Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
bf0f6f24 IM |
21 | */ |
22 | ||
9745512c | 23 | #include <linux/latencytop.h> |
1983a922 | 24 | #include <linux/sched.h> |
3436ae12 | 25 | #include <linux/cpumask.h> |
029632fb PZ |
26 | #include <linux/slab.h> |
27 | #include <linux/profile.h> | |
28 | #include <linux/interrupt.h> | |
cbee9f88 | 29 | #include <linux/mempolicy.h> |
e14808b4 | 30 | #include <linux/migrate.h> |
cbee9f88 | 31 | #include <linux/task_work.h> |
029632fb PZ |
32 | |
33 | #include <trace/events/sched.h> | |
34 | ||
35 | #include "sched.h" | |
9745512c | 36 | |
bf0f6f24 | 37 | /* |
21805085 | 38 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 39 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 40 | * |
21805085 | 41 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
42 | * 'timeslice length' - timeslices in CFS are of variable length |
43 | * and have no persistent notion like in traditional, time-slice | |
44 | * based scheduling concepts. | |
bf0f6f24 | 45 | * |
d274a4ce IM |
46 | * (to see the precise effective timeslice length of your workload, |
47 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 48 | */ |
21406928 MG |
49 | unsigned int sysctl_sched_latency = 6000000ULL; |
50 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 51 | |
1983a922 CE |
52 | /* |
53 | * The initial- and re-scaling of tunables is configurable | |
54 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
55 | * | |
56 | * Options are: | |
57 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
58 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
59 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
60 | */ | |
61 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
62 | = SCHED_TUNABLESCALING_LOG; | |
63 | ||
2bd8e6d4 | 64 | /* |
b2be5e96 | 65 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 66 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 67 | */ |
0bf377bb IM |
68 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
69 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
70 | |
71 | /* | |
b2be5e96 PZ |
72 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
73 | */ | |
0bf377bb | 74 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
75 | |
76 | /* | |
2bba22c5 | 77 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 78 | * parent will (try to) run first. |
21805085 | 79 | */ |
2bba22c5 | 80 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 81 | |
bf0f6f24 IM |
82 | /* |
83 | * SCHED_OTHER wake-up granularity. | |
172e082a | 84 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
85 | * |
86 | * This option delays the preemption effects of decoupled workloads | |
87 | * and reduces their over-scheduling. Synchronous workloads will still | |
88 | * have immediate wakeup/sleep latencies. | |
89 | */ | |
172e082a | 90 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 91 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 92 | |
da84d961 IM |
93 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
94 | ||
a7a4f8a7 PT |
95 | /* |
96 | * The exponential sliding window over which load is averaged for shares | |
97 | * distribution. | |
98 | * (default: 10msec) | |
99 | */ | |
100 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
101 | ||
ec12cb7f PT |
102 | #ifdef CONFIG_CFS_BANDWIDTH |
103 | /* | |
104 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
105 | * each time a cfs_rq requests quota. | |
106 | * | |
107 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
108 | * to consumption or the quota being specified to be smaller than the slice) | |
109 | * we will always only issue the remaining available time. | |
110 | * | |
111 | * default: 5 msec, units: microseconds | |
112 | */ | |
113 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
114 | #endif | |
115 | ||
8527632d PG |
116 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
117 | { | |
118 | lw->weight += inc; | |
119 | lw->inv_weight = 0; | |
120 | } | |
121 | ||
122 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
123 | { | |
124 | lw->weight -= dec; | |
125 | lw->inv_weight = 0; | |
126 | } | |
127 | ||
128 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
129 | { | |
130 | lw->weight = w; | |
131 | lw->inv_weight = 0; | |
132 | } | |
133 | ||
029632fb PZ |
134 | /* |
135 | * Increase the granularity value when there are more CPUs, | |
136 | * because with more CPUs the 'effective latency' as visible | |
137 | * to users decreases. But the relationship is not linear, | |
138 | * so pick a second-best guess by going with the log2 of the | |
139 | * number of CPUs. | |
140 | * | |
141 | * This idea comes from the SD scheduler of Con Kolivas: | |
142 | */ | |
143 | static int get_update_sysctl_factor(void) | |
144 | { | |
145 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
146 | unsigned int factor; | |
147 | ||
148 | switch (sysctl_sched_tunable_scaling) { | |
149 | case SCHED_TUNABLESCALING_NONE: | |
150 | factor = 1; | |
151 | break; | |
152 | case SCHED_TUNABLESCALING_LINEAR: | |
153 | factor = cpus; | |
154 | break; | |
155 | case SCHED_TUNABLESCALING_LOG: | |
156 | default: | |
157 | factor = 1 + ilog2(cpus); | |
158 | break; | |
159 | } | |
160 | ||
161 | return factor; | |
162 | } | |
163 | ||
164 | static void update_sysctl(void) | |
165 | { | |
166 | unsigned int factor = get_update_sysctl_factor(); | |
167 | ||
168 | #define SET_SYSCTL(name) \ | |
169 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
170 | SET_SYSCTL(sched_min_granularity); | |
171 | SET_SYSCTL(sched_latency); | |
172 | SET_SYSCTL(sched_wakeup_granularity); | |
173 | #undef SET_SYSCTL | |
174 | } | |
175 | ||
176 | void sched_init_granularity(void) | |
177 | { | |
178 | update_sysctl(); | |
179 | } | |
180 | ||
181 | #if BITS_PER_LONG == 32 | |
182 | # define WMULT_CONST (~0UL) | |
183 | #else | |
184 | # define WMULT_CONST (1UL << 32) | |
185 | #endif | |
186 | ||
187 | #define WMULT_SHIFT 32 | |
188 | ||
189 | /* | |
190 | * Shift right and round: | |
191 | */ | |
192 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) | |
193 | ||
194 | /* | |
195 | * delta *= weight / lw | |
196 | */ | |
197 | static unsigned long | |
198 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, | |
199 | struct load_weight *lw) | |
200 | { | |
201 | u64 tmp; | |
202 | ||
203 | /* | |
204 | * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched | |
205 | * entities since MIN_SHARES = 2. Treat weight as 1 if less than | |
206 | * 2^SCHED_LOAD_RESOLUTION. | |
207 | */ | |
208 | if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) | |
209 | tmp = (u64)delta_exec * scale_load_down(weight); | |
210 | else | |
211 | tmp = (u64)delta_exec; | |
212 | ||
213 | if (!lw->inv_weight) { | |
214 | unsigned long w = scale_load_down(lw->weight); | |
215 | ||
216 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
217 | lw->inv_weight = 1; | |
218 | else if (unlikely(!w)) | |
219 | lw->inv_weight = WMULT_CONST; | |
220 | else | |
221 | lw->inv_weight = WMULT_CONST / w; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Check whether we'd overflow the 64-bit multiplication: | |
226 | */ | |
227 | if (unlikely(tmp > WMULT_CONST)) | |
228 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, | |
229 | WMULT_SHIFT/2); | |
230 | else | |
231 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); | |
232 | ||
233 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); | |
234 | } | |
235 | ||
236 | ||
237 | const struct sched_class fair_sched_class; | |
a4c2f00f | 238 | |
bf0f6f24 IM |
239 | /************************************************************** |
240 | * CFS operations on generic schedulable entities: | |
241 | */ | |
242 | ||
62160e3f | 243 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 244 | |
62160e3f | 245 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
246 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
247 | { | |
62160e3f | 248 | return cfs_rq->rq; |
bf0f6f24 IM |
249 | } |
250 | ||
62160e3f IM |
251 | /* An entity is a task if it doesn't "own" a runqueue */ |
252 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 253 | |
8f48894f PZ |
254 | static inline struct task_struct *task_of(struct sched_entity *se) |
255 | { | |
256 | #ifdef CONFIG_SCHED_DEBUG | |
257 | WARN_ON_ONCE(!entity_is_task(se)); | |
258 | #endif | |
259 | return container_of(se, struct task_struct, se); | |
260 | } | |
261 | ||
b758149c PZ |
262 | /* Walk up scheduling entities hierarchy */ |
263 | #define for_each_sched_entity(se) \ | |
264 | for (; se; se = se->parent) | |
265 | ||
266 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
267 | { | |
268 | return p->se.cfs_rq; | |
269 | } | |
270 | ||
271 | /* runqueue on which this entity is (to be) queued */ | |
272 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
273 | { | |
274 | return se->cfs_rq; | |
275 | } | |
276 | ||
277 | /* runqueue "owned" by this group */ | |
278 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
279 | { | |
280 | return grp->my_q; | |
281 | } | |
282 | ||
aff3e498 PT |
283 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
284 | int force_update); | |
9ee474f5 | 285 | |
3d4b47b4 PZ |
286 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
287 | { | |
288 | if (!cfs_rq->on_list) { | |
67e86250 PT |
289 | /* |
290 | * Ensure we either appear before our parent (if already | |
291 | * enqueued) or force our parent to appear after us when it is | |
292 | * enqueued. The fact that we always enqueue bottom-up | |
293 | * reduces this to two cases. | |
294 | */ | |
295 | if (cfs_rq->tg->parent && | |
296 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
297 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
298 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
299 | } else { | |
300 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 302 | } |
3d4b47b4 PZ |
303 | |
304 | cfs_rq->on_list = 1; | |
9ee474f5 | 305 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 306 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
307 | } |
308 | } | |
309 | ||
310 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
311 | { | |
312 | if (cfs_rq->on_list) { | |
313 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
314 | cfs_rq->on_list = 0; | |
315 | } | |
316 | } | |
317 | ||
b758149c PZ |
318 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
319 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
320 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
321 | ||
322 | /* Do the two (enqueued) entities belong to the same group ? */ | |
323 | static inline int | |
324 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
325 | { | |
326 | if (se->cfs_rq == pse->cfs_rq) | |
327 | return 1; | |
328 | ||
329 | return 0; | |
330 | } | |
331 | ||
332 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
333 | { | |
334 | return se->parent; | |
335 | } | |
336 | ||
464b7527 PZ |
337 | /* return depth at which a sched entity is present in the hierarchy */ |
338 | static inline int depth_se(struct sched_entity *se) | |
339 | { | |
340 | int depth = 0; | |
341 | ||
342 | for_each_sched_entity(se) | |
343 | depth++; | |
344 | ||
345 | return depth; | |
346 | } | |
347 | ||
348 | static void | |
349 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
350 | { | |
351 | int se_depth, pse_depth; | |
352 | ||
353 | /* | |
354 | * preemption test can be made between sibling entities who are in the | |
355 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
356 | * both tasks until we find their ancestors who are siblings of common | |
357 | * parent. | |
358 | */ | |
359 | ||
360 | /* First walk up until both entities are at same depth */ | |
361 | se_depth = depth_se(*se); | |
362 | pse_depth = depth_se(*pse); | |
363 | ||
364 | while (se_depth > pse_depth) { | |
365 | se_depth--; | |
366 | *se = parent_entity(*se); | |
367 | } | |
368 | ||
369 | while (pse_depth > se_depth) { | |
370 | pse_depth--; | |
371 | *pse = parent_entity(*pse); | |
372 | } | |
373 | ||
374 | while (!is_same_group(*se, *pse)) { | |
375 | *se = parent_entity(*se); | |
376 | *pse = parent_entity(*pse); | |
377 | } | |
378 | } | |
379 | ||
8f48894f PZ |
380 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
381 | ||
382 | static inline struct task_struct *task_of(struct sched_entity *se) | |
383 | { | |
384 | return container_of(se, struct task_struct, se); | |
385 | } | |
bf0f6f24 | 386 | |
62160e3f IM |
387 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
388 | { | |
389 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
390 | } |
391 | ||
392 | #define entity_is_task(se) 1 | |
393 | ||
b758149c PZ |
394 | #define for_each_sched_entity(se) \ |
395 | for (; se; se = NULL) | |
bf0f6f24 | 396 | |
b758149c | 397 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 398 | { |
b758149c | 399 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
400 | } |
401 | ||
b758149c PZ |
402 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
403 | { | |
404 | struct task_struct *p = task_of(se); | |
405 | struct rq *rq = task_rq(p); | |
406 | ||
407 | return &rq->cfs; | |
408 | } | |
409 | ||
410 | /* runqueue "owned" by this group */ | |
411 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
412 | { | |
413 | return NULL; | |
414 | } | |
415 | ||
3d4b47b4 PZ |
416 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
417 | { | |
418 | } | |
419 | ||
420 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
421 | { | |
422 | } | |
423 | ||
b758149c PZ |
424 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
425 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
426 | ||
427 | static inline int | |
428 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | |
429 | { | |
430 | return 1; | |
431 | } | |
432 | ||
433 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
434 | { | |
435 | return NULL; | |
436 | } | |
437 | ||
464b7527 PZ |
438 | static inline void |
439 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
440 | { | |
441 | } | |
442 | ||
b758149c PZ |
443 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
444 | ||
6c16a6dc PZ |
445 | static __always_inline |
446 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); | |
bf0f6f24 IM |
447 | |
448 | /************************************************************** | |
449 | * Scheduling class tree data structure manipulation methods: | |
450 | */ | |
451 | ||
1bf08230 | 452 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 453 | { |
1bf08230 | 454 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 455 | if (delta > 0) |
1bf08230 | 456 | max_vruntime = vruntime; |
02e0431a | 457 | |
1bf08230 | 458 | return max_vruntime; |
02e0431a PZ |
459 | } |
460 | ||
0702e3eb | 461 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
462 | { |
463 | s64 delta = (s64)(vruntime - min_vruntime); | |
464 | if (delta < 0) | |
465 | min_vruntime = vruntime; | |
466 | ||
467 | return min_vruntime; | |
468 | } | |
469 | ||
54fdc581 FC |
470 | static inline int entity_before(struct sched_entity *a, |
471 | struct sched_entity *b) | |
472 | { | |
473 | return (s64)(a->vruntime - b->vruntime) < 0; | |
474 | } | |
475 | ||
1af5f730 PZ |
476 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
477 | { | |
478 | u64 vruntime = cfs_rq->min_vruntime; | |
479 | ||
480 | if (cfs_rq->curr) | |
481 | vruntime = cfs_rq->curr->vruntime; | |
482 | ||
483 | if (cfs_rq->rb_leftmost) { | |
484 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
485 | struct sched_entity, | |
486 | run_node); | |
487 | ||
e17036da | 488 | if (!cfs_rq->curr) |
1af5f730 PZ |
489 | vruntime = se->vruntime; |
490 | else | |
491 | vruntime = min_vruntime(vruntime, se->vruntime); | |
492 | } | |
493 | ||
1bf08230 | 494 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 495 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
496 | #ifndef CONFIG_64BIT |
497 | smp_wmb(); | |
498 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
499 | #endif | |
1af5f730 PZ |
500 | } |
501 | ||
bf0f6f24 IM |
502 | /* |
503 | * Enqueue an entity into the rb-tree: | |
504 | */ | |
0702e3eb | 505 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
506 | { |
507 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
508 | struct rb_node *parent = NULL; | |
509 | struct sched_entity *entry; | |
bf0f6f24 IM |
510 | int leftmost = 1; |
511 | ||
512 | /* | |
513 | * Find the right place in the rbtree: | |
514 | */ | |
515 | while (*link) { | |
516 | parent = *link; | |
517 | entry = rb_entry(parent, struct sched_entity, run_node); | |
518 | /* | |
519 | * We dont care about collisions. Nodes with | |
520 | * the same key stay together. | |
521 | */ | |
2bd2d6f2 | 522 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
523 | link = &parent->rb_left; |
524 | } else { | |
525 | link = &parent->rb_right; | |
526 | leftmost = 0; | |
527 | } | |
528 | } | |
529 | ||
530 | /* | |
531 | * Maintain a cache of leftmost tree entries (it is frequently | |
532 | * used): | |
533 | */ | |
1af5f730 | 534 | if (leftmost) |
57cb499d | 535 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
536 | |
537 | rb_link_node(&se->run_node, parent, link); | |
538 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
539 | } |
540 | ||
0702e3eb | 541 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 542 | { |
3fe69747 PZ |
543 | if (cfs_rq->rb_leftmost == &se->run_node) { |
544 | struct rb_node *next_node; | |
3fe69747 PZ |
545 | |
546 | next_node = rb_next(&se->run_node); | |
547 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 548 | } |
e9acbff6 | 549 | |
bf0f6f24 | 550 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
551 | } |
552 | ||
029632fb | 553 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 554 | { |
f4b6755f PZ |
555 | struct rb_node *left = cfs_rq->rb_leftmost; |
556 | ||
557 | if (!left) | |
558 | return NULL; | |
559 | ||
560 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
561 | } |
562 | ||
ac53db59 RR |
563 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
564 | { | |
565 | struct rb_node *next = rb_next(&se->run_node); | |
566 | ||
567 | if (!next) | |
568 | return NULL; | |
569 | ||
570 | return rb_entry(next, struct sched_entity, run_node); | |
571 | } | |
572 | ||
573 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 574 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 575 | { |
7eee3e67 | 576 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 577 | |
70eee74b BS |
578 | if (!last) |
579 | return NULL; | |
7eee3e67 IM |
580 | |
581 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
582 | } |
583 | ||
bf0f6f24 IM |
584 | /************************************************************** |
585 | * Scheduling class statistics methods: | |
586 | */ | |
587 | ||
acb4a848 | 588 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 589 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
590 | loff_t *ppos) |
591 | { | |
8d65af78 | 592 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 593 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
594 | |
595 | if (ret || !write) | |
596 | return ret; | |
597 | ||
598 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
599 | sysctl_sched_min_granularity); | |
600 | ||
acb4a848 CE |
601 | #define WRT_SYSCTL(name) \ |
602 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
603 | WRT_SYSCTL(sched_min_granularity); | |
604 | WRT_SYSCTL(sched_latency); | |
605 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
606 | #undef WRT_SYSCTL |
607 | ||
b2be5e96 PZ |
608 | return 0; |
609 | } | |
610 | #endif | |
647e7cac | 611 | |
a7be37ac | 612 | /* |
f9c0b095 | 613 | * delta /= w |
a7be37ac PZ |
614 | */ |
615 | static inline unsigned long | |
616 | calc_delta_fair(unsigned long delta, struct sched_entity *se) | |
617 | { | |
f9c0b095 PZ |
618 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
619 | delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); | |
a7be37ac PZ |
620 | |
621 | return delta; | |
622 | } | |
623 | ||
647e7cac IM |
624 | /* |
625 | * The idea is to set a period in which each task runs once. | |
626 | * | |
532b1858 | 627 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
628 | * this period because otherwise the slices get too small. |
629 | * | |
630 | * p = (nr <= nl) ? l : l*nr/nl | |
631 | */ | |
4d78e7b6 PZ |
632 | static u64 __sched_period(unsigned long nr_running) |
633 | { | |
634 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 635 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
636 | |
637 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 638 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 639 | period *= nr_running; |
4d78e7b6 PZ |
640 | } |
641 | ||
642 | return period; | |
643 | } | |
644 | ||
647e7cac IM |
645 | /* |
646 | * We calculate the wall-time slice from the period by taking a part | |
647 | * proportional to the weight. | |
648 | * | |
f9c0b095 | 649 | * s = p*P[w/rw] |
647e7cac | 650 | */ |
6d0f0ebd | 651 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 652 | { |
0a582440 | 653 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 654 | |
0a582440 | 655 | for_each_sched_entity(se) { |
6272d68c | 656 | struct load_weight *load; |
3104bf03 | 657 | struct load_weight lw; |
6272d68c LM |
658 | |
659 | cfs_rq = cfs_rq_of(se); | |
660 | load = &cfs_rq->load; | |
f9c0b095 | 661 | |
0a582440 | 662 | if (unlikely(!se->on_rq)) { |
3104bf03 | 663 | lw = cfs_rq->load; |
0a582440 MG |
664 | |
665 | update_load_add(&lw, se->load.weight); | |
666 | load = &lw; | |
667 | } | |
668 | slice = calc_delta_mine(slice, se->load.weight, load); | |
669 | } | |
670 | return slice; | |
bf0f6f24 IM |
671 | } |
672 | ||
647e7cac | 673 | /* |
660cc00f | 674 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 675 | * |
f9c0b095 | 676 | * vs = s/w |
647e7cac | 677 | */ |
f9c0b095 | 678 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 679 | { |
f9c0b095 | 680 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
681 | } |
682 | ||
a75cdaa9 | 683 | #ifdef CONFIG_SMP |
fb13c7ee MG |
684 | static unsigned long task_h_load(struct task_struct *p); |
685 | ||
a75cdaa9 AS |
686 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
687 | ||
688 | /* Give new task start runnable values to heavy its load in infant time */ | |
689 | void init_task_runnable_average(struct task_struct *p) | |
690 | { | |
691 | u32 slice; | |
692 | ||
693 | p->se.avg.decay_count = 0; | |
694 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | |
695 | p->se.avg.runnable_avg_sum = slice; | |
696 | p->se.avg.runnable_avg_period = slice; | |
697 | __update_task_entity_contrib(&p->se); | |
698 | } | |
699 | #else | |
700 | void init_task_runnable_average(struct task_struct *p) | |
701 | { | |
702 | } | |
703 | #endif | |
704 | ||
bf0f6f24 IM |
705 | /* |
706 | * Update the current task's runtime statistics. Skip current tasks that | |
707 | * are not in our scheduling class. | |
708 | */ | |
709 | static inline void | |
8ebc91d9 IM |
710 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
711 | unsigned long delta_exec) | |
bf0f6f24 | 712 | { |
bbdba7c0 | 713 | unsigned long delta_exec_weighted; |
bf0f6f24 | 714 | |
41acab88 LDM |
715 | schedstat_set(curr->statistics.exec_max, |
716 | max((u64)delta_exec, curr->statistics.exec_max)); | |
bf0f6f24 IM |
717 | |
718 | curr->sum_exec_runtime += delta_exec; | |
7a62eabc | 719 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
a7be37ac | 720 | delta_exec_weighted = calc_delta_fair(delta_exec, curr); |
88ec22d3 | 721 | |
e9acbff6 | 722 | curr->vruntime += delta_exec_weighted; |
1af5f730 | 723 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
724 | } |
725 | ||
b7cc0896 | 726 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 727 | { |
429d43bc | 728 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 729 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
730 | unsigned long delta_exec; |
731 | ||
732 | if (unlikely(!curr)) | |
733 | return; | |
734 | ||
735 | /* | |
736 | * Get the amount of time the current task was running | |
737 | * since the last time we changed load (this cannot | |
738 | * overflow on 32 bits): | |
739 | */ | |
8ebc91d9 | 740 | delta_exec = (unsigned long)(now - curr->exec_start); |
34f28ecd PZ |
741 | if (!delta_exec) |
742 | return; | |
bf0f6f24 | 743 | |
8ebc91d9 IM |
744 | __update_curr(cfs_rq, curr, delta_exec); |
745 | curr->exec_start = now; | |
d842de87 SV |
746 | |
747 | if (entity_is_task(curr)) { | |
748 | struct task_struct *curtask = task_of(curr); | |
749 | ||
f977bb49 | 750 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 751 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 752 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 753 | } |
ec12cb7f PT |
754 | |
755 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
756 | } |
757 | ||
758 | static inline void | |
5870db5b | 759 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 760 | { |
78becc27 | 761 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
762 | } |
763 | ||
bf0f6f24 IM |
764 | /* |
765 | * Task is being enqueued - update stats: | |
766 | */ | |
d2417e5a | 767 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 768 | { |
bf0f6f24 IM |
769 | /* |
770 | * Are we enqueueing a waiting task? (for current tasks | |
771 | * a dequeue/enqueue event is a NOP) | |
772 | */ | |
429d43bc | 773 | if (se != cfs_rq->curr) |
5870db5b | 774 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
775 | } |
776 | ||
bf0f6f24 | 777 | static void |
9ef0a961 | 778 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 779 | { |
41acab88 | 780 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 781 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
782 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
783 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 784 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
785 | #ifdef CONFIG_SCHEDSTATS |
786 | if (entity_is_task(se)) { | |
787 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 788 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
789 | } |
790 | #endif | |
41acab88 | 791 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
792 | } |
793 | ||
794 | static inline void | |
19b6a2e3 | 795 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 796 | { |
bf0f6f24 IM |
797 | /* |
798 | * Mark the end of the wait period if dequeueing a | |
799 | * waiting task: | |
800 | */ | |
429d43bc | 801 | if (se != cfs_rq->curr) |
9ef0a961 | 802 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
803 | } |
804 | ||
805 | /* | |
806 | * We are picking a new current task - update its stats: | |
807 | */ | |
808 | static inline void | |
79303e9e | 809 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
810 | { |
811 | /* | |
812 | * We are starting a new run period: | |
813 | */ | |
78becc27 | 814 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
815 | } |
816 | ||
bf0f6f24 IM |
817 | /************************************************** |
818 | * Scheduling class queueing methods: | |
819 | */ | |
820 | ||
cbee9f88 PZ |
821 | #ifdef CONFIG_NUMA_BALANCING |
822 | /* | |
598f0ec0 MG |
823 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
824 | * calculated based on the tasks virtual memory size and | |
825 | * numa_balancing_scan_size. | |
cbee9f88 | 826 | */ |
598f0ec0 MG |
827 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
828 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
829 | unsigned int sysctl_numa_balancing_scan_period_reset = 60000; | |
6e5fb223 PZ |
830 | |
831 | /* Portion of address space to scan in MB */ | |
832 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 833 | |
4b96a29b PZ |
834 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
835 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
836 | ||
598f0ec0 MG |
837 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
838 | { | |
839 | unsigned long rss = 0; | |
840 | unsigned long nr_scan_pages; | |
841 | ||
842 | /* | |
843 | * Calculations based on RSS as non-present and empty pages are skipped | |
844 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
845 | * on resident pages | |
846 | */ | |
847 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
848 | rss = get_mm_rss(p->mm); | |
849 | if (!rss) | |
850 | rss = nr_scan_pages; | |
851 | ||
852 | rss = round_up(rss, nr_scan_pages); | |
853 | return rss / nr_scan_pages; | |
854 | } | |
855 | ||
856 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
857 | #define MAX_SCAN_WINDOW 2560 | |
858 | ||
859 | static unsigned int task_scan_min(struct task_struct *p) | |
860 | { | |
861 | unsigned int scan, floor; | |
862 | unsigned int windows = 1; | |
863 | ||
864 | if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW) | |
865 | windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size; | |
866 | floor = 1000 / windows; | |
867 | ||
868 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
869 | return max_t(unsigned int, floor, scan); | |
870 | } | |
871 | ||
872 | static unsigned int task_scan_max(struct task_struct *p) | |
873 | { | |
874 | unsigned int smin = task_scan_min(p); | |
875 | unsigned int smax; | |
876 | ||
877 | /* Watch for min being lower than max due to floor calculations */ | |
878 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
879 | return max(smin, smax); | |
880 | } | |
881 | ||
3a7053b3 MG |
882 | /* |
883 | * Once a preferred node is selected the scheduler balancer will prefer moving | |
884 | * a task to that node for sysctl_numa_balancing_settle_count number of PTE | |
885 | * scans. This will give the process the chance to accumulate more faults on | |
886 | * the preferred node but still allow the scheduler to move the task again if | |
887 | * the nodes CPUs are overloaded. | |
888 | */ | |
6fe6b2d6 | 889 | unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4; |
3a7053b3 | 890 | |
8c8a743c PZ |
891 | struct numa_group { |
892 | atomic_t refcount; | |
893 | ||
894 | spinlock_t lock; /* nr_tasks, tasks */ | |
895 | int nr_tasks; | |
896 | struct list_head task_list; | |
897 | ||
898 | struct rcu_head rcu; | |
899 | atomic_long_t faults[0]; | |
900 | }; | |
901 | ||
ac8e895b MG |
902 | static inline int task_faults_idx(int nid, int priv) |
903 | { | |
904 | return 2 * nid + priv; | |
905 | } | |
906 | ||
907 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
908 | { | |
909 | if (!p->numa_faults) | |
910 | return 0; | |
911 | ||
912 | return p->numa_faults[task_faults_idx(nid, 0)] + | |
913 | p->numa_faults[task_faults_idx(nid, 1)]; | |
914 | } | |
915 | ||
e6628d5b | 916 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
917 | static unsigned long source_load(int cpu, int type); |
918 | static unsigned long target_load(int cpu, int type); | |
919 | static unsigned long power_of(int cpu); | |
920 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
921 | ||
fb13c7ee | 922 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 923 | struct numa_stats { |
fb13c7ee | 924 | unsigned long nr_running; |
58d081b5 | 925 | unsigned long load; |
fb13c7ee MG |
926 | |
927 | /* Total compute capacity of CPUs on a node */ | |
928 | unsigned long power; | |
929 | ||
930 | /* Approximate capacity in terms of runnable tasks on a node */ | |
931 | unsigned long capacity; | |
932 | int has_capacity; | |
58d081b5 | 933 | }; |
e6628d5b | 934 | |
fb13c7ee MG |
935 | /* |
936 | * XXX borrowed from update_sg_lb_stats | |
937 | */ | |
938 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
939 | { | |
940 | int cpu; | |
941 | ||
942 | memset(ns, 0, sizeof(*ns)); | |
943 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
944 | struct rq *rq = cpu_rq(cpu); | |
945 | ||
946 | ns->nr_running += rq->nr_running; | |
947 | ns->load += weighted_cpuload(cpu); | |
948 | ns->power += power_of(cpu); | |
949 | } | |
950 | ||
951 | ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power; | |
952 | ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE); | |
953 | ns->has_capacity = (ns->nr_running < ns->capacity); | |
954 | } | |
955 | ||
58d081b5 MG |
956 | struct task_numa_env { |
957 | struct task_struct *p; | |
e6628d5b | 958 | |
58d081b5 MG |
959 | int src_cpu, src_nid; |
960 | int dst_cpu, dst_nid; | |
e6628d5b | 961 | |
58d081b5 | 962 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 963 | |
fb13c7ee MG |
964 | int imbalance_pct, idx; |
965 | ||
966 | struct task_struct *best_task; | |
967 | long best_imp; | |
58d081b5 MG |
968 | int best_cpu; |
969 | }; | |
970 | ||
fb13c7ee MG |
971 | static void task_numa_assign(struct task_numa_env *env, |
972 | struct task_struct *p, long imp) | |
973 | { | |
974 | if (env->best_task) | |
975 | put_task_struct(env->best_task); | |
976 | if (p) | |
977 | get_task_struct(p); | |
978 | ||
979 | env->best_task = p; | |
980 | env->best_imp = imp; | |
981 | env->best_cpu = env->dst_cpu; | |
982 | } | |
983 | ||
984 | /* | |
985 | * This checks if the overall compute and NUMA accesses of the system would | |
986 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
987 | * into account that it might be best if task running on the dst_cpu should | |
988 | * be exchanged with the source task | |
989 | */ | |
990 | static void task_numa_compare(struct task_numa_env *env, long imp) | |
991 | { | |
992 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
993 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
994 | struct task_struct *cur; | |
995 | long dst_load, src_load; | |
996 | long load; | |
997 | ||
998 | rcu_read_lock(); | |
999 | cur = ACCESS_ONCE(dst_rq->curr); | |
1000 | if (cur->pid == 0) /* idle */ | |
1001 | cur = NULL; | |
1002 | ||
1003 | /* | |
1004 | * "imp" is the fault differential for the source task between the | |
1005 | * source and destination node. Calculate the total differential for | |
1006 | * the source task and potential destination task. The more negative | |
1007 | * the value is, the more rmeote accesses that would be expected to | |
1008 | * be incurred if the tasks were swapped. | |
1009 | */ | |
1010 | if (cur) { | |
1011 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1012 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1013 | goto unlock; | |
1014 | ||
1015 | imp += task_faults(cur, env->src_nid) - | |
1016 | task_faults(cur, env->dst_nid); | |
1017 | } | |
1018 | ||
1019 | if (imp < env->best_imp) | |
1020 | goto unlock; | |
1021 | ||
1022 | if (!cur) { | |
1023 | /* Is there capacity at our destination? */ | |
1024 | if (env->src_stats.has_capacity && | |
1025 | !env->dst_stats.has_capacity) | |
1026 | goto unlock; | |
1027 | ||
1028 | goto balance; | |
1029 | } | |
1030 | ||
1031 | /* Balance doesn't matter much if we're running a task per cpu */ | |
1032 | if (src_rq->nr_running == 1 && dst_rq->nr_running == 1) | |
1033 | goto assign; | |
1034 | ||
1035 | /* | |
1036 | * In the overloaded case, try and keep the load balanced. | |
1037 | */ | |
1038 | balance: | |
1039 | dst_load = env->dst_stats.load; | |
1040 | src_load = env->src_stats.load; | |
1041 | ||
1042 | /* XXX missing power terms */ | |
1043 | load = task_h_load(env->p); | |
1044 | dst_load += load; | |
1045 | src_load -= load; | |
1046 | ||
1047 | if (cur) { | |
1048 | load = task_h_load(cur); | |
1049 | dst_load -= load; | |
1050 | src_load += load; | |
1051 | } | |
1052 | ||
1053 | /* make src_load the smaller */ | |
1054 | if (dst_load < src_load) | |
1055 | swap(dst_load, src_load); | |
1056 | ||
1057 | if (src_load * env->imbalance_pct < dst_load * 100) | |
1058 | goto unlock; | |
1059 | ||
1060 | assign: | |
1061 | task_numa_assign(env, cur, imp); | |
1062 | unlock: | |
1063 | rcu_read_unlock(); | |
1064 | } | |
1065 | ||
2c8a50aa MG |
1066 | static void task_numa_find_cpu(struct task_numa_env *env, long imp) |
1067 | { | |
1068 | int cpu; | |
1069 | ||
1070 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1071 | /* Skip this CPU if the source task cannot migrate */ | |
1072 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1073 | continue; | |
1074 | ||
1075 | env->dst_cpu = cpu; | |
1076 | task_numa_compare(env, imp); | |
1077 | } | |
1078 | } | |
1079 | ||
58d081b5 MG |
1080 | static int task_numa_migrate(struct task_struct *p) |
1081 | { | |
58d081b5 MG |
1082 | struct task_numa_env env = { |
1083 | .p = p, | |
fb13c7ee | 1084 | |
58d081b5 MG |
1085 | .src_cpu = task_cpu(p), |
1086 | .src_nid = cpu_to_node(task_cpu(p)), | |
fb13c7ee MG |
1087 | |
1088 | .imbalance_pct = 112, | |
1089 | ||
1090 | .best_task = NULL, | |
1091 | .best_imp = 0, | |
1092 | .best_cpu = -1 | |
58d081b5 MG |
1093 | }; |
1094 | struct sched_domain *sd; | |
fb13c7ee | 1095 | unsigned long faults; |
2c8a50aa MG |
1096 | int nid, ret; |
1097 | long imp; | |
e6628d5b | 1098 | |
58d081b5 | 1099 | /* |
fb13c7ee MG |
1100 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1101 | * imbalance and would be the first to start moving tasks about. | |
1102 | * | |
1103 | * And we want to avoid any moving of tasks about, as that would create | |
1104 | * random movement of tasks -- counter the numa conditions we're trying | |
1105 | * to satisfy here. | |
58d081b5 MG |
1106 | */ |
1107 | rcu_read_lock(); | |
fb13c7ee MG |
1108 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
1109 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1110 | rcu_read_unlock(); |
1111 | ||
fb13c7ee MG |
1112 | faults = task_faults(p, env.src_nid); |
1113 | update_numa_stats(&env.src_stats, env.src_nid); | |
2c8a50aa MG |
1114 | env.dst_nid = p->numa_preferred_nid; |
1115 | imp = task_faults(env.p, env.dst_nid) - faults; | |
1116 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
58d081b5 | 1117 | |
e1dda8a7 RR |
1118 | /* If the preferred nid has capacity, try to use it. */ |
1119 | if (env.dst_stats.has_capacity) | |
2c8a50aa | 1120 | task_numa_find_cpu(&env, imp); |
e1dda8a7 RR |
1121 | |
1122 | /* No space available on the preferred nid. Look elsewhere. */ | |
1123 | if (env.best_cpu == -1) { | |
2c8a50aa MG |
1124 | for_each_online_node(nid) { |
1125 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1126 | continue; | |
58d081b5 | 1127 | |
2c8a50aa MG |
1128 | /* Only consider nodes that recorded more faults */ |
1129 | imp = task_faults(env.p, nid) - faults; | |
1130 | if (imp < 0) | |
fb13c7ee MG |
1131 | continue; |
1132 | ||
2c8a50aa MG |
1133 | env.dst_nid = nid; |
1134 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
1135 | task_numa_find_cpu(&env, imp); | |
58d081b5 MG |
1136 | } |
1137 | } | |
1138 | ||
fb13c7ee MG |
1139 | /* No better CPU than the current one was found. */ |
1140 | if (env.best_cpu == -1) | |
1141 | return -EAGAIN; | |
1142 | ||
1143 | if (env.best_task == NULL) { | |
1144 | int ret = migrate_task_to(p, env.best_cpu); | |
1145 | return ret; | |
1146 | } | |
1147 | ||
1148 | ret = migrate_swap(p, env.best_task); | |
1149 | put_task_struct(env.best_task); | |
1150 | return ret; | |
e6628d5b MG |
1151 | } |
1152 | ||
6b9a7460 MG |
1153 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1154 | static void numa_migrate_preferred(struct task_struct *p) | |
1155 | { | |
1156 | /* Success if task is already running on preferred CPU */ | |
1157 | p->numa_migrate_retry = 0; | |
06ea5e03 RR |
1158 | if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) { |
1159 | /* | |
1160 | * If migration is temporarily disabled due to a task migration | |
1161 | * then re-enable it now as the task is running on its | |
1162 | * preferred node and memory should migrate locally | |
1163 | */ | |
1164 | if (!p->numa_migrate_seq) | |
1165 | p->numa_migrate_seq++; | |
6b9a7460 | 1166 | return; |
06ea5e03 | 1167 | } |
6b9a7460 MG |
1168 | |
1169 | /* This task has no NUMA fault statistics yet */ | |
1170 | if (unlikely(p->numa_preferred_nid == -1)) | |
1171 | return; | |
1172 | ||
1173 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
1174 | if (task_numa_migrate(p) != 0) | |
1175 | p->numa_migrate_retry = jiffies + HZ*5; | |
1176 | } | |
1177 | ||
cbee9f88 PZ |
1178 | static void task_numa_placement(struct task_struct *p) |
1179 | { | |
688b7585 MG |
1180 | int seq, nid, max_nid = -1; |
1181 | unsigned long max_faults = 0; | |
cbee9f88 | 1182 | |
2832bc19 | 1183 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1184 | if (p->numa_scan_seq == seq) |
1185 | return; | |
1186 | p->numa_scan_seq = seq; | |
3a7053b3 | 1187 | p->numa_migrate_seq++; |
598f0ec0 | 1188 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1189 | |
688b7585 MG |
1190 | /* Find the node with the highest number of faults */ |
1191 | for_each_online_node(nid) { | |
fb13c7ee | 1192 | unsigned long faults = 0; |
ac8e895b | 1193 | int priv, i; |
745d6147 | 1194 | |
ac8e895b | 1195 | for (priv = 0; priv < 2; priv++) { |
8c8a743c PZ |
1196 | long diff; |
1197 | ||
ac8e895b | 1198 | i = task_faults_idx(nid, priv); |
8c8a743c | 1199 | diff = -p->numa_faults[i]; |
745d6147 | 1200 | |
ac8e895b MG |
1201 | /* Decay existing window, copy faults since last scan */ |
1202 | p->numa_faults[i] >>= 1; | |
1203 | p->numa_faults[i] += p->numa_faults_buffer[i]; | |
1204 | p->numa_faults_buffer[i] = 0; | |
fb13c7ee MG |
1205 | |
1206 | faults += p->numa_faults[i]; | |
8c8a743c PZ |
1207 | diff += p->numa_faults[i]; |
1208 | if (p->numa_group) { | |
1209 | /* safe because we can only change our own group */ | |
1210 | atomic_long_add(diff, &p->numa_group->faults[i]); | |
1211 | } | |
ac8e895b MG |
1212 | } |
1213 | ||
688b7585 MG |
1214 | if (faults > max_faults) { |
1215 | max_faults = faults; | |
1216 | max_nid = nid; | |
1217 | } | |
1218 | } | |
1219 | ||
6b9a7460 | 1220 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1221 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1222 | /* Update the preferred nid and migrate task if possible */ |
688b7585 | 1223 | p->numa_preferred_nid = max_nid; |
6fe6b2d6 | 1224 | p->numa_migrate_seq = 1; |
6b9a7460 | 1225 | numa_migrate_preferred(p); |
3a7053b3 | 1226 | } |
cbee9f88 PZ |
1227 | } |
1228 | ||
8c8a743c PZ |
1229 | static inline int get_numa_group(struct numa_group *grp) |
1230 | { | |
1231 | return atomic_inc_not_zero(&grp->refcount); | |
1232 | } | |
1233 | ||
1234 | static inline void put_numa_group(struct numa_group *grp) | |
1235 | { | |
1236 | if (atomic_dec_and_test(&grp->refcount)) | |
1237 | kfree_rcu(grp, rcu); | |
1238 | } | |
1239 | ||
1240 | static void double_lock(spinlock_t *l1, spinlock_t *l2) | |
1241 | { | |
1242 | if (l1 > l2) | |
1243 | swap(l1, l2); | |
1244 | ||
1245 | spin_lock(l1); | |
1246 | spin_lock_nested(l2, SINGLE_DEPTH_NESTING); | |
1247 | } | |
1248 | ||
1249 | static void task_numa_group(struct task_struct *p, int cpupid) | |
1250 | { | |
1251 | struct numa_group *grp, *my_grp; | |
1252 | struct task_struct *tsk; | |
1253 | bool join = false; | |
1254 | int cpu = cpupid_to_cpu(cpupid); | |
1255 | int i; | |
1256 | ||
1257 | if (unlikely(!p->numa_group)) { | |
1258 | unsigned int size = sizeof(struct numa_group) + | |
1259 | 2*nr_node_ids*sizeof(atomic_long_t); | |
1260 | ||
1261 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1262 | if (!grp) | |
1263 | return; | |
1264 | ||
1265 | atomic_set(&grp->refcount, 1); | |
1266 | spin_lock_init(&grp->lock); | |
1267 | INIT_LIST_HEAD(&grp->task_list); | |
1268 | ||
1269 | for (i = 0; i < 2*nr_node_ids; i++) | |
1270 | atomic_long_set(&grp->faults[i], p->numa_faults[i]); | |
1271 | ||
1272 | list_add(&p->numa_entry, &grp->task_list); | |
1273 | grp->nr_tasks++; | |
1274 | rcu_assign_pointer(p->numa_group, grp); | |
1275 | } | |
1276 | ||
1277 | rcu_read_lock(); | |
1278 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1279 | ||
1280 | if (!cpupid_match_pid(tsk, cpupid)) | |
1281 | goto unlock; | |
1282 | ||
1283 | grp = rcu_dereference(tsk->numa_group); | |
1284 | if (!grp) | |
1285 | goto unlock; | |
1286 | ||
1287 | my_grp = p->numa_group; | |
1288 | if (grp == my_grp) | |
1289 | goto unlock; | |
1290 | ||
1291 | /* | |
1292 | * Only join the other group if its bigger; if we're the bigger group, | |
1293 | * the other task will join us. | |
1294 | */ | |
1295 | if (my_grp->nr_tasks > grp->nr_tasks) | |
1296 | goto unlock; | |
1297 | ||
1298 | /* | |
1299 | * Tie-break on the grp address. | |
1300 | */ | |
1301 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
1302 | goto unlock; | |
1303 | ||
1304 | if (!get_numa_group(grp)) | |
1305 | goto unlock; | |
1306 | ||
1307 | join = true; | |
1308 | ||
1309 | unlock: | |
1310 | rcu_read_unlock(); | |
1311 | ||
1312 | if (!join) | |
1313 | return; | |
1314 | ||
1315 | for (i = 0; i < 2*nr_node_ids; i++) { | |
1316 | atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]); | |
1317 | atomic_long_add(p->numa_faults[i], &grp->faults[i]); | |
1318 | } | |
1319 | ||
1320 | double_lock(&my_grp->lock, &grp->lock); | |
1321 | ||
1322 | list_move(&p->numa_entry, &grp->task_list); | |
1323 | my_grp->nr_tasks--; | |
1324 | grp->nr_tasks++; | |
1325 | ||
1326 | spin_unlock(&my_grp->lock); | |
1327 | spin_unlock(&grp->lock); | |
1328 | ||
1329 | rcu_assign_pointer(p->numa_group, grp); | |
1330 | ||
1331 | put_numa_group(my_grp); | |
1332 | } | |
1333 | ||
1334 | void task_numa_free(struct task_struct *p) | |
1335 | { | |
1336 | struct numa_group *grp = p->numa_group; | |
1337 | int i; | |
1338 | ||
1339 | if (grp) { | |
1340 | for (i = 0; i < 2*nr_node_ids; i++) | |
1341 | atomic_long_sub(p->numa_faults[i], &grp->faults[i]); | |
1342 | ||
1343 | spin_lock(&grp->lock); | |
1344 | list_del(&p->numa_entry); | |
1345 | grp->nr_tasks--; | |
1346 | spin_unlock(&grp->lock); | |
1347 | rcu_assign_pointer(p->numa_group, NULL); | |
1348 | put_numa_group(grp); | |
1349 | } | |
1350 | ||
1351 | kfree(p->numa_faults); | |
1352 | } | |
1353 | ||
cbee9f88 PZ |
1354 | /* |
1355 | * Got a PROT_NONE fault for a page on @node. | |
1356 | */ | |
90572890 | 1357 | void task_numa_fault(int last_cpupid, int node, int pages, bool migrated) |
cbee9f88 PZ |
1358 | { |
1359 | struct task_struct *p = current; | |
ac8e895b | 1360 | int priv; |
cbee9f88 | 1361 | |
10e84b97 | 1362 | if (!numabalancing_enabled) |
1a687c2e MG |
1363 | return; |
1364 | ||
9ff1d9ff MG |
1365 | /* for example, ksmd faulting in a user's mm */ |
1366 | if (!p->mm) | |
1367 | return; | |
1368 | ||
f809ca9a MG |
1369 | /* Allocate buffer to track faults on a per-node basis */ |
1370 | if (unlikely(!p->numa_faults)) { | |
ac8e895b | 1371 | int size = sizeof(*p->numa_faults) * 2 * nr_node_ids; |
f809ca9a | 1372 | |
745d6147 MG |
1373 | /* numa_faults and numa_faults_buffer share the allocation */ |
1374 | p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN); | |
f809ca9a MG |
1375 | if (!p->numa_faults) |
1376 | return; | |
745d6147 MG |
1377 | |
1378 | BUG_ON(p->numa_faults_buffer); | |
ac8e895b | 1379 | p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids); |
f809ca9a | 1380 | } |
cbee9f88 | 1381 | |
8c8a743c PZ |
1382 | /* |
1383 | * First accesses are treated as private, otherwise consider accesses | |
1384 | * to be private if the accessing pid has not changed | |
1385 | */ | |
1386 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
1387 | priv = 1; | |
1388 | } else { | |
1389 | priv = cpupid_match_pid(p, last_cpupid); | |
1390 | if (!priv) | |
1391 | task_numa_group(p, last_cpupid); | |
1392 | } | |
1393 | ||
fb003b80 | 1394 | /* |
b8593bfd MG |
1395 | * If pages are properly placed (did not migrate) then scan slower. |
1396 | * This is reset periodically in case of phase changes | |
fb003b80 | 1397 | */ |
598f0ec0 MG |
1398 | if (!migrated) { |
1399 | /* Initialise if necessary */ | |
1400 | if (!p->numa_scan_period_max) | |
1401 | p->numa_scan_period_max = task_scan_max(p); | |
1402 | ||
1403 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1404 | p->numa_scan_period + 10); | |
1405 | } | |
fb003b80 | 1406 | |
cbee9f88 | 1407 | task_numa_placement(p); |
f809ca9a | 1408 | |
6b9a7460 MG |
1409 | /* Retry task to preferred node migration if it previously failed */ |
1410 | if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry)) | |
1411 | numa_migrate_preferred(p); | |
1412 | ||
ac8e895b | 1413 | p->numa_faults_buffer[task_faults_idx(node, priv)] += pages; |
cbee9f88 PZ |
1414 | } |
1415 | ||
6e5fb223 PZ |
1416 | static void reset_ptenuma_scan(struct task_struct *p) |
1417 | { | |
1418 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1419 | p->mm->numa_scan_offset = 0; | |
1420 | } | |
1421 | ||
cbee9f88 PZ |
1422 | /* |
1423 | * The expensive part of numa migration is done from task_work context. | |
1424 | * Triggered from task_tick_numa(). | |
1425 | */ | |
1426 | void task_numa_work(struct callback_head *work) | |
1427 | { | |
1428 | unsigned long migrate, next_scan, now = jiffies; | |
1429 | struct task_struct *p = current; | |
1430 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1431 | struct vm_area_struct *vma; |
9f40604c | 1432 | unsigned long start, end; |
598f0ec0 | 1433 | unsigned long nr_pte_updates = 0; |
9f40604c | 1434 | long pages; |
cbee9f88 PZ |
1435 | |
1436 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1437 | ||
1438 | work->next = work; /* protect against double add */ | |
1439 | /* | |
1440 | * Who cares about NUMA placement when they're dying. | |
1441 | * | |
1442 | * NOTE: make sure not to dereference p->mm before this check, | |
1443 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1444 | * without p->mm even though we still had it when we enqueued this | |
1445 | * work. | |
1446 | */ | |
1447 | if (p->flags & PF_EXITING) | |
1448 | return; | |
1449 | ||
7e8d16b6 MG |
1450 | if (!mm->numa_next_reset || !mm->numa_next_scan) { |
1451 | mm->numa_next_scan = now + | |
1452 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
1453 | mm->numa_next_reset = now + | |
1454 | msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
1455 | } | |
1456 | ||
b8593bfd MG |
1457 | /* |
1458 | * Reset the scan period if enough time has gone by. Objective is that | |
1459 | * scanning will be reduced if pages are properly placed. As tasks | |
1460 | * can enter different phases this needs to be re-examined. Lacking | |
1461 | * proper tracking of reference behaviour, this blunt hammer is used. | |
1462 | */ | |
1463 | migrate = mm->numa_next_reset; | |
1464 | if (time_after(now, migrate)) { | |
598f0ec0 | 1465 | p->numa_scan_period = task_scan_min(p); |
b8593bfd MG |
1466 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); |
1467 | xchg(&mm->numa_next_reset, next_scan); | |
1468 | } | |
1469 | ||
cbee9f88 PZ |
1470 | /* |
1471 | * Enforce maximal scan/migration frequency.. | |
1472 | */ | |
1473 | migrate = mm->numa_next_scan; | |
1474 | if (time_before(now, migrate)) | |
1475 | return; | |
1476 | ||
598f0ec0 MG |
1477 | if (p->numa_scan_period == 0) { |
1478 | p->numa_scan_period_max = task_scan_max(p); | |
1479 | p->numa_scan_period = task_scan_min(p); | |
1480 | } | |
cbee9f88 | 1481 | |
fb003b80 | 1482 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1483 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1484 | return; | |
1485 | ||
19a78d11 PZ |
1486 | /* |
1487 | * Delay this task enough that another task of this mm will likely win | |
1488 | * the next time around. | |
1489 | */ | |
1490 | p->node_stamp += 2 * TICK_NSEC; | |
1491 | ||
9f40604c MG |
1492 | start = mm->numa_scan_offset; |
1493 | pages = sysctl_numa_balancing_scan_size; | |
1494 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1495 | if (!pages) | |
1496 | return; | |
cbee9f88 | 1497 | |
6e5fb223 | 1498 | down_read(&mm->mmap_sem); |
9f40604c | 1499 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1500 | if (!vma) { |
1501 | reset_ptenuma_scan(p); | |
9f40604c | 1502 | start = 0; |
6e5fb223 PZ |
1503 | vma = mm->mmap; |
1504 | } | |
9f40604c | 1505 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1506 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1507 | continue; |
1508 | ||
4591ce4f MG |
1509 | /* |
1510 | * Shared library pages mapped by multiple processes are not | |
1511 | * migrated as it is expected they are cache replicated. Avoid | |
1512 | * hinting faults in read-only file-backed mappings or the vdso | |
1513 | * as migrating the pages will be of marginal benefit. | |
1514 | */ | |
1515 | if (!vma->vm_mm || | |
1516 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1517 | continue; | |
1518 | ||
9f40604c MG |
1519 | do { |
1520 | start = max(start, vma->vm_start); | |
1521 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1522 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1523 | nr_pte_updates += change_prot_numa(vma, start, end); |
1524 | ||
1525 | /* | |
1526 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1527 | * at least one PTE is updated so that unused virtual | |
1528 | * address space is quickly skipped. | |
1529 | */ | |
1530 | if (nr_pte_updates) | |
1531 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1532 | |
9f40604c MG |
1533 | start = end; |
1534 | if (pages <= 0) | |
1535 | goto out; | |
1536 | } while (end != vma->vm_end); | |
cbee9f88 | 1537 | } |
6e5fb223 | 1538 | |
9f40604c | 1539 | out: |
f307cd1a MG |
1540 | /* |
1541 | * If the whole process was scanned without updates then no NUMA | |
1542 | * hinting faults are being recorded and scan rate should be lower. | |
1543 | */ | |
1544 | if (mm->numa_scan_offset == 0 && !nr_pte_updates) { | |
1545 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1546 | p->numa_scan_period << 1); | |
1547 | ||
1548 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); | |
1549 | mm->numa_next_scan = next_scan; | |
1550 | } | |
1551 | ||
6e5fb223 | 1552 | /* |
c69307d5 PZ |
1553 | * It is possible to reach the end of the VMA list but the last few |
1554 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1555 | * would find the !migratable VMA on the next scan but not reset the | |
1556 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1557 | */ |
1558 | if (vma) | |
9f40604c | 1559 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1560 | else |
1561 | reset_ptenuma_scan(p); | |
1562 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1563 | } |
1564 | ||
1565 | /* | |
1566 | * Drive the periodic memory faults.. | |
1567 | */ | |
1568 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1569 | { | |
1570 | struct callback_head *work = &curr->numa_work; | |
1571 | u64 period, now; | |
1572 | ||
1573 | /* | |
1574 | * We don't care about NUMA placement if we don't have memory. | |
1575 | */ | |
1576 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1577 | return; | |
1578 | ||
1579 | /* | |
1580 | * Using runtime rather than walltime has the dual advantage that | |
1581 | * we (mostly) drive the selection from busy threads and that the | |
1582 | * task needs to have done some actual work before we bother with | |
1583 | * NUMA placement. | |
1584 | */ | |
1585 | now = curr->se.sum_exec_runtime; | |
1586 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1587 | ||
1588 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1589 | if (!curr->node_stamp) |
598f0ec0 | 1590 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1591 | curr->node_stamp += period; |
cbee9f88 PZ |
1592 | |
1593 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1594 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1595 | task_work_add(curr, work, true); | |
1596 | } | |
1597 | } | |
1598 | } | |
1599 | #else | |
1600 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1601 | { | |
1602 | } | |
1603 | #endif /* CONFIG_NUMA_BALANCING */ | |
1604 | ||
30cfdcfc DA |
1605 | static void |
1606 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1607 | { | |
1608 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1609 | if (!parent_entity(se)) |
029632fb | 1610 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
1611 | #ifdef CONFIG_SMP |
1612 | if (entity_is_task(se)) | |
eb95308e | 1613 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 1614 | #endif |
30cfdcfc | 1615 | cfs_rq->nr_running++; |
30cfdcfc DA |
1616 | } |
1617 | ||
1618 | static void | |
1619 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1620 | { | |
1621 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1622 | if (!parent_entity(se)) |
029632fb | 1623 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1624 | if (entity_is_task(se)) |
b87f1724 | 1625 | list_del_init(&se->group_node); |
30cfdcfc | 1626 | cfs_rq->nr_running--; |
30cfdcfc DA |
1627 | } |
1628 | ||
3ff6dcac YZ |
1629 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1630 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1631 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1632 | { | |
1633 | long tg_weight; | |
1634 | ||
1635 | /* | |
1636 | * Use this CPU's actual weight instead of the last load_contribution | |
1637 | * to gain a more accurate current total weight. See | |
1638 | * update_cfs_rq_load_contribution(). | |
1639 | */ | |
bf5b986e | 1640 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 1641 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
1642 | tg_weight += cfs_rq->load.weight; |
1643 | ||
1644 | return tg_weight; | |
1645 | } | |
1646 | ||
6d5ab293 | 1647 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1648 | { |
cf5f0acf | 1649 | long tg_weight, load, shares; |
3ff6dcac | 1650 | |
cf5f0acf | 1651 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1652 | load = cfs_rq->load.weight; |
3ff6dcac | 1653 | |
3ff6dcac | 1654 | shares = (tg->shares * load); |
cf5f0acf PZ |
1655 | if (tg_weight) |
1656 | shares /= tg_weight; | |
3ff6dcac YZ |
1657 | |
1658 | if (shares < MIN_SHARES) | |
1659 | shares = MIN_SHARES; | |
1660 | if (shares > tg->shares) | |
1661 | shares = tg->shares; | |
1662 | ||
1663 | return shares; | |
1664 | } | |
3ff6dcac | 1665 | # else /* CONFIG_SMP */ |
6d5ab293 | 1666 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1667 | { |
1668 | return tg->shares; | |
1669 | } | |
3ff6dcac | 1670 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1671 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1672 | unsigned long weight) | |
1673 | { | |
19e5eebb PT |
1674 | if (se->on_rq) { |
1675 | /* commit outstanding execution time */ | |
1676 | if (cfs_rq->curr == se) | |
1677 | update_curr(cfs_rq); | |
2069dd75 | 1678 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1679 | } |
2069dd75 PZ |
1680 | |
1681 | update_load_set(&se->load, weight); | |
1682 | ||
1683 | if (se->on_rq) | |
1684 | account_entity_enqueue(cfs_rq, se); | |
1685 | } | |
1686 | ||
82958366 PT |
1687 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1688 | ||
6d5ab293 | 1689 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1690 | { |
1691 | struct task_group *tg; | |
1692 | struct sched_entity *se; | |
3ff6dcac | 1693 | long shares; |
2069dd75 | 1694 | |
2069dd75 PZ |
1695 | tg = cfs_rq->tg; |
1696 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1697 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1698 | return; |
3ff6dcac YZ |
1699 | #ifndef CONFIG_SMP |
1700 | if (likely(se->load.weight == tg->shares)) | |
1701 | return; | |
1702 | #endif | |
6d5ab293 | 1703 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1704 | |
1705 | reweight_entity(cfs_rq_of(se), se, shares); | |
1706 | } | |
1707 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1708 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1709 | { |
1710 | } | |
1711 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1712 | ||
141965c7 | 1713 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
1714 | /* |
1715 | * We choose a half-life close to 1 scheduling period. | |
1716 | * Note: The tables below are dependent on this value. | |
1717 | */ | |
1718 | #define LOAD_AVG_PERIOD 32 | |
1719 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1720 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1721 | ||
1722 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1723 | static const u32 runnable_avg_yN_inv[] = { | |
1724 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1725 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1726 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1727 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1728 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1729 | 0x85aac367, 0x82cd8698, | |
1730 | }; | |
1731 | ||
1732 | /* | |
1733 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1734 | * over-estimates when re-combining. | |
1735 | */ | |
1736 | static const u32 runnable_avg_yN_sum[] = { | |
1737 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1738 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1739 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1740 | }; | |
1741 | ||
9d85f21c PT |
1742 | /* |
1743 | * Approximate: | |
1744 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1745 | */ | |
1746 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1747 | { | |
5b51f2f8 PT |
1748 | unsigned int local_n; |
1749 | ||
1750 | if (!n) | |
1751 | return val; | |
1752 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1753 | return 0; | |
1754 | ||
1755 | /* after bounds checking we can collapse to 32-bit */ | |
1756 | local_n = n; | |
1757 | ||
1758 | /* | |
1759 | * As y^PERIOD = 1/2, we can combine | |
1760 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
1761 | * With a look-up table which covers k^n (n<PERIOD) | |
1762 | * | |
1763 | * To achieve constant time decay_load. | |
1764 | */ | |
1765 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
1766 | val >>= local_n / LOAD_AVG_PERIOD; | |
1767 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
1768 | } |
1769 | ||
5b51f2f8 PT |
1770 | val *= runnable_avg_yN_inv[local_n]; |
1771 | /* We don't use SRR here since we always want to round down. */ | |
1772 | return val >> 32; | |
1773 | } | |
1774 | ||
1775 | /* | |
1776 | * For updates fully spanning n periods, the contribution to runnable | |
1777 | * average will be: \Sum 1024*y^n | |
1778 | * | |
1779 | * We can compute this reasonably efficiently by combining: | |
1780 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
1781 | */ | |
1782 | static u32 __compute_runnable_contrib(u64 n) | |
1783 | { | |
1784 | u32 contrib = 0; | |
1785 | ||
1786 | if (likely(n <= LOAD_AVG_PERIOD)) | |
1787 | return runnable_avg_yN_sum[n]; | |
1788 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
1789 | return LOAD_AVG_MAX; | |
1790 | ||
1791 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
1792 | do { | |
1793 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
1794 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
1795 | ||
1796 | n -= LOAD_AVG_PERIOD; | |
1797 | } while (n > LOAD_AVG_PERIOD); | |
1798 | ||
1799 | contrib = decay_load(contrib, n); | |
1800 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
1801 | } |
1802 | ||
1803 | /* | |
1804 | * We can represent the historical contribution to runnable average as the | |
1805 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
1806 | * history into segments of approximately 1ms (1024us); label the segment that | |
1807 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
1808 | * | |
1809 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
1810 | * p0 p1 p2 | |
1811 | * (now) (~1ms ago) (~2ms ago) | |
1812 | * | |
1813 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
1814 | * | |
1815 | * We then designate the fractions u_i as our co-efficients, yielding the | |
1816 | * following representation of historical load: | |
1817 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
1818 | * | |
1819 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
1820 | * y^32 = 0.5 | |
1821 | * | |
1822 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
1823 | * approximately half as much as the contribution to load within the last ms | |
1824 | * (u_0). | |
1825 | * | |
1826 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
1827 | * sum again by y is sufficient to update: | |
1828 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1829 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1830 | */ | |
1831 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1832 | struct sched_avg *sa, | |
1833 | int runnable) | |
1834 | { | |
5b51f2f8 PT |
1835 | u64 delta, periods; |
1836 | u32 runnable_contrib; | |
9d85f21c PT |
1837 | int delta_w, decayed = 0; |
1838 | ||
1839 | delta = now - sa->last_runnable_update; | |
1840 | /* | |
1841 | * This should only happen when time goes backwards, which it | |
1842 | * unfortunately does during sched clock init when we swap over to TSC. | |
1843 | */ | |
1844 | if ((s64)delta < 0) { | |
1845 | sa->last_runnable_update = now; | |
1846 | return 0; | |
1847 | } | |
1848 | ||
1849 | /* | |
1850 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1851 | * approximation of 1us and fast to compute. | |
1852 | */ | |
1853 | delta >>= 10; | |
1854 | if (!delta) | |
1855 | return 0; | |
1856 | sa->last_runnable_update = now; | |
1857 | ||
1858 | /* delta_w is the amount already accumulated against our next period */ | |
1859 | delta_w = sa->runnable_avg_period % 1024; | |
1860 | if (delta + delta_w >= 1024) { | |
1861 | /* period roll-over */ | |
1862 | decayed = 1; | |
1863 | ||
1864 | /* | |
1865 | * Now that we know we're crossing a period boundary, figure | |
1866 | * out how much from delta we need to complete the current | |
1867 | * period and accrue it. | |
1868 | */ | |
1869 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
1870 | if (runnable) |
1871 | sa->runnable_avg_sum += delta_w; | |
1872 | sa->runnable_avg_period += delta_w; | |
1873 | ||
1874 | delta -= delta_w; | |
1875 | ||
1876 | /* Figure out how many additional periods this update spans */ | |
1877 | periods = delta / 1024; | |
1878 | delta %= 1024; | |
1879 | ||
1880 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
1881 | periods + 1); | |
1882 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
1883 | periods + 1); | |
1884 | ||
1885 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
1886 | runnable_contrib = __compute_runnable_contrib(periods); | |
1887 | if (runnable) | |
1888 | sa->runnable_avg_sum += runnable_contrib; | |
1889 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
1890 | } |
1891 | ||
1892 | /* Remainder of delta accrued against u_0` */ | |
1893 | if (runnable) | |
1894 | sa->runnable_avg_sum += delta; | |
1895 | sa->runnable_avg_period += delta; | |
1896 | ||
1897 | return decayed; | |
1898 | } | |
1899 | ||
9ee474f5 | 1900 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 1901 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
1902 | { |
1903 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1904 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1905 | ||
1906 | decays -= se->avg.decay_count; | |
1907 | if (!decays) | |
aff3e498 | 1908 | return 0; |
9ee474f5 PT |
1909 | |
1910 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1911 | se->avg.decay_count = 0; | |
aff3e498 PT |
1912 | |
1913 | return decays; | |
9ee474f5 PT |
1914 | } |
1915 | ||
c566e8e9 PT |
1916 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1917 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1918 | int force_update) | |
1919 | { | |
1920 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 1921 | long tg_contrib; |
c566e8e9 PT |
1922 | |
1923 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
1924 | tg_contrib -= cfs_rq->tg_load_contrib; | |
1925 | ||
bf5b986e AS |
1926 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
1927 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
1928 | cfs_rq->tg_load_contrib += tg_contrib; |
1929 | } | |
1930 | } | |
8165e145 | 1931 | |
bb17f655 PT |
1932 | /* |
1933 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
1934 | * representation for computing load contributions. | |
1935 | */ | |
1936 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
1937 | struct cfs_rq *cfs_rq) | |
1938 | { | |
1939 | struct task_group *tg = cfs_rq->tg; | |
1940 | long contrib; | |
1941 | ||
1942 | /* The fraction of a cpu used by this cfs_rq */ | |
1943 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
1944 | sa->runnable_avg_period + 1); | |
1945 | contrib -= cfs_rq->tg_runnable_contrib; | |
1946 | ||
1947 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
1948 | atomic_add(contrib, &tg->runnable_avg); | |
1949 | cfs_rq->tg_runnable_contrib += contrib; | |
1950 | } | |
1951 | } | |
1952 | ||
8165e145 PT |
1953 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
1954 | { | |
1955 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
1956 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
1957 | int runnable_avg; |
1958 | ||
8165e145 PT |
1959 | u64 contrib; |
1960 | ||
1961 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
1962 | se->avg.load_avg_contrib = div_u64(contrib, |
1963 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
1964 | |
1965 | /* | |
1966 | * For group entities we need to compute a correction term in the case | |
1967 | * that they are consuming <1 cpu so that we would contribute the same | |
1968 | * load as a task of equal weight. | |
1969 | * | |
1970 | * Explicitly co-ordinating this measurement would be expensive, but | |
1971 | * fortunately the sum of each cpus contribution forms a usable | |
1972 | * lower-bound on the true value. | |
1973 | * | |
1974 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
1975 | * (and the sum represents true value) or they are disjoint and we are | |
1976 | * understating by the aggregate of their overlap. | |
1977 | * | |
1978 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
1979 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
1980 | * cpus that overlap for this interval and w_i is the interval width. | |
1981 | * | |
1982 | * On a small machine; the first term is well-bounded which bounds the | |
1983 | * total error since w_i is a subset of the period. Whereas on a | |
1984 | * larger machine, while this first term can be larger, if w_i is the | |
1985 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
1986 | * our upper bound of 1-cpu. | |
1987 | */ | |
1988 | runnable_avg = atomic_read(&tg->runnable_avg); | |
1989 | if (runnable_avg < NICE_0_LOAD) { | |
1990 | se->avg.load_avg_contrib *= runnable_avg; | |
1991 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
1992 | } | |
8165e145 | 1993 | } |
c566e8e9 PT |
1994 | #else |
1995 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1996 | int force_update) {} | |
bb17f655 PT |
1997 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
1998 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 1999 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
2000 | #endif |
2001 | ||
8165e145 PT |
2002 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2003 | { | |
2004 | u32 contrib; | |
2005 | ||
2006 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2007 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2008 | contrib /= (se->avg.runnable_avg_period + 1); | |
2009 | se->avg.load_avg_contrib = scale_load(contrib); | |
2010 | } | |
2011 | ||
2dac754e PT |
2012 | /* Compute the current contribution to load_avg by se, return any delta */ |
2013 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2014 | { | |
2015 | long old_contrib = se->avg.load_avg_contrib; | |
2016 | ||
8165e145 PT |
2017 | if (entity_is_task(se)) { |
2018 | __update_task_entity_contrib(se); | |
2019 | } else { | |
bb17f655 | 2020 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2021 | __update_group_entity_contrib(se); |
2022 | } | |
2dac754e PT |
2023 | |
2024 | return se->avg.load_avg_contrib - old_contrib; | |
2025 | } | |
2026 | ||
9ee474f5 PT |
2027 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2028 | long load_contrib) | |
2029 | { | |
2030 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2031 | cfs_rq->blocked_load_avg -= load_contrib; | |
2032 | else | |
2033 | cfs_rq->blocked_load_avg = 0; | |
2034 | } | |
2035 | ||
f1b17280 PT |
2036 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2037 | ||
9d85f21c | 2038 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2039 | static inline void update_entity_load_avg(struct sched_entity *se, |
2040 | int update_cfs_rq) | |
9d85f21c | 2041 | { |
2dac754e PT |
2042 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2043 | long contrib_delta; | |
f1b17280 | 2044 | u64 now; |
2dac754e | 2045 | |
f1b17280 PT |
2046 | /* |
2047 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2048 | * case they are the parent of a throttled hierarchy. | |
2049 | */ | |
2050 | if (entity_is_task(se)) | |
2051 | now = cfs_rq_clock_task(cfs_rq); | |
2052 | else | |
2053 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2054 | ||
2055 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2056 | return; |
2057 | ||
2058 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2059 | |
2060 | if (!update_cfs_rq) | |
2061 | return; | |
2062 | ||
2dac754e PT |
2063 | if (se->on_rq) |
2064 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2065 | else |
2066 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2067 | } | |
2068 | ||
2069 | /* | |
2070 | * Decay the load contributed by all blocked children and account this so that | |
2071 | * their contribution may appropriately discounted when they wake up. | |
2072 | */ | |
aff3e498 | 2073 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2074 | { |
f1b17280 | 2075 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2076 | u64 decays; |
2077 | ||
2078 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2079 | if (!decays && !force_update) |
9ee474f5 PT |
2080 | return; |
2081 | ||
2509940f AS |
2082 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2083 | unsigned long removed_load; | |
2084 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2085 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2086 | } | |
9ee474f5 | 2087 | |
aff3e498 PT |
2088 | if (decays) { |
2089 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2090 | decays); | |
2091 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2092 | cfs_rq->last_decay = now; | |
2093 | } | |
c566e8e9 PT |
2094 | |
2095 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2096 | } |
18bf2805 BS |
2097 | |
2098 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2099 | { | |
78becc27 | 2100 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 2101 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 2102 | } |
2dac754e PT |
2103 | |
2104 | /* Add the load generated by se into cfs_rq's child load-average */ | |
2105 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2106 | struct sched_entity *se, |
2107 | int wakeup) | |
2dac754e | 2108 | { |
aff3e498 PT |
2109 | /* |
2110 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2111 | * migration we use a negative decay count to track the remote decays | |
2112 | * accumulated while sleeping. | |
a75cdaa9 AS |
2113 | * |
2114 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2115 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2116 | * constructed load_avg_contrib. | |
aff3e498 PT |
2117 | */ |
2118 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2119 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2120 | if (se->avg.decay_count) { |
2121 | /* | |
2122 | * In a wake-up migration we have to approximate the | |
2123 | * time sleeping. This is because we can't synchronize | |
2124 | * clock_task between the two cpus, and it is not | |
2125 | * guaranteed to be read-safe. Instead, we can | |
2126 | * approximate this using our carried decays, which are | |
2127 | * explicitly atomically readable. | |
2128 | */ | |
2129 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2130 | << 20; | |
2131 | update_entity_load_avg(se, 0); | |
2132 | /* Indicate that we're now synchronized and on-rq */ | |
2133 | se->avg.decay_count = 0; | |
2134 | } | |
9ee474f5 PT |
2135 | wakeup = 0; |
2136 | } else { | |
282cf499 AS |
2137 | /* |
2138 | * Task re-woke on same cpu (or else migrate_task_rq_fair() | |
2139 | * would have made count negative); we must be careful to avoid | |
2140 | * double-accounting blocked time after synchronizing decays. | |
2141 | */ | |
2142 | se->avg.last_runnable_update += __synchronize_entity_decay(se) | |
2143 | << 20; | |
9ee474f5 PT |
2144 | } |
2145 | ||
aff3e498 PT |
2146 | /* migrated tasks did not contribute to our blocked load */ |
2147 | if (wakeup) { | |
9ee474f5 | 2148 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2149 | update_entity_load_avg(se, 0); |
2150 | } | |
9ee474f5 | 2151 | |
2dac754e | 2152 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2153 | /* we force update consideration on load-balancer moves */ |
2154 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2155 | } |
2156 | ||
9ee474f5 PT |
2157 | /* |
2158 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2159 | * transitioning to a blocked state we track its projected decay using | |
2160 | * blocked_load_avg. | |
2161 | */ | |
2dac754e | 2162 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2163 | struct sched_entity *se, |
2164 | int sleep) | |
2dac754e | 2165 | { |
9ee474f5 | 2166 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2167 | /* we force update consideration on load-balancer moves */ |
2168 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2169 | |
2dac754e | 2170 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2171 | if (sleep) { |
2172 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2173 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2174 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2175 | } |
642dbc39 VG |
2176 | |
2177 | /* | |
2178 | * Update the rq's load with the elapsed running time before entering | |
2179 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2180 | * be the only way to update the runnable statistic. | |
2181 | */ | |
2182 | void idle_enter_fair(struct rq *this_rq) | |
2183 | { | |
2184 | update_rq_runnable_avg(this_rq, 1); | |
2185 | } | |
2186 | ||
2187 | /* | |
2188 | * Update the rq's load with the elapsed idle time before a task is | |
2189 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2190 | * be the only way to update the runnable statistic. | |
2191 | */ | |
2192 | void idle_exit_fair(struct rq *this_rq) | |
2193 | { | |
2194 | update_rq_runnable_avg(this_rq, 0); | |
2195 | } | |
2196 | ||
9d85f21c | 2197 | #else |
9ee474f5 PT |
2198 | static inline void update_entity_load_avg(struct sched_entity *se, |
2199 | int update_cfs_rq) {} | |
18bf2805 | 2200 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2201 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2202 | struct sched_entity *se, |
2203 | int wakeup) {} | |
2dac754e | 2204 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2205 | struct sched_entity *se, |
2206 | int sleep) {} | |
aff3e498 PT |
2207 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2208 | int force_update) {} | |
9d85f21c PT |
2209 | #endif |
2210 | ||
2396af69 | 2211 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2212 | { |
bf0f6f24 | 2213 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2214 | struct task_struct *tsk = NULL; |
2215 | ||
2216 | if (entity_is_task(se)) | |
2217 | tsk = task_of(se); | |
2218 | ||
41acab88 | 2219 | if (se->statistics.sleep_start) { |
78becc27 | 2220 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2221 | |
2222 | if ((s64)delta < 0) | |
2223 | delta = 0; | |
2224 | ||
41acab88 LDM |
2225 | if (unlikely(delta > se->statistics.sleep_max)) |
2226 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2227 | |
8c79a045 | 2228 | se->statistics.sleep_start = 0; |
41acab88 | 2229 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2230 | |
768d0c27 | 2231 | if (tsk) { |
e414314c | 2232 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2233 | trace_sched_stat_sleep(tsk, delta); |
2234 | } | |
bf0f6f24 | 2235 | } |
41acab88 | 2236 | if (se->statistics.block_start) { |
78becc27 | 2237 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2238 | |
2239 | if ((s64)delta < 0) | |
2240 | delta = 0; | |
2241 | ||
41acab88 LDM |
2242 | if (unlikely(delta > se->statistics.block_max)) |
2243 | se->statistics.block_max = delta; | |
bf0f6f24 | 2244 | |
8c79a045 | 2245 | se->statistics.block_start = 0; |
41acab88 | 2246 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2247 | |
e414314c | 2248 | if (tsk) { |
8f0dfc34 | 2249 | if (tsk->in_iowait) { |
41acab88 LDM |
2250 | se->statistics.iowait_sum += delta; |
2251 | se->statistics.iowait_count++; | |
768d0c27 | 2252 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2253 | } |
2254 | ||
b781a602 AV |
2255 | trace_sched_stat_blocked(tsk, delta); |
2256 | ||
e414314c PZ |
2257 | /* |
2258 | * Blocking time is in units of nanosecs, so shift by | |
2259 | * 20 to get a milliseconds-range estimation of the | |
2260 | * amount of time that the task spent sleeping: | |
2261 | */ | |
2262 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2263 | profile_hits(SLEEP_PROFILING, | |
2264 | (void *)get_wchan(tsk), | |
2265 | delta >> 20); | |
2266 | } | |
2267 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2268 | } |
bf0f6f24 IM |
2269 | } |
2270 | #endif | |
2271 | } | |
2272 | ||
ddc97297 PZ |
2273 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2274 | { | |
2275 | #ifdef CONFIG_SCHED_DEBUG | |
2276 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2277 | ||
2278 | if (d < 0) | |
2279 | d = -d; | |
2280 | ||
2281 | if (d > 3*sysctl_sched_latency) | |
2282 | schedstat_inc(cfs_rq, nr_spread_over); | |
2283 | #endif | |
2284 | } | |
2285 | ||
aeb73b04 PZ |
2286 | static void |
2287 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2288 | { | |
1af5f730 | 2289 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2290 | |
2cb8600e PZ |
2291 | /* |
2292 | * The 'current' period is already promised to the current tasks, | |
2293 | * however the extra weight of the new task will slow them down a | |
2294 | * little, place the new task so that it fits in the slot that | |
2295 | * stays open at the end. | |
2296 | */ | |
94dfb5e7 | 2297 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2298 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2299 | |
a2e7a7eb | 2300 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2301 | if (!initial) { |
a2e7a7eb | 2302 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2303 | |
a2e7a7eb MG |
2304 | /* |
2305 | * Halve their sleep time's effect, to allow | |
2306 | * for a gentler effect of sleepers: | |
2307 | */ | |
2308 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2309 | thresh >>= 1; | |
51e0304c | 2310 | |
a2e7a7eb | 2311 | vruntime -= thresh; |
aeb73b04 PZ |
2312 | } |
2313 | ||
b5d9d734 | 2314 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2315 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2316 | } |
2317 | ||
d3d9dc33 PT |
2318 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2319 | ||
bf0f6f24 | 2320 | static void |
88ec22d3 | 2321 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2322 | { |
88ec22d3 PZ |
2323 | /* |
2324 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2325 | * through calling update_curr(). |
88ec22d3 | 2326 | */ |
371fd7e7 | 2327 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2328 | se->vruntime += cfs_rq->min_vruntime; |
2329 | ||
bf0f6f24 | 2330 | /* |
a2a2d680 | 2331 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2332 | */ |
b7cc0896 | 2333 | update_curr(cfs_rq); |
f269ae04 | 2334 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2335 | account_entity_enqueue(cfs_rq, se); |
2336 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2337 | |
88ec22d3 | 2338 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2339 | place_entity(cfs_rq, se, 0); |
2396af69 | 2340 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2341 | } |
bf0f6f24 | 2342 | |
d2417e5a | 2343 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2344 | check_spread(cfs_rq, se); |
83b699ed SV |
2345 | if (se != cfs_rq->curr) |
2346 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2347 | se->on_rq = 1; |
3d4b47b4 | 2348 | |
d3d9dc33 | 2349 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2350 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2351 | check_enqueue_throttle(cfs_rq); |
2352 | } | |
bf0f6f24 IM |
2353 | } |
2354 | ||
2c13c919 | 2355 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2356 | { |
2c13c919 RR |
2357 | for_each_sched_entity(se) { |
2358 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2359 | if (cfs_rq->last == se) | |
2360 | cfs_rq->last = NULL; | |
2361 | else | |
2362 | break; | |
2363 | } | |
2364 | } | |
2002c695 | 2365 | |
2c13c919 RR |
2366 | static void __clear_buddies_next(struct sched_entity *se) |
2367 | { | |
2368 | for_each_sched_entity(se) { | |
2369 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2370 | if (cfs_rq->next == se) | |
2371 | cfs_rq->next = NULL; | |
2372 | else | |
2373 | break; | |
2374 | } | |
2002c695 PZ |
2375 | } |
2376 | ||
ac53db59 RR |
2377 | static void __clear_buddies_skip(struct sched_entity *se) |
2378 | { | |
2379 | for_each_sched_entity(se) { | |
2380 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2381 | if (cfs_rq->skip == se) | |
2382 | cfs_rq->skip = NULL; | |
2383 | else | |
2384 | break; | |
2385 | } | |
2386 | } | |
2387 | ||
a571bbea PZ |
2388 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2389 | { | |
2c13c919 RR |
2390 | if (cfs_rq->last == se) |
2391 | __clear_buddies_last(se); | |
2392 | ||
2393 | if (cfs_rq->next == se) | |
2394 | __clear_buddies_next(se); | |
ac53db59 RR |
2395 | |
2396 | if (cfs_rq->skip == se) | |
2397 | __clear_buddies_skip(se); | |
a571bbea PZ |
2398 | } |
2399 | ||
6c16a6dc | 2400 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2401 | |
bf0f6f24 | 2402 | static void |
371fd7e7 | 2403 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2404 | { |
a2a2d680 DA |
2405 | /* |
2406 | * Update run-time statistics of the 'current'. | |
2407 | */ | |
2408 | update_curr(cfs_rq); | |
17bc14b7 | 2409 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2410 | |
19b6a2e3 | 2411 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2412 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2413 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2414 | if (entity_is_task(se)) { |
2415 | struct task_struct *tsk = task_of(se); | |
2416 | ||
2417 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2418 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2419 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2420 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2421 | } |
db36cc7d | 2422 | #endif |
67e9fb2a PZ |
2423 | } |
2424 | ||
2002c695 | 2425 | clear_buddies(cfs_rq, se); |
4793241b | 2426 | |
83b699ed | 2427 | if (se != cfs_rq->curr) |
30cfdcfc | 2428 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2429 | se->on_rq = 0; |
30cfdcfc | 2430 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2431 | |
2432 | /* | |
2433 | * Normalize the entity after updating the min_vruntime because the | |
2434 | * update can refer to the ->curr item and we need to reflect this | |
2435 | * movement in our normalized position. | |
2436 | */ | |
371fd7e7 | 2437 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2438 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2439 | |
d8b4986d PT |
2440 | /* return excess runtime on last dequeue */ |
2441 | return_cfs_rq_runtime(cfs_rq); | |
2442 | ||
1e876231 | 2443 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2444 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2445 | } |
2446 | ||
2447 | /* | |
2448 | * Preempt the current task with a newly woken task if needed: | |
2449 | */ | |
7c92e54f | 2450 | static void |
2e09bf55 | 2451 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2452 | { |
11697830 | 2453 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2454 | struct sched_entity *se; |
2455 | s64 delta; | |
11697830 | 2456 | |
6d0f0ebd | 2457 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2458 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2459 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2460 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2461 | /* |
2462 | * The current task ran long enough, ensure it doesn't get | |
2463 | * re-elected due to buddy favours. | |
2464 | */ | |
2465 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2466 | return; |
2467 | } | |
2468 | ||
2469 | /* | |
2470 | * Ensure that a task that missed wakeup preemption by a | |
2471 | * narrow margin doesn't have to wait for a full slice. | |
2472 | * This also mitigates buddy induced latencies under load. | |
2473 | */ | |
f685ceac MG |
2474 | if (delta_exec < sysctl_sched_min_granularity) |
2475 | return; | |
2476 | ||
f4cfb33e WX |
2477 | se = __pick_first_entity(cfs_rq); |
2478 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2479 | |
f4cfb33e WX |
2480 | if (delta < 0) |
2481 | return; | |
d7d82944 | 2482 | |
f4cfb33e WX |
2483 | if (delta > ideal_runtime) |
2484 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2485 | } |
2486 | ||
83b699ed | 2487 | static void |
8494f412 | 2488 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2489 | { |
83b699ed SV |
2490 | /* 'current' is not kept within the tree. */ |
2491 | if (se->on_rq) { | |
2492 | /* | |
2493 | * Any task has to be enqueued before it get to execute on | |
2494 | * a CPU. So account for the time it spent waiting on the | |
2495 | * runqueue. | |
2496 | */ | |
2497 | update_stats_wait_end(cfs_rq, se); | |
2498 | __dequeue_entity(cfs_rq, se); | |
2499 | } | |
2500 | ||
79303e9e | 2501 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2502 | cfs_rq->curr = se; |
eba1ed4b IM |
2503 | #ifdef CONFIG_SCHEDSTATS |
2504 | /* | |
2505 | * Track our maximum slice length, if the CPU's load is at | |
2506 | * least twice that of our own weight (i.e. dont track it | |
2507 | * when there are only lesser-weight tasks around): | |
2508 | */ | |
495eca49 | 2509 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2510 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2511 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2512 | } | |
2513 | #endif | |
4a55b450 | 2514 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2515 | } |
2516 | ||
3f3a4904 PZ |
2517 | static int |
2518 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2519 | ||
ac53db59 RR |
2520 | /* |
2521 | * Pick the next process, keeping these things in mind, in this order: | |
2522 | * 1) keep things fair between processes/task groups | |
2523 | * 2) pick the "next" process, since someone really wants that to run | |
2524 | * 3) pick the "last" process, for cache locality | |
2525 | * 4) do not run the "skip" process, if something else is available | |
2526 | */ | |
f4b6755f | 2527 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 2528 | { |
ac53db59 | 2529 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 2530 | struct sched_entity *left = se; |
f4b6755f | 2531 | |
ac53db59 RR |
2532 | /* |
2533 | * Avoid running the skip buddy, if running something else can | |
2534 | * be done without getting too unfair. | |
2535 | */ | |
2536 | if (cfs_rq->skip == se) { | |
2537 | struct sched_entity *second = __pick_next_entity(se); | |
2538 | if (second && wakeup_preempt_entity(second, left) < 1) | |
2539 | se = second; | |
2540 | } | |
aa2ac252 | 2541 | |
f685ceac MG |
2542 | /* |
2543 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2544 | */ | |
2545 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2546 | se = cfs_rq->last; | |
2547 | ||
ac53db59 RR |
2548 | /* |
2549 | * Someone really wants this to run. If it's not unfair, run it. | |
2550 | */ | |
2551 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2552 | se = cfs_rq->next; | |
2553 | ||
f685ceac | 2554 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2555 | |
2556 | return se; | |
aa2ac252 PZ |
2557 | } |
2558 | ||
d3d9dc33 PT |
2559 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
2560 | ||
ab6cde26 | 2561 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2562 | { |
2563 | /* | |
2564 | * If still on the runqueue then deactivate_task() | |
2565 | * was not called and update_curr() has to be done: | |
2566 | */ | |
2567 | if (prev->on_rq) | |
b7cc0896 | 2568 | update_curr(cfs_rq); |
bf0f6f24 | 2569 | |
d3d9dc33 PT |
2570 | /* throttle cfs_rqs exceeding runtime */ |
2571 | check_cfs_rq_runtime(cfs_rq); | |
2572 | ||
ddc97297 | 2573 | check_spread(cfs_rq, prev); |
30cfdcfc | 2574 | if (prev->on_rq) { |
5870db5b | 2575 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2576 | /* Put 'current' back into the tree. */ |
2577 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2578 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2579 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2580 | } |
429d43bc | 2581 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2582 | } |
2583 | ||
8f4d37ec PZ |
2584 | static void |
2585 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2586 | { |
bf0f6f24 | 2587 | /* |
30cfdcfc | 2588 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2589 | */ |
30cfdcfc | 2590 | update_curr(cfs_rq); |
bf0f6f24 | 2591 | |
9d85f21c PT |
2592 | /* |
2593 | * Ensure that runnable average is periodically updated. | |
2594 | */ | |
9ee474f5 | 2595 | update_entity_load_avg(curr, 1); |
aff3e498 | 2596 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2597 | update_cfs_shares(cfs_rq); |
9d85f21c | 2598 | |
8f4d37ec PZ |
2599 | #ifdef CONFIG_SCHED_HRTICK |
2600 | /* | |
2601 | * queued ticks are scheduled to match the slice, so don't bother | |
2602 | * validating it and just reschedule. | |
2603 | */ | |
983ed7a6 HH |
2604 | if (queued) { |
2605 | resched_task(rq_of(cfs_rq)->curr); | |
2606 | return; | |
2607 | } | |
8f4d37ec PZ |
2608 | /* |
2609 | * don't let the period tick interfere with the hrtick preemption | |
2610 | */ | |
2611 | if (!sched_feat(DOUBLE_TICK) && | |
2612 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2613 | return; | |
2614 | #endif | |
2615 | ||
2c2efaed | 2616 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2617 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2618 | } |
2619 | ||
ab84d31e PT |
2620 | |
2621 | /************************************************** | |
2622 | * CFS bandwidth control machinery | |
2623 | */ | |
2624 | ||
2625 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2626 | |
2627 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2628 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2629 | |
2630 | static inline bool cfs_bandwidth_used(void) | |
2631 | { | |
c5905afb | 2632 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2633 | } |
2634 | ||
2635 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
2636 | { | |
2637 | /* only need to count groups transitioning between enabled/!enabled */ | |
2638 | if (enabled && !was_enabled) | |
c5905afb | 2639 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 2640 | else if (!enabled && was_enabled) |
c5905afb | 2641 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
2642 | } |
2643 | #else /* HAVE_JUMP_LABEL */ | |
2644 | static bool cfs_bandwidth_used(void) | |
2645 | { | |
2646 | return true; | |
2647 | } | |
2648 | ||
2649 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
2650 | #endif /* HAVE_JUMP_LABEL */ | |
2651 | ||
ab84d31e PT |
2652 | /* |
2653 | * default period for cfs group bandwidth. | |
2654 | * default: 0.1s, units: nanoseconds | |
2655 | */ | |
2656 | static inline u64 default_cfs_period(void) | |
2657 | { | |
2658 | return 100000000ULL; | |
2659 | } | |
ec12cb7f PT |
2660 | |
2661 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2662 | { | |
2663 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2664 | } | |
2665 | ||
a9cf55b2 PT |
2666 | /* |
2667 | * Replenish runtime according to assigned quota and update expiration time. | |
2668 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2669 | * additional synchronization around rq->lock. | |
2670 | * | |
2671 | * requires cfs_b->lock | |
2672 | */ | |
029632fb | 2673 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2674 | { |
2675 | u64 now; | |
2676 | ||
2677 | if (cfs_b->quota == RUNTIME_INF) | |
2678 | return; | |
2679 | ||
2680 | now = sched_clock_cpu(smp_processor_id()); | |
2681 | cfs_b->runtime = cfs_b->quota; | |
2682 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2683 | } | |
2684 | ||
029632fb PZ |
2685 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2686 | { | |
2687 | return &tg->cfs_bandwidth; | |
2688 | } | |
2689 | ||
f1b17280 PT |
2690 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2691 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2692 | { | |
2693 | if (unlikely(cfs_rq->throttle_count)) | |
2694 | return cfs_rq->throttled_clock_task; | |
2695 | ||
78becc27 | 2696 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2697 | } |
2698 | ||
85dac906 PT |
2699 | /* returns 0 on failure to allocate runtime */ |
2700 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2701 | { |
2702 | struct task_group *tg = cfs_rq->tg; | |
2703 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2704 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2705 | |
2706 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2707 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2708 | ||
2709 | raw_spin_lock(&cfs_b->lock); | |
2710 | if (cfs_b->quota == RUNTIME_INF) | |
2711 | amount = min_amount; | |
58088ad0 | 2712 | else { |
a9cf55b2 PT |
2713 | /* |
2714 | * If the bandwidth pool has become inactive, then at least one | |
2715 | * period must have elapsed since the last consumption. | |
2716 | * Refresh the global state and ensure bandwidth timer becomes | |
2717 | * active. | |
2718 | */ | |
2719 | if (!cfs_b->timer_active) { | |
2720 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2721 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2722 | } |
58088ad0 PT |
2723 | |
2724 | if (cfs_b->runtime > 0) { | |
2725 | amount = min(cfs_b->runtime, min_amount); | |
2726 | cfs_b->runtime -= amount; | |
2727 | cfs_b->idle = 0; | |
2728 | } | |
ec12cb7f | 2729 | } |
a9cf55b2 | 2730 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2731 | raw_spin_unlock(&cfs_b->lock); |
2732 | ||
2733 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2734 | /* |
2735 | * we may have advanced our local expiration to account for allowed | |
2736 | * spread between our sched_clock and the one on which runtime was | |
2737 | * issued. | |
2738 | */ | |
2739 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2740 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2741 | |
2742 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2743 | } |
2744 | ||
a9cf55b2 PT |
2745 | /* |
2746 | * Note: This depends on the synchronization provided by sched_clock and the | |
2747 | * fact that rq->clock snapshots this value. | |
2748 | */ | |
2749 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2750 | { |
a9cf55b2 | 2751 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2752 | |
2753 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2754 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2755 | return; |
2756 | ||
a9cf55b2 PT |
2757 | if (cfs_rq->runtime_remaining < 0) |
2758 | return; | |
2759 | ||
2760 | /* | |
2761 | * If the local deadline has passed we have to consider the | |
2762 | * possibility that our sched_clock is 'fast' and the global deadline | |
2763 | * has not truly expired. | |
2764 | * | |
2765 | * Fortunately we can check determine whether this the case by checking | |
2766 | * whether the global deadline has advanced. | |
2767 | */ | |
2768 | ||
2769 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
2770 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
2771 | cfs_rq->runtime_expires += TICK_NSEC; | |
2772 | } else { | |
2773 | /* global deadline is ahead, expiration has passed */ | |
2774 | cfs_rq->runtime_remaining = 0; | |
2775 | } | |
2776 | } | |
2777 | ||
2778 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2779 | unsigned long delta_exec) | |
2780 | { | |
2781 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 2782 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
2783 | expire_cfs_rq_runtime(cfs_rq); |
2784 | ||
2785 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
2786 | return; |
2787 | ||
85dac906 PT |
2788 | /* |
2789 | * if we're unable to extend our runtime we resched so that the active | |
2790 | * hierarchy can be throttled | |
2791 | */ | |
2792 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
2793 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
2794 | } |
2795 | ||
6c16a6dc PZ |
2796 | static __always_inline |
2797 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 2798 | { |
56f570e5 | 2799 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
2800 | return; |
2801 | ||
2802 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
2803 | } | |
2804 | ||
85dac906 PT |
2805 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
2806 | { | |
56f570e5 | 2807 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
2808 | } |
2809 | ||
64660c86 PT |
2810 | /* check whether cfs_rq, or any parent, is throttled */ |
2811 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2812 | { | |
56f570e5 | 2813 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
2814 | } |
2815 | ||
2816 | /* | |
2817 | * Ensure that neither of the group entities corresponding to src_cpu or | |
2818 | * dest_cpu are members of a throttled hierarchy when performing group | |
2819 | * load-balance operations. | |
2820 | */ | |
2821 | static inline int throttled_lb_pair(struct task_group *tg, | |
2822 | int src_cpu, int dest_cpu) | |
2823 | { | |
2824 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
2825 | ||
2826 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
2827 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
2828 | ||
2829 | return throttled_hierarchy(src_cfs_rq) || | |
2830 | throttled_hierarchy(dest_cfs_rq); | |
2831 | } | |
2832 | ||
2833 | /* updated child weight may affect parent so we have to do this bottom up */ | |
2834 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
2835 | { | |
2836 | struct rq *rq = data; | |
2837 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2838 | ||
2839 | cfs_rq->throttle_count--; | |
2840 | #ifdef CONFIG_SMP | |
2841 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 2842 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 2843 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 2844 | cfs_rq->throttled_clock_task; |
64660c86 PT |
2845 | } |
2846 | #endif | |
2847 | ||
2848 | return 0; | |
2849 | } | |
2850 | ||
2851 | static int tg_throttle_down(struct task_group *tg, void *data) | |
2852 | { | |
2853 | struct rq *rq = data; | |
2854 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2855 | ||
82958366 PT |
2856 | /* group is entering throttled state, stop time */ |
2857 | if (!cfs_rq->throttle_count) | |
78becc27 | 2858 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
2859 | cfs_rq->throttle_count++; |
2860 | ||
2861 | return 0; | |
2862 | } | |
2863 | ||
d3d9dc33 | 2864 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
2865 | { |
2866 | struct rq *rq = rq_of(cfs_rq); | |
2867 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2868 | struct sched_entity *se; | |
2869 | long task_delta, dequeue = 1; | |
2870 | ||
2871 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2872 | ||
f1b17280 | 2873 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
2874 | rcu_read_lock(); |
2875 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
2876 | rcu_read_unlock(); | |
85dac906 PT |
2877 | |
2878 | task_delta = cfs_rq->h_nr_running; | |
2879 | for_each_sched_entity(se) { | |
2880 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
2881 | /* throttled entity or throttle-on-deactivate */ | |
2882 | if (!se->on_rq) | |
2883 | break; | |
2884 | ||
2885 | if (dequeue) | |
2886 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
2887 | qcfs_rq->h_nr_running -= task_delta; | |
2888 | ||
2889 | if (qcfs_rq->load.weight) | |
2890 | dequeue = 0; | |
2891 | } | |
2892 | ||
2893 | if (!se) | |
2894 | rq->nr_running -= task_delta; | |
2895 | ||
2896 | cfs_rq->throttled = 1; | |
78becc27 | 2897 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
2898 | raw_spin_lock(&cfs_b->lock); |
2899 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
2900 | raw_spin_unlock(&cfs_b->lock); | |
2901 | } | |
2902 | ||
029632fb | 2903 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
2904 | { |
2905 | struct rq *rq = rq_of(cfs_rq); | |
2906 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2907 | struct sched_entity *se; | |
2908 | int enqueue = 1; | |
2909 | long task_delta; | |
2910 | ||
22b958d8 | 2911 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
2912 | |
2913 | cfs_rq->throttled = 0; | |
1a55af2e FW |
2914 | |
2915 | update_rq_clock(rq); | |
2916 | ||
671fd9da | 2917 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 2918 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
2919 | list_del_rcu(&cfs_rq->throttled_list); |
2920 | raw_spin_unlock(&cfs_b->lock); | |
2921 | ||
64660c86 PT |
2922 | /* update hierarchical throttle state */ |
2923 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
2924 | ||
671fd9da PT |
2925 | if (!cfs_rq->load.weight) |
2926 | return; | |
2927 | ||
2928 | task_delta = cfs_rq->h_nr_running; | |
2929 | for_each_sched_entity(se) { | |
2930 | if (se->on_rq) | |
2931 | enqueue = 0; | |
2932 | ||
2933 | cfs_rq = cfs_rq_of(se); | |
2934 | if (enqueue) | |
2935 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
2936 | cfs_rq->h_nr_running += task_delta; | |
2937 | ||
2938 | if (cfs_rq_throttled(cfs_rq)) | |
2939 | break; | |
2940 | } | |
2941 | ||
2942 | if (!se) | |
2943 | rq->nr_running += task_delta; | |
2944 | ||
2945 | /* determine whether we need to wake up potentially idle cpu */ | |
2946 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
2947 | resched_task(rq->curr); | |
2948 | } | |
2949 | ||
2950 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
2951 | u64 remaining, u64 expires) | |
2952 | { | |
2953 | struct cfs_rq *cfs_rq; | |
2954 | u64 runtime = remaining; | |
2955 | ||
2956 | rcu_read_lock(); | |
2957 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
2958 | throttled_list) { | |
2959 | struct rq *rq = rq_of(cfs_rq); | |
2960 | ||
2961 | raw_spin_lock(&rq->lock); | |
2962 | if (!cfs_rq_throttled(cfs_rq)) | |
2963 | goto next; | |
2964 | ||
2965 | runtime = -cfs_rq->runtime_remaining + 1; | |
2966 | if (runtime > remaining) | |
2967 | runtime = remaining; | |
2968 | remaining -= runtime; | |
2969 | ||
2970 | cfs_rq->runtime_remaining += runtime; | |
2971 | cfs_rq->runtime_expires = expires; | |
2972 | ||
2973 | /* we check whether we're throttled above */ | |
2974 | if (cfs_rq->runtime_remaining > 0) | |
2975 | unthrottle_cfs_rq(cfs_rq); | |
2976 | ||
2977 | next: | |
2978 | raw_spin_unlock(&rq->lock); | |
2979 | ||
2980 | if (!remaining) | |
2981 | break; | |
2982 | } | |
2983 | rcu_read_unlock(); | |
2984 | ||
2985 | return remaining; | |
2986 | } | |
2987 | ||
58088ad0 PT |
2988 | /* |
2989 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2990 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2991 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2992 | * used to track this state. | |
2993 | */ | |
2994 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
2995 | { | |
671fd9da PT |
2996 | u64 runtime, runtime_expires; |
2997 | int idle = 1, throttled; | |
58088ad0 PT |
2998 | |
2999 | raw_spin_lock(&cfs_b->lock); | |
3000 | /* no need to continue the timer with no bandwidth constraint */ | |
3001 | if (cfs_b->quota == RUNTIME_INF) | |
3002 | goto out_unlock; | |
3003 | ||
671fd9da PT |
3004 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3005 | /* idle depends on !throttled (for the case of a large deficit) */ | |
3006 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 3007 | cfs_b->nr_periods += overrun; |
671fd9da | 3008 | |
a9cf55b2 PT |
3009 | /* if we're going inactive then everything else can be deferred */ |
3010 | if (idle) | |
3011 | goto out_unlock; | |
3012 | ||
3013 | __refill_cfs_bandwidth_runtime(cfs_b); | |
3014 | ||
671fd9da PT |
3015 | if (!throttled) { |
3016 | /* mark as potentially idle for the upcoming period */ | |
3017 | cfs_b->idle = 1; | |
3018 | goto out_unlock; | |
3019 | } | |
3020 | ||
e8da1b18 NR |
3021 | /* account preceding periods in which throttling occurred */ |
3022 | cfs_b->nr_throttled += overrun; | |
3023 | ||
671fd9da PT |
3024 | /* |
3025 | * There are throttled entities so we must first use the new bandwidth | |
3026 | * to unthrottle them before making it generally available. This | |
3027 | * ensures that all existing debts will be paid before a new cfs_rq is | |
3028 | * allowed to run. | |
3029 | */ | |
3030 | runtime = cfs_b->runtime; | |
3031 | runtime_expires = cfs_b->runtime_expires; | |
3032 | cfs_b->runtime = 0; | |
3033 | ||
3034 | /* | |
3035 | * This check is repeated as we are holding onto the new bandwidth | |
3036 | * while we unthrottle. This can potentially race with an unthrottled | |
3037 | * group trying to acquire new bandwidth from the global pool. | |
3038 | */ | |
3039 | while (throttled && runtime > 0) { | |
3040 | raw_spin_unlock(&cfs_b->lock); | |
3041 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3042 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3043 | runtime_expires); | |
3044 | raw_spin_lock(&cfs_b->lock); | |
3045 | ||
3046 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
3047 | } | |
58088ad0 | 3048 | |
671fd9da PT |
3049 | /* return (any) remaining runtime */ |
3050 | cfs_b->runtime = runtime; | |
3051 | /* | |
3052 | * While we are ensured activity in the period following an | |
3053 | * unthrottle, this also covers the case in which the new bandwidth is | |
3054 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3055 | * timer to remain active while there are any throttled entities.) | |
3056 | */ | |
3057 | cfs_b->idle = 0; | |
58088ad0 PT |
3058 | out_unlock: |
3059 | if (idle) | |
3060 | cfs_b->timer_active = 0; | |
3061 | raw_spin_unlock(&cfs_b->lock); | |
3062 | ||
3063 | return idle; | |
3064 | } | |
d3d9dc33 | 3065 | |
d8b4986d PT |
3066 | /* a cfs_rq won't donate quota below this amount */ |
3067 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3068 | /* minimum remaining period time to redistribute slack quota */ | |
3069 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3070 | /* how long we wait to gather additional slack before distributing */ | |
3071 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3072 | ||
3073 | /* are we near the end of the current quota period? */ | |
3074 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
3075 | { | |
3076 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3077 | u64 remaining; | |
3078 | ||
3079 | /* if the call-back is running a quota refresh is already occurring */ | |
3080 | if (hrtimer_callback_running(refresh_timer)) | |
3081 | return 1; | |
3082 | ||
3083 | /* is a quota refresh about to occur? */ | |
3084 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3085 | if (remaining < min_expire) | |
3086 | return 1; | |
3087 | ||
3088 | return 0; | |
3089 | } | |
3090 | ||
3091 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3092 | { | |
3093 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3094 | ||
3095 | /* if there's a quota refresh soon don't bother with slack */ | |
3096 | if (runtime_refresh_within(cfs_b, min_left)) | |
3097 | return; | |
3098 | ||
3099 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3100 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3101 | } | |
3102 | ||
3103 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3104 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3105 | { | |
3106 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3107 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3108 | ||
3109 | if (slack_runtime <= 0) | |
3110 | return; | |
3111 | ||
3112 | raw_spin_lock(&cfs_b->lock); | |
3113 | if (cfs_b->quota != RUNTIME_INF && | |
3114 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3115 | cfs_b->runtime += slack_runtime; | |
3116 | ||
3117 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3118 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3119 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3120 | start_cfs_slack_bandwidth(cfs_b); | |
3121 | } | |
3122 | raw_spin_unlock(&cfs_b->lock); | |
3123 | ||
3124 | /* even if it's not valid for return we don't want to try again */ | |
3125 | cfs_rq->runtime_remaining -= slack_runtime; | |
3126 | } | |
3127 | ||
3128 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3129 | { | |
56f570e5 PT |
3130 | if (!cfs_bandwidth_used()) |
3131 | return; | |
3132 | ||
fccfdc6f | 3133 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3134 | return; |
3135 | ||
3136 | __return_cfs_rq_runtime(cfs_rq); | |
3137 | } | |
3138 | ||
3139 | /* | |
3140 | * This is done with a timer (instead of inline with bandwidth return) since | |
3141 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3142 | */ | |
3143 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3144 | { | |
3145 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3146 | u64 expires; | |
3147 | ||
3148 | /* confirm we're still not at a refresh boundary */ | |
3149 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
3150 | return; | |
3151 | ||
3152 | raw_spin_lock(&cfs_b->lock); | |
3153 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
3154 | runtime = cfs_b->runtime; | |
3155 | cfs_b->runtime = 0; | |
3156 | } | |
3157 | expires = cfs_b->runtime_expires; | |
3158 | raw_spin_unlock(&cfs_b->lock); | |
3159 | ||
3160 | if (!runtime) | |
3161 | return; | |
3162 | ||
3163 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3164 | ||
3165 | raw_spin_lock(&cfs_b->lock); | |
3166 | if (expires == cfs_b->runtime_expires) | |
3167 | cfs_b->runtime = runtime; | |
3168 | raw_spin_unlock(&cfs_b->lock); | |
3169 | } | |
3170 | ||
d3d9dc33 PT |
3171 | /* |
3172 | * When a group wakes up we want to make sure that its quota is not already | |
3173 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3174 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3175 | */ | |
3176 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3177 | { | |
56f570e5 PT |
3178 | if (!cfs_bandwidth_used()) |
3179 | return; | |
3180 | ||
d3d9dc33 PT |
3181 | /* an active group must be handled by the update_curr()->put() path */ |
3182 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3183 | return; | |
3184 | ||
3185 | /* ensure the group is not already throttled */ | |
3186 | if (cfs_rq_throttled(cfs_rq)) | |
3187 | return; | |
3188 | ||
3189 | /* update runtime allocation */ | |
3190 | account_cfs_rq_runtime(cfs_rq, 0); | |
3191 | if (cfs_rq->runtime_remaining <= 0) | |
3192 | throttle_cfs_rq(cfs_rq); | |
3193 | } | |
3194 | ||
3195 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
3196 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3197 | { | |
56f570e5 PT |
3198 | if (!cfs_bandwidth_used()) |
3199 | return; | |
3200 | ||
d3d9dc33 PT |
3201 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
3202 | return; | |
3203 | ||
3204 | /* | |
3205 | * it's possible for a throttled entity to be forced into a running | |
3206 | * state (e.g. set_curr_task), in this case we're finished. | |
3207 | */ | |
3208 | if (cfs_rq_throttled(cfs_rq)) | |
3209 | return; | |
3210 | ||
3211 | throttle_cfs_rq(cfs_rq); | |
3212 | } | |
029632fb | 3213 | |
029632fb PZ |
3214 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3215 | { | |
3216 | struct cfs_bandwidth *cfs_b = | |
3217 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3218 | do_sched_cfs_slack_timer(cfs_b); | |
3219 | ||
3220 | return HRTIMER_NORESTART; | |
3221 | } | |
3222 | ||
3223 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3224 | { | |
3225 | struct cfs_bandwidth *cfs_b = | |
3226 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3227 | ktime_t now; | |
3228 | int overrun; | |
3229 | int idle = 0; | |
3230 | ||
3231 | for (;;) { | |
3232 | now = hrtimer_cb_get_time(timer); | |
3233 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3234 | ||
3235 | if (!overrun) | |
3236 | break; | |
3237 | ||
3238 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3239 | } | |
3240 | ||
3241 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3242 | } | |
3243 | ||
3244 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3245 | { | |
3246 | raw_spin_lock_init(&cfs_b->lock); | |
3247 | cfs_b->runtime = 0; | |
3248 | cfs_b->quota = RUNTIME_INF; | |
3249 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3250 | ||
3251 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3252 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3253 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3254 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3255 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3256 | } | |
3257 | ||
3258 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3259 | { | |
3260 | cfs_rq->runtime_enabled = 0; | |
3261 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3262 | } | |
3263 | ||
3264 | /* requires cfs_b->lock, may release to reprogram timer */ | |
3265 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3266 | { | |
3267 | /* | |
3268 | * The timer may be active because we're trying to set a new bandwidth | |
3269 | * period or because we're racing with the tear-down path | |
3270 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3271 | * terminates). In either case we ensure that it's re-programmed | |
3272 | */ | |
3273 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
3274 | raw_spin_unlock(&cfs_b->lock); | |
3275 | /* ensure cfs_b->lock is available while we wait */ | |
3276 | hrtimer_cancel(&cfs_b->period_timer); | |
3277 | ||
3278 | raw_spin_lock(&cfs_b->lock); | |
3279 | /* if someone else restarted the timer then we're done */ | |
3280 | if (cfs_b->timer_active) | |
3281 | return; | |
3282 | } | |
3283 | ||
3284 | cfs_b->timer_active = 1; | |
3285 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3286 | } | |
3287 | ||
3288 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3289 | { | |
3290 | hrtimer_cancel(&cfs_b->period_timer); | |
3291 | hrtimer_cancel(&cfs_b->slack_timer); | |
3292 | } | |
3293 | ||
38dc3348 | 3294 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3295 | { |
3296 | struct cfs_rq *cfs_rq; | |
3297 | ||
3298 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3299 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3300 | ||
3301 | if (!cfs_rq->runtime_enabled) | |
3302 | continue; | |
3303 | ||
3304 | /* | |
3305 | * clock_task is not advancing so we just need to make sure | |
3306 | * there's some valid quota amount | |
3307 | */ | |
3308 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3309 | if (cfs_rq_throttled(cfs_rq)) | |
3310 | unthrottle_cfs_rq(cfs_rq); | |
3311 | } | |
3312 | } | |
3313 | ||
3314 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3315 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3316 | { | |
78becc27 | 3317 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3318 | } |
3319 | ||
3320 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
3321 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
3322 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
3323 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 3324 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3325 | |
3326 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3327 | { | |
3328 | return 0; | |
3329 | } | |
64660c86 PT |
3330 | |
3331 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3332 | { | |
3333 | return 0; | |
3334 | } | |
3335 | ||
3336 | static inline int throttled_lb_pair(struct task_group *tg, | |
3337 | int src_cpu, int dest_cpu) | |
3338 | { | |
3339 | return 0; | |
3340 | } | |
029632fb PZ |
3341 | |
3342 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3343 | ||
3344 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3345 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3346 | #endif |
3347 | ||
029632fb PZ |
3348 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3349 | { | |
3350 | return NULL; | |
3351 | } | |
3352 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3353 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3354 | |
3355 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3356 | ||
bf0f6f24 IM |
3357 | /************************************************** |
3358 | * CFS operations on tasks: | |
3359 | */ | |
3360 | ||
8f4d37ec PZ |
3361 | #ifdef CONFIG_SCHED_HRTICK |
3362 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3363 | { | |
8f4d37ec PZ |
3364 | struct sched_entity *se = &p->se; |
3365 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3366 | ||
3367 | WARN_ON(task_rq(p) != rq); | |
3368 | ||
b39e66ea | 3369 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3370 | u64 slice = sched_slice(cfs_rq, se); |
3371 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3372 | s64 delta = slice - ran; | |
3373 | ||
3374 | if (delta < 0) { | |
3375 | if (rq->curr == p) | |
3376 | resched_task(p); | |
3377 | return; | |
3378 | } | |
3379 | ||
3380 | /* | |
3381 | * Don't schedule slices shorter than 10000ns, that just | |
3382 | * doesn't make sense. Rely on vruntime for fairness. | |
3383 | */ | |
31656519 | 3384 | if (rq->curr != p) |
157124c1 | 3385 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3386 | |
31656519 | 3387 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3388 | } |
3389 | } | |
a4c2f00f PZ |
3390 | |
3391 | /* | |
3392 | * called from enqueue/dequeue and updates the hrtick when the | |
3393 | * current task is from our class and nr_running is low enough | |
3394 | * to matter. | |
3395 | */ | |
3396 | static void hrtick_update(struct rq *rq) | |
3397 | { | |
3398 | struct task_struct *curr = rq->curr; | |
3399 | ||
b39e66ea | 3400 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3401 | return; |
3402 | ||
3403 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3404 | hrtick_start_fair(rq, curr); | |
3405 | } | |
55e12e5e | 3406 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3407 | static inline void |
3408 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3409 | { | |
3410 | } | |
a4c2f00f PZ |
3411 | |
3412 | static inline void hrtick_update(struct rq *rq) | |
3413 | { | |
3414 | } | |
8f4d37ec PZ |
3415 | #endif |
3416 | ||
bf0f6f24 IM |
3417 | /* |
3418 | * The enqueue_task method is called before nr_running is | |
3419 | * increased. Here we update the fair scheduling stats and | |
3420 | * then put the task into the rbtree: | |
3421 | */ | |
ea87bb78 | 3422 | static void |
371fd7e7 | 3423 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3424 | { |
3425 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3426 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3427 | |
3428 | for_each_sched_entity(se) { | |
62fb1851 | 3429 | if (se->on_rq) |
bf0f6f24 IM |
3430 | break; |
3431 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3432 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3433 | |
3434 | /* | |
3435 | * end evaluation on encountering a throttled cfs_rq | |
3436 | * | |
3437 | * note: in the case of encountering a throttled cfs_rq we will | |
3438 | * post the final h_nr_running increment below. | |
3439 | */ | |
3440 | if (cfs_rq_throttled(cfs_rq)) | |
3441 | break; | |
953bfcd1 | 3442 | cfs_rq->h_nr_running++; |
85dac906 | 3443 | |
88ec22d3 | 3444 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3445 | } |
8f4d37ec | 3446 | |
2069dd75 | 3447 | for_each_sched_entity(se) { |
0f317143 | 3448 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3449 | cfs_rq->h_nr_running++; |
2069dd75 | 3450 | |
85dac906 PT |
3451 | if (cfs_rq_throttled(cfs_rq)) |
3452 | break; | |
3453 | ||
17bc14b7 | 3454 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3455 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3456 | } |
3457 | ||
18bf2805 BS |
3458 | if (!se) { |
3459 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3460 | inc_nr_running(rq); |
18bf2805 | 3461 | } |
a4c2f00f | 3462 | hrtick_update(rq); |
bf0f6f24 IM |
3463 | } |
3464 | ||
2f36825b VP |
3465 | static void set_next_buddy(struct sched_entity *se); |
3466 | ||
bf0f6f24 IM |
3467 | /* |
3468 | * The dequeue_task method is called before nr_running is | |
3469 | * decreased. We remove the task from the rbtree and | |
3470 | * update the fair scheduling stats: | |
3471 | */ | |
371fd7e7 | 3472 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3473 | { |
3474 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3475 | struct sched_entity *se = &p->se; |
2f36825b | 3476 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3477 | |
3478 | for_each_sched_entity(se) { | |
3479 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3480 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3481 | |
3482 | /* | |
3483 | * end evaluation on encountering a throttled cfs_rq | |
3484 | * | |
3485 | * note: in the case of encountering a throttled cfs_rq we will | |
3486 | * post the final h_nr_running decrement below. | |
3487 | */ | |
3488 | if (cfs_rq_throttled(cfs_rq)) | |
3489 | break; | |
953bfcd1 | 3490 | cfs_rq->h_nr_running--; |
2069dd75 | 3491 | |
bf0f6f24 | 3492 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3493 | if (cfs_rq->load.weight) { |
3494 | /* | |
3495 | * Bias pick_next to pick a task from this cfs_rq, as | |
3496 | * p is sleeping when it is within its sched_slice. | |
3497 | */ | |
3498 | if (task_sleep && parent_entity(se)) | |
3499 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3500 | |
3501 | /* avoid re-evaluating load for this entity */ | |
3502 | se = parent_entity(se); | |
bf0f6f24 | 3503 | break; |
2f36825b | 3504 | } |
371fd7e7 | 3505 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3506 | } |
8f4d37ec | 3507 | |
2069dd75 | 3508 | for_each_sched_entity(se) { |
0f317143 | 3509 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3510 | cfs_rq->h_nr_running--; |
2069dd75 | 3511 | |
85dac906 PT |
3512 | if (cfs_rq_throttled(cfs_rq)) |
3513 | break; | |
3514 | ||
17bc14b7 | 3515 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3516 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3517 | } |
3518 | ||
18bf2805 | 3519 | if (!se) { |
85dac906 | 3520 | dec_nr_running(rq); |
18bf2805 BS |
3521 | update_rq_runnable_avg(rq, 1); |
3522 | } | |
a4c2f00f | 3523 | hrtick_update(rq); |
bf0f6f24 IM |
3524 | } |
3525 | ||
e7693a36 | 3526 | #ifdef CONFIG_SMP |
029632fb PZ |
3527 | /* Used instead of source_load when we know the type == 0 */ |
3528 | static unsigned long weighted_cpuload(const int cpu) | |
3529 | { | |
b92486cb | 3530 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3531 | } |
3532 | ||
3533 | /* | |
3534 | * Return a low guess at the load of a migration-source cpu weighted | |
3535 | * according to the scheduling class and "nice" value. | |
3536 | * | |
3537 | * We want to under-estimate the load of migration sources, to | |
3538 | * balance conservatively. | |
3539 | */ | |
3540 | static unsigned long source_load(int cpu, int type) | |
3541 | { | |
3542 | struct rq *rq = cpu_rq(cpu); | |
3543 | unsigned long total = weighted_cpuload(cpu); | |
3544 | ||
3545 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3546 | return total; | |
3547 | ||
3548 | return min(rq->cpu_load[type-1], total); | |
3549 | } | |
3550 | ||
3551 | /* | |
3552 | * Return a high guess at the load of a migration-target cpu weighted | |
3553 | * according to the scheduling class and "nice" value. | |
3554 | */ | |
3555 | static unsigned long target_load(int cpu, int type) | |
3556 | { | |
3557 | struct rq *rq = cpu_rq(cpu); | |
3558 | unsigned long total = weighted_cpuload(cpu); | |
3559 | ||
3560 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3561 | return total; | |
3562 | ||
3563 | return max(rq->cpu_load[type-1], total); | |
3564 | } | |
3565 | ||
3566 | static unsigned long power_of(int cpu) | |
3567 | { | |
3568 | return cpu_rq(cpu)->cpu_power; | |
3569 | } | |
3570 | ||
3571 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3572 | { | |
3573 | struct rq *rq = cpu_rq(cpu); | |
3574 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3575 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3576 | |
3577 | if (nr_running) | |
b92486cb | 3578 | return load_avg / nr_running; |
029632fb PZ |
3579 | |
3580 | return 0; | |
3581 | } | |
3582 | ||
62470419 MW |
3583 | static void record_wakee(struct task_struct *p) |
3584 | { | |
3585 | /* | |
3586 | * Rough decay (wiping) for cost saving, don't worry | |
3587 | * about the boundary, really active task won't care | |
3588 | * about the loss. | |
3589 | */ | |
3590 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
3591 | current->wakee_flips = 0; | |
3592 | current->wakee_flip_decay_ts = jiffies; | |
3593 | } | |
3594 | ||
3595 | if (current->last_wakee != p) { | |
3596 | current->last_wakee = p; | |
3597 | current->wakee_flips++; | |
3598 | } | |
3599 | } | |
098fb9db | 3600 | |
74f8e4b2 | 3601 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
3602 | { |
3603 | struct sched_entity *se = &p->se; | |
3604 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
3605 | u64 min_vruntime; |
3606 | ||
3607 | #ifndef CONFIG_64BIT | |
3608 | u64 min_vruntime_copy; | |
88ec22d3 | 3609 | |
3fe1698b PZ |
3610 | do { |
3611 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3612 | smp_rmb(); | |
3613 | min_vruntime = cfs_rq->min_vruntime; | |
3614 | } while (min_vruntime != min_vruntime_copy); | |
3615 | #else | |
3616 | min_vruntime = cfs_rq->min_vruntime; | |
3617 | #endif | |
88ec22d3 | 3618 | |
3fe1698b | 3619 | se->vruntime -= min_vruntime; |
62470419 | 3620 | record_wakee(p); |
88ec22d3 PZ |
3621 | } |
3622 | ||
bb3469ac | 3623 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3624 | /* |
3625 | * effective_load() calculates the load change as seen from the root_task_group | |
3626 | * | |
3627 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3628 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3629 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3630 | * |
3631 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3632 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3633 | * total group weight. | |
3634 | * | |
3635 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3636 | * distribution (s_i) using: | |
3637 | * | |
3638 | * s_i = rw_i / \Sum rw_j (1) | |
3639 | * | |
3640 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3641 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3642 | * shares distribution (s_i): | |
3643 | * | |
3644 | * rw_i = { 2, 4, 1, 0 } | |
3645 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3646 | * | |
3647 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3648 | * task used to run on and the CPU the waker is running on), we need to | |
3649 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3650 | * wakeup, compute the effect of the current task going to sleep. | |
3651 | * | |
3652 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3653 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3654 | * | |
3655 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3656 | * | |
3657 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3658 | * differences in waking a task to CPU 0. The additional task changes the | |
3659 | * weight and shares distributions like: | |
3660 | * | |
3661 | * rw'_i = { 3, 4, 1, 0 } | |
3662 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3663 | * | |
3664 | * We can then compute the difference in effective weight by using: | |
3665 | * | |
3666 | * dw_i = S * (s'_i - s_i) (3) | |
3667 | * | |
3668 | * Where 'S' is the group weight as seen by its parent. | |
3669 | * | |
3670 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3671 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3672 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3673 | */ |
2069dd75 | 3674 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3675 | { |
4be9daaa | 3676 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3677 | |
58d081b5 | 3678 | if (!tg->parent || !wl) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3679 | return wl; |
3680 | ||
4be9daaa | 3681 | for_each_sched_entity(se) { |
cf5f0acf | 3682 | long w, W; |
4be9daaa | 3683 | |
977dda7c | 3684 | tg = se->my_q->tg; |
bb3469ac | 3685 | |
cf5f0acf PZ |
3686 | /* |
3687 | * W = @wg + \Sum rw_j | |
3688 | */ | |
3689 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3690 | |
cf5f0acf PZ |
3691 | /* |
3692 | * w = rw_i + @wl | |
3693 | */ | |
3694 | w = se->my_q->load.weight + wl; | |
940959e9 | 3695 | |
cf5f0acf PZ |
3696 | /* |
3697 | * wl = S * s'_i; see (2) | |
3698 | */ | |
3699 | if (W > 0 && w < W) | |
3700 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3701 | else |
3702 | wl = tg->shares; | |
940959e9 | 3703 | |
cf5f0acf PZ |
3704 | /* |
3705 | * Per the above, wl is the new se->load.weight value; since | |
3706 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3707 | * calc_cfs_shares(). | |
3708 | */ | |
977dda7c PT |
3709 | if (wl < MIN_SHARES) |
3710 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3711 | |
3712 | /* | |
3713 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3714 | */ | |
977dda7c | 3715 | wl -= se->load.weight; |
cf5f0acf PZ |
3716 | |
3717 | /* | |
3718 | * Recursively apply this logic to all parent groups to compute | |
3719 | * the final effective load change on the root group. Since | |
3720 | * only the @tg group gets extra weight, all parent groups can | |
3721 | * only redistribute existing shares. @wl is the shift in shares | |
3722 | * resulting from this level per the above. | |
3723 | */ | |
4be9daaa | 3724 | wg = 0; |
4be9daaa | 3725 | } |
bb3469ac | 3726 | |
4be9daaa | 3727 | return wl; |
bb3469ac PZ |
3728 | } |
3729 | #else | |
4be9daaa | 3730 | |
58d081b5 | 3731 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 3732 | { |
83378269 | 3733 | return wl; |
bb3469ac | 3734 | } |
4be9daaa | 3735 | |
bb3469ac PZ |
3736 | #endif |
3737 | ||
62470419 MW |
3738 | static int wake_wide(struct task_struct *p) |
3739 | { | |
7d9ffa89 | 3740 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
3741 | |
3742 | /* | |
3743 | * Yeah, it's the switching-frequency, could means many wakee or | |
3744 | * rapidly switch, use factor here will just help to automatically | |
3745 | * adjust the loose-degree, so bigger node will lead to more pull. | |
3746 | */ | |
3747 | if (p->wakee_flips > factor) { | |
3748 | /* | |
3749 | * wakee is somewhat hot, it needs certain amount of cpu | |
3750 | * resource, so if waker is far more hot, prefer to leave | |
3751 | * it alone. | |
3752 | */ | |
3753 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
3754 | return 1; | |
3755 | } | |
3756 | ||
3757 | return 0; | |
3758 | } | |
3759 | ||
c88d5910 | 3760 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 3761 | { |
e37b6a7b | 3762 | s64 this_load, load; |
c88d5910 | 3763 | int idx, this_cpu, prev_cpu; |
098fb9db | 3764 | unsigned long tl_per_task; |
c88d5910 | 3765 | struct task_group *tg; |
83378269 | 3766 | unsigned long weight; |
b3137bc8 | 3767 | int balanced; |
098fb9db | 3768 | |
62470419 MW |
3769 | /* |
3770 | * If we wake multiple tasks be careful to not bounce | |
3771 | * ourselves around too much. | |
3772 | */ | |
3773 | if (wake_wide(p)) | |
3774 | return 0; | |
3775 | ||
c88d5910 PZ |
3776 | idx = sd->wake_idx; |
3777 | this_cpu = smp_processor_id(); | |
3778 | prev_cpu = task_cpu(p); | |
3779 | load = source_load(prev_cpu, idx); | |
3780 | this_load = target_load(this_cpu, idx); | |
098fb9db | 3781 | |
b3137bc8 MG |
3782 | /* |
3783 | * If sync wakeup then subtract the (maximum possible) | |
3784 | * effect of the currently running task from the load | |
3785 | * of the current CPU: | |
3786 | */ | |
83378269 PZ |
3787 | if (sync) { |
3788 | tg = task_group(current); | |
3789 | weight = current->se.load.weight; | |
3790 | ||
c88d5910 | 3791 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
3792 | load += effective_load(tg, prev_cpu, 0, -weight); |
3793 | } | |
b3137bc8 | 3794 | |
83378269 PZ |
3795 | tg = task_group(p); |
3796 | weight = p->se.load.weight; | |
b3137bc8 | 3797 | |
71a29aa7 PZ |
3798 | /* |
3799 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
3800 | * due to the sync cause above having dropped this_load to 0, we'll |
3801 | * always have an imbalance, but there's really nothing you can do | |
3802 | * about that, so that's good too. | |
71a29aa7 PZ |
3803 | * |
3804 | * Otherwise check if either cpus are near enough in load to allow this | |
3805 | * task to be woken on this_cpu. | |
3806 | */ | |
e37b6a7b PT |
3807 | if (this_load > 0) { |
3808 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
3809 | |
3810 | this_eff_load = 100; | |
3811 | this_eff_load *= power_of(prev_cpu); | |
3812 | this_eff_load *= this_load + | |
3813 | effective_load(tg, this_cpu, weight, weight); | |
3814 | ||
3815 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
3816 | prev_eff_load *= power_of(this_cpu); | |
3817 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
3818 | ||
3819 | balanced = this_eff_load <= prev_eff_load; | |
3820 | } else | |
3821 | balanced = true; | |
b3137bc8 | 3822 | |
098fb9db | 3823 | /* |
4ae7d5ce IM |
3824 | * If the currently running task will sleep within |
3825 | * a reasonable amount of time then attract this newly | |
3826 | * woken task: | |
098fb9db | 3827 | */ |
2fb7635c PZ |
3828 | if (sync && balanced) |
3829 | return 1; | |
098fb9db | 3830 | |
41acab88 | 3831 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
3832 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
3833 | ||
c88d5910 PZ |
3834 | if (balanced || |
3835 | (this_load <= load && | |
3836 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
3837 | /* |
3838 | * This domain has SD_WAKE_AFFINE and | |
3839 | * p is cache cold in this domain, and | |
3840 | * there is no bad imbalance. | |
3841 | */ | |
c88d5910 | 3842 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 3843 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
3844 | |
3845 | return 1; | |
3846 | } | |
3847 | return 0; | |
3848 | } | |
3849 | ||
aaee1203 PZ |
3850 | /* |
3851 | * find_idlest_group finds and returns the least busy CPU group within the | |
3852 | * domain. | |
3853 | */ | |
3854 | static struct sched_group * | |
78e7ed53 | 3855 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 3856 | int this_cpu, int load_idx) |
e7693a36 | 3857 | { |
b3bd3de6 | 3858 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 3859 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 3860 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 3861 | |
aaee1203 PZ |
3862 | do { |
3863 | unsigned long load, avg_load; | |
3864 | int local_group; | |
3865 | int i; | |
e7693a36 | 3866 | |
aaee1203 PZ |
3867 | /* Skip over this group if it has no CPUs allowed */ |
3868 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 3869 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
3870 | continue; |
3871 | ||
3872 | local_group = cpumask_test_cpu(this_cpu, | |
3873 | sched_group_cpus(group)); | |
3874 | ||
3875 | /* Tally up the load of all CPUs in the group */ | |
3876 | avg_load = 0; | |
3877 | ||
3878 | for_each_cpu(i, sched_group_cpus(group)) { | |
3879 | /* Bias balancing toward cpus of our domain */ | |
3880 | if (local_group) | |
3881 | load = source_load(i, load_idx); | |
3882 | else | |
3883 | load = target_load(i, load_idx); | |
3884 | ||
3885 | avg_load += load; | |
3886 | } | |
3887 | ||
3888 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3889 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
3890 | |
3891 | if (local_group) { | |
3892 | this_load = avg_load; | |
aaee1203 PZ |
3893 | } else if (avg_load < min_load) { |
3894 | min_load = avg_load; | |
3895 | idlest = group; | |
3896 | } | |
3897 | } while (group = group->next, group != sd->groups); | |
3898 | ||
3899 | if (!idlest || 100*this_load < imbalance*min_load) | |
3900 | return NULL; | |
3901 | return idlest; | |
3902 | } | |
3903 | ||
3904 | /* | |
3905 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
3906 | */ | |
3907 | static int | |
3908 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
3909 | { | |
3910 | unsigned long load, min_load = ULONG_MAX; | |
3911 | int idlest = -1; | |
3912 | int i; | |
3913 | ||
3914 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 3915 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
3916 | load = weighted_cpuload(i); |
3917 | ||
3918 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
3919 | min_load = load; | |
3920 | idlest = i; | |
e7693a36 GH |
3921 | } |
3922 | } | |
3923 | ||
aaee1203 PZ |
3924 | return idlest; |
3925 | } | |
e7693a36 | 3926 | |
a50bde51 PZ |
3927 | /* |
3928 | * Try and locate an idle CPU in the sched_domain. | |
3929 | */ | |
99bd5e2f | 3930 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 3931 | { |
99bd5e2f | 3932 | struct sched_domain *sd; |
37407ea7 | 3933 | struct sched_group *sg; |
e0a79f52 | 3934 | int i = task_cpu(p); |
a50bde51 | 3935 | |
e0a79f52 MG |
3936 | if (idle_cpu(target)) |
3937 | return target; | |
99bd5e2f SS |
3938 | |
3939 | /* | |
e0a79f52 | 3940 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 3941 | */ |
e0a79f52 MG |
3942 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
3943 | return i; | |
a50bde51 PZ |
3944 | |
3945 | /* | |
37407ea7 | 3946 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 3947 | */ |
518cd623 | 3948 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 3949 | for_each_lower_domain(sd) { |
37407ea7 LT |
3950 | sg = sd->groups; |
3951 | do { | |
3952 | if (!cpumask_intersects(sched_group_cpus(sg), | |
3953 | tsk_cpus_allowed(p))) | |
3954 | goto next; | |
3955 | ||
3956 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 3957 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
3958 | goto next; |
3959 | } | |
970e1789 | 3960 | |
37407ea7 LT |
3961 | target = cpumask_first_and(sched_group_cpus(sg), |
3962 | tsk_cpus_allowed(p)); | |
3963 | goto done; | |
3964 | next: | |
3965 | sg = sg->next; | |
3966 | } while (sg != sd->groups); | |
3967 | } | |
3968 | done: | |
a50bde51 PZ |
3969 | return target; |
3970 | } | |
3971 | ||
aaee1203 PZ |
3972 | /* |
3973 | * sched_balance_self: balance the current task (running on cpu) in domains | |
3974 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
3975 | * SD_BALANCE_EXEC. | |
3976 | * | |
3977 | * Balance, ie. select the least loaded group. | |
3978 | * | |
3979 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
3980 | * | |
3981 | * preempt must be disabled. | |
3982 | */ | |
0017d735 | 3983 | static int |
ac66f547 | 3984 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 3985 | { |
29cd8bae | 3986 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 3987 | int cpu = smp_processor_id(); |
c88d5910 | 3988 | int new_cpu = cpu; |
99bd5e2f | 3989 | int want_affine = 0; |
5158f4e4 | 3990 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 3991 | |
29baa747 | 3992 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
3993 | return prev_cpu; |
3994 | ||
0763a660 | 3995 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 3996 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
3997 | want_affine = 1; |
3998 | new_cpu = prev_cpu; | |
3999 | } | |
aaee1203 | 4000 | |
dce840a0 | 4001 | rcu_read_lock(); |
aaee1203 | 4002 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4003 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4004 | continue; | |
4005 | ||
fe3bcfe1 | 4006 | /* |
99bd5e2f SS |
4007 | * If both cpu and prev_cpu are part of this domain, |
4008 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4009 | */ |
99bd5e2f SS |
4010 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4011 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4012 | affine_sd = tmp; | |
29cd8bae | 4013 | break; |
f03542a7 | 4014 | } |
29cd8bae | 4015 | |
f03542a7 | 4016 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4017 | sd = tmp; |
4018 | } | |
4019 | ||
8b911acd | 4020 | if (affine_sd) { |
f03542a7 | 4021 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
4022 | prev_cpu = cpu; |
4023 | ||
4024 | new_cpu = select_idle_sibling(p, prev_cpu); | |
4025 | goto unlock; | |
8b911acd | 4026 | } |
e7693a36 | 4027 | |
aaee1203 | 4028 | while (sd) { |
5158f4e4 | 4029 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4030 | struct sched_group *group; |
c88d5910 | 4031 | int weight; |
098fb9db | 4032 | |
0763a660 | 4033 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4034 | sd = sd->child; |
4035 | continue; | |
4036 | } | |
098fb9db | 4037 | |
5158f4e4 PZ |
4038 | if (sd_flag & SD_BALANCE_WAKE) |
4039 | load_idx = sd->wake_idx; | |
098fb9db | 4040 | |
5158f4e4 | 4041 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
4042 | if (!group) { |
4043 | sd = sd->child; | |
4044 | continue; | |
4045 | } | |
4ae7d5ce | 4046 | |
d7c33c49 | 4047 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4048 | if (new_cpu == -1 || new_cpu == cpu) { |
4049 | /* Now try balancing at a lower domain level of cpu */ | |
4050 | sd = sd->child; | |
4051 | continue; | |
e7693a36 | 4052 | } |
aaee1203 PZ |
4053 | |
4054 | /* Now try balancing at a lower domain level of new_cpu */ | |
4055 | cpu = new_cpu; | |
669c55e9 | 4056 | weight = sd->span_weight; |
aaee1203 PZ |
4057 | sd = NULL; |
4058 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4059 | if (weight <= tmp->span_weight) |
aaee1203 | 4060 | break; |
0763a660 | 4061 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4062 | sd = tmp; |
4063 | } | |
4064 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4065 | } |
dce840a0 PZ |
4066 | unlock: |
4067 | rcu_read_unlock(); | |
e7693a36 | 4068 | |
c88d5910 | 4069 | return new_cpu; |
e7693a36 | 4070 | } |
0a74bef8 PT |
4071 | |
4072 | /* | |
4073 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4074 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4075 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4076 | * other assumptions, including the state of rq->lock, should be made. | |
4077 | */ | |
4078 | static void | |
4079 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4080 | { | |
aff3e498 PT |
4081 | struct sched_entity *se = &p->se; |
4082 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4083 | ||
4084 | /* | |
4085 | * Load tracking: accumulate removed load so that it can be processed | |
4086 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4087 | * to blocked load iff they have a positive decay-count. It can never | |
4088 | * be negative here since on-rq tasks have decay-count == 0. | |
4089 | */ | |
4090 | if (se->avg.decay_count) { | |
4091 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4092 | atomic_long_add(se->avg.load_avg_contrib, |
4093 | &cfs_rq->removed_load); | |
aff3e498 | 4094 | } |
0a74bef8 | 4095 | } |
e7693a36 GH |
4096 | #endif /* CONFIG_SMP */ |
4097 | ||
e52fb7c0 PZ |
4098 | static unsigned long |
4099 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4100 | { |
4101 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4102 | ||
4103 | /* | |
e52fb7c0 PZ |
4104 | * Since its curr running now, convert the gran from real-time |
4105 | * to virtual-time in his units. | |
13814d42 MG |
4106 | * |
4107 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4108 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4109 | * the resulting gran will be larger, therefore penalizing the | |
4110 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4111 | * be smaller, again penalizing the lighter task. | |
4112 | * | |
4113 | * This is especially important for buddies when the leftmost | |
4114 | * task is higher priority than the buddy. | |
0bbd3336 | 4115 | */ |
f4ad9bd2 | 4116 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4117 | } |
4118 | ||
464b7527 PZ |
4119 | /* |
4120 | * Should 'se' preempt 'curr'. | |
4121 | * | |
4122 | * |s1 | |
4123 | * |s2 | |
4124 | * |s3 | |
4125 | * g | |
4126 | * |<--->|c | |
4127 | * | |
4128 | * w(c, s1) = -1 | |
4129 | * w(c, s2) = 0 | |
4130 | * w(c, s3) = 1 | |
4131 | * | |
4132 | */ | |
4133 | static int | |
4134 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4135 | { | |
4136 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4137 | ||
4138 | if (vdiff <= 0) | |
4139 | return -1; | |
4140 | ||
e52fb7c0 | 4141 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4142 | if (vdiff > gran) |
4143 | return 1; | |
4144 | ||
4145 | return 0; | |
4146 | } | |
4147 | ||
02479099 PZ |
4148 | static void set_last_buddy(struct sched_entity *se) |
4149 | { | |
69c80f3e VP |
4150 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4151 | return; | |
4152 | ||
4153 | for_each_sched_entity(se) | |
4154 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4155 | } |
4156 | ||
4157 | static void set_next_buddy(struct sched_entity *se) | |
4158 | { | |
69c80f3e VP |
4159 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4160 | return; | |
4161 | ||
4162 | for_each_sched_entity(se) | |
4163 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4164 | } |
4165 | ||
ac53db59 RR |
4166 | static void set_skip_buddy(struct sched_entity *se) |
4167 | { | |
69c80f3e VP |
4168 | for_each_sched_entity(se) |
4169 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4170 | } |
4171 | ||
bf0f6f24 IM |
4172 | /* |
4173 | * Preempt the current task with a newly woken task if needed: | |
4174 | */ | |
5a9b86f6 | 4175 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4176 | { |
4177 | struct task_struct *curr = rq->curr; | |
8651a86c | 4178 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4179 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4180 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4181 | int next_buddy_marked = 0; |
bf0f6f24 | 4182 | |
4ae7d5ce IM |
4183 | if (unlikely(se == pse)) |
4184 | return; | |
4185 | ||
5238cdd3 | 4186 | /* |
ddcdf6e7 | 4187 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
4188 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4189 | * lead to a throttle). This both saves work and prevents false | |
4190 | * next-buddy nomination below. | |
4191 | */ | |
4192 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4193 | return; | |
4194 | ||
2f36825b | 4195 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4196 | set_next_buddy(pse); |
2f36825b VP |
4197 | next_buddy_marked = 1; |
4198 | } | |
57fdc26d | 4199 | |
aec0a514 BR |
4200 | /* |
4201 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4202 | * wake up path. | |
5238cdd3 PT |
4203 | * |
4204 | * Note: this also catches the edge-case of curr being in a throttled | |
4205 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4206 | * enqueue of curr) will have resulted in resched being set. This | |
4207 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4208 | * below. | |
aec0a514 BR |
4209 | */ |
4210 | if (test_tsk_need_resched(curr)) | |
4211 | return; | |
4212 | ||
a2f5c9ab DH |
4213 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4214 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4215 | likely(p->policy != SCHED_IDLE)) | |
4216 | goto preempt; | |
4217 | ||
91c234b4 | 4218 | /* |
a2f5c9ab DH |
4219 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4220 | * is driven by the tick): | |
91c234b4 | 4221 | */ |
8ed92e51 | 4222 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4223 | return; |
bf0f6f24 | 4224 | |
464b7527 | 4225 | find_matching_se(&se, &pse); |
9bbd7374 | 4226 | update_curr(cfs_rq_of(se)); |
002f128b | 4227 | BUG_ON(!pse); |
2f36825b VP |
4228 | if (wakeup_preempt_entity(se, pse) == 1) { |
4229 | /* | |
4230 | * Bias pick_next to pick the sched entity that is | |
4231 | * triggering this preemption. | |
4232 | */ | |
4233 | if (!next_buddy_marked) | |
4234 | set_next_buddy(pse); | |
3a7e73a2 | 4235 | goto preempt; |
2f36825b | 4236 | } |
464b7527 | 4237 | |
3a7e73a2 | 4238 | return; |
a65ac745 | 4239 | |
3a7e73a2 PZ |
4240 | preempt: |
4241 | resched_task(curr); | |
4242 | /* | |
4243 | * Only set the backward buddy when the current task is still | |
4244 | * on the rq. This can happen when a wakeup gets interleaved | |
4245 | * with schedule on the ->pre_schedule() or idle_balance() | |
4246 | * point, either of which can * drop the rq lock. | |
4247 | * | |
4248 | * Also, during early boot the idle thread is in the fair class, | |
4249 | * for obvious reasons its a bad idea to schedule back to it. | |
4250 | */ | |
4251 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
4252 | return; | |
4253 | ||
4254 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
4255 | set_last_buddy(se); | |
bf0f6f24 IM |
4256 | } |
4257 | ||
fb8d4724 | 4258 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 4259 | { |
8f4d37ec | 4260 | struct task_struct *p; |
bf0f6f24 IM |
4261 | struct cfs_rq *cfs_rq = &rq->cfs; |
4262 | struct sched_entity *se; | |
4263 | ||
36ace27e | 4264 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
4265 | return NULL; |
4266 | ||
4267 | do { | |
9948f4b2 | 4268 | se = pick_next_entity(cfs_rq); |
f4b6755f | 4269 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
4270 | cfs_rq = group_cfs_rq(se); |
4271 | } while (cfs_rq); | |
4272 | ||
8f4d37ec | 4273 | p = task_of(se); |
b39e66ea MG |
4274 | if (hrtick_enabled(rq)) |
4275 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4276 | |
4277 | return p; | |
bf0f6f24 IM |
4278 | } |
4279 | ||
4280 | /* | |
4281 | * Account for a descheduled task: | |
4282 | */ | |
31ee529c | 4283 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4284 | { |
4285 | struct sched_entity *se = &prev->se; | |
4286 | struct cfs_rq *cfs_rq; | |
4287 | ||
4288 | for_each_sched_entity(se) { | |
4289 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4290 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4291 | } |
4292 | } | |
4293 | ||
ac53db59 RR |
4294 | /* |
4295 | * sched_yield() is very simple | |
4296 | * | |
4297 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4298 | */ | |
4299 | static void yield_task_fair(struct rq *rq) | |
4300 | { | |
4301 | struct task_struct *curr = rq->curr; | |
4302 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4303 | struct sched_entity *se = &curr->se; | |
4304 | ||
4305 | /* | |
4306 | * Are we the only task in the tree? | |
4307 | */ | |
4308 | if (unlikely(rq->nr_running == 1)) | |
4309 | return; | |
4310 | ||
4311 | clear_buddies(cfs_rq, se); | |
4312 | ||
4313 | if (curr->policy != SCHED_BATCH) { | |
4314 | update_rq_clock(rq); | |
4315 | /* | |
4316 | * Update run-time statistics of the 'current'. | |
4317 | */ | |
4318 | update_curr(cfs_rq); | |
916671c0 MG |
4319 | /* |
4320 | * Tell update_rq_clock() that we've just updated, | |
4321 | * so we don't do microscopic update in schedule() | |
4322 | * and double the fastpath cost. | |
4323 | */ | |
4324 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4325 | } |
4326 | ||
4327 | set_skip_buddy(se); | |
4328 | } | |
4329 | ||
d95f4122 MG |
4330 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4331 | { | |
4332 | struct sched_entity *se = &p->se; | |
4333 | ||
5238cdd3 PT |
4334 | /* throttled hierarchies are not runnable */ |
4335 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4336 | return false; |
4337 | ||
4338 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4339 | set_next_buddy(se); | |
4340 | ||
d95f4122 MG |
4341 | yield_task_fair(rq); |
4342 | ||
4343 | return true; | |
4344 | } | |
4345 | ||
681f3e68 | 4346 | #ifdef CONFIG_SMP |
bf0f6f24 | 4347 | /************************************************** |
e9c84cb8 PZ |
4348 | * Fair scheduling class load-balancing methods. |
4349 | * | |
4350 | * BASICS | |
4351 | * | |
4352 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4353 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4354 | * time to each task. This is expressed in the following equation: | |
4355 | * | |
4356 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4357 | * | |
4358 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4359 | * W_i,0 is defined as: | |
4360 | * | |
4361 | * W_i,0 = \Sum_j w_i,j (2) | |
4362 | * | |
4363 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4364 | * is derived from the nice value as per prio_to_weight[]. | |
4365 | * | |
4366 | * The weight average is an exponential decay average of the instantaneous | |
4367 | * weight: | |
4368 | * | |
4369 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4370 | * | |
4371 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4372 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4373 | * can also include other factors [XXX]. | |
4374 | * | |
4375 | * To achieve this balance we define a measure of imbalance which follows | |
4376 | * directly from (1): | |
4377 | * | |
4378 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4379 | * | |
4380 | * We them move tasks around to minimize the imbalance. In the continuous | |
4381 | * function space it is obvious this converges, in the discrete case we get | |
4382 | * a few fun cases generally called infeasible weight scenarios. | |
4383 | * | |
4384 | * [XXX expand on: | |
4385 | * - infeasible weights; | |
4386 | * - local vs global optima in the discrete case. ] | |
4387 | * | |
4388 | * | |
4389 | * SCHED DOMAINS | |
4390 | * | |
4391 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4392 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4393 | * topology where each level pairs two lower groups (or better). This results | |
4394 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4395 | * tree to only the first of the previous level and we decrease the frequency | |
4396 | * of load-balance at each level inv. proportional to the number of cpus in | |
4397 | * the groups. | |
4398 | * | |
4399 | * This yields: | |
4400 | * | |
4401 | * log_2 n 1 n | |
4402 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4403 | * i = 0 2^i 2^i | |
4404 | * `- size of each group | |
4405 | * | | `- number of cpus doing load-balance | |
4406 | * | `- freq | |
4407 | * `- sum over all levels | |
4408 | * | |
4409 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4410 | * this makes (5) the runtime complexity of the balancer. | |
4411 | * | |
4412 | * An important property here is that each CPU is still (indirectly) connected | |
4413 | * to every other cpu in at most O(log n) steps: | |
4414 | * | |
4415 | * The adjacency matrix of the resulting graph is given by: | |
4416 | * | |
4417 | * log_2 n | |
4418 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4419 | * k = 0 | |
4420 | * | |
4421 | * And you'll find that: | |
4422 | * | |
4423 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
4424 | * | |
4425 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
4426 | * The task movement gives a factor of O(m), giving a convergence complexity | |
4427 | * of: | |
4428 | * | |
4429 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
4430 | * | |
4431 | * | |
4432 | * WORK CONSERVING | |
4433 | * | |
4434 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
4435 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
4436 | * tree itself instead of relying on other CPUs to bring it work. | |
4437 | * | |
4438 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
4439 | * time. | |
4440 | * | |
4441 | * [XXX more?] | |
4442 | * | |
4443 | * | |
4444 | * CGROUPS | |
4445 | * | |
4446 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
4447 | * | |
4448 | * s_k,i | |
4449 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
4450 | * S_k | |
4451 | * | |
4452 | * Where | |
4453 | * | |
4454 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4455 | * | |
4456 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4457 | * | |
4458 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4459 | * property. | |
4460 | * | |
4461 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4462 | * rewrite all of this once again.] | |
4463 | */ | |
bf0f6f24 | 4464 | |
ed387b78 HS |
4465 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4466 | ||
ddcdf6e7 | 4467 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4468 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4469 | #define LBF_DST_PINNED 0x04 |
4470 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4471 | |
4472 | struct lb_env { | |
4473 | struct sched_domain *sd; | |
4474 | ||
ddcdf6e7 | 4475 | struct rq *src_rq; |
85c1e7da | 4476 | int src_cpu; |
ddcdf6e7 PZ |
4477 | |
4478 | int dst_cpu; | |
4479 | struct rq *dst_rq; | |
4480 | ||
88b8dac0 SV |
4481 | struct cpumask *dst_grpmask; |
4482 | int new_dst_cpu; | |
ddcdf6e7 | 4483 | enum cpu_idle_type idle; |
bd939f45 | 4484 | long imbalance; |
b9403130 MW |
4485 | /* The set of CPUs under consideration for load-balancing */ |
4486 | struct cpumask *cpus; | |
4487 | ||
ddcdf6e7 | 4488 | unsigned int flags; |
367456c7 PZ |
4489 | |
4490 | unsigned int loop; | |
4491 | unsigned int loop_break; | |
4492 | unsigned int loop_max; | |
ddcdf6e7 PZ |
4493 | }; |
4494 | ||
1e3c88bd | 4495 | /* |
ddcdf6e7 | 4496 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
4497 | * Both runqueues must be locked. |
4498 | */ | |
ddcdf6e7 | 4499 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 4500 | { |
ddcdf6e7 PZ |
4501 | deactivate_task(env->src_rq, p, 0); |
4502 | set_task_cpu(p, env->dst_cpu); | |
4503 | activate_task(env->dst_rq, p, 0); | |
4504 | check_preempt_curr(env->dst_rq, p, 0); | |
6fe6b2d6 RR |
4505 | #ifdef CONFIG_NUMA_BALANCING |
4506 | if (p->numa_preferred_nid != -1) { | |
4507 | int src_nid = cpu_to_node(env->src_cpu); | |
4508 | int dst_nid = cpu_to_node(env->dst_cpu); | |
4509 | ||
4510 | /* | |
4511 | * If the load balancer has moved the task then limit | |
4512 | * migrations from taking place in the short term in | |
4513 | * case this is a short-lived migration. | |
4514 | */ | |
4515 | if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid) | |
4516 | p->numa_migrate_seq = 0; | |
4517 | } | |
4518 | #endif | |
1e3c88bd PZ |
4519 | } |
4520 | ||
029632fb PZ |
4521 | /* |
4522 | * Is this task likely cache-hot: | |
4523 | */ | |
4524 | static int | |
4525 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
4526 | { | |
4527 | s64 delta; | |
4528 | ||
4529 | if (p->sched_class != &fair_sched_class) | |
4530 | return 0; | |
4531 | ||
4532 | if (unlikely(p->policy == SCHED_IDLE)) | |
4533 | return 0; | |
4534 | ||
4535 | /* | |
4536 | * Buddy candidates are cache hot: | |
4537 | */ | |
4538 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
4539 | (&p->se == cfs_rq_of(&p->se)->next || | |
4540 | &p->se == cfs_rq_of(&p->se)->last)) | |
4541 | return 1; | |
4542 | ||
4543 | if (sysctl_sched_migration_cost == -1) | |
4544 | return 1; | |
4545 | if (sysctl_sched_migration_cost == 0) | |
4546 | return 0; | |
4547 | ||
4548 | delta = now - p->se.exec_start; | |
4549 | ||
4550 | return delta < (s64)sysctl_sched_migration_cost; | |
4551 | } | |
4552 | ||
3a7053b3 MG |
4553 | #ifdef CONFIG_NUMA_BALANCING |
4554 | /* Returns true if the destination node has incurred more faults */ | |
4555 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
4556 | { | |
4557 | int src_nid, dst_nid; | |
4558 | ||
4559 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || | |
4560 | !(env->sd->flags & SD_NUMA)) { | |
4561 | return false; | |
4562 | } | |
4563 | ||
4564 | src_nid = cpu_to_node(env->src_cpu); | |
4565 | dst_nid = cpu_to_node(env->dst_cpu); | |
4566 | ||
4567 | if (src_nid == dst_nid || | |
4568 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4569 | return false; | |
4570 | ||
4571 | if (dst_nid == p->numa_preferred_nid || | |
ac8e895b | 4572 | task_faults(p, dst_nid) > task_faults(p, src_nid)) |
3a7053b3 MG |
4573 | return true; |
4574 | ||
4575 | return false; | |
4576 | } | |
7a0f3083 MG |
4577 | |
4578 | ||
4579 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
4580 | { | |
4581 | int src_nid, dst_nid; | |
4582 | ||
4583 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
4584 | return false; | |
4585 | ||
4586 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) | |
4587 | return false; | |
4588 | ||
4589 | src_nid = cpu_to_node(env->src_cpu); | |
4590 | dst_nid = cpu_to_node(env->dst_cpu); | |
4591 | ||
4592 | if (src_nid == dst_nid || | |
4593 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4594 | return false; | |
4595 | ||
ac8e895b | 4596 | if (task_faults(p, dst_nid) < task_faults(p, src_nid)) |
7a0f3083 MG |
4597 | return true; |
4598 | ||
4599 | return false; | |
4600 | } | |
4601 | ||
3a7053b3 MG |
4602 | #else |
4603 | static inline bool migrate_improves_locality(struct task_struct *p, | |
4604 | struct lb_env *env) | |
4605 | { | |
4606 | return false; | |
4607 | } | |
7a0f3083 MG |
4608 | |
4609 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
4610 | struct lb_env *env) | |
4611 | { | |
4612 | return false; | |
4613 | } | |
3a7053b3 MG |
4614 | #endif |
4615 | ||
1e3c88bd PZ |
4616 | /* |
4617 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
4618 | */ | |
4619 | static | |
8e45cb54 | 4620 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
4621 | { |
4622 | int tsk_cache_hot = 0; | |
4623 | /* | |
4624 | * We do not migrate tasks that are: | |
d3198084 | 4625 | * 1) throttled_lb_pair, or |
1e3c88bd | 4626 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
4627 | * 3) running (obviously), or |
4628 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 4629 | */ |
d3198084 JK |
4630 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
4631 | return 0; | |
4632 | ||
ddcdf6e7 | 4633 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 4634 | int cpu; |
88b8dac0 | 4635 | |
41acab88 | 4636 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 4637 | |
6263322c PZ |
4638 | env->flags |= LBF_SOME_PINNED; |
4639 | ||
88b8dac0 SV |
4640 | /* |
4641 | * Remember if this task can be migrated to any other cpu in | |
4642 | * our sched_group. We may want to revisit it if we couldn't | |
4643 | * meet load balance goals by pulling other tasks on src_cpu. | |
4644 | * | |
4645 | * Also avoid computing new_dst_cpu if we have already computed | |
4646 | * one in current iteration. | |
4647 | */ | |
6263322c | 4648 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
4649 | return 0; |
4650 | ||
e02e60c1 JK |
4651 | /* Prevent to re-select dst_cpu via env's cpus */ |
4652 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
4653 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 4654 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
4655 | env->new_dst_cpu = cpu; |
4656 | break; | |
4657 | } | |
88b8dac0 | 4658 | } |
e02e60c1 | 4659 | |
1e3c88bd PZ |
4660 | return 0; |
4661 | } | |
88b8dac0 SV |
4662 | |
4663 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 4664 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 4665 | |
ddcdf6e7 | 4666 | if (task_running(env->src_rq, p)) { |
41acab88 | 4667 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
4668 | return 0; |
4669 | } | |
4670 | ||
4671 | /* | |
4672 | * Aggressive migration if: | |
3a7053b3 MG |
4673 | * 1) destination numa is preferred |
4674 | * 2) task is cache cold, or | |
4675 | * 3) too many balance attempts have failed. | |
1e3c88bd | 4676 | */ |
78becc27 | 4677 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
7a0f3083 MG |
4678 | if (!tsk_cache_hot) |
4679 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
4680 | |
4681 | if (migrate_improves_locality(p, env)) { | |
4682 | #ifdef CONFIG_SCHEDSTATS | |
4683 | if (tsk_cache_hot) { | |
4684 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
4685 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
4686 | } | |
4687 | #endif | |
4688 | return 1; | |
4689 | } | |
4690 | ||
1e3c88bd | 4691 | if (!tsk_cache_hot || |
8e45cb54 | 4692 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 4693 | |
1e3c88bd | 4694 | if (tsk_cache_hot) { |
8e45cb54 | 4695 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 4696 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 4697 | } |
4e2dcb73 | 4698 | |
1e3c88bd PZ |
4699 | return 1; |
4700 | } | |
4701 | ||
4e2dcb73 ZH |
4702 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
4703 | return 0; | |
1e3c88bd PZ |
4704 | } |
4705 | ||
897c395f PZ |
4706 | /* |
4707 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
4708 | * part of active balancing operations within "domain". | |
4709 | * Returns 1 if successful and 0 otherwise. | |
4710 | * | |
4711 | * Called with both runqueues locked. | |
4712 | */ | |
8e45cb54 | 4713 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
4714 | { |
4715 | struct task_struct *p, *n; | |
897c395f | 4716 | |
367456c7 | 4717 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
4718 | if (!can_migrate_task(p, env)) |
4719 | continue; | |
897c395f | 4720 | |
367456c7 PZ |
4721 | move_task(p, env); |
4722 | /* | |
4723 | * Right now, this is only the second place move_task() | |
4724 | * is called, so we can safely collect move_task() | |
4725 | * stats here rather than inside move_task(). | |
4726 | */ | |
4727 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4728 | return 1; | |
897c395f | 4729 | } |
897c395f PZ |
4730 | return 0; |
4731 | } | |
4732 | ||
eb95308e PZ |
4733 | static const unsigned int sched_nr_migrate_break = 32; |
4734 | ||
5d6523eb | 4735 | /* |
bd939f45 | 4736 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4737 | * this_rq, as part of a balancing operation within domain "sd". |
4738 | * Returns 1 if successful and 0 otherwise. | |
4739 | * | |
4740 | * Called with both runqueues locked. | |
4741 | */ | |
4742 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4743 | { |
5d6523eb PZ |
4744 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4745 | struct task_struct *p; | |
367456c7 PZ |
4746 | unsigned long load; |
4747 | int pulled = 0; | |
1e3c88bd | 4748 | |
bd939f45 | 4749 | if (env->imbalance <= 0) |
5d6523eb | 4750 | return 0; |
1e3c88bd | 4751 | |
5d6523eb PZ |
4752 | while (!list_empty(tasks)) { |
4753 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 4754 | |
367456c7 PZ |
4755 | env->loop++; |
4756 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 4757 | if (env->loop > env->loop_max) |
367456c7 | 4758 | break; |
5d6523eb PZ |
4759 | |
4760 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 4761 | if (env->loop > env->loop_break) { |
eb95308e | 4762 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 4763 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 4764 | break; |
a195f004 | 4765 | } |
1e3c88bd | 4766 | |
d3198084 | 4767 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
4768 | goto next; |
4769 | ||
4770 | load = task_h_load(p); | |
5d6523eb | 4771 | |
eb95308e | 4772 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
4773 | goto next; |
4774 | ||
bd939f45 | 4775 | if ((load / 2) > env->imbalance) |
367456c7 | 4776 | goto next; |
1e3c88bd | 4777 | |
ddcdf6e7 | 4778 | move_task(p, env); |
ee00e66f | 4779 | pulled++; |
bd939f45 | 4780 | env->imbalance -= load; |
1e3c88bd PZ |
4781 | |
4782 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
4783 | /* |
4784 | * NEWIDLE balancing is a source of latency, so preemptible | |
4785 | * kernels will stop after the first task is pulled to minimize | |
4786 | * the critical section. | |
4787 | */ | |
5d6523eb | 4788 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 4789 | break; |
1e3c88bd PZ |
4790 | #endif |
4791 | ||
ee00e66f PZ |
4792 | /* |
4793 | * We only want to steal up to the prescribed amount of | |
4794 | * weighted load. | |
4795 | */ | |
bd939f45 | 4796 | if (env->imbalance <= 0) |
ee00e66f | 4797 | break; |
367456c7 PZ |
4798 | |
4799 | continue; | |
4800 | next: | |
5d6523eb | 4801 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 4802 | } |
5d6523eb | 4803 | |
1e3c88bd | 4804 | /* |
ddcdf6e7 PZ |
4805 | * Right now, this is one of only two places move_task() is called, |
4806 | * so we can safely collect move_task() stats here rather than | |
4807 | * inside move_task(). | |
1e3c88bd | 4808 | */ |
8e45cb54 | 4809 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 4810 | |
5d6523eb | 4811 | return pulled; |
1e3c88bd PZ |
4812 | } |
4813 | ||
230059de | 4814 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
4815 | /* |
4816 | * update tg->load_weight by folding this cpu's load_avg | |
4817 | */ | |
48a16753 | 4818 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 4819 | { |
48a16753 PT |
4820 | struct sched_entity *se = tg->se[cpu]; |
4821 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 4822 | |
48a16753 PT |
4823 | /* throttled entities do not contribute to load */ |
4824 | if (throttled_hierarchy(cfs_rq)) | |
4825 | return; | |
9e3081ca | 4826 | |
aff3e498 | 4827 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 4828 | |
82958366 PT |
4829 | if (se) { |
4830 | update_entity_load_avg(se, 1); | |
4831 | /* | |
4832 | * We pivot on our runnable average having decayed to zero for | |
4833 | * list removal. This generally implies that all our children | |
4834 | * have also been removed (modulo rounding error or bandwidth | |
4835 | * control); however, such cases are rare and we can fix these | |
4836 | * at enqueue. | |
4837 | * | |
4838 | * TODO: fix up out-of-order children on enqueue. | |
4839 | */ | |
4840 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
4841 | list_del_leaf_cfs_rq(cfs_rq); | |
4842 | } else { | |
48a16753 | 4843 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
4844 | update_rq_runnable_avg(rq, rq->nr_running); |
4845 | } | |
9e3081ca PZ |
4846 | } |
4847 | ||
48a16753 | 4848 | static void update_blocked_averages(int cpu) |
9e3081ca | 4849 | { |
9e3081ca | 4850 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
4851 | struct cfs_rq *cfs_rq; |
4852 | unsigned long flags; | |
9e3081ca | 4853 | |
48a16753 PT |
4854 | raw_spin_lock_irqsave(&rq->lock, flags); |
4855 | update_rq_clock(rq); | |
9763b67f PZ |
4856 | /* |
4857 | * Iterates the task_group tree in a bottom up fashion, see | |
4858 | * list_add_leaf_cfs_rq() for details. | |
4859 | */ | |
64660c86 | 4860 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
4861 | /* |
4862 | * Note: We may want to consider periodically releasing | |
4863 | * rq->lock about these updates so that creating many task | |
4864 | * groups does not result in continually extending hold time. | |
4865 | */ | |
4866 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 4867 | } |
48a16753 PT |
4868 | |
4869 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
4870 | } |
4871 | ||
9763b67f | 4872 | /* |
68520796 | 4873 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
4874 | * This needs to be done in a top-down fashion because the load of a child |
4875 | * group is a fraction of its parents load. | |
4876 | */ | |
68520796 | 4877 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 4878 | { |
68520796 VD |
4879 | struct rq *rq = rq_of(cfs_rq); |
4880 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 4881 | unsigned long now = jiffies; |
68520796 | 4882 | unsigned long load; |
a35b6466 | 4883 | |
68520796 | 4884 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
4885 | return; |
4886 | ||
68520796 VD |
4887 | cfs_rq->h_load_next = NULL; |
4888 | for_each_sched_entity(se) { | |
4889 | cfs_rq = cfs_rq_of(se); | |
4890 | cfs_rq->h_load_next = se; | |
4891 | if (cfs_rq->last_h_load_update == now) | |
4892 | break; | |
4893 | } | |
a35b6466 | 4894 | |
68520796 | 4895 | if (!se) { |
7e3115ef | 4896 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
4897 | cfs_rq->last_h_load_update = now; |
4898 | } | |
4899 | ||
4900 | while ((se = cfs_rq->h_load_next) != NULL) { | |
4901 | load = cfs_rq->h_load; | |
4902 | load = div64_ul(load * se->avg.load_avg_contrib, | |
4903 | cfs_rq->runnable_load_avg + 1); | |
4904 | cfs_rq = group_cfs_rq(se); | |
4905 | cfs_rq->h_load = load; | |
4906 | cfs_rq->last_h_load_update = now; | |
4907 | } | |
9763b67f PZ |
4908 | } |
4909 | ||
367456c7 | 4910 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 4911 | { |
367456c7 | 4912 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 4913 | |
68520796 | 4914 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
4915 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
4916 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
4917 | } |
4918 | #else | |
48a16753 | 4919 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
4920 | { |
4921 | } | |
4922 | ||
367456c7 | 4923 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 4924 | { |
a003a25b | 4925 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 4926 | } |
230059de | 4927 | #endif |
1e3c88bd | 4928 | |
1e3c88bd | 4929 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
4930 | /* |
4931 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
4932 | */ | |
4933 | struct sg_lb_stats { | |
4934 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
4935 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 4936 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 4937 | unsigned long load_per_task; |
3ae11c90 | 4938 | unsigned long group_power; |
147c5fc2 PZ |
4939 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
4940 | unsigned int group_capacity; | |
4941 | unsigned int idle_cpus; | |
4942 | unsigned int group_weight; | |
1e3c88bd | 4943 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 4944 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
4945 | }; |
4946 | ||
56cf515b JK |
4947 | /* |
4948 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
4949 | * during load balancing. | |
4950 | */ | |
4951 | struct sd_lb_stats { | |
4952 | struct sched_group *busiest; /* Busiest group in this sd */ | |
4953 | struct sched_group *local; /* Local group in this sd */ | |
4954 | unsigned long total_load; /* Total load of all groups in sd */ | |
4955 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
4956 | unsigned long avg_load; /* Average load across all groups in sd */ | |
4957 | ||
56cf515b | 4958 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 4959 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
4960 | }; |
4961 | ||
147c5fc2 PZ |
4962 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
4963 | { | |
4964 | /* | |
4965 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
4966 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
4967 | * We must however clear busiest_stat::avg_load because | |
4968 | * update_sd_pick_busiest() reads this before assignment. | |
4969 | */ | |
4970 | *sds = (struct sd_lb_stats){ | |
4971 | .busiest = NULL, | |
4972 | .local = NULL, | |
4973 | .total_load = 0UL, | |
4974 | .total_pwr = 0UL, | |
4975 | .busiest_stat = { | |
4976 | .avg_load = 0UL, | |
4977 | }, | |
4978 | }; | |
4979 | } | |
4980 | ||
1e3c88bd PZ |
4981 | /** |
4982 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
4983 | * @sd: The sched_domain whose load_idx is to be obtained. | |
4984 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
e69f6186 YB |
4985 | * |
4986 | * Return: The load index. | |
1e3c88bd PZ |
4987 | */ |
4988 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
4989 | enum cpu_idle_type idle) | |
4990 | { | |
4991 | int load_idx; | |
4992 | ||
4993 | switch (idle) { | |
4994 | case CPU_NOT_IDLE: | |
4995 | load_idx = sd->busy_idx; | |
4996 | break; | |
4997 | ||
4998 | case CPU_NEWLY_IDLE: | |
4999 | load_idx = sd->newidle_idx; | |
5000 | break; | |
5001 | default: | |
5002 | load_idx = sd->idle_idx; | |
5003 | break; | |
5004 | } | |
5005 | ||
5006 | return load_idx; | |
5007 | } | |
5008 | ||
15f803c9 | 5009 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5010 | { |
1399fa78 | 5011 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
5012 | } |
5013 | ||
5014 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
5015 | { | |
5016 | return default_scale_freq_power(sd, cpu); | |
5017 | } | |
5018 | ||
15f803c9 | 5019 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5020 | { |
669c55e9 | 5021 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
5022 | unsigned long smt_gain = sd->smt_gain; |
5023 | ||
5024 | smt_gain /= weight; | |
5025 | ||
5026 | return smt_gain; | |
5027 | } | |
5028 | ||
5029 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
5030 | { | |
5031 | return default_scale_smt_power(sd, cpu); | |
5032 | } | |
5033 | ||
15f803c9 | 5034 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
5035 | { |
5036 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5037 | u64 total, available, age_stamp, avg; |
1e3c88bd | 5038 | |
b654f7de PZ |
5039 | /* |
5040 | * Since we're reading these variables without serialization make sure | |
5041 | * we read them once before doing sanity checks on them. | |
5042 | */ | |
5043 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5044 | avg = ACCESS_ONCE(rq->rt_avg); | |
5045 | ||
78becc27 | 5046 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 5047 | |
b654f7de | 5048 | if (unlikely(total < avg)) { |
aa483808 VP |
5049 | /* Ensures that power won't end up being negative */ |
5050 | available = 0; | |
5051 | } else { | |
b654f7de | 5052 | available = total - avg; |
aa483808 | 5053 | } |
1e3c88bd | 5054 | |
1399fa78 NR |
5055 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
5056 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 5057 | |
1399fa78 | 5058 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5059 | |
5060 | return div_u64(available, total); | |
5061 | } | |
5062 | ||
5063 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
5064 | { | |
669c55e9 | 5065 | unsigned long weight = sd->span_weight; |
1399fa78 | 5066 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
5067 | struct sched_group *sdg = sd->groups; |
5068 | ||
1e3c88bd PZ |
5069 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
5070 | if (sched_feat(ARCH_POWER)) | |
5071 | power *= arch_scale_smt_power(sd, cpu); | |
5072 | else | |
5073 | power *= default_scale_smt_power(sd, cpu); | |
5074 | ||
1399fa78 | 5075 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5076 | } |
5077 | ||
9c3f75cb | 5078 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
5079 | |
5080 | if (sched_feat(ARCH_POWER)) | |
5081 | power *= arch_scale_freq_power(sd, cpu); | |
5082 | else | |
5083 | power *= default_scale_freq_power(sd, cpu); | |
5084 | ||
1399fa78 | 5085 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 5086 | |
1e3c88bd | 5087 | power *= scale_rt_power(cpu); |
1399fa78 | 5088 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5089 | |
5090 | if (!power) | |
5091 | power = 1; | |
5092 | ||
e51fd5e2 | 5093 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 5094 | sdg->sgp->power = power; |
1e3c88bd PZ |
5095 | } |
5096 | ||
029632fb | 5097 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
5098 | { |
5099 | struct sched_domain *child = sd->child; | |
5100 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 5101 | unsigned long power, power_orig; |
4ec4412e VG |
5102 | unsigned long interval; |
5103 | ||
5104 | interval = msecs_to_jiffies(sd->balance_interval); | |
5105 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
5106 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
5107 | |
5108 | if (!child) { | |
5109 | update_cpu_power(sd, cpu); | |
5110 | return; | |
5111 | } | |
5112 | ||
863bffc8 | 5113 | power_orig = power = 0; |
1e3c88bd | 5114 | |
74a5ce20 PZ |
5115 | if (child->flags & SD_OVERLAP) { |
5116 | /* | |
5117 | * SD_OVERLAP domains cannot assume that child groups | |
5118 | * span the current group. | |
5119 | */ | |
5120 | ||
863bffc8 PZ |
5121 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
5122 | struct sched_group *sg = cpu_rq(cpu)->sd->groups; | |
5123 | ||
5124 | power_orig += sg->sgp->power_orig; | |
5125 | power += sg->sgp->power; | |
5126 | } | |
74a5ce20 PZ |
5127 | } else { |
5128 | /* | |
5129 | * !SD_OVERLAP domains can assume that child groups | |
5130 | * span the current group. | |
5131 | */ | |
5132 | ||
5133 | group = child->groups; | |
5134 | do { | |
863bffc8 | 5135 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
5136 | power += group->sgp->power; |
5137 | group = group->next; | |
5138 | } while (group != child->groups); | |
5139 | } | |
1e3c88bd | 5140 | |
863bffc8 PZ |
5141 | sdg->sgp->power_orig = power_orig; |
5142 | sdg->sgp->power = power; | |
1e3c88bd PZ |
5143 | } |
5144 | ||
9d5efe05 SV |
5145 | /* |
5146 | * Try and fix up capacity for tiny siblings, this is needed when | |
5147 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
5148 | * which on its own isn't powerful enough. | |
5149 | * | |
5150 | * See update_sd_pick_busiest() and check_asym_packing(). | |
5151 | */ | |
5152 | static inline int | |
5153 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
5154 | { | |
5155 | /* | |
1399fa78 | 5156 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 5157 | */ |
a6c75f2f | 5158 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
5159 | return 0; |
5160 | ||
5161 | /* | |
5162 | * If ~90% of the cpu_power is still there, we're good. | |
5163 | */ | |
9c3f75cb | 5164 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
5165 | return 1; |
5166 | ||
5167 | return 0; | |
5168 | } | |
5169 | ||
30ce5dab PZ |
5170 | /* |
5171 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
5172 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
5173 | * | |
5174 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
5175 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
5176 | * Something like: | |
5177 | * | |
5178 | * { 0 1 2 3 } { 4 5 6 7 } | |
5179 | * * * * * | |
5180 | * | |
5181 | * If we were to balance group-wise we'd place two tasks in the first group and | |
5182 | * two tasks in the second group. Clearly this is undesired as it will overload | |
5183 | * cpu 3 and leave one of the cpus in the second group unused. | |
5184 | * | |
5185 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
5186 | * by noticing the lower domain failed to reach balance and had difficulty |
5187 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
5188 | * |
5189 | * When this is so detected; this group becomes a candidate for busiest; see | |
5190 | * update_sd_pick_busiest(). And calculcate_imbalance() and | |
6263322c | 5191 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
5192 | * to create an effective group imbalance. |
5193 | * | |
5194 | * This is a somewhat tricky proposition since the next run might not find the | |
5195 | * group imbalance and decide the groups need to be balanced again. A most | |
5196 | * subtle and fragile situation. | |
5197 | */ | |
5198 | ||
6263322c | 5199 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 5200 | { |
6263322c | 5201 | return group->sgp->imbalance; |
30ce5dab PZ |
5202 | } |
5203 | ||
b37d9316 PZ |
5204 | /* |
5205 | * Compute the group capacity. | |
5206 | * | |
c61037e9 PZ |
5207 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
5208 | * first dividing out the smt factor and computing the actual number of cores | |
5209 | * and limit power unit capacity with that. | |
b37d9316 PZ |
5210 | */ |
5211 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
5212 | { | |
c61037e9 PZ |
5213 | unsigned int capacity, smt, cpus; |
5214 | unsigned int power, power_orig; | |
5215 | ||
5216 | power = group->sgp->power; | |
5217 | power_orig = group->sgp->power_orig; | |
5218 | cpus = group->group_weight; | |
b37d9316 | 5219 | |
c61037e9 PZ |
5220 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
5221 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
5222 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 5223 | |
c61037e9 | 5224 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
5225 | if (!capacity) |
5226 | capacity = fix_small_capacity(env->sd, group); | |
5227 | ||
5228 | return capacity; | |
5229 | } | |
5230 | ||
1e3c88bd PZ |
5231 | /** |
5232 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 5233 | * @env: The load balancing environment. |
1e3c88bd | 5234 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 5235 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 5236 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
5237 | * @sgs: variable to hold the statistics for this group. |
5238 | */ | |
bd939f45 PZ |
5239 | static inline void update_sg_lb_stats(struct lb_env *env, |
5240 | struct sched_group *group, int load_idx, | |
23f0d209 | 5241 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 5242 | { |
30ce5dab PZ |
5243 | unsigned long nr_running; |
5244 | unsigned long load; | |
bd939f45 | 5245 | int i; |
1e3c88bd | 5246 | |
b72ff13c PZ |
5247 | memset(sgs, 0, sizeof(*sgs)); |
5248 | ||
b9403130 | 5249 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
5250 | struct rq *rq = cpu_rq(i); |
5251 | ||
e44bc5c5 PZ |
5252 | nr_running = rq->nr_running; |
5253 | ||
1e3c88bd | 5254 | /* Bias balancing toward cpus of our domain */ |
6263322c | 5255 | if (local_group) |
04f733b4 | 5256 | load = target_load(i, load_idx); |
6263322c | 5257 | else |
1e3c88bd | 5258 | load = source_load(i, load_idx); |
1e3c88bd PZ |
5259 | |
5260 | sgs->group_load += load; | |
e44bc5c5 | 5261 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 5262 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
5263 | if (idle_cpu(i)) |
5264 | sgs->idle_cpus++; | |
1e3c88bd PZ |
5265 | } |
5266 | ||
1e3c88bd | 5267 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
5268 | sgs->group_power = group->sgp->power; |
5269 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 5270 | |
dd5feea1 | 5271 | if (sgs->sum_nr_running) |
38d0f770 | 5272 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5273 | |
aae6d3dd | 5274 | sgs->group_weight = group->group_weight; |
fab47622 | 5275 | |
b37d9316 PZ |
5276 | sgs->group_imb = sg_imbalanced(group); |
5277 | sgs->group_capacity = sg_capacity(env, group); | |
5278 | ||
fab47622 NR |
5279 | if (sgs->group_capacity > sgs->sum_nr_running) |
5280 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5281 | } |
5282 | ||
532cb4c4 MN |
5283 | /** |
5284 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5285 | * @env: The load balancing environment. |
532cb4c4 MN |
5286 | * @sds: sched_domain statistics |
5287 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5288 | * @sgs: sched_group statistics |
532cb4c4 MN |
5289 | * |
5290 | * Determine if @sg is a busier group than the previously selected | |
5291 | * busiest group. | |
e69f6186 YB |
5292 | * |
5293 | * Return: %true if @sg is a busier group than the previously selected | |
5294 | * busiest group. %false otherwise. | |
532cb4c4 | 5295 | */ |
bd939f45 | 5296 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5297 | struct sd_lb_stats *sds, |
5298 | struct sched_group *sg, | |
bd939f45 | 5299 | struct sg_lb_stats *sgs) |
532cb4c4 | 5300 | { |
56cf515b | 5301 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5302 | return false; |
5303 | ||
5304 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5305 | return true; | |
5306 | ||
5307 | if (sgs->group_imb) | |
5308 | return true; | |
5309 | ||
5310 | /* | |
5311 | * ASYM_PACKING needs to move all the work to the lowest | |
5312 | * numbered CPUs in the group, therefore mark all groups | |
5313 | * higher than ourself as busy. | |
5314 | */ | |
bd939f45 PZ |
5315 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5316 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5317 | if (!sds->busiest) |
5318 | return true; | |
5319 | ||
5320 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5321 | return true; | |
5322 | } | |
5323 | ||
5324 | return false; | |
5325 | } | |
5326 | ||
1e3c88bd | 5327 | /** |
461819ac | 5328 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5329 | * @env: The load balancing environment. |
1e3c88bd PZ |
5330 | * @balance: Should we balance. |
5331 | * @sds: variable to hold the statistics for this sched_domain. | |
5332 | */ | |
bd939f45 | 5333 | static inline void update_sd_lb_stats(struct lb_env *env, |
23f0d209 | 5334 | struct sd_lb_stats *sds) |
1e3c88bd | 5335 | { |
bd939f45 PZ |
5336 | struct sched_domain *child = env->sd->child; |
5337 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5338 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5339 | int load_idx, prefer_sibling = 0; |
5340 | ||
5341 | if (child && child->flags & SD_PREFER_SIBLING) | |
5342 | prefer_sibling = 1; | |
5343 | ||
bd939f45 | 5344 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5345 | |
5346 | do { | |
56cf515b | 5347 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5348 | int local_group; |
5349 | ||
bd939f45 | 5350 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5351 | if (local_group) { |
5352 | sds->local = sg; | |
5353 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5354 | |
5355 | if (env->idle != CPU_NEWLY_IDLE || | |
5356 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5357 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5358 | } |
1e3c88bd | 5359 | |
56cf515b | 5360 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5361 | |
b72ff13c PZ |
5362 | if (local_group) |
5363 | goto next_group; | |
5364 | ||
1e3c88bd PZ |
5365 | /* |
5366 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 5367 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
5368 | * and move all the excess tasks away. We lower the capacity |
5369 | * of a group only if the local group has the capacity to fit | |
5370 | * these excess tasks, i.e. nr_running < group_capacity. The | |
5371 | * extra check prevents the case where you always pull from the | |
5372 | * heaviest group when it is already under-utilized (possible | |
5373 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 5374 | */ |
b72ff13c PZ |
5375 | if (prefer_sibling && sds->local && |
5376 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 5377 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 5378 | |
b72ff13c | 5379 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 5380 | sds->busiest = sg; |
56cf515b | 5381 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
5382 | } |
5383 | ||
b72ff13c PZ |
5384 | next_group: |
5385 | /* Now, start updating sd_lb_stats */ | |
5386 | sds->total_load += sgs->group_load; | |
5387 | sds->total_pwr += sgs->group_power; | |
5388 | ||
532cb4c4 | 5389 | sg = sg->next; |
bd939f45 | 5390 | } while (sg != env->sd->groups); |
532cb4c4 MN |
5391 | } |
5392 | ||
532cb4c4 MN |
5393 | /** |
5394 | * check_asym_packing - Check to see if the group is packed into the | |
5395 | * sched doman. | |
5396 | * | |
5397 | * This is primarily intended to used at the sibling level. Some | |
5398 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
5399 | * case of POWER7, it can move to lower SMT modes only when higher | |
5400 | * threads are idle. When in lower SMT modes, the threads will | |
5401 | * perform better since they share less core resources. Hence when we | |
5402 | * have idle threads, we want them to be the higher ones. | |
5403 | * | |
5404 | * This packing function is run on idle threads. It checks to see if | |
5405 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
5406 | * CPU number than the packing function is being run on. Here we are | |
5407 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
5408 | * number. | |
5409 | * | |
e69f6186 | 5410 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
5411 | * this CPU. The amount of the imbalance is returned in *imbalance. |
5412 | * | |
cd96891d | 5413 | * @env: The load balancing environment. |
532cb4c4 | 5414 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 5415 | */ |
bd939f45 | 5416 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
5417 | { |
5418 | int busiest_cpu; | |
5419 | ||
bd939f45 | 5420 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
5421 | return 0; |
5422 | ||
5423 | if (!sds->busiest) | |
5424 | return 0; | |
5425 | ||
5426 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 5427 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
5428 | return 0; |
5429 | ||
bd939f45 | 5430 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
5431 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
5432 | SCHED_POWER_SCALE); | |
bd939f45 | 5433 | |
532cb4c4 | 5434 | return 1; |
1e3c88bd PZ |
5435 | } |
5436 | ||
5437 | /** | |
5438 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
5439 | * amongst the groups of a sched_domain, during | |
5440 | * load balancing. | |
cd96891d | 5441 | * @env: The load balancing environment. |
1e3c88bd | 5442 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5443 | */ |
bd939f45 PZ |
5444 | static inline |
5445 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
5446 | { |
5447 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
5448 | unsigned int imbn = 2; | |
dd5feea1 | 5449 | unsigned long scaled_busy_load_per_task; |
56cf515b | 5450 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 5451 | |
56cf515b JK |
5452 | local = &sds->local_stat; |
5453 | busiest = &sds->busiest_stat; | |
1e3c88bd | 5454 | |
56cf515b JK |
5455 | if (!local->sum_nr_running) |
5456 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
5457 | else if (busiest->load_per_task > local->load_per_task) | |
5458 | imbn = 1; | |
dd5feea1 | 5459 | |
56cf515b JK |
5460 | scaled_busy_load_per_task = |
5461 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5462 | busiest->group_power; |
56cf515b | 5463 | |
3029ede3 VD |
5464 | if (busiest->avg_load + scaled_busy_load_per_task >= |
5465 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 5466 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5467 | return; |
5468 | } | |
5469 | ||
5470 | /* | |
5471 | * OK, we don't have enough imbalance to justify moving tasks, | |
5472 | * however we may be able to increase total CPU power used by | |
5473 | * moving them. | |
5474 | */ | |
5475 | ||
3ae11c90 | 5476 | pwr_now += busiest->group_power * |
56cf515b | 5477 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 5478 | pwr_now += local->group_power * |
56cf515b | 5479 | min(local->load_per_task, local->avg_load); |
1399fa78 | 5480 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5481 | |
5482 | /* Amount of load we'd subtract */ | |
56cf515b | 5483 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 5484 | busiest->group_power; |
56cf515b | 5485 | if (busiest->avg_load > tmp) { |
3ae11c90 | 5486 | pwr_move += busiest->group_power * |
56cf515b JK |
5487 | min(busiest->load_per_task, |
5488 | busiest->avg_load - tmp); | |
5489 | } | |
1e3c88bd PZ |
5490 | |
5491 | /* Amount of load we'd add */ | |
3ae11c90 | 5492 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 5493 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
5494 | tmp = (busiest->avg_load * busiest->group_power) / |
5495 | local->group_power; | |
56cf515b JK |
5496 | } else { |
5497 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5498 | local->group_power; |
56cf515b | 5499 | } |
3ae11c90 PZ |
5500 | pwr_move += local->group_power * |
5501 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 5502 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5503 | |
5504 | /* Move if we gain throughput */ | |
5505 | if (pwr_move > pwr_now) | |
56cf515b | 5506 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5507 | } |
5508 | ||
5509 | /** | |
5510 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
5511 | * groups of a given sched_domain during load balance. | |
bd939f45 | 5512 | * @env: load balance environment |
1e3c88bd | 5513 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5514 | */ |
bd939f45 | 5515 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5516 | { |
dd5feea1 | 5517 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
5518 | struct sg_lb_stats *local, *busiest; |
5519 | ||
5520 | local = &sds->local_stat; | |
56cf515b | 5521 | busiest = &sds->busiest_stat; |
dd5feea1 | 5522 | |
56cf515b | 5523 | if (busiest->group_imb) { |
30ce5dab PZ |
5524 | /* |
5525 | * In the group_imb case we cannot rely on group-wide averages | |
5526 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
5527 | */ | |
56cf515b JK |
5528 | busiest->load_per_task = |
5529 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
5530 | } |
5531 | ||
1e3c88bd PZ |
5532 | /* |
5533 | * In the presence of smp nice balancing, certain scenarios can have | |
5534 | * max load less than avg load(as we skip the groups at or below | |
5535 | * its cpu_power, while calculating max_load..) | |
5536 | */ | |
b1885550 VD |
5537 | if (busiest->avg_load <= sds->avg_load || |
5538 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
5539 | env->imbalance = 0; |
5540 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
5541 | } |
5542 | ||
56cf515b | 5543 | if (!busiest->group_imb) { |
dd5feea1 SS |
5544 | /* |
5545 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
5546 | * Except of course for the group_imb case, since then we might |
5547 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 5548 | */ |
56cf515b JK |
5549 | load_above_capacity = |
5550 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 5551 | |
1399fa78 | 5552 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 5553 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
5554 | } |
5555 | ||
5556 | /* | |
5557 | * We're trying to get all the cpus to the average_load, so we don't | |
5558 | * want to push ourselves above the average load, nor do we wish to | |
5559 | * reduce the max loaded cpu below the average load. At the same time, | |
5560 | * we also don't want to reduce the group load below the group capacity | |
5561 | * (so that we can implement power-savings policies etc). Thus we look | |
5562 | * for the minimum possible imbalance. | |
dd5feea1 | 5563 | */ |
30ce5dab | 5564 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
5565 | |
5566 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 5567 | env->imbalance = min( |
3ae11c90 PZ |
5568 | max_pull * busiest->group_power, |
5569 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 5570 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
5571 | |
5572 | /* | |
5573 | * if *imbalance is less than the average load per runnable task | |
25985edc | 5574 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
5575 | * a think about bumping its value to force at least one task to be |
5576 | * moved | |
5577 | */ | |
56cf515b | 5578 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 5579 | return fix_small_imbalance(env, sds); |
1e3c88bd | 5580 | } |
fab47622 | 5581 | |
1e3c88bd PZ |
5582 | /******* find_busiest_group() helpers end here *********************/ |
5583 | ||
5584 | /** | |
5585 | * find_busiest_group - Returns the busiest group within the sched_domain | |
5586 | * if there is an imbalance. If there isn't an imbalance, and | |
5587 | * the user has opted for power-savings, it returns a group whose | |
5588 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
5589 | * such a group exists. | |
5590 | * | |
5591 | * Also calculates the amount of weighted load which should be moved | |
5592 | * to restore balance. | |
5593 | * | |
cd96891d | 5594 | * @env: The load balancing environment. |
1e3c88bd | 5595 | * |
e69f6186 | 5596 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
5597 | * - If no imbalance and user has opted for power-savings balance, |
5598 | * return the least loaded group whose CPUs can be | |
5599 | * put to idle by rebalancing its tasks onto our group. | |
5600 | */ | |
56cf515b | 5601 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 5602 | { |
56cf515b | 5603 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
5604 | struct sd_lb_stats sds; |
5605 | ||
147c5fc2 | 5606 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
5607 | |
5608 | /* | |
5609 | * Compute the various statistics relavent for load balancing at | |
5610 | * this level. | |
5611 | */ | |
23f0d209 | 5612 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
5613 | local = &sds.local_stat; |
5614 | busiest = &sds.busiest_stat; | |
1e3c88bd | 5615 | |
bd939f45 PZ |
5616 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
5617 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
5618 | return sds.busiest; |
5619 | ||
cc57aa8f | 5620 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 5621 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
5622 | goto out_balanced; |
5623 | ||
1399fa78 | 5624 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 5625 | |
866ab43e PZ |
5626 | /* |
5627 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 5628 | * work because they assume all things are equal, which typically |
866ab43e PZ |
5629 | * isn't true due to cpus_allowed constraints and the like. |
5630 | */ | |
56cf515b | 5631 | if (busiest->group_imb) |
866ab43e PZ |
5632 | goto force_balance; |
5633 | ||
cc57aa8f | 5634 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
5635 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
5636 | !busiest->group_has_capacity) | |
fab47622 NR |
5637 | goto force_balance; |
5638 | ||
cc57aa8f PZ |
5639 | /* |
5640 | * If the local group is more busy than the selected busiest group | |
5641 | * don't try and pull any tasks. | |
5642 | */ | |
56cf515b | 5643 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
5644 | goto out_balanced; |
5645 | ||
cc57aa8f PZ |
5646 | /* |
5647 | * Don't pull any tasks if this group is already above the domain | |
5648 | * average load. | |
5649 | */ | |
56cf515b | 5650 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
5651 | goto out_balanced; |
5652 | ||
bd939f45 | 5653 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
5654 | /* |
5655 | * This cpu is idle. If the busiest group load doesn't | |
5656 | * have more tasks than the number of available cpu's and | |
5657 | * there is no imbalance between this and busiest group | |
5658 | * wrt to idle cpu's, it is balanced. | |
5659 | */ | |
56cf515b JK |
5660 | if ((local->idle_cpus < busiest->idle_cpus) && |
5661 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 5662 | goto out_balanced; |
c186fafe PZ |
5663 | } else { |
5664 | /* | |
5665 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
5666 | * imbalance_pct to be conservative. | |
5667 | */ | |
56cf515b JK |
5668 | if (100 * busiest->avg_load <= |
5669 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 5670 | goto out_balanced; |
aae6d3dd | 5671 | } |
1e3c88bd | 5672 | |
fab47622 | 5673 | force_balance: |
1e3c88bd | 5674 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 5675 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
5676 | return sds.busiest; |
5677 | ||
5678 | out_balanced: | |
bd939f45 | 5679 | env->imbalance = 0; |
1e3c88bd PZ |
5680 | return NULL; |
5681 | } | |
5682 | ||
5683 | /* | |
5684 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
5685 | */ | |
bd939f45 | 5686 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 5687 | struct sched_group *group) |
1e3c88bd PZ |
5688 | { |
5689 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 5690 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
5691 | int i; |
5692 | ||
6906a408 | 5693 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd | 5694 | unsigned long power = power_of(i); |
1399fa78 NR |
5695 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
5696 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
5697 | unsigned long wl; |
5698 | ||
9d5efe05 | 5699 | if (!capacity) |
bd939f45 | 5700 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 5701 | |
1e3c88bd | 5702 | rq = cpu_rq(i); |
6e40f5bb | 5703 | wl = weighted_cpuload(i); |
1e3c88bd | 5704 | |
6e40f5bb TG |
5705 | /* |
5706 | * When comparing with imbalance, use weighted_cpuload() | |
5707 | * which is not scaled with the cpu power. | |
5708 | */ | |
bd939f45 | 5709 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
5710 | continue; |
5711 | ||
6e40f5bb TG |
5712 | /* |
5713 | * For the load comparisons with the other cpu's, consider | |
5714 | * the weighted_cpuload() scaled with the cpu power, so that | |
5715 | * the load can be moved away from the cpu that is potentially | |
5716 | * running at a lower capacity. | |
95a79b80 JK |
5717 | * |
5718 | * Thus we're looking for max(wl_i / power_i), crosswise | |
5719 | * multiplication to rid ourselves of the division works out | |
5720 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
5721 | * previous maximum. | |
6e40f5bb | 5722 | */ |
95a79b80 JK |
5723 | if (wl * busiest_power > busiest_load * power) { |
5724 | busiest_load = wl; | |
5725 | busiest_power = power; | |
1e3c88bd PZ |
5726 | busiest = rq; |
5727 | } | |
5728 | } | |
5729 | ||
5730 | return busiest; | |
5731 | } | |
5732 | ||
5733 | /* | |
5734 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
5735 | * so long as it is large enough. | |
5736 | */ | |
5737 | #define MAX_PINNED_INTERVAL 512 | |
5738 | ||
5739 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 5740 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 5741 | |
bd939f45 | 5742 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 5743 | { |
bd939f45 PZ |
5744 | struct sched_domain *sd = env->sd; |
5745 | ||
5746 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
5747 | |
5748 | /* | |
5749 | * ASYM_PACKING needs to force migrate tasks from busy but | |
5750 | * higher numbered CPUs in order to pack all tasks in the | |
5751 | * lowest numbered CPUs. | |
5752 | */ | |
bd939f45 | 5753 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 5754 | return 1; |
1af3ed3d PZ |
5755 | } |
5756 | ||
5757 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
5758 | } | |
5759 | ||
969c7921 TH |
5760 | static int active_load_balance_cpu_stop(void *data); |
5761 | ||
23f0d209 JK |
5762 | static int should_we_balance(struct lb_env *env) |
5763 | { | |
5764 | struct sched_group *sg = env->sd->groups; | |
5765 | struct cpumask *sg_cpus, *sg_mask; | |
5766 | int cpu, balance_cpu = -1; | |
5767 | ||
5768 | /* | |
5769 | * In the newly idle case, we will allow all the cpu's | |
5770 | * to do the newly idle load balance. | |
5771 | */ | |
5772 | if (env->idle == CPU_NEWLY_IDLE) | |
5773 | return 1; | |
5774 | ||
5775 | sg_cpus = sched_group_cpus(sg); | |
5776 | sg_mask = sched_group_mask(sg); | |
5777 | /* Try to find first idle cpu */ | |
5778 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
5779 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
5780 | continue; | |
5781 | ||
5782 | balance_cpu = cpu; | |
5783 | break; | |
5784 | } | |
5785 | ||
5786 | if (balance_cpu == -1) | |
5787 | balance_cpu = group_balance_cpu(sg); | |
5788 | ||
5789 | /* | |
5790 | * First idle cpu or the first cpu(busiest) in this sched group | |
5791 | * is eligible for doing load balancing at this and above domains. | |
5792 | */ | |
b0cff9d8 | 5793 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
5794 | } |
5795 | ||
1e3c88bd PZ |
5796 | /* |
5797 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
5798 | * tasks if there is an imbalance. | |
5799 | */ | |
5800 | static int load_balance(int this_cpu, struct rq *this_rq, | |
5801 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 5802 | int *continue_balancing) |
1e3c88bd | 5803 | { |
88b8dac0 | 5804 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 5805 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 5806 | struct sched_group *group; |
1e3c88bd PZ |
5807 | struct rq *busiest; |
5808 | unsigned long flags; | |
e6252c3e | 5809 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 5810 | |
8e45cb54 PZ |
5811 | struct lb_env env = { |
5812 | .sd = sd, | |
ddcdf6e7 PZ |
5813 | .dst_cpu = this_cpu, |
5814 | .dst_rq = this_rq, | |
88b8dac0 | 5815 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 5816 | .idle = idle, |
eb95308e | 5817 | .loop_break = sched_nr_migrate_break, |
b9403130 | 5818 | .cpus = cpus, |
8e45cb54 PZ |
5819 | }; |
5820 | ||
cfc03118 JK |
5821 | /* |
5822 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
5823 | * other cpus in our group | |
5824 | */ | |
e02e60c1 | 5825 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 5826 | env.dst_grpmask = NULL; |
cfc03118 | 5827 | |
1e3c88bd PZ |
5828 | cpumask_copy(cpus, cpu_active_mask); |
5829 | ||
1e3c88bd PZ |
5830 | schedstat_inc(sd, lb_count[idle]); |
5831 | ||
5832 | redo: | |
23f0d209 JK |
5833 | if (!should_we_balance(&env)) { |
5834 | *continue_balancing = 0; | |
1e3c88bd | 5835 | goto out_balanced; |
23f0d209 | 5836 | } |
1e3c88bd | 5837 | |
23f0d209 | 5838 | group = find_busiest_group(&env); |
1e3c88bd PZ |
5839 | if (!group) { |
5840 | schedstat_inc(sd, lb_nobusyg[idle]); | |
5841 | goto out_balanced; | |
5842 | } | |
5843 | ||
b9403130 | 5844 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
5845 | if (!busiest) { |
5846 | schedstat_inc(sd, lb_nobusyq[idle]); | |
5847 | goto out_balanced; | |
5848 | } | |
5849 | ||
78feefc5 | 5850 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 5851 | |
bd939f45 | 5852 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
5853 | |
5854 | ld_moved = 0; | |
5855 | if (busiest->nr_running > 1) { | |
5856 | /* | |
5857 | * Attempt to move tasks. If find_busiest_group has found | |
5858 | * an imbalance but busiest->nr_running <= 1, the group is | |
5859 | * still unbalanced. ld_moved simply stays zero, so it is | |
5860 | * correctly treated as an imbalance. | |
5861 | */ | |
8e45cb54 | 5862 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
5863 | env.src_cpu = busiest->cpu; |
5864 | env.src_rq = busiest; | |
5865 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 5866 | |
5d6523eb | 5867 | more_balance: |
1e3c88bd | 5868 | local_irq_save(flags); |
78feefc5 | 5869 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
5870 | |
5871 | /* | |
5872 | * cur_ld_moved - load moved in current iteration | |
5873 | * ld_moved - cumulative load moved across iterations | |
5874 | */ | |
5875 | cur_ld_moved = move_tasks(&env); | |
5876 | ld_moved += cur_ld_moved; | |
78feefc5 | 5877 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
5878 | local_irq_restore(flags); |
5879 | ||
5880 | /* | |
5881 | * some other cpu did the load balance for us. | |
5882 | */ | |
88b8dac0 SV |
5883 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
5884 | resched_cpu(env.dst_cpu); | |
5885 | ||
f1cd0858 JK |
5886 | if (env.flags & LBF_NEED_BREAK) { |
5887 | env.flags &= ~LBF_NEED_BREAK; | |
5888 | goto more_balance; | |
5889 | } | |
5890 | ||
88b8dac0 SV |
5891 | /* |
5892 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
5893 | * us and move them to an alternate dst_cpu in our sched_group | |
5894 | * where they can run. The upper limit on how many times we | |
5895 | * iterate on same src_cpu is dependent on number of cpus in our | |
5896 | * sched_group. | |
5897 | * | |
5898 | * This changes load balance semantics a bit on who can move | |
5899 | * load to a given_cpu. In addition to the given_cpu itself | |
5900 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
5901 | * nohz-idle), we now have balance_cpu in a position to move | |
5902 | * load to given_cpu. In rare situations, this may cause | |
5903 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
5904 | * _independently_ and at _same_ time to move some load to | |
5905 | * given_cpu) causing exceess load to be moved to given_cpu. | |
5906 | * This however should not happen so much in practice and | |
5907 | * moreover subsequent load balance cycles should correct the | |
5908 | * excess load moved. | |
5909 | */ | |
6263322c | 5910 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 5911 | |
7aff2e3a VD |
5912 | /* Prevent to re-select dst_cpu via env's cpus */ |
5913 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
5914 | ||
78feefc5 | 5915 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 5916 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 5917 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
5918 | env.loop = 0; |
5919 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 5920 | |
88b8dac0 SV |
5921 | /* |
5922 | * Go back to "more_balance" rather than "redo" since we | |
5923 | * need to continue with same src_cpu. | |
5924 | */ | |
5925 | goto more_balance; | |
5926 | } | |
1e3c88bd | 5927 | |
6263322c PZ |
5928 | /* |
5929 | * We failed to reach balance because of affinity. | |
5930 | */ | |
5931 | if (sd_parent) { | |
5932 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
5933 | ||
5934 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
5935 | *group_imbalance = 1; | |
5936 | } else if (*group_imbalance) | |
5937 | *group_imbalance = 0; | |
5938 | } | |
5939 | ||
1e3c88bd | 5940 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 5941 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 5942 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
5943 | if (!cpumask_empty(cpus)) { |
5944 | env.loop = 0; | |
5945 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 5946 | goto redo; |
bbf18b19 | 5947 | } |
1e3c88bd PZ |
5948 | goto out_balanced; |
5949 | } | |
5950 | } | |
5951 | ||
5952 | if (!ld_moved) { | |
5953 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
5954 | /* |
5955 | * Increment the failure counter only on periodic balance. | |
5956 | * We do not want newidle balance, which can be very | |
5957 | * frequent, pollute the failure counter causing | |
5958 | * excessive cache_hot migrations and active balances. | |
5959 | */ | |
5960 | if (idle != CPU_NEWLY_IDLE) | |
5961 | sd->nr_balance_failed++; | |
1e3c88bd | 5962 | |
bd939f45 | 5963 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
5964 | raw_spin_lock_irqsave(&busiest->lock, flags); |
5965 | ||
969c7921 TH |
5966 | /* don't kick the active_load_balance_cpu_stop, |
5967 | * if the curr task on busiest cpu can't be | |
5968 | * moved to this_cpu | |
1e3c88bd PZ |
5969 | */ |
5970 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 5971 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
5972 | raw_spin_unlock_irqrestore(&busiest->lock, |
5973 | flags); | |
8e45cb54 | 5974 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
5975 | goto out_one_pinned; |
5976 | } | |
5977 | ||
969c7921 TH |
5978 | /* |
5979 | * ->active_balance synchronizes accesses to | |
5980 | * ->active_balance_work. Once set, it's cleared | |
5981 | * only after active load balance is finished. | |
5982 | */ | |
1e3c88bd PZ |
5983 | if (!busiest->active_balance) { |
5984 | busiest->active_balance = 1; | |
5985 | busiest->push_cpu = this_cpu; | |
5986 | active_balance = 1; | |
5987 | } | |
5988 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 5989 | |
bd939f45 | 5990 | if (active_balance) { |
969c7921 TH |
5991 | stop_one_cpu_nowait(cpu_of(busiest), |
5992 | active_load_balance_cpu_stop, busiest, | |
5993 | &busiest->active_balance_work); | |
bd939f45 | 5994 | } |
1e3c88bd PZ |
5995 | |
5996 | /* | |
5997 | * We've kicked active balancing, reset the failure | |
5998 | * counter. | |
5999 | */ | |
6000 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
6001 | } | |
6002 | } else | |
6003 | sd->nr_balance_failed = 0; | |
6004 | ||
6005 | if (likely(!active_balance)) { | |
6006 | /* We were unbalanced, so reset the balancing interval */ | |
6007 | sd->balance_interval = sd->min_interval; | |
6008 | } else { | |
6009 | /* | |
6010 | * If we've begun active balancing, start to back off. This | |
6011 | * case may not be covered by the all_pinned logic if there | |
6012 | * is only 1 task on the busy runqueue (because we don't call | |
6013 | * move_tasks). | |
6014 | */ | |
6015 | if (sd->balance_interval < sd->max_interval) | |
6016 | sd->balance_interval *= 2; | |
6017 | } | |
6018 | ||
1e3c88bd PZ |
6019 | goto out; |
6020 | ||
6021 | out_balanced: | |
6022 | schedstat_inc(sd, lb_balanced[idle]); | |
6023 | ||
6024 | sd->nr_balance_failed = 0; | |
6025 | ||
6026 | out_one_pinned: | |
6027 | /* tune up the balancing interval */ | |
8e45cb54 | 6028 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 6029 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
6030 | (sd->balance_interval < sd->max_interval)) |
6031 | sd->balance_interval *= 2; | |
6032 | ||
46e49b38 | 6033 | ld_moved = 0; |
1e3c88bd | 6034 | out: |
1e3c88bd PZ |
6035 | return ld_moved; |
6036 | } | |
6037 | ||
1e3c88bd PZ |
6038 | /* |
6039 | * idle_balance is called by schedule() if this_cpu is about to become | |
6040 | * idle. Attempts to pull tasks from other CPUs. | |
6041 | */ | |
029632fb | 6042 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
6043 | { |
6044 | struct sched_domain *sd; | |
6045 | int pulled_task = 0; | |
6046 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 6047 | u64 curr_cost = 0; |
1e3c88bd | 6048 | |
78becc27 | 6049 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
6050 | |
6051 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
6052 | return; | |
6053 | ||
f492e12e PZ |
6054 | /* |
6055 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
6056 | */ | |
6057 | raw_spin_unlock(&this_rq->lock); | |
6058 | ||
48a16753 | 6059 | update_blocked_averages(this_cpu); |
dce840a0 | 6060 | rcu_read_lock(); |
1e3c88bd PZ |
6061 | for_each_domain(this_cpu, sd) { |
6062 | unsigned long interval; | |
23f0d209 | 6063 | int continue_balancing = 1; |
9bd721c5 | 6064 | u64 t0, domain_cost; |
1e3c88bd PZ |
6065 | |
6066 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
6067 | continue; | |
6068 | ||
9bd721c5 JL |
6069 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
6070 | break; | |
6071 | ||
f492e12e | 6072 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
6073 | t0 = sched_clock_cpu(this_cpu); |
6074 | ||
1e3c88bd | 6075 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 6076 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
6077 | sd, CPU_NEWLY_IDLE, |
6078 | &continue_balancing); | |
9bd721c5 JL |
6079 | |
6080 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
6081 | if (domain_cost > sd->max_newidle_lb_cost) | |
6082 | sd->max_newidle_lb_cost = domain_cost; | |
6083 | ||
6084 | curr_cost += domain_cost; | |
f492e12e | 6085 | } |
1e3c88bd PZ |
6086 | |
6087 | interval = msecs_to_jiffies(sd->balance_interval); | |
6088 | if (time_after(next_balance, sd->last_balance + interval)) | |
6089 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
6090 | if (pulled_task) { |
6091 | this_rq->idle_stamp = 0; | |
1e3c88bd | 6092 | break; |
d5ad140b | 6093 | } |
1e3c88bd | 6094 | } |
dce840a0 | 6095 | rcu_read_unlock(); |
f492e12e PZ |
6096 | |
6097 | raw_spin_lock(&this_rq->lock); | |
6098 | ||
1e3c88bd PZ |
6099 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
6100 | /* | |
6101 | * We are going idle. next_balance may be set based on | |
6102 | * a busy processor. So reset next_balance. | |
6103 | */ | |
6104 | this_rq->next_balance = next_balance; | |
6105 | } | |
9bd721c5 JL |
6106 | |
6107 | if (curr_cost > this_rq->max_idle_balance_cost) | |
6108 | this_rq->max_idle_balance_cost = curr_cost; | |
1e3c88bd PZ |
6109 | } |
6110 | ||
6111 | /* | |
969c7921 TH |
6112 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
6113 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
6114 | * least 1 task to be running on each physical CPU where possible, and | |
6115 | * avoids physical / logical imbalances. | |
1e3c88bd | 6116 | */ |
969c7921 | 6117 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 6118 | { |
969c7921 TH |
6119 | struct rq *busiest_rq = data; |
6120 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 6121 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 6122 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 6123 | struct sched_domain *sd; |
969c7921 TH |
6124 | |
6125 | raw_spin_lock_irq(&busiest_rq->lock); | |
6126 | ||
6127 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
6128 | if (unlikely(busiest_cpu != smp_processor_id() || | |
6129 | !busiest_rq->active_balance)) | |
6130 | goto out_unlock; | |
1e3c88bd PZ |
6131 | |
6132 | /* Is there any task to move? */ | |
6133 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 6134 | goto out_unlock; |
1e3c88bd PZ |
6135 | |
6136 | /* | |
6137 | * This condition is "impossible", if it occurs | |
6138 | * we need to fix it. Originally reported by | |
6139 | * Bjorn Helgaas on a 128-cpu setup. | |
6140 | */ | |
6141 | BUG_ON(busiest_rq == target_rq); | |
6142 | ||
6143 | /* move a task from busiest_rq to target_rq */ | |
6144 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
6145 | |
6146 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 6147 | rcu_read_lock(); |
1e3c88bd PZ |
6148 | for_each_domain(target_cpu, sd) { |
6149 | if ((sd->flags & SD_LOAD_BALANCE) && | |
6150 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
6151 | break; | |
6152 | } | |
6153 | ||
6154 | if (likely(sd)) { | |
8e45cb54 PZ |
6155 | struct lb_env env = { |
6156 | .sd = sd, | |
ddcdf6e7 PZ |
6157 | .dst_cpu = target_cpu, |
6158 | .dst_rq = target_rq, | |
6159 | .src_cpu = busiest_rq->cpu, | |
6160 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
6161 | .idle = CPU_IDLE, |
6162 | }; | |
6163 | ||
1e3c88bd PZ |
6164 | schedstat_inc(sd, alb_count); |
6165 | ||
8e45cb54 | 6166 | if (move_one_task(&env)) |
1e3c88bd PZ |
6167 | schedstat_inc(sd, alb_pushed); |
6168 | else | |
6169 | schedstat_inc(sd, alb_failed); | |
6170 | } | |
dce840a0 | 6171 | rcu_read_unlock(); |
1e3c88bd | 6172 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
6173 | out_unlock: |
6174 | busiest_rq->active_balance = 0; | |
6175 | raw_spin_unlock_irq(&busiest_rq->lock); | |
6176 | return 0; | |
1e3c88bd PZ |
6177 | } |
6178 | ||
3451d024 | 6179 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
6180 | /* |
6181 | * idle load balancing details | |
83cd4fe2 VP |
6182 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
6183 | * needed, they will kick the idle load balancer, which then does idle | |
6184 | * load balancing for all the idle CPUs. | |
6185 | */ | |
1e3c88bd | 6186 | static struct { |
83cd4fe2 | 6187 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 6188 | atomic_t nr_cpus; |
83cd4fe2 VP |
6189 | unsigned long next_balance; /* in jiffy units */ |
6190 | } nohz ____cacheline_aligned; | |
1e3c88bd | 6191 | |
8e7fbcbc | 6192 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 6193 | { |
0b005cf5 | 6194 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 6195 | |
786d6dc7 SS |
6196 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
6197 | return ilb; | |
6198 | ||
6199 | return nr_cpu_ids; | |
1e3c88bd | 6200 | } |
1e3c88bd | 6201 | |
83cd4fe2 VP |
6202 | /* |
6203 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
6204 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
6205 | * CPU (if there is one). | |
6206 | */ | |
6207 | static void nohz_balancer_kick(int cpu) | |
6208 | { | |
6209 | int ilb_cpu; | |
6210 | ||
6211 | nohz.next_balance++; | |
6212 | ||
0b005cf5 | 6213 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 6214 | |
0b005cf5 SS |
6215 | if (ilb_cpu >= nr_cpu_ids) |
6216 | return; | |
83cd4fe2 | 6217 | |
cd490c5b | 6218 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
6219 | return; |
6220 | /* | |
6221 | * Use smp_send_reschedule() instead of resched_cpu(). | |
6222 | * This way we generate a sched IPI on the target cpu which | |
6223 | * is idle. And the softirq performing nohz idle load balance | |
6224 | * will be run before returning from the IPI. | |
6225 | */ | |
6226 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
6227 | return; |
6228 | } | |
6229 | ||
c1cc017c | 6230 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
6231 | { |
6232 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
6233 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
6234 | atomic_dec(&nohz.nr_cpus); | |
6235 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
6236 | } | |
6237 | } | |
6238 | ||
69e1e811 SS |
6239 | static inline void set_cpu_sd_state_busy(void) |
6240 | { | |
6241 | struct sched_domain *sd; | |
69e1e811 | 6242 | |
69e1e811 | 6243 | rcu_read_lock(); |
424c93fe | 6244 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
6245 | |
6246 | if (!sd || !sd->nohz_idle) | |
6247 | goto unlock; | |
6248 | sd->nohz_idle = 0; | |
6249 | ||
6250 | for (; sd; sd = sd->parent) | |
69e1e811 | 6251 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6252 | unlock: |
69e1e811 SS |
6253 | rcu_read_unlock(); |
6254 | } | |
6255 | ||
6256 | void set_cpu_sd_state_idle(void) | |
6257 | { | |
6258 | struct sched_domain *sd; | |
69e1e811 | 6259 | |
69e1e811 | 6260 | rcu_read_lock(); |
424c93fe | 6261 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
6262 | |
6263 | if (!sd || sd->nohz_idle) | |
6264 | goto unlock; | |
6265 | sd->nohz_idle = 1; | |
6266 | ||
6267 | for (; sd; sd = sd->parent) | |
69e1e811 | 6268 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6269 | unlock: |
69e1e811 SS |
6270 | rcu_read_unlock(); |
6271 | } | |
6272 | ||
1e3c88bd | 6273 | /* |
c1cc017c | 6274 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 6275 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 6276 | */ |
c1cc017c | 6277 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 6278 | { |
71325960 SS |
6279 | /* |
6280 | * If this cpu is going down, then nothing needs to be done. | |
6281 | */ | |
6282 | if (!cpu_active(cpu)) | |
6283 | return; | |
6284 | ||
c1cc017c AS |
6285 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
6286 | return; | |
1e3c88bd | 6287 | |
c1cc017c AS |
6288 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
6289 | atomic_inc(&nohz.nr_cpus); | |
6290 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 6291 | } |
71325960 | 6292 | |
0db0628d | 6293 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
6294 | unsigned long action, void *hcpu) |
6295 | { | |
6296 | switch (action & ~CPU_TASKS_FROZEN) { | |
6297 | case CPU_DYING: | |
c1cc017c | 6298 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
6299 | return NOTIFY_OK; |
6300 | default: | |
6301 | return NOTIFY_DONE; | |
6302 | } | |
6303 | } | |
1e3c88bd PZ |
6304 | #endif |
6305 | ||
6306 | static DEFINE_SPINLOCK(balancing); | |
6307 | ||
49c022e6 PZ |
6308 | /* |
6309 | * Scale the max load_balance interval with the number of CPUs in the system. | |
6310 | * This trades load-balance latency on larger machines for less cross talk. | |
6311 | */ | |
029632fb | 6312 | void update_max_interval(void) |
49c022e6 PZ |
6313 | { |
6314 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
6315 | } | |
6316 | ||
1e3c88bd PZ |
6317 | /* |
6318 | * It checks each scheduling domain to see if it is due to be balanced, | |
6319 | * and initiates a balancing operation if so. | |
6320 | * | |
b9b0853a | 6321 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd PZ |
6322 | */ |
6323 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
6324 | { | |
23f0d209 | 6325 | int continue_balancing = 1; |
1e3c88bd PZ |
6326 | struct rq *rq = cpu_rq(cpu); |
6327 | unsigned long interval; | |
04f733b4 | 6328 | struct sched_domain *sd; |
1e3c88bd PZ |
6329 | /* Earliest time when we have to do rebalance again */ |
6330 | unsigned long next_balance = jiffies + 60*HZ; | |
6331 | int update_next_balance = 0; | |
f48627e6 JL |
6332 | int need_serialize, need_decay = 0; |
6333 | u64 max_cost = 0; | |
1e3c88bd | 6334 | |
48a16753 | 6335 | update_blocked_averages(cpu); |
2069dd75 | 6336 | |
dce840a0 | 6337 | rcu_read_lock(); |
1e3c88bd | 6338 | for_each_domain(cpu, sd) { |
f48627e6 JL |
6339 | /* |
6340 | * Decay the newidle max times here because this is a regular | |
6341 | * visit to all the domains. Decay ~1% per second. | |
6342 | */ | |
6343 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
6344 | sd->max_newidle_lb_cost = | |
6345 | (sd->max_newidle_lb_cost * 253) / 256; | |
6346 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
6347 | need_decay = 1; | |
6348 | } | |
6349 | max_cost += sd->max_newidle_lb_cost; | |
6350 | ||
1e3c88bd PZ |
6351 | if (!(sd->flags & SD_LOAD_BALANCE)) |
6352 | continue; | |
6353 | ||
f48627e6 JL |
6354 | /* |
6355 | * Stop the load balance at this level. There is another | |
6356 | * CPU in our sched group which is doing load balancing more | |
6357 | * actively. | |
6358 | */ | |
6359 | if (!continue_balancing) { | |
6360 | if (need_decay) | |
6361 | continue; | |
6362 | break; | |
6363 | } | |
6364 | ||
1e3c88bd PZ |
6365 | interval = sd->balance_interval; |
6366 | if (idle != CPU_IDLE) | |
6367 | interval *= sd->busy_factor; | |
6368 | ||
6369 | /* scale ms to jiffies */ | |
6370 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 6371 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
6372 | |
6373 | need_serialize = sd->flags & SD_SERIALIZE; | |
6374 | ||
6375 | if (need_serialize) { | |
6376 | if (!spin_trylock(&balancing)) | |
6377 | goto out; | |
6378 | } | |
6379 | ||
6380 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 6381 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 6382 | /* |
6263322c | 6383 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
6384 | * env->dst_cpu, so we can't know our idle |
6385 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 6386 | */ |
de5eb2dd | 6387 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
6388 | } |
6389 | sd->last_balance = jiffies; | |
6390 | } | |
6391 | if (need_serialize) | |
6392 | spin_unlock(&balancing); | |
6393 | out: | |
6394 | if (time_after(next_balance, sd->last_balance + interval)) { | |
6395 | next_balance = sd->last_balance + interval; | |
6396 | update_next_balance = 1; | |
6397 | } | |
f48627e6 JL |
6398 | } |
6399 | if (need_decay) { | |
1e3c88bd | 6400 | /* |
f48627e6 JL |
6401 | * Ensure the rq-wide value also decays but keep it at a |
6402 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 6403 | */ |
f48627e6 JL |
6404 | rq->max_idle_balance_cost = |
6405 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 6406 | } |
dce840a0 | 6407 | rcu_read_unlock(); |
1e3c88bd PZ |
6408 | |
6409 | /* | |
6410 | * next_balance will be updated only when there is a need. | |
6411 | * When the cpu is attached to null domain for ex, it will not be | |
6412 | * updated. | |
6413 | */ | |
6414 | if (likely(update_next_balance)) | |
6415 | rq->next_balance = next_balance; | |
6416 | } | |
6417 | ||
3451d024 | 6418 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 6419 | /* |
3451d024 | 6420 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
6421 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
6422 | */ | |
83cd4fe2 VP |
6423 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
6424 | { | |
6425 | struct rq *this_rq = cpu_rq(this_cpu); | |
6426 | struct rq *rq; | |
6427 | int balance_cpu; | |
6428 | ||
1c792db7 SS |
6429 | if (idle != CPU_IDLE || |
6430 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
6431 | goto end; | |
83cd4fe2 VP |
6432 | |
6433 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 6434 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
6435 | continue; |
6436 | ||
6437 | /* | |
6438 | * If this cpu gets work to do, stop the load balancing | |
6439 | * work being done for other cpus. Next load | |
6440 | * balancing owner will pick it up. | |
6441 | */ | |
1c792db7 | 6442 | if (need_resched()) |
83cd4fe2 | 6443 | break; |
83cd4fe2 | 6444 | |
5ed4f1d9 VG |
6445 | rq = cpu_rq(balance_cpu); |
6446 | ||
6447 | raw_spin_lock_irq(&rq->lock); | |
6448 | update_rq_clock(rq); | |
6449 | update_idle_cpu_load(rq); | |
6450 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
6451 | |
6452 | rebalance_domains(balance_cpu, CPU_IDLE); | |
6453 | ||
83cd4fe2 VP |
6454 | if (time_after(this_rq->next_balance, rq->next_balance)) |
6455 | this_rq->next_balance = rq->next_balance; | |
6456 | } | |
6457 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
6458 | end: |
6459 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
6460 | } |
6461 | ||
6462 | /* | |
0b005cf5 SS |
6463 | * Current heuristic for kicking the idle load balancer in the presence |
6464 | * of an idle cpu is the system. | |
6465 | * - This rq has more than one task. | |
6466 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
6467 | * busy cpu's exceeding the group's power. | |
6468 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
6469 | * domain span are idle. | |
83cd4fe2 VP |
6470 | */ |
6471 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
6472 | { | |
6473 | unsigned long now = jiffies; | |
0b005cf5 | 6474 | struct sched_domain *sd; |
83cd4fe2 | 6475 | |
1c792db7 | 6476 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
6477 | return 0; |
6478 | ||
1c792db7 SS |
6479 | /* |
6480 | * We may be recently in ticked or tickless idle mode. At the first | |
6481 | * busy tick after returning from idle, we will update the busy stats. | |
6482 | */ | |
69e1e811 | 6483 | set_cpu_sd_state_busy(); |
c1cc017c | 6484 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
6485 | |
6486 | /* | |
6487 | * None are in tickless mode and hence no need for NOHZ idle load | |
6488 | * balancing. | |
6489 | */ | |
6490 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
6491 | return 0; | |
1c792db7 SS |
6492 | |
6493 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
6494 | return 0; |
6495 | ||
0b005cf5 SS |
6496 | if (rq->nr_running >= 2) |
6497 | goto need_kick; | |
83cd4fe2 | 6498 | |
067491b7 | 6499 | rcu_read_lock(); |
0b005cf5 SS |
6500 | for_each_domain(cpu, sd) { |
6501 | struct sched_group *sg = sd->groups; | |
6502 | struct sched_group_power *sgp = sg->sgp; | |
6503 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 6504 | |
0b005cf5 | 6505 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 6506 | goto need_kick_unlock; |
0b005cf5 SS |
6507 | |
6508 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
6509 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
6510 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 6511 | goto need_kick_unlock; |
0b005cf5 SS |
6512 | |
6513 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
6514 | break; | |
83cd4fe2 | 6515 | } |
067491b7 | 6516 | rcu_read_unlock(); |
83cd4fe2 | 6517 | return 0; |
067491b7 PZ |
6518 | |
6519 | need_kick_unlock: | |
6520 | rcu_read_unlock(); | |
0b005cf5 SS |
6521 | need_kick: |
6522 | return 1; | |
83cd4fe2 VP |
6523 | } |
6524 | #else | |
6525 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
6526 | #endif | |
6527 | ||
6528 | /* | |
6529 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
6530 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
6531 | */ | |
1e3c88bd PZ |
6532 | static void run_rebalance_domains(struct softirq_action *h) |
6533 | { | |
6534 | int this_cpu = smp_processor_id(); | |
6535 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 6536 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
6537 | CPU_IDLE : CPU_NOT_IDLE; |
6538 | ||
6539 | rebalance_domains(this_cpu, idle); | |
6540 | ||
1e3c88bd | 6541 | /* |
83cd4fe2 | 6542 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
6543 | * balancing on behalf of the other idle cpus whose ticks are |
6544 | * stopped. | |
6545 | */ | |
83cd4fe2 | 6546 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
6547 | } |
6548 | ||
6549 | static inline int on_null_domain(int cpu) | |
6550 | { | |
90a6501f | 6551 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
6552 | } |
6553 | ||
6554 | /* | |
6555 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 6556 | */ |
029632fb | 6557 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 6558 | { |
1e3c88bd PZ |
6559 | /* Don't need to rebalance while attached to NULL domain */ |
6560 | if (time_after_eq(jiffies, rq->next_balance) && | |
6561 | likely(!on_null_domain(cpu))) | |
6562 | raise_softirq(SCHED_SOFTIRQ); | |
3451d024 | 6563 | #ifdef CONFIG_NO_HZ_COMMON |
1c792db7 | 6564 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
6565 | nohz_balancer_kick(cpu); |
6566 | #endif | |
1e3c88bd PZ |
6567 | } |
6568 | ||
0bcdcf28 CE |
6569 | static void rq_online_fair(struct rq *rq) |
6570 | { | |
6571 | update_sysctl(); | |
6572 | } | |
6573 | ||
6574 | static void rq_offline_fair(struct rq *rq) | |
6575 | { | |
6576 | update_sysctl(); | |
a4c96ae3 PB |
6577 | |
6578 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
6579 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
6580 | } |
6581 | ||
55e12e5e | 6582 | #endif /* CONFIG_SMP */ |
e1d1484f | 6583 | |
bf0f6f24 IM |
6584 | /* |
6585 | * scheduler tick hitting a task of our scheduling class: | |
6586 | */ | |
8f4d37ec | 6587 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
6588 | { |
6589 | struct cfs_rq *cfs_rq; | |
6590 | struct sched_entity *se = &curr->se; | |
6591 | ||
6592 | for_each_sched_entity(se) { | |
6593 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 6594 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 6595 | } |
18bf2805 | 6596 | |
10e84b97 | 6597 | if (numabalancing_enabled) |
cbee9f88 | 6598 | task_tick_numa(rq, curr); |
3d59eebc | 6599 | |
18bf2805 | 6600 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
6601 | } |
6602 | ||
6603 | /* | |
cd29fe6f PZ |
6604 | * called on fork with the child task as argument from the parent's context |
6605 | * - child not yet on the tasklist | |
6606 | * - preemption disabled | |
bf0f6f24 | 6607 | */ |
cd29fe6f | 6608 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 6609 | { |
4fc420c9 DN |
6610 | struct cfs_rq *cfs_rq; |
6611 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 6612 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
6613 | struct rq *rq = this_rq(); |
6614 | unsigned long flags; | |
6615 | ||
05fa785c | 6616 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 6617 | |
861d034e PZ |
6618 | update_rq_clock(rq); |
6619 | ||
4fc420c9 DN |
6620 | cfs_rq = task_cfs_rq(current); |
6621 | curr = cfs_rq->curr; | |
6622 | ||
6c9a27f5 DN |
6623 | /* |
6624 | * Not only the cpu but also the task_group of the parent might have | |
6625 | * been changed after parent->se.parent,cfs_rq were copied to | |
6626 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
6627 | * of child point to valid ones. | |
6628 | */ | |
6629 | rcu_read_lock(); | |
6630 | __set_task_cpu(p, this_cpu); | |
6631 | rcu_read_unlock(); | |
bf0f6f24 | 6632 | |
7109c442 | 6633 | update_curr(cfs_rq); |
cd29fe6f | 6634 | |
b5d9d734 MG |
6635 | if (curr) |
6636 | se->vruntime = curr->vruntime; | |
aeb73b04 | 6637 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 6638 | |
cd29fe6f | 6639 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 6640 | /* |
edcb60a3 IM |
6641 | * Upon rescheduling, sched_class::put_prev_task() will place |
6642 | * 'current' within the tree based on its new key value. | |
6643 | */ | |
4d78e7b6 | 6644 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 6645 | resched_task(rq->curr); |
4d78e7b6 | 6646 | } |
bf0f6f24 | 6647 | |
88ec22d3 PZ |
6648 | se->vruntime -= cfs_rq->min_vruntime; |
6649 | ||
05fa785c | 6650 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
6651 | } |
6652 | ||
cb469845 SR |
6653 | /* |
6654 | * Priority of the task has changed. Check to see if we preempt | |
6655 | * the current task. | |
6656 | */ | |
da7a735e PZ |
6657 | static void |
6658 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 6659 | { |
da7a735e PZ |
6660 | if (!p->se.on_rq) |
6661 | return; | |
6662 | ||
cb469845 SR |
6663 | /* |
6664 | * Reschedule if we are currently running on this runqueue and | |
6665 | * our priority decreased, or if we are not currently running on | |
6666 | * this runqueue and our priority is higher than the current's | |
6667 | */ | |
da7a735e | 6668 | if (rq->curr == p) { |
cb469845 SR |
6669 | if (p->prio > oldprio) |
6670 | resched_task(rq->curr); | |
6671 | } else | |
15afe09b | 6672 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6673 | } |
6674 | ||
da7a735e PZ |
6675 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
6676 | { | |
6677 | struct sched_entity *se = &p->se; | |
6678 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6679 | ||
6680 | /* | |
6681 | * Ensure the task's vruntime is normalized, so that when its | |
6682 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
6683 | * do the right thing. | |
6684 | * | |
6685 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
6686 | * have normalized the vruntime, if it was !on_rq, then only when | |
6687 | * the task is sleeping will it still have non-normalized vruntime. | |
6688 | */ | |
6689 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
6690 | /* | |
6691 | * Fix up our vruntime so that the current sleep doesn't | |
6692 | * cause 'unlimited' sleep bonus. | |
6693 | */ | |
6694 | place_entity(cfs_rq, se, 0); | |
6695 | se->vruntime -= cfs_rq->min_vruntime; | |
6696 | } | |
9ee474f5 | 6697 | |
141965c7 | 6698 | #ifdef CONFIG_SMP |
9ee474f5 PT |
6699 | /* |
6700 | * Remove our load from contribution when we leave sched_fair | |
6701 | * and ensure we don't carry in an old decay_count if we | |
6702 | * switch back. | |
6703 | */ | |
87e3c8ae KT |
6704 | if (se->avg.decay_count) { |
6705 | __synchronize_entity_decay(se); | |
6706 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
6707 | } |
6708 | #endif | |
da7a735e PZ |
6709 | } |
6710 | ||
cb469845 SR |
6711 | /* |
6712 | * We switched to the sched_fair class. | |
6713 | */ | |
da7a735e | 6714 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 6715 | { |
da7a735e PZ |
6716 | if (!p->se.on_rq) |
6717 | return; | |
6718 | ||
cb469845 SR |
6719 | /* |
6720 | * We were most likely switched from sched_rt, so | |
6721 | * kick off the schedule if running, otherwise just see | |
6722 | * if we can still preempt the current task. | |
6723 | */ | |
da7a735e | 6724 | if (rq->curr == p) |
cb469845 SR |
6725 | resched_task(rq->curr); |
6726 | else | |
15afe09b | 6727 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6728 | } |
6729 | ||
83b699ed SV |
6730 | /* Account for a task changing its policy or group. |
6731 | * | |
6732 | * This routine is mostly called to set cfs_rq->curr field when a task | |
6733 | * migrates between groups/classes. | |
6734 | */ | |
6735 | static void set_curr_task_fair(struct rq *rq) | |
6736 | { | |
6737 | struct sched_entity *se = &rq->curr->se; | |
6738 | ||
ec12cb7f PT |
6739 | for_each_sched_entity(se) { |
6740 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6741 | ||
6742 | set_next_entity(cfs_rq, se); | |
6743 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
6744 | account_cfs_rq_runtime(cfs_rq, 0); | |
6745 | } | |
83b699ed SV |
6746 | } |
6747 | ||
029632fb PZ |
6748 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
6749 | { | |
6750 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
6751 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
6752 | #ifndef CONFIG_64BIT | |
6753 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
6754 | #endif | |
141965c7 | 6755 | #ifdef CONFIG_SMP |
9ee474f5 | 6756 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 6757 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 6758 | #endif |
029632fb PZ |
6759 | } |
6760 | ||
810b3817 | 6761 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6762 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 6763 | { |
aff3e498 | 6764 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
6765 | /* |
6766 | * If the task was not on the rq at the time of this cgroup movement | |
6767 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
6768 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
6769 | * bonus in place_entity()). | |
6770 | * | |
6771 | * If it was on the rq, we've just 'preempted' it, which does convert | |
6772 | * ->vruntime to a relative base. | |
6773 | * | |
6774 | * Make sure both cases convert their relative position when migrating | |
6775 | * to another cgroup's rq. This does somewhat interfere with the | |
6776 | * fair sleeper stuff for the first placement, but who cares. | |
6777 | */ | |
7ceff013 DN |
6778 | /* |
6779 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
6780 | * But there are some cases where it has already been normalized: | |
6781 | * | |
6782 | * - Moving a forked child which is waiting for being woken up by | |
6783 | * wake_up_new_task(). | |
62af3783 DN |
6784 | * - Moving a task which has been woken up by try_to_wake_up() and |
6785 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
6786 | * |
6787 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
6788 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
6789 | */ | |
62af3783 | 6790 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
6791 | on_rq = 1; |
6792 | ||
b2b5ce02 PZ |
6793 | if (!on_rq) |
6794 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
6795 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
6796 | if (!on_rq) { |
6797 | cfs_rq = cfs_rq_of(&p->se); | |
6798 | p->se.vruntime += cfs_rq->min_vruntime; | |
6799 | #ifdef CONFIG_SMP | |
6800 | /* | |
6801 | * migrate_task_rq_fair() will have removed our previous | |
6802 | * contribution, but we must synchronize for ongoing future | |
6803 | * decay. | |
6804 | */ | |
6805 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
6806 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
6807 | #endif | |
6808 | } | |
810b3817 | 6809 | } |
029632fb PZ |
6810 | |
6811 | void free_fair_sched_group(struct task_group *tg) | |
6812 | { | |
6813 | int i; | |
6814 | ||
6815 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6816 | ||
6817 | for_each_possible_cpu(i) { | |
6818 | if (tg->cfs_rq) | |
6819 | kfree(tg->cfs_rq[i]); | |
6820 | if (tg->se) | |
6821 | kfree(tg->se[i]); | |
6822 | } | |
6823 | ||
6824 | kfree(tg->cfs_rq); | |
6825 | kfree(tg->se); | |
6826 | } | |
6827 | ||
6828 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6829 | { | |
6830 | struct cfs_rq *cfs_rq; | |
6831 | struct sched_entity *se; | |
6832 | int i; | |
6833 | ||
6834 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
6835 | if (!tg->cfs_rq) | |
6836 | goto err; | |
6837 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
6838 | if (!tg->se) | |
6839 | goto err; | |
6840 | ||
6841 | tg->shares = NICE_0_LOAD; | |
6842 | ||
6843 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6844 | ||
6845 | for_each_possible_cpu(i) { | |
6846 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
6847 | GFP_KERNEL, cpu_to_node(i)); | |
6848 | if (!cfs_rq) | |
6849 | goto err; | |
6850 | ||
6851 | se = kzalloc_node(sizeof(struct sched_entity), | |
6852 | GFP_KERNEL, cpu_to_node(i)); | |
6853 | if (!se) | |
6854 | goto err_free_rq; | |
6855 | ||
6856 | init_cfs_rq(cfs_rq); | |
6857 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
6858 | } | |
6859 | ||
6860 | return 1; | |
6861 | ||
6862 | err_free_rq: | |
6863 | kfree(cfs_rq); | |
6864 | err: | |
6865 | return 0; | |
6866 | } | |
6867 | ||
6868 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
6869 | { | |
6870 | struct rq *rq = cpu_rq(cpu); | |
6871 | unsigned long flags; | |
6872 | ||
6873 | /* | |
6874 | * Only empty task groups can be destroyed; so we can speculatively | |
6875 | * check on_list without danger of it being re-added. | |
6876 | */ | |
6877 | if (!tg->cfs_rq[cpu]->on_list) | |
6878 | return; | |
6879 | ||
6880 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6881 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
6882 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6883 | } | |
6884 | ||
6885 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
6886 | struct sched_entity *se, int cpu, | |
6887 | struct sched_entity *parent) | |
6888 | { | |
6889 | struct rq *rq = cpu_rq(cpu); | |
6890 | ||
6891 | cfs_rq->tg = tg; | |
6892 | cfs_rq->rq = rq; | |
029632fb PZ |
6893 | init_cfs_rq_runtime(cfs_rq); |
6894 | ||
6895 | tg->cfs_rq[cpu] = cfs_rq; | |
6896 | tg->se[cpu] = se; | |
6897 | ||
6898 | /* se could be NULL for root_task_group */ | |
6899 | if (!se) | |
6900 | return; | |
6901 | ||
6902 | if (!parent) | |
6903 | se->cfs_rq = &rq->cfs; | |
6904 | else | |
6905 | se->cfs_rq = parent->my_q; | |
6906 | ||
6907 | se->my_q = cfs_rq; | |
6908 | update_load_set(&se->load, 0); | |
6909 | se->parent = parent; | |
6910 | } | |
6911 | ||
6912 | static DEFINE_MUTEX(shares_mutex); | |
6913 | ||
6914 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
6915 | { | |
6916 | int i; | |
6917 | unsigned long flags; | |
6918 | ||
6919 | /* | |
6920 | * We can't change the weight of the root cgroup. | |
6921 | */ | |
6922 | if (!tg->se[0]) | |
6923 | return -EINVAL; | |
6924 | ||
6925 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
6926 | ||
6927 | mutex_lock(&shares_mutex); | |
6928 | if (tg->shares == shares) | |
6929 | goto done; | |
6930 | ||
6931 | tg->shares = shares; | |
6932 | for_each_possible_cpu(i) { | |
6933 | struct rq *rq = cpu_rq(i); | |
6934 | struct sched_entity *se; | |
6935 | ||
6936 | se = tg->se[i]; | |
6937 | /* Propagate contribution to hierarchy */ | |
6938 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
6939 | |
6940 | /* Possible calls to update_curr() need rq clock */ | |
6941 | update_rq_clock(rq); | |
17bc14b7 | 6942 | for_each_sched_entity(se) |
029632fb PZ |
6943 | update_cfs_shares(group_cfs_rq(se)); |
6944 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6945 | } | |
6946 | ||
6947 | done: | |
6948 | mutex_unlock(&shares_mutex); | |
6949 | return 0; | |
6950 | } | |
6951 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6952 | ||
6953 | void free_fair_sched_group(struct task_group *tg) { } | |
6954 | ||
6955 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6956 | { | |
6957 | return 1; | |
6958 | } | |
6959 | ||
6960 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
6961 | ||
6962 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
6963 | ||
810b3817 | 6964 | |
6d686f45 | 6965 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
6966 | { |
6967 | struct sched_entity *se = &task->se; | |
0d721cea PW |
6968 | unsigned int rr_interval = 0; |
6969 | ||
6970 | /* | |
6971 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
6972 | * idle runqueue: | |
6973 | */ | |
0d721cea | 6974 | if (rq->cfs.load.weight) |
a59f4e07 | 6975 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
6976 | |
6977 | return rr_interval; | |
6978 | } | |
6979 | ||
bf0f6f24 IM |
6980 | /* |
6981 | * All the scheduling class methods: | |
6982 | */ | |
029632fb | 6983 | const struct sched_class fair_sched_class = { |
5522d5d5 | 6984 | .next = &idle_sched_class, |
bf0f6f24 IM |
6985 | .enqueue_task = enqueue_task_fair, |
6986 | .dequeue_task = dequeue_task_fair, | |
6987 | .yield_task = yield_task_fair, | |
d95f4122 | 6988 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 6989 | |
2e09bf55 | 6990 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
6991 | |
6992 | .pick_next_task = pick_next_task_fair, | |
6993 | .put_prev_task = put_prev_task_fair, | |
6994 | ||
681f3e68 | 6995 | #ifdef CONFIG_SMP |
4ce72a2c | 6996 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 6997 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 6998 | |
0bcdcf28 CE |
6999 | .rq_online = rq_online_fair, |
7000 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
7001 | |
7002 | .task_waking = task_waking_fair, | |
681f3e68 | 7003 | #endif |
bf0f6f24 | 7004 | |
83b699ed | 7005 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 7006 | .task_tick = task_tick_fair, |
cd29fe6f | 7007 | .task_fork = task_fork_fair, |
cb469845 SR |
7008 | |
7009 | .prio_changed = prio_changed_fair, | |
da7a735e | 7010 | .switched_from = switched_from_fair, |
cb469845 | 7011 | .switched_to = switched_to_fair, |
810b3817 | 7012 | |
0d721cea PW |
7013 | .get_rr_interval = get_rr_interval_fair, |
7014 | ||
810b3817 | 7015 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7016 | .task_move_group = task_move_group_fair, |
810b3817 | 7017 | #endif |
bf0f6f24 IM |
7018 | }; |
7019 | ||
7020 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 7021 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 7022 | { |
bf0f6f24 IM |
7023 | struct cfs_rq *cfs_rq; |
7024 | ||
5973e5b9 | 7025 | rcu_read_lock(); |
c3b64f1e | 7026 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 7027 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 7028 | rcu_read_unlock(); |
bf0f6f24 IM |
7029 | } |
7030 | #endif | |
029632fb PZ |
7031 | |
7032 | __init void init_sched_fair_class(void) | |
7033 | { | |
7034 | #ifdef CONFIG_SMP | |
7035 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
7036 | ||
3451d024 | 7037 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 7038 | nohz.next_balance = jiffies; |
029632fb | 7039 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 7040 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
7041 | #endif |
7042 | #endif /* SMP */ | |
7043 | ||
7044 | } |