
4440 | { |
4441 | struct task_struct *p, *n; | |
897c395f | 4442 | |
367456c7 | 4443 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
4444 | if (!can_migrate_task(p, env)) |
4445 | continue; | |
897c395f | 4446 | |
367456c7 PZ |
4447 | move_task(p, env); |
4448 | /* | |
4449 | * Right now, this is only the second place move_task() | |
4450 | * is called, so we can safely collect move_task() | |
4451 | * stats here rather than inside move_task(). | |
4452 | */ | |
4453 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4454 | return 1; | |
897c395f | 4455 | } |
897c395f PZ |
4456 | return 0; |
4457 | } | |
4458 | ||
367456c7 PZ |
4459 | static unsigned long task_h_load(struct task_struct *p); |
4460 | ||
eb95308e PZ |
4461 | static const unsigned int sched_nr_migrate_break = 32; |
4462 | ||
5d6523eb | 4463 | /* |
bd939f45 | 4464 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4465 | * this_rq, as part of a balancing operation within domain "sd". |
4466 | * Returns 1 if successful and 0 otherwise. | |
4467 | * | |
4468 | * Called with both runqueues locked. | |
4469 | */ | |
4470 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4471 | { |
5d6523eb PZ |
4472 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4473 | struct task_struct *p; | |
367456c7 PZ |
4474 | unsigned long load; |
4475 | int pulled = 0; | |
1e3c88bd | 4476 | |
bd939f45 | 4477 | if (env->imbalance <= 0) |
5d6523eb | 4478 | return 0; |
1e3c88bd | 4479 | |
5d6523eb PZ |
4480 | while (!list_empty(tasks)) { |
4481 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 4482 | |
367456c7 PZ |
4483 | env->loop++; |
4484 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 4485 | if (env->loop > env->loop_max) |
367456c7 | 4486 | break; |
5d6523eb PZ |
4487 | |
4488 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 4489 | if (env->loop > env->loop_break) { |
eb95308e | 4490 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 4491 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 4492 | break; |
a195f004 | 4493 | } |
1e3c88bd | 4494 | |
d3198084 | 4495 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
4496 | goto next; |
4497 | ||
4498 | load = task_h_load(p); | |
5d6523eb | 4499 | |
eb95308e | 4500 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
4501 | goto next; |
4502 | ||
bd939f45 | 4503 | if ((load / 2) > env->imbalance) |
367456c7 | 4504 | goto next; |
1e3c88bd | 4505 | |
ddcdf6e7 | 4506 | move_task(p, env); |
ee00e66f | 4507 | pulled++; |
bd939f45 | 4508 | env->imbalance -= load; |
1e3c88bd PZ |
4509 | |
4510 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
4511 | /* |
4512 | * NEWIDLE balancing is a source of latency, so preemptible | |
4513 | * kernels will stop after the first task is pulled to minimize | |
4514 | * the critical section. | |
4515 | */ | |
5d6523eb | 4516 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 4517 | break; |
1e3c88bd PZ |
4518 | #endif |
4519 | ||
ee00e66f PZ |
4520 | /* |
4521 | * We only want to steal up to the prescribed amount of | |
4522 | * weighted load. | |
4523 | */ | |
bd939f45 | 4524 | if (env->imbalance <= 0) |
ee00e66f | 4525 | break; |
367456c7 PZ |
4526 | |
4527 | continue; | |
4528 | next: | |
5d6523eb | 4529 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 4530 | } |
5d6523eb | 4531 | |
1e3c88bd | 4532 | /* |
ddcdf6e7 PZ |
4533 | * Right now, this is one of only two places move_task() is called, |
4534 | * so we can safely collect move_task() stats here rather than | |
4535 | * inside move_task(). | |
1e3c88bd | 4536 | */ |
8e45cb54 | 4537 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 4538 | |
5d6523eb | 4539 | return pulled; |
1e3c88bd PZ |
4540 | } |
4541 | ||
230059de | 4542 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
4543 | /* |
4544 | * update tg->load_weight by folding this cpu's load_avg | |
4545 | */ | |
48a16753 | 4546 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 4547 | { |
48a16753 PT |
4548 | struct sched_entity *se = tg->se[cpu]; |
4549 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 4550 | |
48a16753 PT |
4551 | /* throttled entities do not contribute to load */ |
4552 | if (throttled_hierarchy(cfs_rq)) | |
4553 | return; | |
9e3081ca | 4554 | |
aff3e498 | 4555 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 4556 | |
82958366 PT |
4557 | if (se) { |
4558 | update_entity_load_avg(se, 1); | |
4559 | /* | |
4560 | * We pivot on our runnable average having decayed to zero for | |
4561 | * list removal. This generally implies that all our children | |
4562 | * have also been removed (modulo rounding error or bandwidth | |
4563 | * control); however, such cases are rare and we can fix these | |
4564 | * at enqueue. | |
4565 | * | |
4566 | * TODO: fix up out-of-order children on enqueue. | |
4567 | */ | |
4568 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
4569 | list_del_leaf_cfs_rq(cfs_rq); | |
4570 | } else { | |
48a16753 | 4571 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
4572 | update_rq_runnable_avg(rq, rq->nr_running); |
4573 | } | |
9e3081ca PZ |
4574 | } |
4575 | ||
48a16753 | 4576 | static void update_blocked_averages(int cpu) |
9e3081ca | 4577 | { |
9e3081ca | 4578 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
4579 | struct cfs_rq *cfs_rq; |
4580 | unsigned long flags; | |
9e3081ca | 4581 | |
48a16753 PT |
4582 | raw_spin_lock_irqsave(&rq->lock, flags); |
4583 | update_rq_clock(rq); | |
9763b67f PZ |
4584 | /* |
4585 | * Iterates the task_group tree in a bottom up fashion, see | |
4586 | * list_add_leaf_cfs_rq() for details. | |
4587 | */ | |
64660c86 | 4588 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
4589 | /* |
4590 | * Note: We may want to consider periodically releasing | |
4591 | * rq->lock about these updates so that creating many task | |
4592 | * groups does not result in continually extending hold time. | |
4593 | */ | |
4594 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 4595 | } |
48a16753 PT |
4596 | |
4597 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
4598 | } |
4599 | ||
9763b67f | 4600 | /* |
68520796 | 4601 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
4602 | * This needs to be done in a top-down fashion because the load of a child |
4603 | * group is a fraction of its parents load. | |
4604 | */ | |
68520796 | 4605 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 4606 | { |
68520796 VD |
4607 | struct rq *rq = rq_of(cfs_rq); |
4608 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 4609 | unsigned long now = jiffies; |
68520796 | 4610 | unsigned long load; |
a35b6466 | 4611 | |
68520796 | 4612 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
4613 | return; |
4614 | ||
68520796 VD |
4615 | cfs_rq->h_load_next = NULL; |
4616 | for_each_sched_entity(se) { | |
4617 | cfs_rq = cfs_rq_of(se); | |
4618 | cfs_rq->h_load_next = se; | |
4619 | if (cfs_rq->last_h_load_update == now) | |
4620 | break; | |
4621 | } | |
a35b6466 | 4622 | |
68520796 | 4623 | if (!se) { |
7e3115ef | 4624 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
4625 | cfs_rq->last_h_load_update = now; |
4626 | } | |
4627 | ||
4628 | while ((se = cfs_rq->h_load_next) != NULL) { | |
4629 | load = cfs_rq->h_load; | |
4630 | load = div64_ul(load * se->avg.load_avg_contrib, | |
4631 | cfs_rq->runnable_load_avg + 1); | |
4632 | cfs_rq = group_cfs_rq(se); | |
4633 | cfs_rq->h_load = load; | |
4634 | cfs_rq->last_h_load_update = now; | |
4635 | } | |
9763b67f PZ |
4636 | } |
4637 | ||
367456c7 | 4638 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 4639 | { |
367456c7 | 4640 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 4641 | |
68520796 | 4642 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
4643 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
4644 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
4645 | } |
4646 | #else | |
48a16753 | 4647 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
4648 | { |
4649 | } | |
4650 | ||
367456c7 | 4651 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 4652 | { |
a003a25b | 4653 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 4654 | } |
230059de | 4655 | #endif |
1e3c88bd | 4656 | |
1e3c88bd | 4657 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
4658 | /* |
4659 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
4660 | */ | |
4661 | struct sg_lb_stats { | |
4662 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
4663 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 4664 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 4665 | unsigned long load_per_task; |
3ae11c90 | 4666 | unsigned long group_power; |
147c5fc2 PZ |
4667 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
4668 | unsigned int group_capacity; | |
4669 | unsigned int idle_cpus; | |
4670 | unsigned int group_weight; | |
1e3c88bd | 4671 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 4672 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
4673 | }; |
4674 | ||
56cf515b JK |
4675 | /* |
4676 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
4677 | * during load balancing. | |
4678 | */ | |
4679 | struct sd_lb_stats { | |
4680 | struct sched_group *busiest; /* Busiest group in this sd */ | |
4681 | struct sched_group *local; /* Local group in this sd */ | |
4682 | unsigned long total_load; /* Total load of all groups in sd */ | |
4683 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
4684 | unsigned long avg_load; /* Average load across all groups in sd */ | |
4685 | ||
56cf515b | 4686 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 4687 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
4688 | }; |
4689 | ||
147c5fc2 PZ |
4690 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
4691 | { | |
4692 | /* | |
4693 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
4694 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
4695 | * We must however clear busiest_stat::avg_load because | |
4696 | * update_sd_pick_busiest() reads this before assignment. | |
4697 | */ | |
4698 | *sds = (struct sd_lb_stats){ | |
4699 | .busiest = NULL, | |
4700 | .local = NULL, | |
4701 | .total_load = 0UL, | |
4702 | .total_pwr = 0UL, | |
4703 | .busiest_stat = { | |
4704 | .avg_load = 0UL, | |
4705 | }, | |
4706 | }; | |
4707 | } | |
4708 | ||
1e3c88bd PZ |
4709 | /** |
4710 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
4711 | * @sd: The sched_domain whose load_idx is to be obtained. | |
4712 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
e69f6186 YB |
4713 | * |
4714 | * Return: The load index. | |
1e3c88bd PZ |
4715 | */ |
4716 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
4717 | enum cpu_idle_type idle) | |
4718 | { | |
4719 | int load_idx; | |
4720 | ||
4721 | switch (idle) { | |
4722 | case CPU_NOT_IDLE: | |
4723 | load_idx = sd->busy_idx; | |
4724 | break; | |
4725 | ||
4726 | case CPU_NEWLY_IDLE: | |
4727 | load_idx = sd->newidle_idx; | |
4728 | break; | |
4729 | default: | |
4730 | load_idx = sd->idle_idx; | |
4731 | break; | |
4732 | } | |
4733 | ||
4734 | return load_idx; | |
4735 | } | |
4736 | ||
15f803c9 | 4737 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 4738 | { |
1399fa78 | 4739 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
4740 | } |
4741 | ||
4742 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
4743 | { | |
4744 | return default_scale_freq_power(sd, cpu); | |
4745 | } | |
4746 | ||
15f803c9 | 4747 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 4748 | { |
669c55e9 | 4749 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
4750 | unsigned long smt_gain = sd->smt_gain; |
4751 | ||
4752 | smt_gain /= weight; | |
4753 | ||
4754 | return smt_gain; | |
4755 | } | |
4756 | ||
4757 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
4758 | { | |
4759 | return default_scale_smt_power(sd, cpu); | |
4760 | } | |
4761 | ||
15f803c9 | 4762 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
4763 | { |
4764 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 4765 | u64 total, available, age_stamp, avg; |
1e3c88bd | 4766 | |
b654f7de PZ |
4767 | /* |
4768 | * Since we're reading these variables without serialization make sure | |
4769 | * we read them once before doing sanity checks on them. | |
4770 | */ | |
4771 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
4772 | avg = ACCESS_ONCE(rq->rt_avg); | |
4773 | ||
78becc27 | 4774 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 4775 | |
b654f7de | 4776 | if (unlikely(total < avg)) { |
aa483808 VP |
4777 | /* Ensures that power won't end up being negative */ |
4778 | available = 0; | |
4779 | } else { | |
b654f7de | 4780 | available = total - avg; |
aa483808 | 4781 | } |
1e3c88bd | 4782 | |
1399fa78 NR |
4783 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
4784 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 4785 | |
1399fa78 | 4786 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4787 | |
4788 | return div_u64(available, total); | |
4789 | } | |
4790 | ||
4791 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
4792 | { | |
669c55e9 | 4793 | unsigned long weight = sd->span_weight; |
1399fa78 | 4794 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
4795 | struct sched_group *sdg = sd->groups; |
4796 | ||
1e3c88bd PZ |
4797 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
4798 | if (sched_feat(ARCH_POWER)) | |
4799 | power *= arch_scale_smt_power(sd, cpu); | |
4800 | else | |
4801 | power *= default_scale_smt_power(sd, cpu); | |
4802 | ||
1399fa78 | 4803 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4804 | } |
4805 | ||
9c3f75cb | 4806 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
4807 | |
4808 | if (sched_feat(ARCH_POWER)) | |
4809 | power *= arch_scale_freq_power(sd, cpu); | |
4810 | else | |
4811 | power *= default_scale_freq_power(sd, cpu); | |
4812 | ||
1399fa78 | 4813 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 4814 | |
1e3c88bd | 4815 | power *= scale_rt_power(cpu); |
1399fa78 | 4816 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
4817 | |
4818 | if (!power) | |
4819 | power = 1; | |
4820 | ||
e51fd5e2 | 4821 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 4822 | sdg->sgp->power = power; |
1e3c88bd PZ |
4823 | } |
4824 | ||
029632fb | 4825 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
4826 | { |
4827 | struct sched_domain *child = sd->child; | |
4828 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 4829 | unsigned long power, power_orig; |
4ec4412e VG |
4830 | unsigned long interval; |
4831 | ||
4832 | interval = msecs_to_jiffies(sd->balance_interval); | |
4833 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
4834 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
4835 | |
4836 | if (!child) { | |
4837 | update_cpu_power(sd, cpu); | |
4838 | return; | |
4839 | } | |
4840 | ||
863bffc8 | 4841 | power_orig = power = 0; |
1e3c88bd | 4842 | |
74a5ce20 PZ |
4843 | if (child->flags & SD_OVERLAP) { |
4844 | /* | |
4845 | * SD_OVERLAP domains cannot assume that child groups | |
4846 | * span the current group. | |
4847 | */ | |
4848 | ||
863bffc8 PZ |
4849 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
4850 | struct sched_group *sg = cpu_rq(cpu)->sd->groups; | |
4851 | ||
4852 | power_orig += sg->sgp->power_orig; | |
4853 | power += sg->sgp->power; | |
4854 | } | |
74a5ce20 PZ |
4855 | } else { |
4856 | /* | |
4857 | * !SD_OVERLAP domains can assume that child groups | |
4858 | * span the current group. | |
4859 | */ | |
4860 | ||
4861 | group = child->groups; | |
4862 | do { | |
863bffc8 | 4863 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
4864 | power += group->sgp->power; |
4865 | group = group->next; | |
4866 | } while (group != child->groups); | |
4867 | } | |
1e3c88bd | 4868 | |
863bffc8 PZ |
4869 | sdg->sgp->power_orig = power_orig; |
4870 | sdg->sgp->power = power; | |
1e3c88bd PZ |
4871 | } |
4872 | ||
9d5efe05 SV |
4873 | /* |
4874 | * Try and fix up capacity for tiny siblings, this is needed when | |
4875 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
4876 | * which on its own isn't powerful enough. | |
4877 | * | |
4878 | * See update_sd_pick_busiest() and check_asym_packing(). | |
4879 | */ | |
4880 | static inline int | |
4881 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
4882 | { | |
4883 | /* | |
1399fa78 | 4884 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 4885 | */ |
a6c75f2f | 4886 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
4887 | return 0; |
4888 | ||
4889 | /* | |
4890 | * If ~90% of the cpu_power is still there, we're good. | |
4891 | */ | |
9c3f75cb | 4892 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
4893 | return 1; |
4894 | ||
4895 | return 0; | |
4896 | } | |
4897 | ||
30ce5dab PZ |
4898 | /* |
4899 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
4900 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
4901 | * | |
4902 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
4903 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
4904 | * Something like: | |
4905 | * | |
4906 | * { 0 1 2 3 } { 4 5 6 7 } | |
4907 | * * * * * | |
4908 | * | |
4909 | * If we were to balance group-wise we'd place two tasks in the first group and | |
4910 | * two tasks in the second group. Clearly this is undesired as it will overload | |
4911 | * cpu 3 and leave one of the cpus in the second group unused. | |
4912 | * | |
4913 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
4914 | * by noticing the lower domain failed to reach balance and had difficulty |
4915 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
4916 | * |
4917 | * When this is so detected; this group becomes a candidate for busiest; see | |
4918 | * update_sd_pick_busiest(). And calculcate_imbalance() and | |
6263322c | 4919 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
4920 | * to create an effective group imbalance. |
4921 | * | |
4922 | * This is a somewhat tricky proposition since the next run might not find the | |
4923 | * group imbalance and decide the groups need to be balanced again. A most | |
4924 | * subtle and fragile situation. | |
4925 | */ | |
4926 | ||
6263322c | 4927 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 4928 | { |
6263322c | 4929 | return group->sgp->imbalance; |
30ce5dab PZ |
4930 | } |
4931 | ||
b37d9316 PZ |
4932 | /* |
4933 | * Compute the group capacity. | |
4934 | * | |
c61037e9 PZ |
4935 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
4936 | * first dividing out the smt factor and computing the actual number of cores | |
4937 | * and limit power unit capacity with that. | |
b37d9316 PZ |
4938 | */ |
4939 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
4940 | { | |
c61037e9 PZ |
4941 | unsigned int capacity, smt, cpus; |
4942 | unsigned int power, power_orig; | |
4943 | ||
4944 | power = group->sgp->power; | |
4945 | power_orig = group->sgp->power_orig; | |
4946 | cpus = group->group_weight; | |
b37d9316 | 4947 | |
c61037e9 PZ |
4948 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
4949 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
4950 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 4951 | |
c61037e9 | 4952 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
4953 | if (!capacity) |
4954 | capacity = fix_small_capacity(env->sd, group); | |
4955 | ||
4956 | return capacity; | |
4957 | } | |
4958 | ||
1e3c88bd PZ |
4959 | /** |
4960 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 4961 | * @env: The load balancing environment. |
1e3c88bd | 4962 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 4963 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 4964 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
4965 | * @sgs: variable to hold the statistics for this group. |
4966 | */ | |
bd939f45 PZ |
4967 | static inline void update_sg_lb_stats(struct lb_env *env, |
4968 | struct sched_group *group, int load_idx, | |
23f0d209 | 4969 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 4970 | { |
30ce5dab PZ |
4971 | unsigned long nr_running; |
4972 | unsigned long load; | |
bd939f45 | 4973 | int i; |
1e3c88bd | 4974 | |
b72ff13c PZ |
4975 | memset(sgs, 0, sizeof(*sgs)); |
4976 | ||
b9403130 | 4977 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
4978 | struct rq *rq = cpu_rq(i); |
4979 | ||
e44bc5c5 PZ |
4980 | nr_running = rq->nr_running; |
4981 | ||
1e3c88bd | 4982 | /* Bias balancing toward cpus of our domain */ |
6263322c | 4983 | if (local_group) |
04f733b4 | 4984 | load = target_load(i, load_idx); |
6263322c | 4985 | else |
1e3c88bd | 4986 | load = source_load(i, load_idx); |
1e3c88bd PZ |
4987 | |
4988 | sgs->group_load += load; | |
e44bc5c5 | 4989 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 4990 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
4991 | if (idle_cpu(i)) |
4992 | sgs->idle_cpus++; | |
1e3c88bd PZ |
4993 | } |
4994 | ||
1e3c88bd | 4995 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
4996 | sgs->group_power = group->sgp->power; |
4997 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 4998 | |
dd5feea1 | 4999 | if (sgs->sum_nr_running) |
38d0f770 | 5000 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5001 | |
aae6d3dd | 5002 | sgs->group_weight = group->group_weight; |
fab47622 | 5003 | |
b37d9316 PZ |
5004 | sgs->group_imb = sg_imbalanced(group); |
5005 | sgs->group_capacity = sg_capacity(env, group); | |
5006 | ||
fab47622 NR |
5007 | if (sgs->group_capacity > sgs->sum_nr_running) |
5008 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5009 | } |
5010 | ||
532cb4c4 MN |
5011 | /** |
5012 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5013 | * @env: The load balancing environment. |
532cb4c4 MN |
5014 | * @sds: sched_domain statistics |
5015 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5016 | * @sgs: sched_group statistics |
532cb4c4 MN |
5017 | * |
5018 | * Determine if @sg is a busier group than the previously selected | |
5019 | * busiest group. | |
e69f6186 YB |
5020 | * |
5021 | * Return: %true if @sg is a busier group than the previously selected | |
5022 | * busiest group. %false otherwise. | |
532cb4c4 | 5023 | */ |
bd939f45 | 5024 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5025 | struct sd_lb_stats *sds, |
5026 | struct sched_group *sg, | |
bd939f45 | 5027 | struct sg_lb_stats *sgs) |
532cb4c4 | 5028 | { |
56cf515b | 5029 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5030 | return false; |
5031 | ||
5032 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5033 | return true; | |
5034 | ||
5035 | if (sgs->group_imb) | |
5036 | return true; | |
5037 | ||
5038 | /* | |
5039 | * ASYM_PACKING needs to move all the work to the lowest | |
5040 | * numbered CPUs in the group, therefore mark all groups | |
5041 | * higher than ourself as busy. | |
5042 | */ | |
bd939f45 PZ |
5043 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5044 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5045 | if (!sds->busiest) |
5046 | return true; | |
5047 | ||
5048 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5049 | return true; | |
5050 | } | |
5051 | ||
5052 | return false; | |
5053 | } | |
5054 | ||
1e3c88bd | 5055 | /** |
461819ac | 5056 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5057 | * @env: The load balancing environment. |
1e3c88bd PZ |
5058 | * @balance: Should we balance. |
5059 | * @sds: variable to hold the statistics for this sched_domain. | |
5060 | */ | |
bd939f45 | 5061 | static inline void update_sd_lb_stats(struct lb_env *env, |
23f0d209 | 5062 | struct sd_lb_stats *sds) |
1e3c88bd | 5063 | { |
bd939f45 PZ |
5064 | struct sched_domain *child = env->sd->child; |
5065 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5066 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5067 | int load_idx, prefer_sibling = 0; |
5068 | ||
5069 | if (child && child->flags & SD_PREFER_SIBLING) | |
5070 | prefer_sibling = 1; | |
5071 | ||
bd939f45 | 5072 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5073 | |
5074 | do { | |
56cf515b | 5075 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5076 | int local_group; |
5077 | ||
bd939f45 | 5078 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5079 | if (local_group) { |
5080 | sds->local = sg; | |
5081 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5082 | |
5083 | if (env->idle != CPU_NEWLY_IDLE || | |
5084 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5085 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5086 | } |
1e3c88bd | 5087 | |
56cf515b | 5088 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5089 | |
b72ff13c PZ |
5090 | if (local_group) |
5091 | goto next_group; | |
5092 | ||
1e3c88bd PZ |
5093 | /* |
5094 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 5095 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
5096 | * and move all the excess tasks away. We lower the capacity |
5097 | * of a group only if the local group has the capacity to fit | |
5098 | * these excess tasks, i.e. nr_running < group_capacity. The | |
5099 | * extra check prevents the case where you always pull from the | |
5100 | * heaviest group when it is already under-utilized (possible | |
5101 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 5102 | */ |
b72ff13c PZ |
5103 | if (prefer_sibling && sds->local && |
5104 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 5105 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 5106 | |
b72ff13c | 5107 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 5108 | sds->busiest = sg; |
56cf515b | 5109 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
5110 | } |
5111 | ||
b72ff13c PZ |
5112 | next_group: |
5113 | /* Now, start updating sd_lb_stats */ | |
5114 | sds->total_load += sgs->group_load; | |
5115 | sds->total_pwr += sgs->group_power; | |
5116 | ||
532cb4c4 | 5117 | sg = sg->next; |
bd939f45 | 5118 | } while (sg != env->sd->groups); |
532cb4c4 MN |
5119 | } |
5120 | ||
532cb4c4 MN |
5121 | /** |
5122 | * check_asym_packing - Check to see if the group is packed into the | |
5123 | * sched doman. | |
5124 | * | |
5125 | * This is primarily intended to used at the sibling level. Some | |
5126 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
5127 | * case of POWER7, it can move to lower SMT modes only when higher | |
5128 | * threads are idle. When in lower SMT modes, the threads will | |
5129 | * perform better since they share less core resources. Hence when we | |
5130 | * have idle threads, we want them to be the higher ones. | |
5131 | * | |
5132 | * This packing function is run on idle threads. It checks to see if | |
5133 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
5134 | * CPU number than the packing function is being run on. Here we are | |
5135 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
5136 | * number. | |
5137 | * | |
e69f6186 | 5138 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
5139 | * this CPU. The amount of the imbalance is returned in *imbalance. |
5140 | * | |
cd96891d | 5141 | * @env: The load balancing environment. |
532cb4c4 | 5142 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 5143 | */ |
bd939f45 | 5144 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
5145 | { |
5146 | int busiest_cpu; | |
5147 | ||
bd939f45 | 5148 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
5149 | return 0; |
5150 | ||
5151 | if (!sds->busiest) | |
5152 | return 0; | |
5153 | ||
5154 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 5155 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
5156 | return 0; |
5157 | ||
bd939f45 | 5158 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
5159 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
5160 | SCHED_POWER_SCALE); | |
bd939f45 | 5161 | |
532cb4c4 | 5162 | return 1; |
1e3c88bd PZ |
5163 | } |
5164 | ||
5165 | /** | |
5166 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
5167 | * amongst the groups of a sched_domain, during | |
5168 | * load balancing. | |
cd96891d | 5169 | * @env: The load balancing environment. |
1e3c88bd | 5170 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5171 | */ |
bd939f45 PZ |
5172 | static inline |
5173 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
5174 | { |
5175 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
5176 | unsigned int imbn = 2; | |
dd5feea1 | 5177 | unsigned long scaled_busy_load_per_task; |
56cf515b | 5178 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 5179 | |
56cf515b JK |
5180 | local = &sds->local_stat; |
5181 | busiest = &sds->busiest_stat; | |
1e3c88bd | 5182 | |
56cf515b JK |
5183 | if (!local->sum_nr_running) |
5184 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
5185 | else if (busiest->load_per_task > local->load_per_task) | |
5186 | imbn = 1; | |
dd5feea1 | 5187 | |
56cf515b JK |
5188 | scaled_busy_load_per_task = |
5189 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5190 | busiest->group_power; |
56cf515b | 5191 | |
3029ede3 VD |
5192 | if (busiest->avg_load + scaled_busy_load_per_task >= |
5193 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 5194 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5195 | return; |
5196 | } | |
5197 | ||
5198 | /* | |
5199 | * OK, we don't have enough imbalance to justify moving tasks, | |
5200 | * however we may be able to increase total CPU power used by | |
5201 | * moving them. | |
5202 | */ | |
5203 | ||
3ae11c90 | 5204 | pwr_now += busiest->group_power * |
56cf515b | 5205 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 5206 | pwr_now += local->group_power * |
56cf515b | 5207 | min(local->load_per_task, local->avg_load); |
1399fa78 | 5208 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5209 | |
5210 | /* Amount of load we'd subtract */ | |
56cf515b | 5211 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 5212 | busiest->group_power; |
56cf515b | 5213 | if (busiest->avg_load > tmp) { |
3ae11c90 | 5214 | pwr_move += busiest->group_power * |
56cf515b JK |
5215 | min(busiest->load_per_task, |
5216 | busiest->avg_load - tmp); | |
5217 | } | |
1e3c88bd PZ |
5218 | |
5219 | /* Amount of load we'd add */ | |
3ae11c90 | 5220 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 5221 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
5222 | tmp = (busiest->avg_load * busiest->group_power) / |
5223 | local->group_power; | |
56cf515b JK |
5224 | } else { |
5225 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5226 | local->group_power; |
56cf515b | 5227 | } |
3ae11c90 PZ |
5228 | pwr_move += local->group_power * |
5229 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 5230 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5231 | |
5232 | /* Move if we gain throughput */ | |
5233 | if (pwr_move > pwr_now) | |
56cf515b | 5234 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5235 | } |
5236 | ||
5237 | /** | |
5238 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
5239 | * groups of a given sched_domain during load balance. | |
bd939f45 | 5240 | * @env: load balance environment |
1e3c88bd | 5241 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5242 | */ |
bd939f45 | 5243 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5244 | { |
dd5feea1 | 5245 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
5246 | struct sg_lb_stats *local, *busiest; |
5247 | ||
5248 | local = &sds->local_stat; | |
56cf515b | 5249 | busiest = &sds->busiest_stat; |
dd5feea1 | 5250 | |
56cf515b | 5251 | if (busiest->group_imb) { |
30ce5dab PZ |
5252 | /* |
5253 | * In the group_imb case we cannot rely on group-wide averages | |
5254 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
5255 | */ | |
56cf515b JK |
5256 | busiest->load_per_task = |
5257 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
5258 | } |
5259 | ||
1e3c88bd PZ |
5260 | /* |
5261 | * In the presence of smp nice balancing, certain scenarios can have | |
5262 | * max load less than avg load(as we skip the groups at or below | |
5263 | * its cpu_power, while calculating max_load..) | |
5264 | */ | |
b1885550 VD |
5265 | if (busiest->avg_load <= sds->avg_load || |
5266 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
5267 | env->imbalance = 0; |
5268 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
5269 | } |
5270 | ||
56cf515b | 5271 | if (!busiest->group_imb) { |
dd5feea1 SS |
5272 | /* |
5273 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
5274 | * Except of course for the group_imb case, since then we might |
5275 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 5276 | */ |
56cf515b JK |
5277 | load_above_capacity = |
5278 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 5279 | |
1399fa78 | 5280 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 5281 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
5282 | } |
5283 | ||
5284 | /* | |
5285 | * We're trying to get all the cpus to the average_load, so we don't | |
5286 | * want to push ourselves above the average load, nor do we wish to | |
5287 | * reduce the max loaded cpu below the average load. At the same time, | |
5288 | * we also don't want to reduce the group load below the group capacity | |
5289 | * (so that we can implement power-savings policies etc). Thus we look | |
5290 | * for the minimum possible imbalance. | |
dd5feea1 | 5291 | */ |
30ce5dab | 5292 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
5293 | |
5294 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 5295 | env->imbalance = min( |
3ae11c90 PZ |
5296 | max_pull * busiest->group_power, |
5297 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 5298 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
5299 | |
5300 | /* | |
5301 | * if *imbalance is less than the average load per runnable task | |
25985edc | 5302 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
5303 | * a think about bumping its value to force at least one task to be |
5304 | * moved | |
5305 | */ | |
56cf515b | 5306 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 5307 | return fix_small_imbalance(env, sds); |
1e3c88bd | 5308 | } |
fab47622 | 5309 | |
1e3c88bd PZ |
5310 | /******* find_busiest_group() helpers end here *********************/ |
5311 | ||
5312 | /** | |
5313 | * find_busiest_group - Returns the busiest group within the sched_domain | |
5314 | * if there is an imbalance. If there isn't an imbalance, and | |
5315 | * the user has opted for power-savings, it returns a group whose | |
5316 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
5317 | * such a group exists. | |
5318 | * | |
5319 | * Also calculates the amount of weighted load which should be moved | |
5320 | * to restore balance. | |
5321 | * | |
cd96891d | 5322 | * @env: The load balancing environment. |
1e3c88bd | 5323 | * |
e69f6186 | 5324 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
5325 | * - If no imbalance and user has opted for power-savings balance, |
5326 | * return the least loaded group whose CPUs can be | |
5327 | * put to idle by rebalancing its tasks onto our group. | |
5328 | */ | |
56cf515b | 5329 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 5330 | { |
56cf515b | 5331 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
5332 | struct sd_lb_stats sds; |
5333 | ||
147c5fc2 | 5334 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
5335 | |
5336 | /* | |
5337 | * Compute the various statistics relavent for load balancing at | |
5338 | * this level. | |
5339 | */ | |
23f0d209 | 5340 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
5341 | local = &sds.local_stat; |
5342 | busiest = &sds.busiest_stat; | |
1e3c88bd | 5343 | |
bd939f45 PZ |
5344 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
5345 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
5346 | return sds.busiest; |
5347 | ||
cc57aa8f | 5348 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 5349 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
5350 | goto out_balanced; |
5351 | ||
1399fa78 | 5352 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 5353 | |
866ab43e PZ |
5354 | /* |
5355 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 5356 | * work because they assume all things are equal, which typically |
866ab43e PZ |
5357 | * isn't true due to cpus_allowed constraints and the like. |
5358 | */ | |
56cf515b | 5359 | if (busiest->group_imb) |
866ab43e PZ |
5360 | goto force_balance; |
5361 | ||
cc57aa8f | 5362 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
5363 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
5364 | !busiest->group_has_capacity) | |
fab47622 NR |
5365 | goto force_balance; |
5366 | ||
cc57aa8f PZ |
5367 | /* |
5368 | * If the local group is more busy than the selected busiest group | |
5369 | * don't try and pull any tasks. | |
5370 | */ | |
56cf515b | 5371 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
5372 | goto out_balanced; |
5373 | ||
cc57aa8f PZ |
5374 | /* |
5375 | * Don't pull any tasks if this group is already above the domain | |
5376 | * average load. | |
5377 | */ | |
56cf515b | 5378 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
5379 | goto out_balanced; |
5380 | ||
bd939f45 | 5381 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
5382 | /* |
5383 | * This cpu is idle. If the busiest group load doesn't | |
5384 | * have more tasks than the number of available cpu's and | |
5385 | * there is no imbalance between this and busiest group | |
5386 | * wrt to idle cpu's, it is balanced. | |
5387 | */ | |
56cf515b JK |
5388 | if ((local->idle_cpus < busiest->idle_cpus) && |
5389 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 5390 | goto out_balanced; |
c186fafe PZ |
5391 | } else { |
5392 | /* | |
5393 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
5394 | * imbalance_pct to be conservative. | |
5395 | */ | |
56cf515b JK |
5396 | if (100 * busiest->avg_load <= |
5397 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 5398 | goto out_balanced; |
aae6d3dd | 5399 | } |
1e3c88bd | 5400 | |
fab47622 | 5401 | force_balance: |
1e3c88bd | 5402 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 5403 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
5404 | return sds.busiest; |
5405 | ||
5406 | out_balanced: | |
bd939f45 | 5407 | env->imbalance = 0; |
1e3c88bd PZ |
5408 | return NULL; |
5409 | } | |
5410 | ||
5411 | /* | |
5412 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
5413 | */ | |
bd939f45 | 5414 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 5415 | struct sched_group *group) |
1e3c88bd PZ |
5416 | { |
5417 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 5418 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
5419 | int i; |
5420 | ||
6906a408 | 5421 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd | 5422 | unsigned long power = power_of(i); |
1399fa78 NR |
5423 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
5424 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
5425 | unsigned long wl; |
5426 | ||
9d5efe05 | 5427 | if (!capacity) |
bd939f45 | 5428 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 5429 | |
1e3c88bd | 5430 | rq = cpu_rq(i); |
6e40f5bb | 5431 | wl = weighted_cpuload(i); |
1e3c88bd | 5432 | |
6e40f5bb TG |
5433 | /* |
5434 | * When comparing with imbalance, use weighted_cpuload() | |
5435 | * which is not scaled with the cpu power. | |
5436 | */ | |
bd939f45 | 5437 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
5438 | continue; |
5439 | ||
6e40f5bb TG |
5440 | /* |
5441 | * For the load comparisons with the other cpu's, consider | |
5442 | * the weighted_cpuload() scaled with the cpu power, so that | |
5443 | * the load can be moved away from the cpu that is potentially | |
5444 | * running at a lower capacity. | |
95a79b80 JK |
5445 | * |
5446 | * Thus we're looking for max(wl_i / power_i), crosswise | |
5447 | * multiplication to rid ourselves of the division works out | |
5448 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
5449 | * previous maximum. | |
6e40f5bb | 5450 | */ |
95a79b80 JK |
5451 | if (wl * busiest_power > busiest_load * power) { |
5452 | busiest_load = wl; | |
5453 | busiest_power = power; | |
1e3c88bd PZ |
5454 | busiest = rq; |
5455 | } | |
5456 | } | |
5457 | ||
5458 | return busiest; | |
5459 | } | |
5460 | ||
5461 | /* | |
5462 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
5463 | * so long as it is large enough. | |
5464 | */ | |
5465 | #define MAX_PINNED_INTERVAL 512 | |
5466 | ||
5467 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 5468 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 5469 | |
bd939f45 | 5470 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 5471 | { |
bd939f45 PZ |
5472 | struct sched_domain *sd = env->sd; |
5473 | ||
5474 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
5475 | |
5476 | /* | |
5477 | * ASYM_PACKING needs to force migrate tasks from busy but | |
5478 | * higher numbered CPUs in order to pack all tasks in the | |
5479 | * lowest numbered CPUs. | |
5480 | */ | |
bd939f45 | 5481 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 5482 | return 1; |
1af3ed3d PZ |
5483 | } |
5484 | ||
5485 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
5486 | } | |
5487 | ||
969c7921 TH |
5488 | static int active_load_balance_cpu_stop(void *data); |
5489 | ||
23f0d209 JK |
5490 | static int should_we_balance(struct lb_env *env) |
5491 | { | |
5492 | struct sched_group *sg = env->sd->groups; | |
5493 | struct cpumask *sg_cpus, *sg_mask; | |
5494 | int cpu, balance_cpu = -1; | |
5495 | ||
5496 | /* | |
5497 | * In the newly idle case, we will allow all the cpu's | |
5498 | * to do the newly idle load balance. | |
5499 | */ | |
5500 | if (env->idle == CPU_NEWLY_IDLE) | |
5501 | return 1; | |
5502 | ||
5503 | sg_cpus = sched_group_cpus(sg); | |
5504 | sg_mask = sched_group_mask(sg); | |
5505 | /* Try to find first idle cpu */ | |
5506 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
5507 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
5508 | continue; | |
5509 | ||
5510 | balance_cpu = cpu; | |
5511 | break; | |
5512 | } | |
5513 | ||
5514 | if (balance_cpu == -1) | |
5515 | balance_cpu = group_balance_cpu(sg); | |
5516 | ||
5517 | /* | |
5518 | * First idle cpu or the first cpu(busiest) in this sched group | |
5519 | * is eligible for doing load balancing at this and above domains. | |
5520 | */ | |
b0cff9d8 | 5521 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
5522 | } |
5523 | ||
1e3c88bd PZ |
5524 | /* |
5525 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
5526 | * tasks if there is an imbalance. | |
5527 | */ | |
5528 | static int load_balance(int this_cpu, struct rq *this_rq, | |
5529 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 5530 | int *continue_balancing) |
1e3c88bd | 5531 | { |
88b8dac0 | 5532 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 5533 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 5534 | struct sched_group *group; |
1e3c88bd PZ |
5535 | struct rq *busiest; |
5536 | unsigned long flags; | |
e6252c3e | 5537 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 5538 | |
8e45cb54 PZ |
5539 | struct lb_env env = { |
5540 | .sd = sd, | |
ddcdf6e7 PZ |
5541 | .dst_cpu = this_cpu, |
5542 | .dst_rq = this_rq, | |
88b8dac0 | 5543 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 5544 | .idle = idle, |
eb95308e | 5545 | .loop_break = sched_nr_migrate_break, |
b9403130 | 5546 | .cpus = cpus, |
8e45cb54 PZ |
5547 | }; |
5548 | ||
cfc03118 JK |
5549 | /* |
5550 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
5551 | * other cpus in our group | |
5552 | */ | |
e02e60c1 | 5553 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 5554 | env.dst_grpmask = NULL; |
cfc03118 | 5555 | |
1e3c88bd PZ |
5556 | cpumask_copy(cpus, cpu_active_mask); |
5557 | ||
1e3c88bd PZ |
5558 | schedstat_inc(sd, lb_count[idle]); |
5559 | ||
5560 | redo: | |
23f0d209 JK |
5561 | if (!should_we_balance(&env)) { |
5562 | *continue_balancing = 0; | |
1e3c88bd | 5563 | goto out_balanced; |
23f0d209 | 5564 | } |
1e3c88bd | 5565 | |
23f0d209 | 5566 | group = find_busiest_group(&env); |
1e3c88bd PZ |
5567 | if (!group) { |
5568 | schedstat_inc(sd, lb_nobusyg[idle]); | |
5569 | goto out_balanced; | |
5570 | } | |
5571 | ||
b9403130 | 5572 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
5573 | if (!busiest) { |
5574 | schedstat_inc(sd, lb_nobusyq[idle]); | |
5575 | goto out_balanced; | |
5576 | } | |
5577 | ||
78feefc5 | 5578 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 5579 | |
bd939f45 | 5580 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
5581 | |
5582 | ld_moved = 0; | |
5583 | if (busiest->nr_running > 1) { | |
5584 | /* | |
5585 | * Attempt to move tasks. If find_busiest_group has found | |
5586 | * an imbalance but busiest->nr_running <= 1, the group is | |
5587 | * still unbalanced. ld_moved simply stays zero, so it is | |
5588 | * correctly treated as an imbalance. | |
5589 | */ | |
8e45cb54 | 5590 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
5591 | env.src_cpu = busiest->cpu; |
5592 | env.src_rq = busiest; | |
5593 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 5594 | |
5d6523eb | 5595 | more_balance: |
1e3c88bd | 5596 | local_irq_save(flags); |
78feefc5 | 5597 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
5598 | |
5599 | /* | |
5600 | * cur_ld_moved - load moved in current iteration | |
5601 | * ld_moved - cumulative load moved across iterations | |
5602 | */ | |
5603 | cur_ld_moved = move_tasks(&env); | |
5604 | ld_moved += cur_ld_moved; | |
78feefc5 | 5605 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
5606 | local_irq_restore(flags); |
5607 | ||
5608 | /* | |
5609 | * some other cpu did the load balance for us. | |
5610 | */ | |
88b8dac0 SV |
5611 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
5612 | resched_cpu(env.dst_cpu); | |
5613 | ||
f1cd0858 JK |
5614 | if (env.flags & LBF_NEED_BREAK) { |
5615 | env.flags &= ~LBF_NEED_BREAK; | |
5616 | goto more_balance; | |
5617 | } | |
5618 | ||
88b8dac0 SV |
5619 | /* |
5620 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
5621 | * us and move them to an alternate dst_cpu in our sched_group | |
5622 | * where they can run. The upper limit on how many times we | |
5623 | * iterate on same src_cpu is dependent on number of cpus in our | |
5624 | * sched_group. | |
5625 | * | |
5626 | * This changes load balance semantics a bit on who can move | |
5627 | * load to a given_cpu. In addition to the given_cpu itself | |
5628 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
5629 | * nohz-idle), we now have balance_cpu in a position to move | |
5630 | * load to given_cpu. In rare situations, this may cause | |
5631 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
5632 | * _independently_ and at _same_ time to move some load to | |
5633 | * given_cpu) causing exceess load to be moved to given_cpu. | |
5634 | * This however should not happen so much in practice and | |
5635 | * moreover subsequent load balance cycles should correct the | |
5636 | * excess load moved. | |
5637 | */ | |
6263322c | 5638 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 5639 | |
7aff2e3a VD |
5640 | /* Prevent to re-select dst_cpu via env's cpus */ |
5641 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
5642 | ||
78feefc5 | 5643 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 5644 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 5645 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
5646 | env.loop = 0; |
5647 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 5648 | |
88b8dac0 SV |
5649 | /* |
5650 | * Go back to "more_balance" rather than "redo" since we | |
5651 | * need to continue with same src_cpu. | |
5652 | */ | |
5653 | goto more_balance; | |
5654 | } | |
1e3c88bd | 5655 | |
6263322c PZ |
5656 | /* |
5657 | * We failed to reach balance because of affinity. | |
5658 | */ | |
5659 | if (sd_parent) { | |
5660 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
5661 | ||
5662 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
5663 | *group_imbalance = 1; | |
5664 | } else if (*group_imbalance) | |
5665 | *group_imbalance = 0; | |
5666 | } | |
5667 | ||
1e3c88bd | 5668 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 5669 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 5670 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
5671 | if (!cpumask_empty(cpus)) { |
5672 | env.loop = 0; | |
5673 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 5674 | goto redo; |
bbf18b19 | 5675 | } |
1e3c88bd PZ |
5676 | goto out_balanced; |
5677 | } | |
5678 | } | |
5679 | ||
5680 | if (!ld_moved) { | |
5681 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
5682 | /* |
5683 | * Increment the failure counter only on periodic balance. | |
5684 | * We do not want newidle balance, which can be very | |
5685 | * frequent, pollute the failure counter causing | |
5686 | * excessive cache_hot migrations and active balances. | |
5687 | */ | |
5688 | if (idle != CPU_NEWLY_IDLE) | |
5689 | sd->nr_balance_failed++; | |
1e3c88bd | 5690 | |
bd939f45 | 5691 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
5692 | raw_spin_lock_irqsave(&busiest->lock, flags); |
5693 | ||
969c7921 TH |
5694 | /* don't kick the active_load_balance_cpu_stop, |
5695 | * if the curr task on busiest cpu can't be | |
5696 | * moved to this_cpu | |
1e3c88bd PZ |
5697 | */ |
5698 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 5699 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
5700 | raw_spin_unlock_irqrestore(&busiest->lock, |
5701 | flags); | |
8e45cb54 | 5702 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
5703 | goto out_one_pinned; |
5704 | } | |
5705 | ||
969c7921 TH |
5706 | /* |
5707 | * ->active_balance synchronizes accesses to | |
5708 | * ->active_balance_work. Once set, it's cleared | |
5709 | * only after active load balance is finished. | |
5710 | */ | |
1e3c88bd PZ |
5711 | if (!busiest->active_balance) { |
5712 | busiest->active_balance = 1; | |
5713 | busiest->push_cpu = this_cpu; | |
5714 | active_balance = 1; | |
5715 | } | |
5716 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 5717 | |
bd939f45 | 5718 | if (active_balance) { |
969c7921 TH |
5719 | stop_one_cpu_nowait(cpu_of(busiest), |
5720 | active_load_balance_cpu_stop, busiest, | |
5721 | &busiest->active_balance_work); | |
bd939f45 | 5722 | } |
1e3c88bd PZ |
5723 | |
5724 | /* | |
5725 | * We've kicked active balancing, reset the failure | |
5726 | * counter. | |
5727 | */ | |
5728 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
5729 | } | |
5730 | } else | |
5731 | sd->nr_balance_failed = 0; | |
5732 | ||
5733 | if (likely(!active_balance)) { | |
5734 | /* We were unbalanced, so reset the balancing interval */ | |
5735 | sd->balance_interval = sd->min_interval; | |
5736 | } else { | |
5737 | /* | |
5738 | * If we've begun active balancing, start to back off. This | |
5739 | * case may not be covered by the all_pinned logic if there | |
5740 | * is only 1 task on the busy runqueue (because we don't call | |
5741 | * move_tasks). | |
5742 | */ | |
5743 | if (sd->balance_interval < sd->max_interval) | |
5744 | sd->balance_interval *= 2; | |
5745 | } | |
5746 | ||
1e3c88bd PZ |
5747 | goto out; |
5748 | ||
5749 | out_balanced: | |
5750 | schedstat_inc(sd, lb_balanced[idle]); | |
5751 | ||
5752 | sd->nr_balance_failed = 0; | |
5753 | ||
5754 | out_one_pinned: | |
5755 | /* tune up the balancing interval */ | |
8e45cb54 | 5756 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 5757 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
5758 | (sd->balance_interval < sd->max_interval)) |
5759 | sd->balance_interval *= 2; | |
5760 | ||
46e49b38 | 5761 | ld_moved = 0; |
1e3c88bd | 5762 | out: |
1e3c88bd PZ |
5763 | return ld_moved; |
5764 | } | |
5765 | ||
1e3c88bd PZ |
5766 | /* |
5767 | * idle_balance is called by schedule() if this_cpu is about to become | |
5768 | * idle. Attempts to pull tasks from other CPUs. | |
5769 | */ | |
029632fb | 5770 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
5771 | { |
5772 | struct sched_domain *sd; | |
5773 | int pulled_task = 0; | |
5774 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 5775 | u64 curr_cost = 0; |
1e3c88bd | 5776 | |
78becc27 | 5777 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
5778 | |
5779 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
5780 | return; | |
5781 | ||
f492e12e PZ |
5782 | /* |
5783 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
5784 | */ | |
5785 | raw_spin_unlock(&this_rq->lock); | |
5786 | ||
48a16753 | 5787 | update_blocked_averages(this_cpu); |
dce840a0 | 5788 | rcu_read_lock(); |
1e3c88bd PZ |
5789 | for_each_domain(this_cpu, sd) { |
5790 | unsigned long interval; | |
23f0d209 | 5791 | int continue_balancing = 1; |
9bd721c5 | 5792 | u64 t0, domain_cost; |
1e3c88bd PZ |
5793 | |
5794 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
5795 | continue; | |
5796 | ||
9bd721c5 JL |
5797 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
5798 | break; | |
5799 | ||
f492e12e | 5800 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
5801 | t0 = sched_clock_cpu(this_cpu); |
5802 | ||
1e3c88bd | 5803 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 5804 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
5805 | sd, CPU_NEWLY_IDLE, |
5806 | &continue_balancing); | |
9bd721c5 JL |
5807 | |
5808 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
5809 | if (domain_cost > sd->max_newidle_lb_cost) | |
5810 | sd->max_newidle_lb_cost = domain_cost; | |
5811 | ||
5812 | curr_cost += domain_cost; | |
f492e12e | 5813 | } |
1e3c88bd PZ |
5814 | |
5815 | interval = msecs_to_jiffies(sd->balance_interval); | |
5816 | if (time_after(next_balance, sd->last_balance + interval)) | |
5817 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
5818 | if (pulled_task) { |
5819 | this_rq->idle_stamp = 0; | |
1e3c88bd | 5820 | break; |
d5ad140b | 5821 | } |
1e3c88bd | 5822 | } |
dce840a0 | 5823 | rcu_read_unlock(); |
f492e12e PZ |
5824 | |
5825 | raw_spin_lock(&this_rq->lock); | |
5826 | ||
1e3c88bd PZ |
5827 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
5828 | /* | |
5829 | * We are going idle. next_balance may be set based on | |
5830 | * a busy processor. So reset next_balance. | |
5831 | */ | |
5832 | this_rq->next_balance = next_balance; | |
5833 | } | |
9bd721c5 JL |
5834 | |
5835 | if (curr_cost > this_rq->max_idle_balance_cost) | |
5836 | this_rq->max_idle_balance_cost = curr_cost; | |
1e3c88bd PZ |
5837 | } |
5838 | ||
5839 | /* | |
969c7921 TH |
5840 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
5841 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
5842 | * least 1 task to be running on each physical CPU where possible, and | |
5843 | * avoids physical / logical imbalances. | |
1e3c88bd | 5844 | */ |
969c7921 | 5845 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 5846 | { |
969c7921 TH |
5847 | struct rq *busiest_rq = data; |
5848 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 5849 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 5850 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 5851 | struct sched_domain *sd; |
969c7921 TH |
5852 | |
5853 | raw_spin_lock_irq(&busiest_rq->lock); | |
5854 | ||
5855 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
5856 | if (unlikely(busiest_cpu != smp_processor_id() || | |
5857 | !busiest_rq->active_balance)) | |
5858 | goto out_unlock; | |
1e3c88bd PZ |
5859 | |
5860 | /* Is there any task to move? */ | |
5861 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 5862 | goto out_unlock; |
1e3c88bd PZ |
5863 | |
5864 | /* | |
5865 | * This condition is "impossible", if it occurs | |
5866 | * we need to fix it. Originally reported by | |
5867 | * Bjorn Helgaas on a 128-cpu setup. | |
5868 | */ | |
5869 | BUG_ON(busiest_rq == target_rq); | |
5870 | ||
5871 | /* move a task from busiest_rq to target_rq */ | |
5872 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
5873 | |
5874 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 5875 | rcu_read_lock(); |
1e3c88bd PZ |
5876 | for_each_domain(target_cpu, sd) { |
5877 | if ((sd->flags & SD_LOAD_BALANCE) && | |
5878 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
5879 | break; | |
5880 | } | |
5881 | ||
5882 | if (likely(sd)) { | |
8e45cb54 PZ |
5883 | struct lb_env env = { |
5884 | .sd = sd, | |
ddcdf6e7 PZ |
5885 | .dst_cpu = target_cpu, |
5886 | .dst_rq = target_rq, | |
5887 | .src_cpu = busiest_rq->cpu, | |
5888 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
5889 | .idle = CPU_IDLE, |
5890 | }; | |
5891 | ||
1e3c88bd PZ |
5892 | schedstat_inc(sd, alb_count); |
5893 | ||
8e45cb54 | 5894 | if (move_one_task(&env)) |
1e3c88bd PZ |
5895 | schedstat_inc(sd, alb_pushed); |
5896 | else | |
5897 | schedstat_inc(sd, alb_failed); | |
5898 | } | |
dce840a0 | 5899 | rcu_read_unlock(); |
1e3c88bd | 5900 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
5901 | out_unlock: |
5902 | busiest_rq->active_balance = 0; | |
5903 | raw_spin_unlock_irq(&busiest_rq->lock); | |
5904 | return 0; | |
1e3c88bd PZ |
5905 | } |
5906 | ||
3451d024 | 5907 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
5908 | /* |
5909 | * idle load balancing details | |
83cd4fe2 VP |
5910 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
5911 | * needed, they will kick the idle load balancer, which then does idle | |
5912 | * load balancing for all the idle CPUs. | |
5913 | */ | |
1e3c88bd | 5914 | static struct { |
83cd4fe2 | 5915 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 5916 | atomic_t nr_cpus; |
83cd4fe2 VP |
5917 | unsigned long next_balance; /* in jiffy units */ |
5918 | } nohz ____cacheline_aligned; | |
1e3c88bd | 5919 | |
8e7fbcbc | 5920 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 5921 | { |
0b005cf5 | 5922 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 5923 | |
786d6dc7 SS |
5924 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
5925 | return ilb; | |
5926 | ||
5927 | return nr_cpu_ids; | |
1e3c88bd | 5928 | } |
1e3c88bd | 5929 | |
83cd4fe2 VP |
5930 | /* |
5931 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
5932 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
5933 | * CPU (if there is one). | |
5934 | */ | |
5935 | static void nohz_balancer_kick(int cpu) | |
5936 | { | |
5937 | int ilb_cpu; | |
5938 | ||
5939 | nohz.next_balance++; | |
5940 | ||
0b005cf5 | 5941 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 5942 | |
0b005cf5 SS |
5943 | if (ilb_cpu >= nr_cpu_ids) |
5944 | return; | |
83cd4fe2 | 5945 | |
cd490c5b | 5946 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
5947 | return; |
5948 | /* | |
5949 | * Use smp_send_reschedule() instead of resched_cpu(). | |
5950 | * This way we generate a sched IPI on the target cpu which | |
5951 | * is idle. And the softirq performing nohz idle load balance | |
5952 | * will be run before returning from the IPI. | |
5953 | */ | |
5954 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
5955 | return; |
5956 | } | |
5957 | ||
c1cc017c | 5958 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
5959 | { |
5960 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
5961 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
5962 | atomic_dec(&nohz.nr_cpus); | |
5963 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
5964 | } | |
5965 | } | |
5966 | ||
69e1e811 SS |
5967 | static inline void set_cpu_sd_state_busy(void) |
5968 | { | |
5969 | struct sched_domain *sd; | |
69e1e811 | 5970 | |
69e1e811 | 5971 | rcu_read_lock(); |
424c93fe | 5972 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
5973 | |
5974 | if (!sd || !sd->nohz_idle) | |
5975 | goto unlock; | |
5976 | sd->nohz_idle = 0; | |
5977 | ||
5978 | for (; sd; sd = sd->parent) | |
69e1e811 | 5979 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 5980 | unlock: |
69e1e811 SS |
5981 | rcu_read_unlock(); |
5982 | } | |
5983 | ||
5984 | void set_cpu_sd_state_idle(void) | |
5985 | { | |
5986 | struct sched_domain *sd; | |
69e1e811 | 5987 | |
69e1e811 | 5988 | rcu_read_lock(); |
424c93fe | 5989 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
5990 | |
5991 | if (!sd || sd->nohz_idle) | |
5992 | goto unlock; | |
5993 | sd->nohz_idle = 1; | |
5994 | ||
5995 | for (; sd; sd = sd->parent) | |
69e1e811 | 5996 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 5997 | unlock: |
69e1e811 SS |
5998 | rcu_read_unlock(); |
5999 | } | |
6000 | ||
1e3c88bd | 6001 | /* |
c1cc017c | 6002 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 6003 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 6004 | */ |
c1cc017c | 6005 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 6006 | { |
71325960 SS |
6007 | /* |
6008 | * If this cpu is going down, then nothing needs to be done. | |
6009 | */ | |
6010 | if (!cpu_active(cpu)) | |
6011 | return; | |
6012 | ||
c1cc017c AS |
6013 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
6014 | return; | |
1e3c88bd | 6015 | |
c1cc017c AS |
6016 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
6017 | atomic_inc(&nohz.nr_cpus); | |
6018 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 6019 | } |
71325960 | 6020 | |
0db0628d | 6021 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
6022 | unsigned long action, void *hcpu) |
6023 | { | |
6024 | switch (action & ~CPU_TASKS_FROZEN) { | |
6025 | case CPU_DYING: | |
c1cc017c | 6026 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
6027 | return NOTIFY_OK; |
6028 | default: | |
6029 | return NOTIFY_DONE; | |
6030 | } | |
6031 | } | |
1e3c88bd PZ |
6032 | #endif |
6033 | ||
6034 | static DEFINE_SPINLOCK(balancing); | |
6035 | ||
49c022e6 PZ |
6036 | /* |
6037 | * Scale the max load_balance interval with the number of CPUs in the system. | |
6038 | * This trades load-balance latency on larger machines for less cross talk. | |
6039 | */ | |
029632fb | 6040 | void update_max_interval(void) |
49c022e6 PZ |
6041 | { |
6042 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
6043 | } | |
6044 | ||
1e3c88bd PZ |
6045 | /* |
6046 | * It checks each scheduling domain to see if it is due to be balanced, | |
6047 | * and initiates a balancing operation if so. | |
6048 | * | |
b9b0853a | 6049 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd PZ |
6050 | */ |
6051 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
6052 | { | |
23f0d209 | 6053 | int continue_balancing = 1; |
1e3c88bd PZ |
6054 | struct rq *rq = cpu_rq(cpu); |
6055 | unsigned long interval; | |
04f733b4 | 6056 | struct sched_domain *sd; |
1e3c88bd PZ |
6057 | /* Earliest time when we have to do rebalance again */ |
6058 | unsigned long next_balance = jiffies + 60*HZ; | |
6059 | int update_next_balance = 0; | |
f48627e6 JL |
6060 | int need_serialize, need_decay = 0; |
6061 | u64 max_cost = 0; | |
1e3c88bd | 6062 | |
48a16753 | 6063 | update_blocked_averages(cpu); |
2069dd75 | 6064 | |
dce840a0 | 6065 | rcu_read_lock(); |
1e3c88bd | 6066 | for_each_domain(cpu, sd) { |
f48627e6 JL |
6067 | /* |
6068 | * Decay the newidle max times here because this is a regular | |
6069 | * visit to all the domains. Decay ~1% per second. | |
6070 | */ | |
6071 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
6072 | sd->max_newidle_lb_cost = | |
6073 | (sd->max_newidle_lb_cost * 253) / 256; | |
6074 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
6075 | need_decay = 1; | |
6076 | } | |
6077 | max_cost += sd->max_newidle_lb_cost; | |
6078 | ||
1e3c88bd PZ |
6079 | if (!(sd->flags & SD_LOAD_BALANCE)) |
6080 | continue; | |
6081 | ||
f48627e6 JL |
6082 | /* |
6083 | * Stop the load balance at this level. There is another | |
6084 | * CPU in our sched group which is doing load balancing more | |
6085 | * actively. | |
6086 | */ | |
6087 | if (!continue_balancing) { | |
6088 | if (need_decay) | |
6089 | continue; | |
6090 | break; | |
6091 | } | |
6092 | ||
1e3c88bd PZ |
6093 | interval = sd->balance_interval; |
6094 | if (idle != CPU_IDLE) | |
6095 | interval *= sd->busy_factor; | |
6096 | ||
6097 | /* scale ms to jiffies */ | |
6098 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 6099 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
6100 | |
6101 | need_serialize = sd->flags & SD_SERIALIZE; | |
6102 | ||
6103 | if (need_serialize) { | |
6104 | if (!spin_trylock(&balancing)) | |
6105 | goto out; | |
6106 | } | |
6107 | ||
6108 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 6109 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 6110 | /* |
6263322c | 6111 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
6112 | * env->dst_cpu, so we can't know our idle |
6113 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 6114 | */ |
de5eb2dd | 6115 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
6116 | } |
6117 | sd->last_balance = jiffies; | |
6118 | } | |
6119 | if (need_serialize) | |
6120 | spin_unlock(&balancing); | |
6121 | out: | |
6122 | if (time_after(next_balance, sd->last_balance + interval)) { | |
6123 | next_balance = sd->last_balance + interval; | |
6124 | update_next_balance = 1; | |
6125 | } | |
f48627e6 JL |
6126 | } |
6127 | if (need_decay) { | |
1e3c88bd | 6128 | /* |
f48627e6 JL |
6129 | * Ensure the rq-wide value also decays but keep it at a |
6130 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 6131 | */ |
f48627e6 JL |
6132 | rq->max_idle_balance_cost = |
6133 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 6134 | } |
dce840a0 | 6135 | rcu_read_unlock(); |
1e3c88bd PZ |
6136 | |
6137 | /* | |
6138 | * next_balance will be updated only when there is a need. | |
6139 | * When the cpu is attached to null domain for ex, it will not be | |
6140 | * updated. | |
6141 | */ | |
6142 | if (likely(update_next_balance)) | |
6143 | rq->next_balance = next_balance; | |
6144 | } | |
6145 | ||
3451d024 | 6146 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 6147 | /* |
3451d024 | 6148 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
6149 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
6150 | */ | |
83cd4fe2 VP |
6151 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
6152 | { | |
6153 | struct rq *this_rq = cpu_rq(this_cpu); | |
6154 | struct rq *rq; | |
6155 | int balance_cpu; | |
6156 | ||
1c792db7 SS |
6157 | if (idle != CPU_IDLE || |
6158 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
6159 | goto end; | |
83cd4fe2 VP |
6160 | |
6161 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 6162 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
6163 | continue; |
6164 | ||
6165 | /* | |
6166 | * If this cpu gets work to do, stop the load balancing | |
6167 | * work being done for other cpus. Next load | |
6168 | * balancing owner will pick it up. | |
6169 | */ | |
1c792db7 | 6170 | if (need_resched()) |
83cd4fe2 | 6171 | break; |
83cd4fe2 | 6172 | |
5ed4f1d9 VG |
6173 | rq = cpu_rq(balance_cpu); |
6174 | ||
6175 | raw_spin_lock_irq(&rq->lock); | |
6176 | update_rq_clock(rq); | |
6177 | update_idle_cpu_load(rq); | |
6178 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
6179 | |
6180 | rebalance_domains(balance_cpu, CPU_IDLE); | |
6181 | ||
83cd4fe2 VP |
6182 | if (time_after(this_rq->next_balance, rq->next_balance)) |
6183 | this_rq->next_balance = rq->next_balance; | |
6184 | } | |
6185 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
6186 | end: |
6187 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
6188 | } |
6189 | ||
6190 | /* | |
0b005cf5 SS |
6191 | * Current heuristic for kicking the idle load balancer in the presence |
6192 | * of an idle cpu is the system. | |
6193 | * - This rq has more than one task. | |
6194 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
6195 | * busy cpu's exceeding the group's power. | |
6196 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
6197 | * domain span are idle. | |
83cd4fe2 VP |
6198 | */ |
6199 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
6200 | { | |
6201 | unsigned long now = jiffies; | |
0b005cf5 | 6202 | struct sched_domain *sd; |
83cd4fe2 | 6203 | |
1c792db7 | 6204 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
6205 | return 0; |
6206 | ||
1c792db7 SS |
6207 | /* |
6208 | * We may be recently in ticked or tickless idle mode. At the first | |
6209 | * busy tick after returning from idle, we will update the busy stats. | |
6210 | */ | |
69e1e811 | 6211 | set_cpu_sd_state_busy(); |
c1cc017c | 6212 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
6213 | |
6214 | /* | |
6215 | * None are in tickless mode and hence no need for NOHZ idle load | |
6216 | * balancing. | |
6217 | */ | |
6218 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
6219 | return 0; | |
1c792db7 SS |
6220 | |
6221 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
6222 | return 0; |
6223 | ||
0b005cf5 SS |
6224 | if (rq->nr_running >= 2) |
6225 | goto need_kick; | |
83cd4fe2 | 6226 | |
067491b7 | 6227 | rcu_read_lock(); |
0b005cf5 SS |
6228 | for_each_domain(cpu, sd) { |
6229 | struct sched_group *sg = sd->groups; | |
6230 | struct sched_group_power *sgp = sg->sgp; | |
6231 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 6232 | |
0b005cf5 | 6233 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 6234 | goto need_kick_unlock; |
0b005cf5 SS |
6235 | |
6236 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
6237 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
6238 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 6239 | goto need_kick_unlock; |
0b005cf5 SS |
6240 | |
6241 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
6242 | break; | |
83cd4fe2 | 6243 | } |
067491b7 | 6244 | rcu_read_unlock(); |
83cd4fe2 | 6245 | return 0; |
067491b7 PZ |
6246 | |
6247 | need_kick_unlock: | |
6248 | rcu_read_unlock(); | |
0b005cf5 SS |
6249 | need_kick: |
6250 | return 1; | |
83cd4fe2 VP |
6251 | } |
6252 | #else | |
6253 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
6254 | #endif | |
6255 | ||
6256 | /* | |
6257 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
6258 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
6259 | */ | |
1e3c88bd PZ |
6260 | static void run_rebalance_domains(struct softirq_action *h) |
6261 | { | |
6262 | int this_cpu = smp_processor_id(); | |
6263 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 6264 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
6265 | CPU_IDLE : CPU_NOT_IDLE; |
6266 | ||
6267 | rebalance_domains(this_cpu, idle); | |
6268 | ||
1e3c88bd | 6269 | /* |
83cd4fe2 | 6270 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
6271 | * balancing on behalf of the other idle cpus whose ticks are |
6272 | * stopped. | |
6273 | */ | |
83cd4fe2 | 6274 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
6275 | } |
6276 | ||
6277 | static inline int on_null_domain(int cpu) | |
6278 | { | |
90a6501f | 6279 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
6280 | } |
6281 | ||
6282 | /* | |
6283 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 6284 | */ |
029632fb | 6285 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 6286 | { |
1e3c88bd PZ |
6287 | /* Don't need to rebalance while attached to NULL domain */ |
6288 | if (time_after_eq(jiffies, rq->next_balance) && | |
6289 | likely(!on_null_domain(cpu))) | |
6290 | raise_softirq(SCHED_SOFTIRQ); | |
3451d024 | 6291 | #ifdef CONFIG_NO_HZ_COMMON |
1c792db7 | 6292 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
6293 | nohz_balancer_kick(cpu); |
6294 | #endif | |
1e3c88bd PZ |
6295 | } |
6296 | ||
0bcdcf28 CE |
6297 | static void rq_online_fair(struct rq *rq) |
6298 | { | |
6299 | update_sysctl(); | |
6300 | } | |
6301 | ||
6302 | static void rq_offline_fair(struct rq *rq) | |
6303 | { | |
6304 | update_sysctl(); | |
a4c96ae3 PB |
6305 | |
6306 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
6307 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
6308 | } |
6309 | ||
55e12e5e | 6310 | #endif /* CONFIG_SMP */ |
e1d1484f | 6311 | |
bf0f6f24 IM |
6312 | /* |
6313 | * scheduler tick hitting a task of our scheduling class: | |
6314 | */ | |
8f4d37ec | 6315 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
6316 | { |
6317 | struct cfs_rq *cfs_rq; | |
6318 | struct sched_entity *se = &curr->se; | |
6319 | ||
6320 | for_each_sched_entity(se) { | |
6321 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 6322 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 6323 | } |
18bf2805 | 6324 | |
10e84b97 | 6325 | if (numabalancing_enabled) |
cbee9f88 | 6326 | task_tick_numa(rq, curr); |
3d59eebc | 6327 | |
18bf2805 | 6328 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
6329 | } |
6330 | ||
6331 | /* | |
cd29fe6f PZ |
6332 | * called on fork with the child task as argument from the parent's context |
6333 | * - child not yet on the tasklist | |
6334 | * - preemption disabled | |
bf0f6f24 | 6335 | */ |
cd29fe6f | 6336 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 6337 | { |
4fc420c9 DN |
6338 | struct cfs_rq *cfs_rq; |
6339 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 6340 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
6341 | struct rq *rq = this_rq(); |
6342 | unsigned long flags; | |
6343 | ||
05fa785c | 6344 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 6345 | |
861d034e PZ |
6346 | update_rq_clock(rq); |
6347 | ||
4fc420c9 DN |
6348 | cfs_rq = task_cfs_rq(current); |
6349 | curr = cfs_rq->curr; | |
6350 | ||
6c9a27f5 DN |
6351 | /* |
6352 | * Not only the cpu but also the task_group of the parent might have | |
6353 | * been changed after parent->se.parent,cfs_rq were copied to | |
6354 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
6355 | * of child point to valid ones. | |
6356 | */ | |
6357 | rcu_read_lock(); | |
6358 | __set_task_cpu(p, this_cpu); | |
6359 | rcu_read_unlock(); | |
bf0f6f24 | 6360 | |
7109c442 | 6361 | update_curr(cfs_rq); |
cd29fe6f | 6362 | |
b5d9d734 MG |
6363 | if (curr) |
6364 | se->vruntime = curr->vruntime; | |
aeb73b04 | 6365 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 6366 | |
cd29fe6f | 6367 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 6368 | /* |
edcb60a3 IM |
6369 | * Upon rescheduling, sched_class::put_prev_task() will place |
6370 | * 'current' within the tree based on its new key value. | |
6371 | */ | |
4d78e7b6 | 6372 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 6373 | resched_task(rq->curr); |
4d78e7b6 | 6374 | } |
bf0f6f24 | 6375 | |
88ec22d3 PZ |
6376 | se->vruntime -= cfs_rq->min_vruntime; |
6377 | ||
05fa785c | 6378 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
6379 | } |
6380 | ||
cb469845 SR |
6381 | /* |
6382 | * Priority of the task has changed. Check to see if we preempt | |
6383 | * the current task. | |
6384 | */ | |
da7a735e PZ |
6385 | static void |
6386 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 6387 | { |
da7a735e PZ |
6388 | if (!p->se.on_rq) |
6389 | return; | |
6390 | ||
cb469845 SR |
6391 | /* |
6392 | * Reschedule if we are currently running on this runqueue and | |
6393 | * our priority decreased, or if we are not currently running on | |
6394 | * this runqueue and our priority is higher than the current's | |
6395 | */ | |
da7a735e | 6396 | if (rq->curr == p) { |
cb469845 SR |
6397 | if (p->prio > oldprio) |
6398 | resched_task(rq->curr); | |
6399 | } else | |
15afe09b | 6400 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6401 | } |
6402 | ||
da7a735e PZ |
6403 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
6404 | { | |
6405 | struct sched_entity *se = &p->se; | |
6406 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6407 | ||
6408 | /* | |
6409 | * Ensure the task's vruntime is normalized, so that when its | |
6410 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
6411 | * do the right thing. | |
6412 | * | |
6413 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
6414 | * have normalized the vruntime, if it was !on_rq, then only when | |
6415 | * the task is sleeping will it still have non-normalized vruntime. | |
6416 | */ | |
6417 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
6418 | /* | |
6419 | * Fix up our vruntime so that the current sleep doesn't | |
6420 | * cause 'unlimited' sleep bonus. | |
6421 | */ | |
6422 | place_entity(cfs_rq, se, 0); | |
6423 | se->vruntime -= cfs_rq->min_vruntime; | |
6424 | } | |
9ee474f5 | 6425 | |
141965c7 | 6426 | #ifdef CONFIG_SMP |
9ee474f5 PT |
6427 | /* |
6428 | * Remove our load from contribution when we leave sched_fair | |
6429 | * and ensure we don't carry in an old decay_count if we | |
6430 | * switch back. | |
6431 | */ | |
87e3c8ae KT |
6432 | if (se->avg.decay_count) { |
6433 | __synchronize_entity_decay(se); | |
6434 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
6435 | } |
6436 | #endif | |
da7a735e PZ |
6437 | } |
6438 | ||
cb469845 SR |
6439 | /* |
6440 | * We switched to the sched_fair class. | |
6441 | */ | |
da7a735e | 6442 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 6443 | { |
da7a735e PZ |
6444 | if (!p->se.on_rq) |
6445 | return; | |
6446 | ||
cb469845 SR |
6447 | /* |
6448 | * We were most likely switched from sched_rt, so | |
6449 | * kick off the schedule if running, otherwise just see | |
6450 | * if we can still preempt the current task. | |
6451 | */ | |
da7a735e | 6452 | if (rq->curr == p) |
cb469845 SR |
6453 | resched_task(rq->curr); |
6454 | else | |
15afe09b | 6455 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6456 | } |
6457 | ||
83b699ed SV |
6458 | /* Account for a task changing its policy or group. |
6459 | * | |
6460 | * This routine is mostly called to set cfs_rq->curr field when a task | |
6461 | * migrates between groups/classes. | |
6462 | */ | |
6463 | static void set_curr_task_fair(struct rq *rq) | |
6464 | { | |
6465 | struct sched_entity *se = &rq->curr->se; | |
6466 | ||
ec12cb7f PT |
6467 | for_each_sched_entity(se) { |
6468 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6469 | ||
6470 | set_next_entity(cfs_rq, se); | |
6471 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
6472 | account_cfs_rq_runtime(cfs_rq, 0); | |
6473 | } | |
83b699ed SV |
6474 | } |
6475 | ||
029632fb PZ |
6476 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
6477 | { | |
6478 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
6479 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
6480 | #ifndef CONFIG_64BIT | |
6481 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
6482 | #endif | |
141965c7 | 6483 | #ifdef CONFIG_SMP |
9ee474f5 | 6484 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 6485 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 6486 | #endif |
029632fb PZ |
6487 | } |
6488 | ||
810b3817 | 6489 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6490 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 6491 | { |
aff3e498 | 6492 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
6493 | /* |
6494 | * If the task was not on the rq at the time of this cgroup movement | |
6495 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
6496 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
6497 | * bonus in place_entity()). | |
6498 | * | |
6499 | * If it was on the rq, we've just 'preempted' it, which does convert | |
6500 | * ->vruntime to a relative base. | |
6501 | * | |
6502 | * Make sure both cases convert their relative position when migrating | |
6503 | * to another cgroup's rq. This does somewhat interfere with the | |
6504 | * fair sleeper stuff for the first placement, but who cares. | |
6505 | */ | |
7ceff013 DN |
6506 | /* |
6507 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
6508 | * But there are some cases where it has already been normalized: | |
6509 | * | |
6510 | * - Moving a forked child which is waiting for being woken up by | |
6511 | * wake_up_new_task(). | |
62af3783 DN |
6512 | * - Moving a task which has been woken up by try_to_wake_up() and |
6513 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
6514 | * |
6515 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
6516 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
6517 | */ | |
62af3783 | 6518 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
6519 | on_rq = 1; |
6520 | ||
b2b5ce02 PZ |
6521 | if (!on_rq) |
6522 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
6523 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
6524 | if (!on_rq) { |
6525 | cfs_rq = cfs_rq_of(&p->se); | |
6526 | p->se.vruntime += cfs_rq->min_vruntime; | |
6527 | #ifdef CONFIG_SMP | |
6528 | /* | |
6529 | * migrate_task_rq_fair() will have removed our previous | |
6530 | * contribution, but we must synchronize for ongoing future | |
6531 | * decay. | |
6532 | */ | |
6533 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
6534 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
6535 | #endif | |
6536 | } | |
810b3817 | 6537 | } |
029632fb PZ |
6538 | |
6539 | void free_fair_sched_group(struct task_group *tg) | |
6540 | { | |
6541 | int i; | |
6542 | ||
6543 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6544 | ||
6545 | for_each_possible_cpu(i) { | |
6546 | if (tg->cfs_rq) | |
6547 | kfree(tg->cfs_rq[i]); | |
6548 | if (tg->se) | |
6549 | kfree(tg->se[i]); | |
6550 | } | |
6551 | ||
6552 | kfree(tg->cfs_rq); | |
6553 | kfree(tg->se); | |
6554 | } | |
6555 | ||
6556 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6557 | { | |
6558 | struct cfs_rq *cfs_rq; | |
6559 | struct sched_entity *se; | |
6560 | int i; | |
6561 | ||
6562 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
6563 | if (!tg->cfs_rq) | |
6564 | goto err; | |
6565 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
6566 | if (!tg->se) | |
6567 | goto err; | |
6568 | ||
6569 | tg->shares = NICE_0_LOAD; | |
6570 | ||
6571 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6572 | ||
6573 | for_each_possible_cpu(i) { | |
6574 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
6575 | GFP_KERNEL, cpu_to_node(i)); | |
6576 | if (!cfs_rq) | |
6577 | goto err; | |
6578 | ||
6579 | se = kzalloc_node(sizeof(struct sched_entity), | |
6580 | GFP_KERNEL, cpu_to_node(i)); | |
6581 | if (!se) | |
6582 | goto err_free_rq; | |
6583 | ||
6584 | init_cfs_rq(cfs_rq); | |
6585 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
6586 | } | |
6587 | ||
6588 | return 1; | |
6589 | ||
6590 | err_free_rq: | |
6591 | kfree(cfs_rq); | |
6592 | err: | |
6593 | return 0; | |
6594 | } | |
6595 | ||
6596 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
6597 | { | |
6598 | struct rq *rq = cpu_rq(cpu); | |
6599 | unsigned long flags; | |
6600 | ||
6601 | /* | |
6602 | * Only empty task groups can be destroyed; so we can speculatively | |
6603 | * check on_list without danger of it being re-added. | |
6604 | */ | |
6605 | if (!tg->cfs_rq[cpu]->on_list) | |
6606 | return; | |
6607 | ||
6608 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6609 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
6610 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6611 | } | |
6612 | ||
6613 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
6614 | struct sched_entity *se, int cpu, | |
6615 | struct sched_entity *parent) | |
6616 | { | |
6617 | struct rq *rq = cpu_rq(cpu); | |
6618 | ||
6619 | cfs_rq->tg = tg; | |
6620 | cfs_rq->rq = rq; | |
029632fb PZ |
6621 | init_cfs_rq_runtime(cfs_rq); |
6622 | ||
6623 | tg->cfs_rq[cpu] = cfs_rq; | |
6624 | tg->se[cpu] = se; | |
6625 | ||
6626 | /* se could be NULL for root_task_group */ | |
6627 | if (!se) | |
6628 | return; | |
6629 | ||
6630 | if (!parent) | |
6631 | se->cfs_rq = &rq->cfs; | |
6632 | else | |
6633 | se->cfs_rq = parent->my_q; | |
6634 | ||
6635 | se->my_q = cfs_rq; | |
6636 | update_load_set(&se->load, 0); | |
6637 | se->parent = parent; | |
6638 | } | |
6639 | ||
6640 | static DEFINE_MUTEX(shares_mutex); | |
6641 | ||
6642 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
6643 | { | |
6644 | int i; | |
6645 | unsigned long flags; | |
6646 | ||
6647 | /* | |
6648 | * We can't change the weight of the root cgroup. | |
6649 | */ | |
6650 | if (!tg->se[0]) | |
6651 | return -EINVAL; | |
6652 | ||
6653 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
6654 | ||
6655 | mutex_lock(&shares_mutex); | |
6656 | if (tg->shares == shares) | |
6657 | goto done; | |
6658 | ||
6659 | tg->shares = shares; | |
6660 | for_each_possible_cpu(i) { | |
6661 | struct rq *rq = cpu_rq(i); | |
6662 | struct sched_entity *se; | |
6663 | ||
6664 | se = tg->se[i]; | |
6665 | /* Propagate contribution to hierarchy */ | |
6666 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
6667 | |
6668 | /* Possible calls to update_curr() need rq clock */ | |
6669 | update_rq_clock(rq); | |
17bc14b7 | 6670 | for_each_sched_entity(se) |
029632fb PZ |
6671 | update_cfs_shares(group_cfs_rq(se)); |
6672 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6673 | } | |
6674 | ||
6675 | done: | |
6676 | mutex_unlock(&shares_mutex); | |
6677 | return 0; | |
6678 | } | |
6679 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6680 | ||
6681 | void free_fair_sched_group(struct task_group *tg) { } | |
6682 | ||
6683 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6684 | { | |
6685 | return 1; | |
6686 | } | |
6687 | ||
6688 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
6689 | ||
6690 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
6691 | ||
810b3817 | 6692 | |
6d686f45 | 6693 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
6694 | { |
6695 | struct sched_entity *se = &task->se; | |
0d721cea PW |
6696 | unsigned int rr_interval = 0; |
6697 | ||
6698 | /* | |
6699 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
6700 | * idle runqueue: | |
6701 | */ | |
0d721cea | 6702 | if (rq->cfs.load.weight) |
a59f4e07 | 6703 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
6704 | |
6705 | return rr_interval; | |
6706 | } | |
6707 | ||
bf0f6f24 IM |
6708 | /* |
6709 | * All the scheduling class methods: | |
6710 | */ | |
029632fb | 6711 | const struct sched_class fair_sched_class = { |
5522d5d5 | 6712 | .next = &idle_sched_class, |
bf0f6f24 IM |
6713 | .enqueue_task = enqueue_task_fair, |
6714 | .dequeue_task = dequeue_task_fair, | |
6715 | .yield_task = yield_task_fair, | |
d95f4122 | 6716 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 6717 | |
2e09bf55 | 6718 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
6719 | |
6720 | .pick_next_task = pick_next_task_fair, | |
6721 | .put_prev_task = put_prev_task_fair, | |
6722 | ||
681f3e68 | 6723 | #ifdef CONFIG_SMP |
4ce72a2c | 6724 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 6725 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 6726 | |
0bcdcf28 CE |
6727 | .rq_online = rq_online_fair, |
6728 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
6729 | |
6730 | .task_waking = task_waking_fair, | |
681f3e68 | 6731 | #endif |
bf0f6f24 | 6732 | |
83b699ed | 6733 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 6734 | .task_tick = task_tick_fair, |
cd29fe6f | 6735 | .task_fork = task_fork_fair, |
cb469845 SR |
6736 | |
6737 | .prio_changed = prio_changed_fair, | |
da7a735e | 6738 | .switched_from = switched_from_fair, |
cb469845 | 6739 | .switched_to = switched_to_fair, |
810b3817 | 6740 | |
0d721cea PW |
6741 | .get_rr_interval = get_rr_interval_fair, |
6742 | ||
810b3817 | 6743 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6744 | .task_move_group = task_move_group_fair, |
810b3817 | 6745 | #endif |
bf0f6f24 IM |
6746 | }; |
6747 | ||
6748 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 6749 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 6750 | { |
bf0f6f24 IM |
6751 | struct cfs_rq *cfs_rq; |
6752 | ||
5973e5b9 | 6753 | rcu_read_lock(); |
c3b64f1e | 6754 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 6755 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 6756 | rcu_read_unlock(); |
bf0f6f24 IM |
6757 | } |
6758 | #endif | |
029632fb PZ |
6759 | |
6760 | __init void init_sched_fair_class(void) | |
6761 | { | |
6762 | #ifdef CONFIG_SMP | |
6763 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
6764 | ||
3451d024 | 6765 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 6766 | nohz.next_balance = jiffies; |
029632fb | 6767 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 6768 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
6769 | #endif |
6770 | #endif /* SMP */ | |
6771 | ||
6772 | } |