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; | |
e29cf08b | 896 | pid_t gid; |
8c8a743c PZ |
897 | struct list_head task_list; |
898 | ||
899 | struct rcu_head rcu; | |
900 | atomic_long_t faults[0]; | |
901 | }; | |
902 | ||
e29cf08b MG |
903 | pid_t task_numa_group_id(struct task_struct *p) |
904 | { | |
905 | return p->numa_group ? p->numa_group->gid : 0; | |
906 | } | |
907 | ||
ac8e895b MG |
908 | static inline int task_faults_idx(int nid, int priv) |
909 | { | |
910 | return 2 * nid + priv; | |
911 | } | |
912 | ||
913 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
914 | { | |
915 | if (!p->numa_faults) | |
916 | return 0; | |
917 | ||
918 | return p->numa_faults[task_faults_idx(nid, 0)] + | |
919 | p->numa_faults[task_faults_idx(nid, 1)]; | |
920 | } | |
921 | ||
e6628d5b | 922 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
923 | static unsigned long source_load(int cpu, int type); |
924 | static unsigned long target_load(int cpu, int type); | |
925 | static unsigned long power_of(int cpu); | |
926 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
927 | ||
fb13c7ee | 928 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 929 | struct numa_stats { |
fb13c7ee | 930 | unsigned long nr_running; |
58d081b5 | 931 | unsigned long load; |
fb13c7ee MG |
932 | |
933 | /* Total compute capacity of CPUs on a node */ | |
934 | unsigned long power; | |
935 | ||
936 | /* Approximate capacity in terms of runnable tasks on a node */ | |
937 | unsigned long capacity; | |
938 | int has_capacity; | |
58d081b5 | 939 | }; |
e6628d5b | 940 | |
fb13c7ee MG |
941 | /* |
942 | * XXX borrowed from update_sg_lb_stats | |
943 | */ | |
944 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
945 | { | |
946 | int cpu; | |
947 | ||
948 | memset(ns, 0, sizeof(*ns)); | |
949 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
950 | struct rq *rq = cpu_rq(cpu); | |
951 | ||
952 | ns->nr_running += rq->nr_running; | |
953 | ns->load += weighted_cpuload(cpu); | |
954 | ns->power += power_of(cpu); | |
955 | } | |
956 | ||
957 | ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power; | |
958 | ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE); | |
959 | ns->has_capacity = (ns->nr_running < ns->capacity); | |
960 | } | |
961 | ||
58d081b5 MG |
962 | struct task_numa_env { |
963 | struct task_struct *p; | |
e6628d5b | 964 | |
58d081b5 MG |
965 | int src_cpu, src_nid; |
966 | int dst_cpu, dst_nid; | |
e6628d5b | 967 | |
58d081b5 | 968 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 969 | |
fb13c7ee MG |
970 | int imbalance_pct, idx; |
971 | ||
972 | struct task_struct *best_task; | |
973 | long best_imp; | |
58d081b5 MG |
974 | int best_cpu; |
975 | }; | |
976 | ||
fb13c7ee MG |
977 | static void task_numa_assign(struct task_numa_env *env, |
978 | struct task_struct *p, long imp) | |
979 | { | |
980 | if (env->best_task) | |
981 | put_task_struct(env->best_task); | |
982 | if (p) | |
983 | get_task_struct(p); | |
984 | ||
985 | env->best_task = p; | |
986 | env->best_imp = imp; | |
987 | env->best_cpu = env->dst_cpu; | |
988 | } | |
989 | ||
990 | /* | |
991 | * This checks if the overall compute and NUMA accesses of the system would | |
992 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
993 | * into account that it might be best if task running on the dst_cpu should | |
994 | * be exchanged with the source task | |
995 | */ | |
996 | static void task_numa_compare(struct task_numa_env *env, long imp) | |
997 | { | |
998 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
999 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1000 | struct task_struct *cur; | |
1001 | long dst_load, src_load; | |
1002 | long load; | |
1003 | ||
1004 | rcu_read_lock(); | |
1005 | cur = ACCESS_ONCE(dst_rq->curr); | |
1006 | if (cur->pid == 0) /* idle */ | |
1007 | cur = NULL; | |
1008 | ||
1009 | /* | |
1010 | * "imp" is the fault differential for the source task between the | |
1011 | * source and destination node. Calculate the total differential for | |
1012 | * the source task and potential destination task. The more negative | |
1013 | * the value is, the more rmeote accesses that would be expected to | |
1014 | * be incurred if the tasks were swapped. | |
1015 | */ | |
1016 | if (cur) { | |
1017 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1018 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1019 | goto unlock; | |
1020 | ||
1021 | imp += task_faults(cur, env->src_nid) - | |
1022 | task_faults(cur, env->dst_nid); | |
1023 | } | |
1024 | ||
1025 | if (imp < env->best_imp) | |
1026 | goto unlock; | |
1027 | ||
1028 | if (!cur) { | |
1029 | /* Is there capacity at our destination? */ | |
1030 | if (env->src_stats.has_capacity && | |
1031 | !env->dst_stats.has_capacity) | |
1032 | goto unlock; | |
1033 | ||
1034 | goto balance; | |
1035 | } | |
1036 | ||
1037 | /* Balance doesn't matter much if we're running a task per cpu */ | |
1038 | if (src_rq->nr_running == 1 && dst_rq->nr_running == 1) | |
1039 | goto assign; | |
1040 | ||
1041 | /* | |
1042 | * In the overloaded case, try and keep the load balanced. | |
1043 | */ | |
1044 | balance: | |
1045 | dst_load = env->dst_stats.load; | |
1046 | src_load = env->src_stats.load; | |
1047 | ||
1048 | /* XXX missing power terms */ | |
1049 | load = task_h_load(env->p); | |
1050 | dst_load += load; | |
1051 | src_load -= load; | |
1052 | ||
1053 | if (cur) { | |
1054 | load = task_h_load(cur); | |
1055 | dst_load -= load; | |
1056 | src_load += load; | |
1057 | } | |
1058 | ||
1059 | /* make src_load the smaller */ | |
1060 | if (dst_load < src_load) | |
1061 | swap(dst_load, src_load); | |
1062 | ||
1063 | if (src_load * env->imbalance_pct < dst_load * 100) | |
1064 | goto unlock; | |
1065 | ||
1066 | assign: | |
1067 | task_numa_assign(env, cur, imp); | |
1068 | unlock: | |
1069 | rcu_read_unlock(); | |
1070 | } | |
1071 | ||
2c8a50aa MG |
1072 | static void task_numa_find_cpu(struct task_numa_env *env, long imp) |
1073 | { | |
1074 | int cpu; | |
1075 | ||
1076 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1077 | /* Skip this CPU if the source task cannot migrate */ | |
1078 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1079 | continue; | |
1080 | ||
1081 | env->dst_cpu = cpu; | |
1082 | task_numa_compare(env, imp); | |
1083 | } | |
1084 | } | |
1085 | ||
58d081b5 MG |
1086 | static int task_numa_migrate(struct task_struct *p) |
1087 | { | |
58d081b5 MG |
1088 | struct task_numa_env env = { |
1089 | .p = p, | |
fb13c7ee | 1090 | |
58d081b5 MG |
1091 | .src_cpu = task_cpu(p), |
1092 | .src_nid = cpu_to_node(task_cpu(p)), | |
fb13c7ee MG |
1093 | |
1094 | .imbalance_pct = 112, | |
1095 | ||
1096 | .best_task = NULL, | |
1097 | .best_imp = 0, | |
1098 | .best_cpu = -1 | |
58d081b5 MG |
1099 | }; |
1100 | struct sched_domain *sd; | |
fb13c7ee | 1101 | unsigned long faults; |
2c8a50aa MG |
1102 | int nid, ret; |
1103 | long imp; | |
e6628d5b | 1104 | |
58d081b5 | 1105 | /* |
fb13c7ee MG |
1106 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1107 | * imbalance and would be the first to start moving tasks about. | |
1108 | * | |
1109 | * And we want to avoid any moving of tasks about, as that would create | |
1110 | * random movement of tasks -- counter the numa conditions we're trying | |
1111 | * to satisfy here. | |
58d081b5 MG |
1112 | */ |
1113 | rcu_read_lock(); | |
fb13c7ee MG |
1114 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
1115 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1116 | rcu_read_unlock(); |
1117 | ||
fb13c7ee MG |
1118 | faults = task_faults(p, env.src_nid); |
1119 | update_numa_stats(&env.src_stats, env.src_nid); | |
2c8a50aa MG |
1120 | env.dst_nid = p->numa_preferred_nid; |
1121 | imp = task_faults(env.p, env.dst_nid) - faults; | |
1122 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
58d081b5 | 1123 | |
e1dda8a7 RR |
1124 | /* If the preferred nid has capacity, try to use it. */ |
1125 | if (env.dst_stats.has_capacity) | |
2c8a50aa | 1126 | task_numa_find_cpu(&env, imp); |
e1dda8a7 RR |
1127 | |
1128 | /* No space available on the preferred nid. Look elsewhere. */ | |
1129 | if (env.best_cpu == -1) { | |
2c8a50aa MG |
1130 | for_each_online_node(nid) { |
1131 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1132 | continue; | |
58d081b5 | 1133 | |
2c8a50aa MG |
1134 | /* Only consider nodes that recorded more faults */ |
1135 | imp = task_faults(env.p, nid) - faults; | |
1136 | if (imp < 0) | |
fb13c7ee MG |
1137 | continue; |
1138 | ||
2c8a50aa MG |
1139 | env.dst_nid = nid; |
1140 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
1141 | task_numa_find_cpu(&env, imp); | |
58d081b5 MG |
1142 | } |
1143 | } | |
1144 | ||
fb13c7ee MG |
1145 | /* No better CPU than the current one was found. */ |
1146 | if (env.best_cpu == -1) | |
1147 | return -EAGAIN; | |
1148 | ||
1149 | if (env.best_task == NULL) { | |
1150 | int ret = migrate_task_to(p, env.best_cpu); | |
1151 | return ret; | |
1152 | } | |
1153 | ||
1154 | ret = migrate_swap(p, env.best_task); | |
1155 | put_task_struct(env.best_task); | |
1156 | return ret; | |
e6628d5b MG |
1157 | } |
1158 | ||
6b9a7460 MG |
1159 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1160 | static void numa_migrate_preferred(struct task_struct *p) | |
1161 | { | |
1162 | /* Success if task is already running on preferred CPU */ | |
1163 | p->numa_migrate_retry = 0; | |
06ea5e03 RR |
1164 | if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) { |
1165 | /* | |
1166 | * If migration is temporarily disabled due to a task migration | |
1167 | * then re-enable it now as the task is running on its | |
1168 | * preferred node and memory should migrate locally | |
1169 | */ | |
1170 | if (!p->numa_migrate_seq) | |
1171 | p->numa_migrate_seq++; | |
6b9a7460 | 1172 | return; |
06ea5e03 | 1173 | } |
6b9a7460 MG |
1174 | |
1175 | /* This task has no NUMA fault statistics yet */ | |
1176 | if (unlikely(p->numa_preferred_nid == -1)) | |
1177 | return; | |
1178 | ||
1179 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
1180 | if (task_numa_migrate(p) != 0) | |
1181 | p->numa_migrate_retry = jiffies + HZ*5; | |
1182 | } | |
1183 | ||
cbee9f88 PZ |
1184 | static void task_numa_placement(struct task_struct *p) |
1185 | { | |
688b7585 MG |
1186 | int seq, nid, max_nid = -1; |
1187 | unsigned long max_faults = 0; | |
cbee9f88 | 1188 | |
2832bc19 | 1189 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1190 | if (p->numa_scan_seq == seq) |
1191 | return; | |
1192 | p->numa_scan_seq = seq; | |
3a7053b3 | 1193 | p->numa_migrate_seq++; |
598f0ec0 | 1194 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1195 | |
688b7585 MG |
1196 | /* Find the node with the highest number of faults */ |
1197 | for_each_online_node(nid) { | |
fb13c7ee | 1198 | unsigned long faults = 0; |
ac8e895b | 1199 | int priv, i; |
745d6147 | 1200 | |
ac8e895b | 1201 | for (priv = 0; priv < 2; priv++) { |
8c8a743c PZ |
1202 | long diff; |
1203 | ||
ac8e895b | 1204 | i = task_faults_idx(nid, priv); |
8c8a743c | 1205 | diff = -p->numa_faults[i]; |
745d6147 | 1206 | |
ac8e895b MG |
1207 | /* Decay existing window, copy faults since last scan */ |
1208 | p->numa_faults[i] >>= 1; | |
1209 | p->numa_faults[i] += p->numa_faults_buffer[i]; | |
1210 | p->numa_faults_buffer[i] = 0; | |
fb13c7ee MG |
1211 | |
1212 | faults += p->numa_faults[i]; | |
8c8a743c PZ |
1213 | diff += p->numa_faults[i]; |
1214 | if (p->numa_group) { | |
1215 | /* safe because we can only change our own group */ | |
1216 | atomic_long_add(diff, &p->numa_group->faults[i]); | |
1217 | } | |
ac8e895b MG |
1218 | } |
1219 | ||
688b7585 MG |
1220 | if (faults > max_faults) { |
1221 | max_faults = faults; | |
1222 | max_nid = nid; | |
1223 | } | |
1224 | } | |
1225 | ||
6b9a7460 | 1226 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1227 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1228 | /* Update the preferred nid and migrate task if possible */ |
688b7585 | 1229 | p->numa_preferred_nid = max_nid; |
6fe6b2d6 | 1230 | p->numa_migrate_seq = 1; |
6b9a7460 | 1231 | numa_migrate_preferred(p); |
3a7053b3 | 1232 | } |
cbee9f88 PZ |
1233 | } |
1234 | ||
8c8a743c PZ |
1235 | static inline int get_numa_group(struct numa_group *grp) |
1236 | { | |
1237 | return atomic_inc_not_zero(&grp->refcount); | |
1238 | } | |
1239 | ||
1240 | static inline void put_numa_group(struct numa_group *grp) | |
1241 | { | |
1242 | if (atomic_dec_and_test(&grp->refcount)) | |
1243 | kfree_rcu(grp, rcu); | |
1244 | } | |
1245 | ||
1246 | static void double_lock(spinlock_t *l1, spinlock_t *l2) | |
1247 | { | |
1248 | if (l1 > l2) | |
1249 | swap(l1, l2); | |
1250 | ||
1251 | spin_lock(l1); | |
1252 | spin_lock_nested(l2, SINGLE_DEPTH_NESTING); | |
1253 | } | |
1254 | ||
1255 | static void task_numa_group(struct task_struct *p, int cpupid) | |
1256 | { | |
1257 | struct numa_group *grp, *my_grp; | |
1258 | struct task_struct *tsk; | |
1259 | bool join = false; | |
1260 | int cpu = cpupid_to_cpu(cpupid); | |
1261 | int i; | |
1262 | ||
1263 | if (unlikely(!p->numa_group)) { | |
1264 | unsigned int size = sizeof(struct numa_group) + | |
1265 | 2*nr_node_ids*sizeof(atomic_long_t); | |
1266 | ||
1267 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1268 | if (!grp) | |
1269 | return; | |
1270 | ||
1271 | atomic_set(&grp->refcount, 1); | |
1272 | spin_lock_init(&grp->lock); | |
1273 | INIT_LIST_HEAD(&grp->task_list); | |
e29cf08b | 1274 | grp->gid = p->pid; |
8c8a743c PZ |
1275 | |
1276 | for (i = 0; i < 2*nr_node_ids; i++) | |
1277 | atomic_long_set(&grp->faults[i], p->numa_faults[i]); | |
1278 | ||
1279 | list_add(&p->numa_entry, &grp->task_list); | |
1280 | grp->nr_tasks++; | |
1281 | rcu_assign_pointer(p->numa_group, grp); | |
1282 | } | |
1283 | ||
1284 | rcu_read_lock(); | |
1285 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1286 | ||
1287 | if (!cpupid_match_pid(tsk, cpupid)) | |
1288 | goto unlock; | |
1289 | ||
1290 | grp = rcu_dereference(tsk->numa_group); | |
1291 | if (!grp) | |
1292 | goto unlock; | |
1293 | ||
1294 | my_grp = p->numa_group; | |
1295 | if (grp == my_grp) | |
1296 | goto unlock; | |
1297 | ||
1298 | /* | |
1299 | * Only join the other group if its bigger; if we're the bigger group, | |
1300 | * the other task will join us. | |
1301 | */ | |
1302 | if (my_grp->nr_tasks > grp->nr_tasks) | |
1303 | goto unlock; | |
1304 | ||
1305 | /* | |
1306 | * Tie-break on the grp address. | |
1307 | */ | |
1308 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
1309 | goto unlock; | |
1310 | ||
1311 | if (!get_numa_group(grp)) | |
1312 | goto unlock; | |
1313 | ||
1314 | join = true; | |
1315 | ||
1316 | unlock: | |
1317 | rcu_read_unlock(); | |
1318 | ||
1319 | if (!join) | |
1320 | return; | |
1321 | ||
1322 | for (i = 0; i < 2*nr_node_ids; i++) { | |
1323 | atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]); | |
1324 | atomic_long_add(p->numa_faults[i], &grp->faults[i]); | |
1325 | } | |
1326 | ||
1327 | double_lock(&my_grp->lock, &grp->lock); | |
1328 | ||
1329 | list_move(&p->numa_entry, &grp->task_list); | |
1330 | my_grp->nr_tasks--; | |
1331 | grp->nr_tasks++; | |
1332 | ||
1333 | spin_unlock(&my_grp->lock); | |
1334 | spin_unlock(&grp->lock); | |
1335 | ||
1336 | rcu_assign_pointer(p->numa_group, grp); | |
1337 | ||
1338 | put_numa_group(my_grp); | |
1339 | } | |
1340 | ||
1341 | void task_numa_free(struct task_struct *p) | |
1342 | { | |
1343 | struct numa_group *grp = p->numa_group; | |
1344 | int i; | |
1345 | ||
1346 | if (grp) { | |
1347 | for (i = 0; i < 2*nr_node_ids; i++) | |
1348 | atomic_long_sub(p->numa_faults[i], &grp->faults[i]); | |
1349 | ||
1350 | spin_lock(&grp->lock); | |
1351 | list_del(&p->numa_entry); | |
1352 | grp->nr_tasks--; | |
1353 | spin_unlock(&grp->lock); | |
1354 | rcu_assign_pointer(p->numa_group, NULL); | |
1355 | put_numa_group(grp); | |
1356 | } | |
1357 | ||
1358 | kfree(p->numa_faults); | |
1359 | } | |
1360 | ||
cbee9f88 PZ |
1361 | /* |
1362 | * Got a PROT_NONE fault for a page on @node. | |
1363 | */ | |
90572890 | 1364 | void task_numa_fault(int last_cpupid, int node, int pages, bool migrated) |
cbee9f88 PZ |
1365 | { |
1366 | struct task_struct *p = current; | |
ac8e895b | 1367 | int priv; |
cbee9f88 | 1368 | |
10e84b97 | 1369 | if (!numabalancing_enabled) |
1a687c2e MG |
1370 | return; |
1371 | ||
9ff1d9ff MG |
1372 | /* for example, ksmd faulting in a user's mm */ |
1373 | if (!p->mm) | |
1374 | return; | |
1375 | ||
f809ca9a MG |
1376 | /* Allocate buffer to track faults on a per-node basis */ |
1377 | if (unlikely(!p->numa_faults)) { | |
ac8e895b | 1378 | int size = sizeof(*p->numa_faults) * 2 * nr_node_ids; |
f809ca9a | 1379 | |
745d6147 MG |
1380 | /* numa_faults and numa_faults_buffer share the allocation */ |
1381 | p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN); | |
f809ca9a MG |
1382 | if (!p->numa_faults) |
1383 | return; | |
745d6147 MG |
1384 | |
1385 | BUG_ON(p->numa_faults_buffer); | |
ac8e895b | 1386 | p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids); |
f809ca9a | 1387 | } |
cbee9f88 | 1388 | |
8c8a743c PZ |
1389 | /* |
1390 | * First accesses are treated as private, otherwise consider accesses | |
1391 | * to be private if the accessing pid has not changed | |
1392 | */ | |
1393 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
1394 | priv = 1; | |
1395 | } else { | |
1396 | priv = cpupid_match_pid(p, last_cpupid); | |
1397 | if (!priv) | |
1398 | task_numa_group(p, last_cpupid); | |
1399 | } | |
1400 | ||
fb003b80 | 1401 | /* |
b8593bfd MG |
1402 | * If pages are properly placed (did not migrate) then scan slower. |
1403 | * This is reset periodically in case of phase changes | |
fb003b80 | 1404 | */ |
598f0ec0 MG |
1405 | if (!migrated) { |
1406 | /* Initialise if necessary */ | |
1407 | if (!p->numa_scan_period_max) | |
1408 | p->numa_scan_period_max = task_scan_max(p); | |
1409 | ||
1410 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1411 | p->numa_scan_period + 10); | |
1412 | } | |
fb003b80 | 1413 | |
cbee9f88 | 1414 | task_numa_placement(p); |
f809ca9a | 1415 | |
6b9a7460 MG |
1416 | /* Retry task to preferred node migration if it previously failed */ |
1417 | if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry)) | |
1418 | numa_migrate_preferred(p); | |
1419 | ||
ac8e895b | 1420 | p->numa_faults_buffer[task_faults_idx(node, priv)] += pages; |
cbee9f88 PZ |
1421 | } |
1422 | ||
6e5fb223 PZ |
1423 | static void reset_ptenuma_scan(struct task_struct *p) |
1424 | { | |
1425 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1426 | p->mm->numa_scan_offset = 0; | |
1427 | } | |
1428 | ||
cbee9f88 PZ |
1429 | /* |
1430 | * The expensive part of numa migration is done from task_work context. | |
1431 | * Triggered from task_tick_numa(). | |
1432 | */ | |
1433 | void task_numa_work(struct callback_head *work) | |
1434 | { | |
1435 | unsigned long migrate, next_scan, now = jiffies; | |
1436 | struct task_struct *p = current; | |
1437 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1438 | struct vm_area_struct *vma; |
9f40604c | 1439 | unsigned long start, end; |
598f0ec0 | 1440 | unsigned long nr_pte_updates = 0; |
9f40604c | 1441 | long pages; |
cbee9f88 PZ |
1442 | |
1443 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1444 | ||
1445 | work->next = work; /* protect against double add */ | |
1446 | /* | |
1447 | * Who cares about NUMA placement when they're dying. | |
1448 | * | |
1449 | * NOTE: make sure not to dereference p->mm before this check, | |
1450 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1451 | * without p->mm even though we still had it when we enqueued this | |
1452 | * work. | |
1453 | */ | |
1454 | if (p->flags & PF_EXITING) | |
1455 | return; | |
1456 | ||
7e8d16b6 MG |
1457 | if (!mm->numa_next_reset || !mm->numa_next_scan) { |
1458 | mm->numa_next_scan = now + | |
1459 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
1460 | mm->numa_next_reset = now + | |
1461 | msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); | |
1462 | } | |
1463 | ||
b8593bfd MG |
1464 | /* |
1465 | * Reset the scan period if enough time has gone by. Objective is that | |
1466 | * scanning will be reduced if pages are properly placed. As tasks | |
1467 | * can enter different phases this needs to be re-examined. Lacking | |
1468 | * proper tracking of reference behaviour, this blunt hammer is used. | |
1469 | */ | |
1470 | migrate = mm->numa_next_reset; | |
1471 | if (time_after(now, migrate)) { | |
598f0ec0 | 1472 | p->numa_scan_period = task_scan_min(p); |
b8593bfd MG |
1473 | next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); |
1474 | xchg(&mm->numa_next_reset, next_scan); | |
1475 | } | |
1476 | ||
cbee9f88 PZ |
1477 | /* |
1478 | * Enforce maximal scan/migration frequency.. | |
1479 | */ | |
1480 | migrate = mm->numa_next_scan; | |
1481 | if (time_before(now, migrate)) | |
1482 | return; | |
1483 | ||
598f0ec0 MG |
1484 | if (p->numa_scan_period == 0) { |
1485 | p->numa_scan_period_max = task_scan_max(p); | |
1486 | p->numa_scan_period = task_scan_min(p); | |
1487 | } | |
cbee9f88 | 1488 | |
fb003b80 | 1489 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1490 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1491 | return; | |
1492 | ||
19a78d11 PZ |
1493 | /* |
1494 | * Delay this task enough that another task of this mm will likely win | |
1495 | * the next time around. | |
1496 | */ | |
1497 | p->node_stamp += 2 * TICK_NSEC; | |
1498 | ||
9f40604c MG |
1499 | start = mm->numa_scan_offset; |
1500 | pages = sysctl_numa_balancing_scan_size; | |
1501 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1502 | if (!pages) | |
1503 | return; | |
cbee9f88 | 1504 | |
6e5fb223 | 1505 | down_read(&mm->mmap_sem); |
9f40604c | 1506 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1507 | if (!vma) { |
1508 | reset_ptenuma_scan(p); | |
9f40604c | 1509 | start = 0; |
6e5fb223 PZ |
1510 | vma = mm->mmap; |
1511 | } | |
9f40604c | 1512 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1513 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1514 | continue; |
1515 | ||
4591ce4f MG |
1516 | /* |
1517 | * Shared library pages mapped by multiple processes are not | |
1518 | * migrated as it is expected they are cache replicated. Avoid | |
1519 | * hinting faults in read-only file-backed mappings or the vdso | |
1520 | * as migrating the pages will be of marginal benefit. | |
1521 | */ | |
1522 | if (!vma->vm_mm || | |
1523 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1524 | continue; | |
1525 | ||
9f40604c MG |
1526 | do { |
1527 | start = max(start, vma->vm_start); | |
1528 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1529 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1530 | nr_pte_updates += change_prot_numa(vma, start, end); |
1531 | ||
1532 | /* | |
1533 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1534 | * at least one PTE is updated so that unused virtual | |
1535 | * address space is quickly skipped. | |
1536 | */ | |
1537 | if (nr_pte_updates) | |
1538 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1539 | |
9f40604c MG |
1540 | start = end; |
1541 | if (pages <= 0) | |
1542 | goto out; | |
1543 | } while (end != vma->vm_end); | |
cbee9f88 | 1544 | } |
6e5fb223 | 1545 | |
9f40604c | 1546 | out: |
f307cd1a MG |
1547 | /* |
1548 | * If the whole process was scanned without updates then no NUMA | |
1549 | * hinting faults are being recorded and scan rate should be lower. | |
1550 | */ | |
1551 | if (mm->numa_scan_offset == 0 && !nr_pte_updates) { | |
1552 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1553 | p->numa_scan_period << 1); | |
1554 | ||
1555 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); | |
1556 | mm->numa_next_scan = next_scan; | |
1557 | } | |
1558 | ||
6e5fb223 | 1559 | /* |
c69307d5 PZ |
1560 | * It is possible to reach the end of the VMA list but the last few |
1561 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1562 | * would find the !migratable VMA on the next scan but not reset the | |
1563 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1564 | */ |
1565 | if (vma) | |
9f40604c | 1566 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1567 | else |
1568 | reset_ptenuma_scan(p); | |
1569 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1570 | } |
1571 | ||
1572 | /* | |
1573 | * Drive the periodic memory faults.. | |
1574 | */ | |
1575 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1576 | { | |
1577 | struct callback_head *work = &curr->numa_work; | |
1578 | u64 period, now; | |
1579 | ||
1580 | /* | |
1581 | * We don't care about NUMA placement if we don't have memory. | |
1582 | */ | |
1583 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1584 | return; | |
1585 | ||
1586 | /* | |
1587 | * Using runtime rather than walltime has the dual advantage that | |
1588 | * we (mostly) drive the selection from busy threads and that the | |
1589 | * task needs to have done some actual work before we bother with | |
1590 | * NUMA placement. | |
1591 | */ | |
1592 | now = curr->se.sum_exec_runtime; | |
1593 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1594 | ||
1595 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1596 | if (!curr->node_stamp) |
598f0ec0 | 1597 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1598 | curr->node_stamp += period; |
cbee9f88 PZ |
1599 | |
1600 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1601 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1602 | task_work_add(curr, work, true); | |
1603 | } | |
1604 | } | |
1605 | } | |
1606 | #else | |
1607 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1608 | { | |
1609 | } | |
1610 | #endif /* CONFIG_NUMA_BALANCING */ | |
1611 | ||
30cfdcfc DA |
1612 | static void |
1613 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1614 | { | |
1615 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1616 | if (!parent_entity(se)) |
029632fb | 1617 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 PZ |
1618 | #ifdef CONFIG_SMP |
1619 | if (entity_is_task(se)) | |
eb95308e | 1620 | list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks); |
367456c7 | 1621 | #endif |
30cfdcfc | 1622 | cfs_rq->nr_running++; |
30cfdcfc DA |
1623 | } |
1624 | ||
1625 | static void | |
1626 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1627 | { | |
1628 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1629 | if (!parent_entity(se)) |
029632fb | 1630 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1631 | if (entity_is_task(se)) |
b87f1724 | 1632 | list_del_init(&se->group_node); |
30cfdcfc | 1633 | cfs_rq->nr_running--; |
30cfdcfc DA |
1634 | } |
1635 | ||
3ff6dcac YZ |
1636 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1637 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
1638 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
1639 | { | |
1640 | long tg_weight; | |
1641 | ||
1642 | /* | |
1643 | * Use this CPU's actual weight instead of the last load_contribution | |
1644 | * to gain a more accurate current total weight. See | |
1645 | * update_cfs_rq_load_contribution(). | |
1646 | */ | |
bf5b986e | 1647 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 1648 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
1649 | tg_weight += cfs_rq->load.weight; |
1650 | ||
1651 | return tg_weight; | |
1652 | } | |
1653 | ||
6d5ab293 | 1654 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 1655 | { |
cf5f0acf | 1656 | long tg_weight, load, shares; |
3ff6dcac | 1657 | |
cf5f0acf | 1658 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 1659 | load = cfs_rq->load.weight; |
3ff6dcac | 1660 | |
3ff6dcac | 1661 | shares = (tg->shares * load); |
cf5f0acf PZ |
1662 | if (tg_weight) |
1663 | shares /= tg_weight; | |
3ff6dcac YZ |
1664 | |
1665 | if (shares < MIN_SHARES) | |
1666 | shares = MIN_SHARES; | |
1667 | if (shares > tg->shares) | |
1668 | shares = tg->shares; | |
1669 | ||
1670 | return shares; | |
1671 | } | |
3ff6dcac | 1672 | # else /* CONFIG_SMP */ |
6d5ab293 | 1673 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
1674 | { |
1675 | return tg->shares; | |
1676 | } | |
3ff6dcac | 1677 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
1678 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1679 | unsigned long weight) | |
1680 | { | |
19e5eebb PT |
1681 | if (se->on_rq) { |
1682 | /* commit outstanding execution time */ | |
1683 | if (cfs_rq->curr == se) | |
1684 | update_curr(cfs_rq); | |
2069dd75 | 1685 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 1686 | } |
2069dd75 PZ |
1687 | |
1688 | update_load_set(&se->load, weight); | |
1689 | ||
1690 | if (se->on_rq) | |
1691 | account_entity_enqueue(cfs_rq, se); | |
1692 | } | |
1693 | ||
82958366 PT |
1694 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
1695 | ||
6d5ab293 | 1696 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1697 | { |
1698 | struct task_group *tg; | |
1699 | struct sched_entity *se; | |
3ff6dcac | 1700 | long shares; |
2069dd75 | 1701 | |
2069dd75 PZ |
1702 | tg = cfs_rq->tg; |
1703 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 1704 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 1705 | return; |
3ff6dcac YZ |
1706 | #ifndef CONFIG_SMP |
1707 | if (likely(se->load.weight == tg->shares)) | |
1708 | return; | |
1709 | #endif | |
6d5ab293 | 1710 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
1711 | |
1712 | reweight_entity(cfs_rq_of(se), se, shares); | |
1713 | } | |
1714 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 1715 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
1716 | { |
1717 | } | |
1718 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
1719 | ||
141965c7 | 1720 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
1721 | /* |
1722 | * We choose a half-life close to 1 scheduling period. | |
1723 | * Note: The tables below are dependent on this value. | |
1724 | */ | |
1725 | #define LOAD_AVG_PERIOD 32 | |
1726 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
1727 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
1728 | ||
1729 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
1730 | static const u32 runnable_avg_yN_inv[] = { | |
1731 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
1732 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
1733 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
1734 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
1735 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
1736 | 0x85aac367, 0x82cd8698, | |
1737 | }; | |
1738 | ||
1739 | /* | |
1740 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
1741 | * over-estimates when re-combining. | |
1742 | */ | |
1743 | static const u32 runnable_avg_yN_sum[] = { | |
1744 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
1745 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
1746 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
1747 | }; | |
1748 | ||
9d85f21c PT |
1749 | /* |
1750 | * Approximate: | |
1751 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
1752 | */ | |
1753 | static __always_inline u64 decay_load(u64 val, u64 n) | |
1754 | { | |
5b51f2f8 PT |
1755 | unsigned int local_n; |
1756 | ||
1757 | if (!n) | |
1758 | return val; | |
1759 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
1760 | return 0; | |
1761 | ||
1762 | /* after bounds checking we can collapse to 32-bit */ | |
1763 | local_n = n; | |
1764 | ||
1765 | /* | |
1766 | * As y^PERIOD = 1/2, we can combine | |
1767 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
1768 | * With a look-up table which covers k^n (n<PERIOD) | |
1769 | * | |
1770 | * To achieve constant time decay_load. | |
1771 | */ | |
1772 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
1773 | val >>= local_n / LOAD_AVG_PERIOD; | |
1774 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
1775 | } |
1776 | ||
5b51f2f8 PT |
1777 | val *= runnable_avg_yN_inv[local_n]; |
1778 | /* We don't use SRR here since we always want to round down. */ | |
1779 | return val >> 32; | |
1780 | } | |
1781 | ||
1782 | /* | |
1783 | * For updates fully spanning n periods, the contribution to runnable | |
1784 | * average will be: \Sum 1024*y^n | |
1785 | * | |
1786 | * We can compute this reasonably efficiently by combining: | |
1787 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
1788 | */ | |
1789 | static u32 __compute_runnable_contrib(u64 n) | |
1790 | { | |
1791 | u32 contrib = 0; | |
1792 | ||
1793 | if (likely(n <= LOAD_AVG_PERIOD)) | |
1794 | return runnable_avg_yN_sum[n]; | |
1795 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
1796 | return LOAD_AVG_MAX; | |
1797 | ||
1798 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
1799 | do { | |
1800 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
1801 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
1802 | ||
1803 | n -= LOAD_AVG_PERIOD; | |
1804 | } while (n > LOAD_AVG_PERIOD); | |
1805 | ||
1806 | contrib = decay_load(contrib, n); | |
1807 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
1808 | } |
1809 | ||
1810 | /* | |
1811 | * We can represent the historical contribution to runnable average as the | |
1812 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
1813 | * history into segments of approximately 1ms (1024us); label the segment that | |
1814 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
1815 | * | |
1816 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
1817 | * p0 p1 p2 | |
1818 | * (now) (~1ms ago) (~2ms ago) | |
1819 | * | |
1820 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
1821 | * | |
1822 | * We then designate the fractions u_i as our co-efficients, yielding the | |
1823 | * following representation of historical load: | |
1824 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
1825 | * | |
1826 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
1827 | * y^32 = 0.5 | |
1828 | * | |
1829 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
1830 | * approximately half as much as the contribution to load within the last ms | |
1831 | * (u_0). | |
1832 | * | |
1833 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
1834 | * sum again by y is sufficient to update: | |
1835 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
1836 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
1837 | */ | |
1838 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
1839 | struct sched_avg *sa, | |
1840 | int runnable) | |
1841 | { | |
5b51f2f8 PT |
1842 | u64 delta, periods; |
1843 | u32 runnable_contrib; | |
9d85f21c PT |
1844 | int delta_w, decayed = 0; |
1845 | ||
1846 | delta = now - sa->last_runnable_update; | |
1847 | /* | |
1848 | * This should only happen when time goes backwards, which it | |
1849 | * unfortunately does during sched clock init when we swap over to TSC. | |
1850 | */ | |
1851 | if ((s64)delta < 0) { | |
1852 | sa->last_runnable_update = now; | |
1853 | return 0; | |
1854 | } | |
1855 | ||
1856 | /* | |
1857 | * Use 1024ns as the unit of measurement since it's a reasonable | |
1858 | * approximation of 1us and fast to compute. | |
1859 | */ | |
1860 | delta >>= 10; | |
1861 | if (!delta) | |
1862 | return 0; | |
1863 | sa->last_runnable_update = now; | |
1864 | ||
1865 | /* delta_w is the amount already accumulated against our next period */ | |
1866 | delta_w = sa->runnable_avg_period % 1024; | |
1867 | if (delta + delta_w >= 1024) { | |
1868 | /* period roll-over */ | |
1869 | decayed = 1; | |
1870 | ||
1871 | /* | |
1872 | * Now that we know we're crossing a period boundary, figure | |
1873 | * out how much from delta we need to complete the current | |
1874 | * period and accrue it. | |
1875 | */ | |
1876 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
1877 | if (runnable) |
1878 | sa->runnable_avg_sum += delta_w; | |
1879 | sa->runnable_avg_period += delta_w; | |
1880 | ||
1881 | delta -= delta_w; | |
1882 | ||
1883 | /* Figure out how many additional periods this update spans */ | |
1884 | periods = delta / 1024; | |
1885 | delta %= 1024; | |
1886 | ||
1887 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
1888 | periods + 1); | |
1889 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
1890 | periods + 1); | |
1891 | ||
1892 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
1893 | runnable_contrib = __compute_runnable_contrib(periods); | |
1894 | if (runnable) | |
1895 | sa->runnable_avg_sum += runnable_contrib; | |
1896 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
1897 | } |
1898 | ||
1899 | /* Remainder of delta accrued against u_0` */ | |
1900 | if (runnable) | |
1901 | sa->runnable_avg_sum += delta; | |
1902 | sa->runnable_avg_period += delta; | |
1903 | ||
1904 | return decayed; | |
1905 | } | |
1906 | ||
9ee474f5 | 1907 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 1908 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
1909 | { |
1910 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
1911 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
1912 | ||
1913 | decays -= se->avg.decay_count; | |
1914 | if (!decays) | |
aff3e498 | 1915 | return 0; |
9ee474f5 PT |
1916 | |
1917 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
1918 | se->avg.decay_count = 0; | |
aff3e498 PT |
1919 | |
1920 | return decays; | |
9ee474f5 PT |
1921 | } |
1922 | ||
c566e8e9 PT |
1923 | #ifdef CONFIG_FAIR_GROUP_SCHED |
1924 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
1925 | int force_update) | |
1926 | { | |
1927 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 1928 | long tg_contrib; |
c566e8e9 PT |
1929 | |
1930 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
1931 | tg_contrib -= cfs_rq->tg_load_contrib; | |
1932 | ||
bf5b986e AS |
1933 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
1934 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
1935 | cfs_rq->tg_load_contrib += tg_contrib; |
1936 | } | |
1937 | } | |
8165e145 | 1938 | |
bb17f655 PT |
1939 | /* |
1940 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
1941 | * representation for computing load contributions. | |
1942 | */ | |
1943 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
1944 | struct cfs_rq *cfs_rq) | |
1945 | { | |
1946 | struct task_group *tg = cfs_rq->tg; | |
1947 | long contrib; | |
1948 | ||
1949 | /* The fraction of a cpu used by this cfs_rq */ | |
1950 | contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT, | |
1951 | sa->runnable_avg_period + 1); | |
1952 | contrib -= cfs_rq->tg_runnable_contrib; | |
1953 | ||
1954 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
1955 | atomic_add(contrib, &tg->runnable_avg); | |
1956 | cfs_rq->tg_runnable_contrib += contrib; | |
1957 | } | |
1958 | } | |
1959 | ||
8165e145 PT |
1960 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
1961 | { | |
1962 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
1963 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
1964 | int runnable_avg; |
1965 | ||
8165e145 PT |
1966 | u64 contrib; |
1967 | ||
1968 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
1969 | se->avg.load_avg_contrib = div_u64(contrib, |
1970 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
1971 | |
1972 | /* | |
1973 | * For group entities we need to compute a correction term in the case | |
1974 | * that they are consuming <1 cpu so that we would contribute the same | |
1975 | * load as a task of equal weight. | |
1976 | * | |
1977 | * Explicitly co-ordinating this measurement would be expensive, but | |
1978 | * fortunately the sum of each cpus contribution forms a usable | |
1979 | * lower-bound on the true value. | |
1980 | * | |
1981 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
1982 | * (and the sum represents true value) or they are disjoint and we are | |
1983 | * understating by the aggregate of their overlap. | |
1984 | * | |
1985 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
1986 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
1987 | * cpus that overlap for this interval and w_i is the interval width. | |
1988 | * | |
1989 | * On a small machine; the first term is well-bounded which bounds the | |
1990 | * total error since w_i is a subset of the period. Whereas on a | |
1991 | * larger machine, while this first term can be larger, if w_i is the | |
1992 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
1993 | * our upper bound of 1-cpu. | |
1994 | */ | |
1995 | runnable_avg = atomic_read(&tg->runnable_avg); | |
1996 | if (runnable_avg < NICE_0_LOAD) { | |
1997 | se->avg.load_avg_contrib *= runnable_avg; | |
1998 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
1999 | } | |
8165e145 | 2000 | } |
c566e8e9 PT |
2001 | #else |
2002 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2003 | int force_update) {} | |
bb17f655 PT |
2004 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2005 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 2006 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
2007 | #endif |
2008 | ||
8165e145 PT |
2009 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2010 | { | |
2011 | u32 contrib; | |
2012 | ||
2013 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2014 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2015 | contrib /= (se->avg.runnable_avg_period + 1); | |
2016 | se->avg.load_avg_contrib = scale_load(contrib); | |
2017 | } | |
2018 | ||
2dac754e PT |
2019 | /* Compute the current contribution to load_avg by se, return any delta */ |
2020 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2021 | { | |
2022 | long old_contrib = se->avg.load_avg_contrib; | |
2023 | ||
8165e145 PT |
2024 | if (entity_is_task(se)) { |
2025 | __update_task_entity_contrib(se); | |
2026 | } else { | |
bb17f655 | 2027 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2028 | __update_group_entity_contrib(se); |
2029 | } | |
2dac754e PT |
2030 | |
2031 | return se->avg.load_avg_contrib - old_contrib; | |
2032 | } | |
2033 | ||
9ee474f5 PT |
2034 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2035 | long load_contrib) | |
2036 | { | |
2037 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2038 | cfs_rq->blocked_load_avg -= load_contrib; | |
2039 | else | |
2040 | cfs_rq->blocked_load_avg = 0; | |
2041 | } | |
2042 | ||
f1b17280 PT |
2043 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2044 | ||
9d85f21c | 2045 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2046 | static inline void update_entity_load_avg(struct sched_entity *se, |
2047 | int update_cfs_rq) | |
9d85f21c | 2048 | { |
2dac754e PT |
2049 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2050 | long contrib_delta; | |
f1b17280 | 2051 | u64 now; |
2dac754e | 2052 | |
f1b17280 PT |
2053 | /* |
2054 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2055 | * case they are the parent of a throttled hierarchy. | |
2056 | */ | |
2057 | if (entity_is_task(se)) | |
2058 | now = cfs_rq_clock_task(cfs_rq); | |
2059 | else | |
2060 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2061 | ||
2062 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2063 | return; |
2064 | ||
2065 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2066 | |
2067 | if (!update_cfs_rq) | |
2068 | return; | |
2069 | ||
2dac754e PT |
2070 | if (se->on_rq) |
2071 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2072 | else |
2073 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2074 | } | |
2075 | ||
2076 | /* | |
2077 | * Decay the load contributed by all blocked children and account this so that | |
2078 | * their contribution may appropriately discounted when they wake up. | |
2079 | */ | |
aff3e498 | 2080 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2081 | { |
f1b17280 | 2082 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2083 | u64 decays; |
2084 | ||
2085 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2086 | if (!decays && !force_update) |
9ee474f5 PT |
2087 | return; |
2088 | ||
2509940f AS |
2089 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2090 | unsigned long removed_load; | |
2091 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2092 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2093 | } | |
9ee474f5 | 2094 | |
aff3e498 PT |
2095 | if (decays) { |
2096 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2097 | decays); | |
2098 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2099 | cfs_rq->last_decay = now; | |
2100 | } | |
c566e8e9 PT |
2101 | |
2102 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2103 | } |
18bf2805 BS |
2104 | |
2105 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2106 | { | |
78becc27 | 2107 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 2108 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 2109 | } |
2dac754e PT |
2110 | |
2111 | /* Add the load generated by se into cfs_rq's child load-average */ | |
2112 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2113 | struct sched_entity *se, |
2114 | int wakeup) | |
2dac754e | 2115 | { |
aff3e498 PT |
2116 | /* |
2117 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2118 | * migration we use a negative decay count to track the remote decays | |
2119 | * accumulated while sleeping. | |
a75cdaa9 AS |
2120 | * |
2121 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2122 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2123 | * constructed load_avg_contrib. | |
aff3e498 PT |
2124 | */ |
2125 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2126 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2127 | if (se->avg.decay_count) { |
2128 | /* | |
2129 | * In a wake-up migration we have to approximate the | |
2130 | * time sleeping. This is because we can't synchronize | |
2131 | * clock_task between the two cpus, and it is not | |
2132 | * guaranteed to be read-safe. Instead, we can | |
2133 | * approximate this using our carried decays, which are | |
2134 | * explicitly atomically readable. | |
2135 | */ | |
2136 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2137 | << 20; | |
2138 | update_entity_load_avg(se, 0); | |
2139 | /* Indicate that we're now synchronized and on-rq */ | |
2140 | se->avg.decay_count = 0; | |
2141 | } | |
9ee474f5 PT |
2142 | wakeup = 0; |
2143 | } else { | |
282cf499 AS |
2144 | /* |
2145 | * Task re-woke on same cpu (or else migrate_task_rq_fair() | |
2146 | * would have made count negative); we must be careful to avoid | |
2147 | * double-accounting blocked time after synchronizing decays. | |
2148 | */ | |
2149 | se->avg.last_runnable_update += __synchronize_entity_decay(se) | |
2150 | << 20; | |
9ee474f5 PT |
2151 | } |
2152 | ||
aff3e498 PT |
2153 | /* migrated tasks did not contribute to our blocked load */ |
2154 | if (wakeup) { | |
9ee474f5 | 2155 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2156 | update_entity_load_avg(se, 0); |
2157 | } | |
9ee474f5 | 2158 | |
2dac754e | 2159 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2160 | /* we force update consideration on load-balancer moves */ |
2161 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2162 | } |
2163 | ||
9ee474f5 PT |
2164 | /* |
2165 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2166 | * transitioning to a blocked state we track its projected decay using | |
2167 | * blocked_load_avg. | |
2168 | */ | |
2dac754e | 2169 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2170 | struct sched_entity *se, |
2171 | int sleep) | |
2dac754e | 2172 | { |
9ee474f5 | 2173 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2174 | /* we force update consideration on load-balancer moves */ |
2175 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2176 | |
2dac754e | 2177 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2178 | if (sleep) { |
2179 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2180 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2181 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2182 | } |
642dbc39 VG |
2183 | |
2184 | /* | |
2185 | * Update the rq's load with the elapsed running time before entering | |
2186 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2187 | * be the only way to update the runnable statistic. | |
2188 | */ | |
2189 | void idle_enter_fair(struct rq *this_rq) | |
2190 | { | |
2191 | update_rq_runnable_avg(this_rq, 1); | |
2192 | } | |
2193 | ||
2194 | /* | |
2195 | * Update the rq's load with the elapsed idle time before a task is | |
2196 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2197 | * be the only way to update the runnable statistic. | |
2198 | */ | |
2199 | void idle_exit_fair(struct rq *this_rq) | |
2200 | { | |
2201 | update_rq_runnable_avg(this_rq, 0); | |
2202 | } | |
2203 | ||
9d85f21c | 2204 | #else |
9ee474f5 PT |
2205 | static inline void update_entity_load_avg(struct sched_entity *se, |
2206 | int update_cfs_rq) {} | |
18bf2805 | 2207 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2208 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2209 | struct sched_entity *se, |
2210 | int wakeup) {} | |
2dac754e | 2211 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2212 | struct sched_entity *se, |
2213 | int sleep) {} | |
aff3e498 PT |
2214 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2215 | int force_update) {} | |
9d85f21c PT |
2216 | #endif |
2217 | ||
2396af69 | 2218 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2219 | { |
bf0f6f24 | 2220 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2221 | struct task_struct *tsk = NULL; |
2222 | ||
2223 | if (entity_is_task(se)) | |
2224 | tsk = task_of(se); | |
2225 | ||
41acab88 | 2226 | if (se->statistics.sleep_start) { |
78becc27 | 2227 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2228 | |
2229 | if ((s64)delta < 0) | |
2230 | delta = 0; | |
2231 | ||
41acab88 LDM |
2232 | if (unlikely(delta > se->statistics.sleep_max)) |
2233 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2234 | |
8c79a045 | 2235 | se->statistics.sleep_start = 0; |
41acab88 | 2236 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2237 | |
768d0c27 | 2238 | if (tsk) { |
e414314c | 2239 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2240 | trace_sched_stat_sleep(tsk, delta); |
2241 | } | |
bf0f6f24 | 2242 | } |
41acab88 | 2243 | if (se->statistics.block_start) { |
78becc27 | 2244 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2245 | |
2246 | if ((s64)delta < 0) | |
2247 | delta = 0; | |
2248 | ||
41acab88 LDM |
2249 | if (unlikely(delta > se->statistics.block_max)) |
2250 | se->statistics.block_max = delta; | |
bf0f6f24 | 2251 | |
8c79a045 | 2252 | se->statistics.block_start = 0; |
41acab88 | 2253 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2254 | |
e414314c | 2255 | if (tsk) { |
8f0dfc34 | 2256 | if (tsk->in_iowait) { |
41acab88 LDM |
2257 | se->statistics.iowait_sum += delta; |
2258 | se->statistics.iowait_count++; | |
768d0c27 | 2259 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2260 | } |
2261 | ||
b781a602 AV |
2262 | trace_sched_stat_blocked(tsk, delta); |
2263 | ||
e414314c PZ |
2264 | /* |
2265 | * Blocking time is in units of nanosecs, so shift by | |
2266 | * 20 to get a milliseconds-range estimation of the | |
2267 | * amount of time that the task spent sleeping: | |
2268 | */ | |
2269 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2270 | profile_hits(SLEEP_PROFILING, | |
2271 | (void *)get_wchan(tsk), | |
2272 | delta >> 20); | |
2273 | } | |
2274 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2275 | } |
bf0f6f24 IM |
2276 | } |
2277 | #endif | |
2278 | } | |
2279 | ||
ddc97297 PZ |
2280 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2281 | { | |
2282 | #ifdef CONFIG_SCHED_DEBUG | |
2283 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2284 | ||
2285 | if (d < 0) | |
2286 | d = -d; | |
2287 | ||
2288 | if (d > 3*sysctl_sched_latency) | |
2289 | schedstat_inc(cfs_rq, nr_spread_over); | |
2290 | #endif | |
2291 | } | |
2292 | ||
aeb73b04 PZ |
2293 | static void |
2294 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2295 | { | |
1af5f730 | 2296 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2297 | |
2cb8600e PZ |
2298 | /* |
2299 | * The 'current' period is already promised to the current tasks, | |
2300 | * however the extra weight of the new task will slow them down a | |
2301 | * little, place the new task so that it fits in the slot that | |
2302 | * stays open at the end. | |
2303 | */ | |
94dfb5e7 | 2304 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2305 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2306 | |
a2e7a7eb | 2307 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2308 | if (!initial) { |
a2e7a7eb | 2309 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2310 | |
a2e7a7eb MG |
2311 | /* |
2312 | * Halve their sleep time's effect, to allow | |
2313 | * for a gentler effect of sleepers: | |
2314 | */ | |
2315 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2316 | thresh >>= 1; | |
51e0304c | 2317 | |
a2e7a7eb | 2318 | vruntime -= thresh; |
aeb73b04 PZ |
2319 | } |
2320 | ||
b5d9d734 | 2321 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2322 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2323 | } |
2324 | ||
d3d9dc33 PT |
2325 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2326 | ||
bf0f6f24 | 2327 | static void |
88ec22d3 | 2328 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2329 | { |
88ec22d3 PZ |
2330 | /* |
2331 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2332 | * through calling update_curr(). |
88ec22d3 | 2333 | */ |
371fd7e7 | 2334 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2335 | se->vruntime += cfs_rq->min_vruntime; |
2336 | ||
bf0f6f24 | 2337 | /* |
a2a2d680 | 2338 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2339 | */ |
b7cc0896 | 2340 | update_curr(cfs_rq); |
f269ae04 | 2341 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2342 | account_entity_enqueue(cfs_rq, se); |
2343 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2344 | |
88ec22d3 | 2345 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2346 | place_entity(cfs_rq, se, 0); |
2396af69 | 2347 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2348 | } |
bf0f6f24 | 2349 | |
d2417e5a | 2350 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2351 | check_spread(cfs_rq, se); |
83b699ed SV |
2352 | if (se != cfs_rq->curr) |
2353 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2354 | se->on_rq = 1; |
3d4b47b4 | 2355 | |
d3d9dc33 | 2356 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2357 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2358 | check_enqueue_throttle(cfs_rq); |
2359 | } | |
bf0f6f24 IM |
2360 | } |
2361 | ||
2c13c919 | 2362 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2363 | { |
2c13c919 RR |
2364 | for_each_sched_entity(se) { |
2365 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2366 | if (cfs_rq->last == se) | |
2367 | cfs_rq->last = NULL; | |
2368 | else | |
2369 | break; | |
2370 | } | |
2371 | } | |
2002c695 | 2372 | |
2c13c919 RR |
2373 | static void __clear_buddies_next(struct sched_entity *se) |
2374 | { | |
2375 | for_each_sched_entity(se) { | |
2376 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2377 | if (cfs_rq->next == se) | |
2378 | cfs_rq->next = NULL; | |
2379 | else | |
2380 | break; | |
2381 | } | |
2002c695 PZ |
2382 | } |
2383 | ||
ac53db59 RR |
2384 | static void __clear_buddies_skip(struct sched_entity *se) |
2385 | { | |
2386 | for_each_sched_entity(se) { | |
2387 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2388 | if (cfs_rq->skip == se) | |
2389 | cfs_rq->skip = NULL; | |
2390 | else | |
2391 | break; | |
2392 | } | |
2393 | } | |
2394 | ||
a571bbea PZ |
2395 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2396 | { | |
2c13c919 RR |
2397 | if (cfs_rq->last == se) |
2398 | __clear_buddies_last(se); | |
2399 | ||
2400 | if (cfs_rq->next == se) | |
2401 | __clear_buddies_next(se); | |
ac53db59 RR |
2402 | |
2403 | if (cfs_rq->skip == se) | |
2404 | __clear_buddies_skip(se); | |
a571bbea PZ |
2405 | } |
2406 | ||
6c16a6dc | 2407 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2408 | |
bf0f6f24 | 2409 | static void |
371fd7e7 | 2410 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2411 | { |
a2a2d680 DA |
2412 | /* |
2413 | * Update run-time statistics of the 'current'. | |
2414 | */ | |
2415 | update_curr(cfs_rq); | |
17bc14b7 | 2416 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2417 | |
19b6a2e3 | 2418 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2419 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2420 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2421 | if (entity_is_task(se)) { |
2422 | struct task_struct *tsk = task_of(se); | |
2423 | ||
2424 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2425 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2426 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2427 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2428 | } |
db36cc7d | 2429 | #endif |
67e9fb2a PZ |
2430 | } |
2431 | ||
2002c695 | 2432 | clear_buddies(cfs_rq, se); |
4793241b | 2433 | |
83b699ed | 2434 | if (se != cfs_rq->curr) |
30cfdcfc | 2435 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2436 | se->on_rq = 0; |
30cfdcfc | 2437 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2438 | |
2439 | /* | |
2440 | * Normalize the entity after updating the min_vruntime because the | |
2441 | * update can refer to the ->curr item and we need to reflect this | |
2442 | * movement in our normalized position. | |
2443 | */ | |
371fd7e7 | 2444 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2445 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2446 | |
d8b4986d PT |
2447 | /* return excess runtime on last dequeue */ |
2448 | return_cfs_rq_runtime(cfs_rq); | |
2449 | ||
1e876231 | 2450 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2451 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2452 | } |
2453 | ||
2454 | /* | |
2455 | * Preempt the current task with a newly woken task if needed: | |
2456 | */ | |
7c92e54f | 2457 | static void |
2e09bf55 | 2458 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2459 | { |
11697830 | 2460 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2461 | struct sched_entity *se; |
2462 | s64 delta; | |
11697830 | 2463 | |
6d0f0ebd | 2464 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2465 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2466 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2467 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2468 | /* |
2469 | * The current task ran long enough, ensure it doesn't get | |
2470 | * re-elected due to buddy favours. | |
2471 | */ | |
2472 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2473 | return; |
2474 | } | |
2475 | ||
2476 | /* | |
2477 | * Ensure that a task that missed wakeup preemption by a | |
2478 | * narrow margin doesn't have to wait for a full slice. | |
2479 | * This also mitigates buddy induced latencies under load. | |
2480 | */ | |
f685ceac MG |
2481 | if (delta_exec < sysctl_sched_min_granularity) |
2482 | return; | |
2483 | ||
f4cfb33e WX |
2484 | se = __pick_first_entity(cfs_rq); |
2485 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2486 | |
f4cfb33e WX |
2487 | if (delta < 0) |
2488 | return; | |
d7d82944 | 2489 | |
f4cfb33e WX |
2490 | if (delta > ideal_runtime) |
2491 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2492 | } |
2493 | ||
83b699ed | 2494 | static void |
8494f412 | 2495 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2496 | { |
83b699ed SV |
2497 | /* 'current' is not kept within the tree. */ |
2498 | if (se->on_rq) { | |
2499 | /* | |
2500 | * Any task has to be enqueued before it get to execute on | |
2501 | * a CPU. So account for the time it spent waiting on the | |
2502 | * runqueue. | |
2503 | */ | |
2504 | update_stats_wait_end(cfs_rq, se); | |
2505 | __dequeue_entity(cfs_rq, se); | |
2506 | } | |
2507 | ||
79303e9e | 2508 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2509 | cfs_rq->curr = se; |
eba1ed4b IM |
2510 | #ifdef CONFIG_SCHEDSTATS |
2511 | /* | |
2512 | * Track our maximum slice length, if the CPU's load is at | |
2513 | * least twice that of our own weight (i.e. dont track it | |
2514 | * when there are only lesser-weight tasks around): | |
2515 | */ | |
495eca49 | 2516 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2517 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2518 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2519 | } | |
2520 | #endif | |
4a55b450 | 2521 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2522 | } |
2523 | ||
3f3a4904 PZ |
2524 | static int |
2525 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2526 | ||
ac53db59 RR |
2527 | /* |
2528 | * Pick the next process, keeping these things in mind, in this order: | |
2529 | * 1) keep things fair between processes/task groups | |
2530 | * 2) pick the "next" process, since someone really wants that to run | |
2531 | * 3) pick the "last" process, for cache locality | |
2532 | * 4) do not run the "skip" process, if something else is available | |
2533 | */ | |
f4b6755f | 2534 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 2535 | { |
ac53db59 | 2536 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
f685ceac | 2537 | struct sched_entity *left = se; |
f4b6755f | 2538 | |
ac53db59 RR |
2539 | /* |
2540 | * Avoid running the skip buddy, if running something else can | |
2541 | * be done without getting too unfair. | |
2542 | */ | |
2543 | if (cfs_rq->skip == se) { | |
2544 | struct sched_entity *second = __pick_next_entity(se); | |
2545 | if (second && wakeup_preempt_entity(second, left) < 1) | |
2546 | se = second; | |
2547 | } | |
aa2ac252 | 2548 | |
f685ceac MG |
2549 | /* |
2550 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2551 | */ | |
2552 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2553 | se = cfs_rq->last; | |
2554 | ||
ac53db59 RR |
2555 | /* |
2556 | * Someone really wants this to run. If it's not unfair, run it. | |
2557 | */ | |
2558 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2559 | se = cfs_rq->next; | |
2560 | ||
f685ceac | 2561 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2562 | |
2563 | return se; | |
aa2ac252 PZ |
2564 | } |
2565 | ||
d3d9dc33 PT |
2566 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
2567 | ||
ab6cde26 | 2568 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2569 | { |
2570 | /* | |
2571 | * If still on the runqueue then deactivate_task() | |
2572 | * was not called and update_curr() has to be done: | |
2573 | */ | |
2574 | if (prev->on_rq) | |
b7cc0896 | 2575 | update_curr(cfs_rq); |
bf0f6f24 | 2576 | |
d3d9dc33 PT |
2577 | /* throttle cfs_rqs exceeding runtime */ |
2578 | check_cfs_rq_runtime(cfs_rq); | |
2579 | ||
ddc97297 | 2580 | check_spread(cfs_rq, prev); |
30cfdcfc | 2581 | if (prev->on_rq) { |
5870db5b | 2582 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2583 | /* Put 'current' back into the tree. */ |
2584 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2585 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2586 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2587 | } |
429d43bc | 2588 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2589 | } |
2590 | ||
8f4d37ec PZ |
2591 | static void |
2592 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2593 | { |
bf0f6f24 | 2594 | /* |
30cfdcfc | 2595 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2596 | */ |
30cfdcfc | 2597 | update_curr(cfs_rq); |
bf0f6f24 | 2598 | |
9d85f21c PT |
2599 | /* |
2600 | * Ensure that runnable average is periodically updated. | |
2601 | */ | |
9ee474f5 | 2602 | update_entity_load_avg(curr, 1); |
aff3e498 | 2603 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2604 | update_cfs_shares(cfs_rq); |
9d85f21c | 2605 | |
8f4d37ec PZ |
2606 | #ifdef CONFIG_SCHED_HRTICK |
2607 | /* | |
2608 | * queued ticks are scheduled to match the slice, so don't bother | |
2609 | * validating it and just reschedule. | |
2610 | */ | |
983ed7a6 HH |
2611 | if (queued) { |
2612 | resched_task(rq_of(cfs_rq)->curr); | |
2613 | return; | |
2614 | } | |
8f4d37ec PZ |
2615 | /* |
2616 | * don't let the period tick interfere with the hrtick preemption | |
2617 | */ | |
2618 | if (!sched_feat(DOUBLE_TICK) && | |
2619 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
2620 | return; | |
2621 | #endif | |
2622 | ||
2c2efaed | 2623 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 2624 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
2625 | } |
2626 | ||
ab84d31e PT |
2627 | |
2628 | /************************************************** | |
2629 | * CFS bandwidth control machinery | |
2630 | */ | |
2631 | ||
2632 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
2633 | |
2634 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 2635 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
2636 | |
2637 | static inline bool cfs_bandwidth_used(void) | |
2638 | { | |
c5905afb | 2639 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
2640 | } |
2641 | ||
2642 | void account_cfs_bandwidth_used(int enabled, int was_enabled) | |
2643 | { | |
2644 | /* only need to count groups transitioning between enabled/!enabled */ | |
2645 | if (enabled && !was_enabled) | |
c5905afb | 2646 | static_key_slow_inc(&__cfs_bandwidth_used); |
029632fb | 2647 | else if (!enabled && was_enabled) |
c5905afb | 2648 | static_key_slow_dec(&__cfs_bandwidth_used); |
029632fb PZ |
2649 | } |
2650 | #else /* HAVE_JUMP_LABEL */ | |
2651 | static bool cfs_bandwidth_used(void) | |
2652 | { | |
2653 | return true; | |
2654 | } | |
2655 | ||
2656 | void account_cfs_bandwidth_used(int enabled, int was_enabled) {} | |
2657 | #endif /* HAVE_JUMP_LABEL */ | |
2658 | ||
ab84d31e PT |
2659 | /* |
2660 | * default period for cfs group bandwidth. | |
2661 | * default: 0.1s, units: nanoseconds | |
2662 | */ | |
2663 | static inline u64 default_cfs_period(void) | |
2664 | { | |
2665 | return 100000000ULL; | |
2666 | } | |
ec12cb7f PT |
2667 | |
2668 | static inline u64 sched_cfs_bandwidth_slice(void) | |
2669 | { | |
2670 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
2671 | } | |
2672 | ||
a9cf55b2 PT |
2673 | /* |
2674 | * Replenish runtime according to assigned quota and update expiration time. | |
2675 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
2676 | * additional synchronization around rq->lock. | |
2677 | * | |
2678 | * requires cfs_b->lock | |
2679 | */ | |
029632fb | 2680 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
2681 | { |
2682 | u64 now; | |
2683 | ||
2684 | if (cfs_b->quota == RUNTIME_INF) | |
2685 | return; | |
2686 | ||
2687 | now = sched_clock_cpu(smp_processor_id()); | |
2688 | cfs_b->runtime = cfs_b->quota; | |
2689 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
2690 | } | |
2691 | ||
029632fb PZ |
2692 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
2693 | { | |
2694 | return &tg->cfs_bandwidth; | |
2695 | } | |
2696 | ||
f1b17280 PT |
2697 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
2698 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
2699 | { | |
2700 | if (unlikely(cfs_rq->throttle_count)) | |
2701 | return cfs_rq->throttled_clock_task; | |
2702 | ||
78becc27 | 2703 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
2704 | } |
2705 | ||
85dac906 PT |
2706 | /* returns 0 on failure to allocate runtime */ |
2707 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
2708 | { |
2709 | struct task_group *tg = cfs_rq->tg; | |
2710 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 2711 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
2712 | |
2713 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
2714 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
2715 | ||
2716 | raw_spin_lock(&cfs_b->lock); | |
2717 | if (cfs_b->quota == RUNTIME_INF) | |
2718 | amount = min_amount; | |
58088ad0 | 2719 | else { |
a9cf55b2 PT |
2720 | /* |
2721 | * If the bandwidth pool has become inactive, then at least one | |
2722 | * period must have elapsed since the last consumption. | |
2723 | * Refresh the global state and ensure bandwidth timer becomes | |
2724 | * active. | |
2725 | */ | |
2726 | if (!cfs_b->timer_active) { | |
2727 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 2728 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 2729 | } |
58088ad0 PT |
2730 | |
2731 | if (cfs_b->runtime > 0) { | |
2732 | amount = min(cfs_b->runtime, min_amount); | |
2733 | cfs_b->runtime -= amount; | |
2734 | cfs_b->idle = 0; | |
2735 | } | |
ec12cb7f | 2736 | } |
a9cf55b2 | 2737 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
2738 | raw_spin_unlock(&cfs_b->lock); |
2739 | ||
2740 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
2741 | /* |
2742 | * we may have advanced our local expiration to account for allowed | |
2743 | * spread between our sched_clock and the one on which runtime was | |
2744 | * issued. | |
2745 | */ | |
2746 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
2747 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
2748 | |
2749 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
2750 | } |
2751 | ||
a9cf55b2 PT |
2752 | /* |
2753 | * Note: This depends on the synchronization provided by sched_clock and the | |
2754 | * fact that rq->clock snapshots this value. | |
2755 | */ | |
2756 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 2757 | { |
a9cf55b2 | 2758 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
2759 | |
2760 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 2761 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
2762 | return; |
2763 | ||
a9cf55b2 PT |
2764 | if (cfs_rq->runtime_remaining < 0) |
2765 | return; | |
2766 | ||
2767 | /* | |
2768 | * If the local deadline has passed we have to consider the | |
2769 | * possibility that our sched_clock is 'fast' and the global deadline | |
2770 | * has not truly expired. | |
2771 | * | |
2772 | * Fortunately we can check determine whether this the case by checking | |
2773 | * whether the global deadline has advanced. | |
2774 | */ | |
2775 | ||
2776 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
2777 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
2778 | cfs_rq->runtime_expires += TICK_NSEC; | |
2779 | } else { | |
2780 | /* global deadline is ahead, expiration has passed */ | |
2781 | cfs_rq->runtime_remaining = 0; | |
2782 | } | |
2783 | } | |
2784 | ||
2785 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
2786 | unsigned long delta_exec) | |
2787 | { | |
2788 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 2789 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
2790 | expire_cfs_rq_runtime(cfs_rq); |
2791 | ||
2792 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
2793 | return; |
2794 | ||
85dac906 PT |
2795 | /* |
2796 | * if we're unable to extend our runtime we resched so that the active | |
2797 | * hierarchy can be throttled | |
2798 | */ | |
2799 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
2800 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
2801 | } |
2802 | ||
6c16a6dc PZ |
2803 | static __always_inline |
2804 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) | |
ec12cb7f | 2805 | { |
56f570e5 | 2806 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
2807 | return; |
2808 | ||
2809 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
2810 | } | |
2811 | ||
85dac906 PT |
2812 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
2813 | { | |
56f570e5 | 2814 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
2815 | } |
2816 | ||
64660c86 PT |
2817 | /* check whether cfs_rq, or any parent, is throttled */ |
2818 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
2819 | { | |
56f570e5 | 2820 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
2821 | } |
2822 | ||
2823 | /* | |
2824 | * Ensure that neither of the group entities corresponding to src_cpu or | |
2825 | * dest_cpu are members of a throttled hierarchy when performing group | |
2826 | * load-balance operations. | |
2827 | */ | |
2828 | static inline int throttled_lb_pair(struct task_group *tg, | |
2829 | int src_cpu, int dest_cpu) | |
2830 | { | |
2831 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
2832 | ||
2833 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
2834 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
2835 | ||
2836 | return throttled_hierarchy(src_cfs_rq) || | |
2837 | throttled_hierarchy(dest_cfs_rq); | |
2838 | } | |
2839 | ||
2840 | /* updated child weight may affect parent so we have to do this bottom up */ | |
2841 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
2842 | { | |
2843 | struct rq *rq = data; | |
2844 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2845 | ||
2846 | cfs_rq->throttle_count--; | |
2847 | #ifdef CONFIG_SMP | |
2848 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 2849 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 2850 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 2851 | cfs_rq->throttled_clock_task; |
64660c86 PT |
2852 | } |
2853 | #endif | |
2854 | ||
2855 | return 0; | |
2856 | } | |
2857 | ||
2858 | static int tg_throttle_down(struct task_group *tg, void *data) | |
2859 | { | |
2860 | struct rq *rq = data; | |
2861 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
2862 | ||
82958366 PT |
2863 | /* group is entering throttled state, stop time */ |
2864 | if (!cfs_rq->throttle_count) | |
78becc27 | 2865 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
2866 | cfs_rq->throttle_count++; |
2867 | ||
2868 | return 0; | |
2869 | } | |
2870 | ||
d3d9dc33 | 2871 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
2872 | { |
2873 | struct rq *rq = rq_of(cfs_rq); | |
2874 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2875 | struct sched_entity *se; | |
2876 | long task_delta, dequeue = 1; | |
2877 | ||
2878 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
2879 | ||
f1b17280 | 2880 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
2881 | rcu_read_lock(); |
2882 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
2883 | rcu_read_unlock(); | |
85dac906 PT |
2884 | |
2885 | task_delta = cfs_rq->h_nr_running; | |
2886 | for_each_sched_entity(se) { | |
2887 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
2888 | /* throttled entity or throttle-on-deactivate */ | |
2889 | if (!se->on_rq) | |
2890 | break; | |
2891 | ||
2892 | if (dequeue) | |
2893 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
2894 | qcfs_rq->h_nr_running -= task_delta; | |
2895 | ||
2896 | if (qcfs_rq->load.weight) | |
2897 | dequeue = 0; | |
2898 | } | |
2899 | ||
2900 | if (!se) | |
2901 | rq->nr_running -= task_delta; | |
2902 | ||
2903 | cfs_rq->throttled = 1; | |
78becc27 | 2904 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
2905 | raw_spin_lock(&cfs_b->lock); |
2906 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
2907 | raw_spin_unlock(&cfs_b->lock); | |
2908 | } | |
2909 | ||
029632fb | 2910 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
2911 | { |
2912 | struct rq *rq = rq_of(cfs_rq); | |
2913 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
2914 | struct sched_entity *se; | |
2915 | int enqueue = 1; | |
2916 | long task_delta; | |
2917 | ||
22b958d8 | 2918 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
2919 | |
2920 | cfs_rq->throttled = 0; | |
1a55af2e FW |
2921 | |
2922 | update_rq_clock(rq); | |
2923 | ||
671fd9da | 2924 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 2925 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
2926 | list_del_rcu(&cfs_rq->throttled_list); |
2927 | raw_spin_unlock(&cfs_b->lock); | |
2928 | ||
64660c86 PT |
2929 | /* update hierarchical throttle state */ |
2930 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
2931 | ||
671fd9da PT |
2932 | if (!cfs_rq->load.weight) |
2933 | return; | |
2934 | ||
2935 | task_delta = cfs_rq->h_nr_running; | |
2936 | for_each_sched_entity(se) { | |
2937 | if (se->on_rq) | |
2938 | enqueue = 0; | |
2939 | ||
2940 | cfs_rq = cfs_rq_of(se); | |
2941 | if (enqueue) | |
2942 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
2943 | cfs_rq->h_nr_running += task_delta; | |
2944 | ||
2945 | if (cfs_rq_throttled(cfs_rq)) | |
2946 | break; | |
2947 | } | |
2948 | ||
2949 | if (!se) | |
2950 | rq->nr_running += task_delta; | |
2951 | ||
2952 | /* determine whether we need to wake up potentially idle cpu */ | |
2953 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
2954 | resched_task(rq->curr); | |
2955 | } | |
2956 | ||
2957 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
2958 | u64 remaining, u64 expires) | |
2959 | { | |
2960 | struct cfs_rq *cfs_rq; | |
2961 | u64 runtime = remaining; | |
2962 | ||
2963 | rcu_read_lock(); | |
2964 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
2965 | throttled_list) { | |
2966 | struct rq *rq = rq_of(cfs_rq); | |
2967 | ||
2968 | raw_spin_lock(&rq->lock); | |
2969 | if (!cfs_rq_throttled(cfs_rq)) | |
2970 | goto next; | |
2971 | ||
2972 | runtime = -cfs_rq->runtime_remaining + 1; | |
2973 | if (runtime > remaining) | |
2974 | runtime = remaining; | |
2975 | remaining -= runtime; | |
2976 | ||
2977 | cfs_rq->runtime_remaining += runtime; | |
2978 | cfs_rq->runtime_expires = expires; | |
2979 | ||
2980 | /* we check whether we're throttled above */ | |
2981 | if (cfs_rq->runtime_remaining > 0) | |
2982 | unthrottle_cfs_rq(cfs_rq); | |
2983 | ||
2984 | next: | |
2985 | raw_spin_unlock(&rq->lock); | |
2986 | ||
2987 | if (!remaining) | |
2988 | break; | |
2989 | } | |
2990 | rcu_read_unlock(); | |
2991 | ||
2992 | return remaining; | |
2993 | } | |
2994 | ||
58088ad0 PT |
2995 | /* |
2996 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
2997 | * cfs_rqs as appropriate. If there has been no activity within the last | |
2998 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
2999 | * used to track this state. | |
3000 | */ | |
3001 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3002 | { | |
671fd9da PT |
3003 | u64 runtime, runtime_expires; |
3004 | int idle = 1, throttled; | |
58088ad0 PT |
3005 | |
3006 | raw_spin_lock(&cfs_b->lock); | |
3007 | /* no need to continue the timer with no bandwidth constraint */ | |
3008 | if (cfs_b->quota == RUNTIME_INF) | |
3009 | goto out_unlock; | |
3010 | ||
671fd9da PT |
3011 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3012 | /* idle depends on !throttled (for the case of a large deficit) */ | |
3013 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 3014 | cfs_b->nr_periods += overrun; |
671fd9da | 3015 | |
a9cf55b2 PT |
3016 | /* if we're going inactive then everything else can be deferred */ |
3017 | if (idle) | |
3018 | goto out_unlock; | |
3019 | ||
3020 | __refill_cfs_bandwidth_runtime(cfs_b); | |
3021 | ||
671fd9da PT |
3022 | if (!throttled) { |
3023 | /* mark as potentially idle for the upcoming period */ | |
3024 | cfs_b->idle = 1; | |
3025 | goto out_unlock; | |
3026 | } | |
3027 | ||
e8da1b18 NR |
3028 | /* account preceding periods in which throttling occurred */ |
3029 | cfs_b->nr_throttled += overrun; | |
3030 | ||
671fd9da PT |
3031 | /* |
3032 | * There are throttled entities so we must first use the new bandwidth | |
3033 | * to unthrottle them before making it generally available. This | |
3034 | * ensures that all existing debts will be paid before a new cfs_rq is | |
3035 | * allowed to run. | |
3036 | */ | |
3037 | runtime = cfs_b->runtime; | |
3038 | runtime_expires = cfs_b->runtime_expires; | |
3039 | cfs_b->runtime = 0; | |
3040 | ||
3041 | /* | |
3042 | * This check is repeated as we are holding onto the new bandwidth | |
3043 | * while we unthrottle. This can potentially race with an unthrottled | |
3044 | * group trying to acquire new bandwidth from the global pool. | |
3045 | */ | |
3046 | while (throttled && runtime > 0) { | |
3047 | raw_spin_unlock(&cfs_b->lock); | |
3048 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3049 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3050 | runtime_expires); | |
3051 | raw_spin_lock(&cfs_b->lock); | |
3052 | ||
3053 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
3054 | } | |
58088ad0 | 3055 | |
671fd9da PT |
3056 | /* return (any) remaining runtime */ |
3057 | cfs_b->runtime = runtime; | |
3058 | /* | |
3059 | * While we are ensured activity in the period following an | |
3060 | * unthrottle, this also covers the case in which the new bandwidth is | |
3061 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3062 | * timer to remain active while there are any throttled entities.) | |
3063 | */ | |
3064 | cfs_b->idle = 0; | |
58088ad0 PT |
3065 | out_unlock: |
3066 | if (idle) | |
3067 | cfs_b->timer_active = 0; | |
3068 | raw_spin_unlock(&cfs_b->lock); | |
3069 | ||
3070 | return idle; | |
3071 | } | |
d3d9dc33 | 3072 | |
d8b4986d PT |
3073 | /* a cfs_rq won't donate quota below this amount */ |
3074 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3075 | /* minimum remaining period time to redistribute slack quota */ | |
3076 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3077 | /* how long we wait to gather additional slack before distributing */ | |
3078 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3079 | ||
3080 | /* are we near the end of the current quota period? */ | |
3081 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | |
3082 | { | |
3083 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3084 | u64 remaining; | |
3085 | ||
3086 | /* if the call-back is running a quota refresh is already occurring */ | |
3087 | if (hrtimer_callback_running(refresh_timer)) | |
3088 | return 1; | |
3089 | ||
3090 | /* is a quota refresh about to occur? */ | |
3091 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3092 | if (remaining < min_expire) | |
3093 | return 1; | |
3094 | ||
3095 | return 0; | |
3096 | } | |
3097 | ||
3098 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3099 | { | |
3100 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3101 | ||
3102 | /* if there's a quota refresh soon don't bother with slack */ | |
3103 | if (runtime_refresh_within(cfs_b, min_left)) | |
3104 | return; | |
3105 | ||
3106 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3107 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3108 | } | |
3109 | ||
3110 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3111 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3112 | { | |
3113 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3114 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3115 | ||
3116 | if (slack_runtime <= 0) | |
3117 | return; | |
3118 | ||
3119 | raw_spin_lock(&cfs_b->lock); | |
3120 | if (cfs_b->quota != RUNTIME_INF && | |
3121 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3122 | cfs_b->runtime += slack_runtime; | |
3123 | ||
3124 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3125 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3126 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3127 | start_cfs_slack_bandwidth(cfs_b); | |
3128 | } | |
3129 | raw_spin_unlock(&cfs_b->lock); | |
3130 | ||
3131 | /* even if it's not valid for return we don't want to try again */ | |
3132 | cfs_rq->runtime_remaining -= slack_runtime; | |
3133 | } | |
3134 | ||
3135 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3136 | { | |
56f570e5 PT |
3137 | if (!cfs_bandwidth_used()) |
3138 | return; | |
3139 | ||
fccfdc6f | 3140 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3141 | return; |
3142 | ||
3143 | __return_cfs_rq_runtime(cfs_rq); | |
3144 | } | |
3145 | ||
3146 | /* | |
3147 | * This is done with a timer (instead of inline with bandwidth return) since | |
3148 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3149 | */ | |
3150 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3151 | { | |
3152 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3153 | u64 expires; | |
3154 | ||
3155 | /* confirm we're still not at a refresh boundary */ | |
3156 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) | |
3157 | return; | |
3158 | ||
3159 | raw_spin_lock(&cfs_b->lock); | |
3160 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { | |
3161 | runtime = cfs_b->runtime; | |
3162 | cfs_b->runtime = 0; | |
3163 | } | |
3164 | expires = cfs_b->runtime_expires; | |
3165 | raw_spin_unlock(&cfs_b->lock); | |
3166 | ||
3167 | if (!runtime) | |
3168 | return; | |
3169 | ||
3170 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3171 | ||
3172 | raw_spin_lock(&cfs_b->lock); | |
3173 | if (expires == cfs_b->runtime_expires) | |
3174 | cfs_b->runtime = runtime; | |
3175 | raw_spin_unlock(&cfs_b->lock); | |
3176 | } | |
3177 | ||
d3d9dc33 PT |
3178 | /* |
3179 | * When a group wakes up we want to make sure that its quota is not already | |
3180 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3181 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3182 | */ | |
3183 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3184 | { | |
56f570e5 PT |
3185 | if (!cfs_bandwidth_used()) |
3186 | return; | |
3187 | ||
d3d9dc33 PT |
3188 | /* an active group must be handled by the update_curr()->put() path */ |
3189 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3190 | return; | |
3191 | ||
3192 | /* ensure the group is not already throttled */ | |
3193 | if (cfs_rq_throttled(cfs_rq)) | |
3194 | return; | |
3195 | ||
3196 | /* update runtime allocation */ | |
3197 | account_cfs_rq_runtime(cfs_rq, 0); | |
3198 | if (cfs_rq->runtime_remaining <= 0) | |
3199 | throttle_cfs_rq(cfs_rq); | |
3200 | } | |
3201 | ||
3202 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
3203 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3204 | { | |
56f570e5 PT |
3205 | if (!cfs_bandwidth_used()) |
3206 | return; | |
3207 | ||
d3d9dc33 PT |
3208 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
3209 | return; | |
3210 | ||
3211 | /* | |
3212 | * it's possible for a throttled entity to be forced into a running | |
3213 | * state (e.g. set_curr_task), in this case we're finished. | |
3214 | */ | |
3215 | if (cfs_rq_throttled(cfs_rq)) | |
3216 | return; | |
3217 | ||
3218 | throttle_cfs_rq(cfs_rq); | |
3219 | } | |
029632fb | 3220 | |
029632fb PZ |
3221 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3222 | { | |
3223 | struct cfs_bandwidth *cfs_b = | |
3224 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3225 | do_sched_cfs_slack_timer(cfs_b); | |
3226 | ||
3227 | return HRTIMER_NORESTART; | |
3228 | } | |
3229 | ||
3230 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3231 | { | |
3232 | struct cfs_bandwidth *cfs_b = | |
3233 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3234 | ktime_t now; | |
3235 | int overrun; | |
3236 | int idle = 0; | |
3237 | ||
3238 | for (;;) { | |
3239 | now = hrtimer_cb_get_time(timer); | |
3240 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3241 | ||
3242 | if (!overrun) | |
3243 | break; | |
3244 | ||
3245 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3246 | } | |
3247 | ||
3248 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3249 | } | |
3250 | ||
3251 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3252 | { | |
3253 | raw_spin_lock_init(&cfs_b->lock); | |
3254 | cfs_b->runtime = 0; | |
3255 | cfs_b->quota = RUNTIME_INF; | |
3256 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3257 | ||
3258 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3259 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3260 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3261 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3262 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3263 | } | |
3264 | ||
3265 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3266 | { | |
3267 | cfs_rq->runtime_enabled = 0; | |
3268 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3269 | } | |
3270 | ||
3271 | /* requires cfs_b->lock, may release to reprogram timer */ | |
3272 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3273 | { | |
3274 | /* | |
3275 | * The timer may be active because we're trying to set a new bandwidth | |
3276 | * period or because we're racing with the tear-down path | |
3277 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3278 | * terminates). In either case we ensure that it's re-programmed | |
3279 | */ | |
3280 | while (unlikely(hrtimer_active(&cfs_b->period_timer))) { | |
3281 | raw_spin_unlock(&cfs_b->lock); | |
3282 | /* ensure cfs_b->lock is available while we wait */ | |
3283 | hrtimer_cancel(&cfs_b->period_timer); | |
3284 | ||
3285 | raw_spin_lock(&cfs_b->lock); | |
3286 | /* if someone else restarted the timer then we're done */ | |
3287 | if (cfs_b->timer_active) | |
3288 | return; | |
3289 | } | |
3290 | ||
3291 | cfs_b->timer_active = 1; | |
3292 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3293 | } | |
3294 | ||
3295 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3296 | { | |
3297 | hrtimer_cancel(&cfs_b->period_timer); | |
3298 | hrtimer_cancel(&cfs_b->slack_timer); | |
3299 | } | |
3300 | ||
38dc3348 | 3301 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3302 | { |
3303 | struct cfs_rq *cfs_rq; | |
3304 | ||
3305 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3306 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3307 | ||
3308 | if (!cfs_rq->runtime_enabled) | |
3309 | continue; | |
3310 | ||
3311 | /* | |
3312 | * clock_task is not advancing so we just need to make sure | |
3313 | * there's some valid quota amount | |
3314 | */ | |
3315 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3316 | if (cfs_rq_throttled(cfs_rq)) | |
3317 | unthrottle_cfs_rq(cfs_rq); | |
3318 | } | |
3319 | } | |
3320 | ||
3321 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3322 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3323 | { | |
78becc27 | 3324 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3325 | } |
3326 | ||
3327 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, | |
3328 | unsigned long delta_exec) {} | |
d3d9dc33 PT |
3329 | static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
3330 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | |
6c16a6dc | 3331 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3332 | |
3333 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3334 | { | |
3335 | return 0; | |
3336 | } | |
64660c86 PT |
3337 | |
3338 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3339 | { | |
3340 | return 0; | |
3341 | } | |
3342 | ||
3343 | static inline int throttled_lb_pair(struct task_group *tg, | |
3344 | int src_cpu, int dest_cpu) | |
3345 | { | |
3346 | return 0; | |
3347 | } | |
029632fb PZ |
3348 | |
3349 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3350 | ||
3351 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3352 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3353 | #endif |
3354 | ||
029632fb PZ |
3355 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3356 | { | |
3357 | return NULL; | |
3358 | } | |
3359 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3360 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3361 | |
3362 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3363 | ||
bf0f6f24 IM |
3364 | /************************************************** |
3365 | * CFS operations on tasks: | |
3366 | */ | |
3367 | ||
8f4d37ec PZ |
3368 | #ifdef CONFIG_SCHED_HRTICK |
3369 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3370 | { | |
8f4d37ec PZ |
3371 | struct sched_entity *se = &p->se; |
3372 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3373 | ||
3374 | WARN_ON(task_rq(p) != rq); | |
3375 | ||
b39e66ea | 3376 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3377 | u64 slice = sched_slice(cfs_rq, se); |
3378 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3379 | s64 delta = slice - ran; | |
3380 | ||
3381 | if (delta < 0) { | |
3382 | if (rq->curr == p) | |
3383 | resched_task(p); | |
3384 | return; | |
3385 | } | |
3386 | ||
3387 | /* | |
3388 | * Don't schedule slices shorter than 10000ns, that just | |
3389 | * doesn't make sense. Rely on vruntime for fairness. | |
3390 | */ | |
31656519 | 3391 | if (rq->curr != p) |
157124c1 | 3392 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3393 | |
31656519 | 3394 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3395 | } |
3396 | } | |
a4c2f00f PZ |
3397 | |
3398 | /* | |
3399 | * called from enqueue/dequeue and updates the hrtick when the | |
3400 | * current task is from our class and nr_running is low enough | |
3401 | * to matter. | |
3402 | */ | |
3403 | static void hrtick_update(struct rq *rq) | |
3404 | { | |
3405 | struct task_struct *curr = rq->curr; | |
3406 | ||
b39e66ea | 3407 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3408 | return; |
3409 | ||
3410 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3411 | hrtick_start_fair(rq, curr); | |
3412 | } | |
55e12e5e | 3413 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3414 | static inline void |
3415 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3416 | { | |
3417 | } | |
a4c2f00f PZ |
3418 | |
3419 | static inline void hrtick_update(struct rq *rq) | |
3420 | { | |
3421 | } | |
8f4d37ec PZ |
3422 | #endif |
3423 | ||
bf0f6f24 IM |
3424 | /* |
3425 | * The enqueue_task method is called before nr_running is | |
3426 | * increased. Here we update the fair scheduling stats and | |
3427 | * then put the task into the rbtree: | |
3428 | */ | |
ea87bb78 | 3429 | static void |
371fd7e7 | 3430 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3431 | { |
3432 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3433 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3434 | |
3435 | for_each_sched_entity(se) { | |
62fb1851 | 3436 | if (se->on_rq) |
bf0f6f24 IM |
3437 | break; |
3438 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3439 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3440 | |
3441 | /* | |
3442 | * end evaluation on encountering a throttled cfs_rq | |
3443 | * | |
3444 | * note: in the case of encountering a throttled cfs_rq we will | |
3445 | * post the final h_nr_running increment below. | |
3446 | */ | |
3447 | if (cfs_rq_throttled(cfs_rq)) | |
3448 | break; | |
953bfcd1 | 3449 | cfs_rq->h_nr_running++; |
85dac906 | 3450 | |
88ec22d3 | 3451 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3452 | } |
8f4d37ec | 3453 | |
2069dd75 | 3454 | for_each_sched_entity(se) { |
0f317143 | 3455 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3456 | cfs_rq->h_nr_running++; |
2069dd75 | 3457 | |
85dac906 PT |
3458 | if (cfs_rq_throttled(cfs_rq)) |
3459 | break; | |
3460 | ||
17bc14b7 | 3461 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3462 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3463 | } |
3464 | ||
18bf2805 BS |
3465 | if (!se) { |
3466 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3467 | inc_nr_running(rq); |
18bf2805 | 3468 | } |
a4c2f00f | 3469 | hrtick_update(rq); |
bf0f6f24 IM |
3470 | } |
3471 | ||
2f36825b VP |
3472 | static void set_next_buddy(struct sched_entity *se); |
3473 | ||
bf0f6f24 IM |
3474 | /* |
3475 | * The dequeue_task method is called before nr_running is | |
3476 | * decreased. We remove the task from the rbtree and | |
3477 | * update the fair scheduling stats: | |
3478 | */ | |
371fd7e7 | 3479 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3480 | { |
3481 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3482 | struct sched_entity *se = &p->se; |
2f36825b | 3483 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3484 | |
3485 | for_each_sched_entity(se) { | |
3486 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3487 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3488 | |
3489 | /* | |
3490 | * end evaluation on encountering a throttled cfs_rq | |
3491 | * | |
3492 | * note: in the case of encountering a throttled cfs_rq we will | |
3493 | * post the final h_nr_running decrement below. | |
3494 | */ | |
3495 | if (cfs_rq_throttled(cfs_rq)) | |
3496 | break; | |
953bfcd1 | 3497 | cfs_rq->h_nr_running--; |
2069dd75 | 3498 | |
bf0f6f24 | 3499 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3500 | if (cfs_rq->load.weight) { |
3501 | /* | |
3502 | * Bias pick_next to pick a task from this cfs_rq, as | |
3503 | * p is sleeping when it is within its sched_slice. | |
3504 | */ | |
3505 | if (task_sleep && parent_entity(se)) | |
3506 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3507 | |
3508 | /* avoid re-evaluating load for this entity */ | |
3509 | se = parent_entity(se); | |
bf0f6f24 | 3510 | break; |
2f36825b | 3511 | } |
371fd7e7 | 3512 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3513 | } |
8f4d37ec | 3514 | |
2069dd75 | 3515 | for_each_sched_entity(se) { |
0f317143 | 3516 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3517 | cfs_rq->h_nr_running--; |
2069dd75 | 3518 | |
85dac906 PT |
3519 | if (cfs_rq_throttled(cfs_rq)) |
3520 | break; | |
3521 | ||
17bc14b7 | 3522 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3523 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3524 | } |
3525 | ||
18bf2805 | 3526 | if (!se) { |
85dac906 | 3527 | dec_nr_running(rq); |
18bf2805 BS |
3528 | update_rq_runnable_avg(rq, 1); |
3529 | } | |
a4c2f00f | 3530 | hrtick_update(rq); |
bf0f6f24 IM |
3531 | } |
3532 | ||
e7693a36 | 3533 | #ifdef CONFIG_SMP |
029632fb PZ |
3534 | /* Used instead of source_load when we know the type == 0 */ |
3535 | static unsigned long weighted_cpuload(const int cpu) | |
3536 | { | |
b92486cb | 3537 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3538 | } |
3539 | ||
3540 | /* | |
3541 | * Return a low guess at the load of a migration-source cpu weighted | |
3542 | * according to the scheduling class and "nice" value. | |
3543 | * | |
3544 | * We want to under-estimate the load of migration sources, to | |
3545 | * balance conservatively. | |
3546 | */ | |
3547 | static unsigned long source_load(int cpu, int type) | |
3548 | { | |
3549 | struct rq *rq = cpu_rq(cpu); | |
3550 | unsigned long total = weighted_cpuload(cpu); | |
3551 | ||
3552 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3553 | return total; | |
3554 | ||
3555 | return min(rq->cpu_load[type-1], total); | |
3556 | } | |
3557 | ||
3558 | /* | |
3559 | * Return a high guess at the load of a migration-target cpu weighted | |
3560 | * according to the scheduling class and "nice" value. | |
3561 | */ | |
3562 | static unsigned long target_load(int cpu, int type) | |
3563 | { | |
3564 | struct rq *rq = cpu_rq(cpu); | |
3565 | unsigned long total = weighted_cpuload(cpu); | |
3566 | ||
3567 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3568 | return total; | |
3569 | ||
3570 | return max(rq->cpu_load[type-1], total); | |
3571 | } | |
3572 | ||
3573 | static unsigned long power_of(int cpu) | |
3574 | { | |
3575 | return cpu_rq(cpu)->cpu_power; | |
3576 | } | |
3577 | ||
3578 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3579 | { | |
3580 | struct rq *rq = cpu_rq(cpu); | |
3581 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3582 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3583 | |
3584 | if (nr_running) | |
b92486cb | 3585 | return load_avg / nr_running; |
029632fb PZ |
3586 | |
3587 | return 0; | |
3588 | } | |
3589 | ||
62470419 MW |
3590 | static void record_wakee(struct task_struct *p) |
3591 | { | |
3592 | /* | |
3593 | * Rough decay (wiping) for cost saving, don't worry | |
3594 | * about the boundary, really active task won't care | |
3595 | * about the loss. | |
3596 | */ | |
3597 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
3598 | current->wakee_flips = 0; | |
3599 | current->wakee_flip_decay_ts = jiffies; | |
3600 | } | |
3601 | ||
3602 | if (current->last_wakee != p) { | |
3603 | current->last_wakee = p; | |
3604 | current->wakee_flips++; | |
3605 | } | |
3606 | } | |
098fb9db | 3607 | |
74f8e4b2 | 3608 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
3609 | { |
3610 | struct sched_entity *se = &p->se; | |
3611 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
3612 | u64 min_vruntime; |
3613 | ||
3614 | #ifndef CONFIG_64BIT | |
3615 | u64 min_vruntime_copy; | |
88ec22d3 | 3616 | |
3fe1698b PZ |
3617 | do { |
3618 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
3619 | smp_rmb(); | |
3620 | min_vruntime = cfs_rq->min_vruntime; | |
3621 | } while (min_vruntime != min_vruntime_copy); | |
3622 | #else | |
3623 | min_vruntime = cfs_rq->min_vruntime; | |
3624 | #endif | |
88ec22d3 | 3625 | |
3fe1698b | 3626 | se->vruntime -= min_vruntime; |
62470419 | 3627 | record_wakee(p); |
88ec22d3 PZ |
3628 | } |
3629 | ||
bb3469ac | 3630 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
3631 | /* |
3632 | * effective_load() calculates the load change as seen from the root_task_group | |
3633 | * | |
3634 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
3635 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
3636 | * can calculate the shift in shares. | |
cf5f0acf PZ |
3637 | * |
3638 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
3639 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
3640 | * total group weight. | |
3641 | * | |
3642 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
3643 | * distribution (s_i) using: | |
3644 | * | |
3645 | * s_i = rw_i / \Sum rw_j (1) | |
3646 | * | |
3647 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
3648 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
3649 | * shares distribution (s_i): | |
3650 | * | |
3651 | * rw_i = { 2, 4, 1, 0 } | |
3652 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
3653 | * | |
3654 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
3655 | * task used to run on and the CPU the waker is running on), we need to | |
3656 | * compute the effect of waking a task on either CPU and, in case of a sync | |
3657 | * wakeup, compute the effect of the current task going to sleep. | |
3658 | * | |
3659 | * So for a change of @wl to the local @cpu with an overall group weight change | |
3660 | * of @wl we can compute the new shares distribution (s'_i) using: | |
3661 | * | |
3662 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
3663 | * | |
3664 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
3665 | * differences in waking a task to CPU 0. The additional task changes the | |
3666 | * weight and shares distributions like: | |
3667 | * | |
3668 | * rw'_i = { 3, 4, 1, 0 } | |
3669 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
3670 | * | |
3671 | * We can then compute the difference in effective weight by using: | |
3672 | * | |
3673 | * dw_i = S * (s'_i - s_i) (3) | |
3674 | * | |
3675 | * Where 'S' is the group weight as seen by its parent. | |
3676 | * | |
3677 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
3678 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
3679 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 3680 | */ |
2069dd75 | 3681 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 3682 | { |
4be9daaa | 3683 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 3684 | |
58d081b5 | 3685 | if (!tg->parent || !wl) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
3686 | return wl; |
3687 | ||
4be9daaa | 3688 | for_each_sched_entity(se) { |
cf5f0acf | 3689 | long w, W; |
4be9daaa | 3690 | |
977dda7c | 3691 | tg = se->my_q->tg; |
bb3469ac | 3692 | |
cf5f0acf PZ |
3693 | /* |
3694 | * W = @wg + \Sum rw_j | |
3695 | */ | |
3696 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 3697 | |
cf5f0acf PZ |
3698 | /* |
3699 | * w = rw_i + @wl | |
3700 | */ | |
3701 | w = se->my_q->load.weight + wl; | |
940959e9 | 3702 | |
cf5f0acf PZ |
3703 | /* |
3704 | * wl = S * s'_i; see (2) | |
3705 | */ | |
3706 | if (W > 0 && w < W) | |
3707 | wl = (w * tg->shares) / W; | |
977dda7c PT |
3708 | else |
3709 | wl = tg->shares; | |
940959e9 | 3710 | |
cf5f0acf PZ |
3711 | /* |
3712 | * Per the above, wl is the new se->load.weight value; since | |
3713 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
3714 | * calc_cfs_shares(). | |
3715 | */ | |
977dda7c PT |
3716 | if (wl < MIN_SHARES) |
3717 | wl = MIN_SHARES; | |
cf5f0acf PZ |
3718 | |
3719 | /* | |
3720 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
3721 | */ | |
977dda7c | 3722 | wl -= se->load.weight; |
cf5f0acf PZ |
3723 | |
3724 | /* | |
3725 | * Recursively apply this logic to all parent groups to compute | |
3726 | * the final effective load change on the root group. Since | |
3727 | * only the @tg group gets extra weight, all parent groups can | |
3728 | * only redistribute existing shares. @wl is the shift in shares | |
3729 | * resulting from this level per the above. | |
3730 | */ | |
4be9daaa | 3731 | wg = 0; |
4be9daaa | 3732 | } |
bb3469ac | 3733 | |
4be9daaa | 3734 | return wl; |
bb3469ac PZ |
3735 | } |
3736 | #else | |
4be9daaa | 3737 | |
58d081b5 | 3738 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 3739 | { |
83378269 | 3740 | return wl; |
bb3469ac | 3741 | } |
4be9daaa | 3742 | |
bb3469ac PZ |
3743 | #endif |
3744 | ||
62470419 MW |
3745 | static int wake_wide(struct task_struct *p) |
3746 | { | |
7d9ffa89 | 3747 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
3748 | |
3749 | /* | |
3750 | * Yeah, it's the switching-frequency, could means many wakee or | |
3751 | * rapidly switch, use factor here will just help to automatically | |
3752 | * adjust the loose-degree, so bigger node will lead to more pull. | |
3753 | */ | |
3754 | if (p->wakee_flips > factor) { | |
3755 | /* | |
3756 | * wakee is somewhat hot, it needs certain amount of cpu | |
3757 | * resource, so if waker is far more hot, prefer to leave | |
3758 | * it alone. | |
3759 | */ | |
3760 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
3761 | return 1; | |
3762 | } | |
3763 | ||
3764 | return 0; | |
3765 | } | |
3766 | ||
c88d5910 | 3767 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 3768 | { |
e37b6a7b | 3769 | s64 this_load, load; |
c88d5910 | 3770 | int idx, this_cpu, prev_cpu; |
098fb9db | 3771 | unsigned long tl_per_task; |
c88d5910 | 3772 | struct task_group *tg; |
83378269 | 3773 | unsigned long weight; |
b3137bc8 | 3774 | int balanced; |
098fb9db | 3775 | |
62470419 MW |
3776 | /* |
3777 | * If we wake multiple tasks be careful to not bounce | |
3778 | * ourselves around too much. | |
3779 | */ | |
3780 | if (wake_wide(p)) | |
3781 | return 0; | |
3782 | ||
c88d5910 PZ |
3783 | idx = sd->wake_idx; |
3784 | this_cpu = smp_processor_id(); | |
3785 | prev_cpu = task_cpu(p); | |
3786 | load = source_load(prev_cpu, idx); | |
3787 | this_load = target_load(this_cpu, idx); | |
098fb9db | 3788 | |
b3137bc8 MG |
3789 | /* |
3790 | * If sync wakeup then subtract the (maximum possible) | |
3791 | * effect of the currently running task from the load | |
3792 | * of the current CPU: | |
3793 | */ | |
83378269 PZ |
3794 | if (sync) { |
3795 | tg = task_group(current); | |
3796 | weight = current->se.load.weight; | |
3797 | ||
c88d5910 | 3798 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
3799 | load += effective_load(tg, prev_cpu, 0, -weight); |
3800 | } | |
b3137bc8 | 3801 | |
83378269 PZ |
3802 | tg = task_group(p); |
3803 | weight = p->se.load.weight; | |
b3137bc8 | 3804 | |
71a29aa7 PZ |
3805 | /* |
3806 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
3807 | * due to the sync cause above having dropped this_load to 0, we'll |
3808 | * always have an imbalance, but there's really nothing you can do | |
3809 | * about that, so that's good too. | |
71a29aa7 PZ |
3810 | * |
3811 | * Otherwise check if either cpus are near enough in load to allow this | |
3812 | * task to be woken on this_cpu. | |
3813 | */ | |
e37b6a7b PT |
3814 | if (this_load > 0) { |
3815 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
3816 | |
3817 | this_eff_load = 100; | |
3818 | this_eff_load *= power_of(prev_cpu); | |
3819 | this_eff_load *= this_load + | |
3820 | effective_load(tg, this_cpu, weight, weight); | |
3821 | ||
3822 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
3823 | prev_eff_load *= power_of(this_cpu); | |
3824 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
3825 | ||
3826 | balanced = this_eff_load <= prev_eff_load; | |
3827 | } else | |
3828 | balanced = true; | |
b3137bc8 | 3829 | |
098fb9db | 3830 | /* |
4ae7d5ce IM |
3831 | * If the currently running task will sleep within |
3832 | * a reasonable amount of time then attract this newly | |
3833 | * woken task: | |
098fb9db | 3834 | */ |
2fb7635c PZ |
3835 | if (sync && balanced) |
3836 | return 1; | |
098fb9db | 3837 | |
41acab88 | 3838 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
3839 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
3840 | ||
c88d5910 PZ |
3841 | if (balanced || |
3842 | (this_load <= load && | |
3843 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
3844 | /* |
3845 | * This domain has SD_WAKE_AFFINE and | |
3846 | * p is cache cold in this domain, and | |
3847 | * there is no bad imbalance. | |
3848 | */ | |
c88d5910 | 3849 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 3850 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
3851 | |
3852 | return 1; | |
3853 | } | |
3854 | return 0; | |
3855 | } | |
3856 | ||
aaee1203 PZ |
3857 | /* |
3858 | * find_idlest_group finds and returns the least busy CPU group within the | |
3859 | * domain. | |
3860 | */ | |
3861 | static struct sched_group * | |
78e7ed53 | 3862 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
5158f4e4 | 3863 | int this_cpu, int load_idx) |
e7693a36 | 3864 | { |
b3bd3de6 | 3865 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 3866 | unsigned long min_load = ULONG_MAX, this_load = 0; |
aaee1203 | 3867 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 3868 | |
aaee1203 PZ |
3869 | do { |
3870 | unsigned long load, avg_load; | |
3871 | int local_group; | |
3872 | int i; | |
e7693a36 | 3873 | |
aaee1203 PZ |
3874 | /* Skip over this group if it has no CPUs allowed */ |
3875 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 3876 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
3877 | continue; |
3878 | ||
3879 | local_group = cpumask_test_cpu(this_cpu, | |
3880 | sched_group_cpus(group)); | |
3881 | ||
3882 | /* Tally up the load of all CPUs in the group */ | |
3883 | avg_load = 0; | |
3884 | ||
3885 | for_each_cpu(i, sched_group_cpus(group)) { | |
3886 | /* Bias balancing toward cpus of our domain */ | |
3887 | if (local_group) | |
3888 | load = source_load(i, load_idx); | |
3889 | else | |
3890 | load = target_load(i, load_idx); | |
3891 | ||
3892 | avg_load += load; | |
3893 | } | |
3894 | ||
3895 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 3896 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
3897 | |
3898 | if (local_group) { | |
3899 | this_load = avg_load; | |
aaee1203 PZ |
3900 | } else if (avg_load < min_load) { |
3901 | min_load = avg_load; | |
3902 | idlest = group; | |
3903 | } | |
3904 | } while (group = group->next, group != sd->groups); | |
3905 | ||
3906 | if (!idlest || 100*this_load < imbalance*min_load) | |
3907 | return NULL; | |
3908 | return idlest; | |
3909 | } | |
3910 | ||
3911 | /* | |
3912 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
3913 | */ | |
3914 | static int | |
3915 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
3916 | { | |
3917 | unsigned long load, min_load = ULONG_MAX; | |
3918 | int idlest = -1; | |
3919 | int i; | |
3920 | ||
3921 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 3922 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
3923 | load = weighted_cpuload(i); |
3924 | ||
3925 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
3926 | min_load = load; | |
3927 | idlest = i; | |
e7693a36 GH |
3928 | } |
3929 | } | |
3930 | ||
aaee1203 PZ |
3931 | return idlest; |
3932 | } | |
e7693a36 | 3933 | |
a50bde51 PZ |
3934 | /* |
3935 | * Try and locate an idle CPU in the sched_domain. | |
3936 | */ | |
99bd5e2f | 3937 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 3938 | { |
99bd5e2f | 3939 | struct sched_domain *sd; |
37407ea7 | 3940 | struct sched_group *sg; |
e0a79f52 | 3941 | int i = task_cpu(p); |
a50bde51 | 3942 | |
e0a79f52 MG |
3943 | if (idle_cpu(target)) |
3944 | return target; | |
99bd5e2f SS |
3945 | |
3946 | /* | |
e0a79f52 | 3947 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 3948 | */ |
e0a79f52 MG |
3949 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
3950 | return i; | |
a50bde51 PZ |
3951 | |
3952 | /* | |
37407ea7 | 3953 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 3954 | */ |
518cd623 | 3955 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 3956 | for_each_lower_domain(sd) { |
37407ea7 LT |
3957 | sg = sd->groups; |
3958 | do { | |
3959 | if (!cpumask_intersects(sched_group_cpus(sg), | |
3960 | tsk_cpus_allowed(p))) | |
3961 | goto next; | |
3962 | ||
3963 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 3964 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
3965 | goto next; |
3966 | } | |
970e1789 | 3967 | |
37407ea7 LT |
3968 | target = cpumask_first_and(sched_group_cpus(sg), |
3969 | tsk_cpus_allowed(p)); | |
3970 | goto done; | |
3971 | next: | |
3972 | sg = sg->next; | |
3973 | } while (sg != sd->groups); | |
3974 | } | |
3975 | done: | |
a50bde51 PZ |
3976 | return target; |
3977 | } | |
3978 | ||
aaee1203 PZ |
3979 | /* |
3980 | * sched_balance_self: balance the current task (running on cpu) in domains | |
3981 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
3982 | * SD_BALANCE_EXEC. | |
3983 | * | |
3984 | * Balance, ie. select the least loaded group. | |
3985 | * | |
3986 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
3987 | * | |
3988 | * preempt must be disabled. | |
3989 | */ | |
0017d735 | 3990 | static int |
ac66f547 | 3991 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 3992 | { |
29cd8bae | 3993 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 3994 | int cpu = smp_processor_id(); |
c88d5910 | 3995 | int new_cpu = cpu; |
99bd5e2f | 3996 | int want_affine = 0; |
5158f4e4 | 3997 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 3998 | |
29baa747 | 3999 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
4000 | return prev_cpu; |
4001 | ||
0763a660 | 4002 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 4003 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
4004 | want_affine = 1; |
4005 | new_cpu = prev_cpu; | |
4006 | } | |
aaee1203 | 4007 | |
dce840a0 | 4008 | rcu_read_lock(); |
aaee1203 | 4009 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4010 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4011 | continue; | |
4012 | ||
fe3bcfe1 | 4013 | /* |
99bd5e2f SS |
4014 | * If both cpu and prev_cpu are part of this domain, |
4015 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4016 | */ |
99bd5e2f SS |
4017 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4018 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4019 | affine_sd = tmp; | |
29cd8bae | 4020 | break; |
f03542a7 | 4021 | } |
29cd8bae | 4022 | |
f03542a7 | 4023 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4024 | sd = tmp; |
4025 | } | |
4026 | ||
8b911acd | 4027 | if (affine_sd) { |
f03542a7 | 4028 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
4029 | prev_cpu = cpu; |
4030 | ||
4031 | new_cpu = select_idle_sibling(p, prev_cpu); | |
4032 | goto unlock; | |
8b911acd | 4033 | } |
e7693a36 | 4034 | |
aaee1203 | 4035 | while (sd) { |
5158f4e4 | 4036 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4037 | struct sched_group *group; |
c88d5910 | 4038 | int weight; |
098fb9db | 4039 | |
0763a660 | 4040 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4041 | sd = sd->child; |
4042 | continue; | |
4043 | } | |
098fb9db | 4044 | |
5158f4e4 PZ |
4045 | if (sd_flag & SD_BALANCE_WAKE) |
4046 | load_idx = sd->wake_idx; | |
098fb9db | 4047 | |
5158f4e4 | 4048 | group = find_idlest_group(sd, p, cpu, load_idx); |
aaee1203 PZ |
4049 | if (!group) { |
4050 | sd = sd->child; | |
4051 | continue; | |
4052 | } | |
4ae7d5ce | 4053 | |
d7c33c49 | 4054 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4055 | if (new_cpu == -1 || new_cpu == cpu) { |
4056 | /* Now try balancing at a lower domain level of cpu */ | |
4057 | sd = sd->child; | |
4058 | continue; | |
e7693a36 | 4059 | } |
aaee1203 PZ |
4060 | |
4061 | /* Now try balancing at a lower domain level of new_cpu */ | |
4062 | cpu = new_cpu; | |
669c55e9 | 4063 | weight = sd->span_weight; |
aaee1203 PZ |
4064 | sd = NULL; |
4065 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4066 | if (weight <= tmp->span_weight) |
aaee1203 | 4067 | break; |
0763a660 | 4068 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4069 | sd = tmp; |
4070 | } | |
4071 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4072 | } |
dce840a0 PZ |
4073 | unlock: |
4074 | rcu_read_unlock(); | |
e7693a36 | 4075 | |
c88d5910 | 4076 | return new_cpu; |
e7693a36 | 4077 | } |
0a74bef8 PT |
4078 | |
4079 | /* | |
4080 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4081 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4082 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4083 | * other assumptions, including the state of rq->lock, should be made. | |
4084 | */ | |
4085 | static void | |
4086 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4087 | { | |
aff3e498 PT |
4088 | struct sched_entity *se = &p->se; |
4089 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4090 | ||
4091 | /* | |
4092 | * Load tracking: accumulate removed load so that it can be processed | |
4093 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4094 | * to blocked load iff they have a positive decay-count. It can never | |
4095 | * be negative here since on-rq tasks have decay-count == 0. | |
4096 | */ | |
4097 | if (se->avg.decay_count) { | |
4098 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4099 | atomic_long_add(se->avg.load_avg_contrib, |
4100 | &cfs_rq->removed_load); | |
aff3e498 | 4101 | } |
0a74bef8 | 4102 | } |
e7693a36 GH |
4103 | #endif /* CONFIG_SMP */ |
4104 | ||
e52fb7c0 PZ |
4105 | static unsigned long |
4106 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4107 | { |
4108 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4109 | ||
4110 | /* | |
e52fb7c0 PZ |
4111 | * Since its curr running now, convert the gran from real-time |
4112 | * to virtual-time in his units. | |
13814d42 MG |
4113 | * |
4114 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4115 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4116 | * the resulting gran will be larger, therefore penalizing the | |
4117 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4118 | * be smaller, again penalizing the lighter task. | |
4119 | * | |
4120 | * This is especially important for buddies when the leftmost | |
4121 | * task is higher priority than the buddy. | |
0bbd3336 | 4122 | */ |
f4ad9bd2 | 4123 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4124 | } |
4125 | ||
464b7527 PZ |
4126 | /* |
4127 | * Should 'se' preempt 'curr'. | |
4128 | * | |
4129 | * |s1 | |
4130 | * |s2 | |
4131 | * |s3 | |
4132 | * g | |
4133 | * |<--->|c | |
4134 | * | |
4135 | * w(c, s1) = -1 | |
4136 | * w(c, s2) = 0 | |
4137 | * w(c, s3) = 1 | |
4138 | * | |
4139 | */ | |
4140 | static int | |
4141 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4142 | { | |
4143 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4144 | ||
4145 | if (vdiff <= 0) | |
4146 | return -1; | |
4147 | ||
e52fb7c0 | 4148 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4149 | if (vdiff > gran) |
4150 | return 1; | |
4151 | ||
4152 | return 0; | |
4153 | } | |
4154 | ||
02479099 PZ |
4155 | static void set_last_buddy(struct sched_entity *se) |
4156 | { | |
69c80f3e VP |
4157 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4158 | return; | |
4159 | ||
4160 | for_each_sched_entity(se) | |
4161 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4162 | } |
4163 | ||
4164 | static void set_next_buddy(struct sched_entity *se) | |
4165 | { | |
69c80f3e VP |
4166 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4167 | return; | |
4168 | ||
4169 | for_each_sched_entity(se) | |
4170 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4171 | } |
4172 | ||
ac53db59 RR |
4173 | static void set_skip_buddy(struct sched_entity *se) |
4174 | { | |
69c80f3e VP |
4175 | for_each_sched_entity(se) |
4176 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4177 | } |
4178 | ||
bf0f6f24 IM |
4179 | /* |
4180 | * Preempt the current task with a newly woken task if needed: | |
4181 | */ | |
5a9b86f6 | 4182 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4183 | { |
4184 | struct task_struct *curr = rq->curr; | |
8651a86c | 4185 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4186 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4187 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4188 | int next_buddy_marked = 0; |
bf0f6f24 | 4189 | |
4ae7d5ce IM |
4190 | if (unlikely(se == pse)) |
4191 | return; | |
4192 | ||
5238cdd3 | 4193 | /* |
ddcdf6e7 | 4194 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
4195 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4196 | * lead to a throttle). This both saves work and prevents false | |
4197 | * next-buddy nomination below. | |
4198 | */ | |
4199 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4200 | return; | |
4201 | ||
2f36825b | 4202 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4203 | set_next_buddy(pse); |
2f36825b VP |
4204 | next_buddy_marked = 1; |
4205 | } | |
57fdc26d | 4206 | |
aec0a514 BR |
4207 | /* |
4208 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4209 | * wake up path. | |
5238cdd3 PT |
4210 | * |
4211 | * Note: this also catches the edge-case of curr being in a throttled | |
4212 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4213 | * enqueue of curr) will have resulted in resched being set. This | |
4214 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4215 | * below. | |
aec0a514 BR |
4216 | */ |
4217 | if (test_tsk_need_resched(curr)) | |
4218 | return; | |
4219 | ||
a2f5c9ab DH |
4220 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4221 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4222 | likely(p->policy != SCHED_IDLE)) | |
4223 | goto preempt; | |
4224 | ||
91c234b4 | 4225 | /* |
a2f5c9ab DH |
4226 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4227 | * is driven by the tick): | |
91c234b4 | 4228 | */ |
8ed92e51 | 4229 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4230 | return; |
bf0f6f24 | 4231 | |
464b7527 | 4232 | find_matching_se(&se, &pse); |
9bbd7374 | 4233 | update_curr(cfs_rq_of(se)); |
002f128b | 4234 | BUG_ON(!pse); |
2f36825b VP |
4235 | if (wakeup_preempt_entity(se, pse) == 1) { |
4236 | /* | |
4237 | * Bias pick_next to pick the sched entity that is | |
4238 | * triggering this preemption. | |
4239 | */ | |
4240 | if (!next_buddy_marked) | |
4241 | set_next_buddy(pse); | |
3a7e73a2 | 4242 | goto preempt; |
2f36825b | 4243 | } |
464b7527 | 4244 | |
3a7e73a2 | 4245 | return; |
a65ac745 | 4246 | |
3a7e73a2 PZ |
4247 | preempt: |
4248 | resched_task(curr); | |
4249 | /* | |
4250 | * Only set the backward buddy when the current task is still | |
4251 | * on the rq. This can happen when a wakeup gets interleaved | |
4252 | * with schedule on the ->pre_schedule() or idle_balance() | |
4253 | * point, either of which can * drop the rq lock. | |
4254 | * | |
4255 | * Also, during early boot the idle thread is in the fair class, | |
4256 | * for obvious reasons its a bad idea to schedule back to it. | |
4257 | */ | |
4258 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
4259 | return; | |
4260 | ||
4261 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
4262 | set_last_buddy(se); | |
bf0f6f24 IM |
4263 | } |
4264 | ||
fb8d4724 | 4265 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
bf0f6f24 | 4266 | { |
8f4d37ec | 4267 | struct task_struct *p; |
bf0f6f24 IM |
4268 | struct cfs_rq *cfs_rq = &rq->cfs; |
4269 | struct sched_entity *se; | |
4270 | ||
36ace27e | 4271 | if (!cfs_rq->nr_running) |
bf0f6f24 IM |
4272 | return NULL; |
4273 | ||
4274 | do { | |
9948f4b2 | 4275 | se = pick_next_entity(cfs_rq); |
f4b6755f | 4276 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
4277 | cfs_rq = group_cfs_rq(se); |
4278 | } while (cfs_rq); | |
4279 | ||
8f4d37ec | 4280 | p = task_of(se); |
b39e66ea MG |
4281 | if (hrtick_enabled(rq)) |
4282 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4283 | |
4284 | return p; | |
bf0f6f24 IM |
4285 | } |
4286 | ||
4287 | /* | |
4288 | * Account for a descheduled task: | |
4289 | */ | |
31ee529c | 4290 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4291 | { |
4292 | struct sched_entity *se = &prev->se; | |
4293 | struct cfs_rq *cfs_rq; | |
4294 | ||
4295 | for_each_sched_entity(se) { | |
4296 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4297 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4298 | } |
4299 | } | |
4300 | ||
ac53db59 RR |
4301 | /* |
4302 | * sched_yield() is very simple | |
4303 | * | |
4304 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4305 | */ | |
4306 | static void yield_task_fair(struct rq *rq) | |
4307 | { | |
4308 | struct task_struct *curr = rq->curr; | |
4309 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4310 | struct sched_entity *se = &curr->se; | |
4311 | ||
4312 | /* | |
4313 | * Are we the only task in the tree? | |
4314 | */ | |
4315 | if (unlikely(rq->nr_running == 1)) | |
4316 | return; | |
4317 | ||
4318 | clear_buddies(cfs_rq, se); | |
4319 | ||
4320 | if (curr->policy != SCHED_BATCH) { | |
4321 | update_rq_clock(rq); | |
4322 | /* | |
4323 | * Update run-time statistics of the 'current'. | |
4324 | */ | |
4325 | update_curr(cfs_rq); | |
916671c0 MG |
4326 | /* |
4327 | * Tell update_rq_clock() that we've just updated, | |
4328 | * so we don't do microscopic update in schedule() | |
4329 | * and double the fastpath cost. | |
4330 | */ | |
4331 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4332 | } |
4333 | ||
4334 | set_skip_buddy(se); | |
4335 | } | |
4336 | ||
d95f4122 MG |
4337 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4338 | { | |
4339 | struct sched_entity *se = &p->se; | |
4340 | ||
5238cdd3 PT |
4341 | /* throttled hierarchies are not runnable */ |
4342 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4343 | return false; |
4344 | ||
4345 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4346 | set_next_buddy(se); | |
4347 | ||
d95f4122 MG |
4348 | yield_task_fair(rq); |
4349 | ||
4350 | return true; | |
4351 | } | |
4352 | ||
681f3e68 | 4353 | #ifdef CONFIG_SMP |
bf0f6f24 | 4354 | /************************************************** |
e9c84cb8 PZ |
4355 | * Fair scheduling class load-balancing methods. |
4356 | * | |
4357 | * BASICS | |
4358 | * | |
4359 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4360 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4361 | * time to each task. This is expressed in the following equation: | |
4362 | * | |
4363 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4364 | * | |
4365 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4366 | * W_i,0 is defined as: | |
4367 | * | |
4368 | * W_i,0 = \Sum_j w_i,j (2) | |
4369 | * | |
4370 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4371 | * is derived from the nice value as per prio_to_weight[]. | |
4372 | * | |
4373 | * The weight average is an exponential decay average of the instantaneous | |
4374 | * weight: | |
4375 | * | |
4376 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4377 | * | |
4378 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4379 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4380 | * can also include other factors [XXX]. | |
4381 | * | |
4382 | * To achieve this balance we define a measure of imbalance which follows | |
4383 | * directly from (1): | |
4384 | * | |
4385 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4386 | * | |
4387 | * We them move tasks around to minimize the imbalance. In the continuous | |
4388 | * function space it is obvious this converges, in the discrete case we get | |
4389 | * a few fun cases generally called infeasible weight scenarios. | |
4390 | * | |
4391 | * [XXX expand on: | |
4392 | * - infeasible weights; | |
4393 | * - local vs global optima in the discrete case. ] | |
4394 | * | |
4395 | * | |
4396 | * SCHED DOMAINS | |
4397 | * | |
4398 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4399 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4400 | * topology where each level pairs two lower groups (or better). This results | |
4401 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4402 | * tree to only the first of the previous level and we decrease the frequency | |
4403 | * of load-balance at each level inv. proportional to the number of cpus in | |
4404 | * the groups. | |
4405 | * | |
4406 | * This yields: | |
4407 | * | |
4408 | * log_2 n 1 n | |
4409 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4410 | * i = 0 2^i 2^i | |
4411 | * `- size of each group | |
4412 | * | | `- number of cpus doing load-balance | |
4413 | * | `- freq | |
4414 | * `- sum over all levels | |
4415 | * | |
4416 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4417 | * this makes (5) the runtime complexity of the balancer. | |
4418 | * | |
4419 | * An important property here is that each CPU is still (indirectly) connected | |
4420 | * to every other cpu in at most O(log n) steps: | |
4421 | * | |
4422 | * The adjacency matrix of the resulting graph is given by: | |
4423 | * | |
4424 | * log_2 n | |
4425 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4426 | * k = 0 | |
4427 | * | |
4428 | * And you'll find that: | |
4429 | * | |
4430 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
4431 | * | |
4432 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
4433 | * The task movement gives a factor of O(m), giving a convergence complexity | |
4434 | * of: | |
4435 | * | |
4436 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
4437 | * | |
4438 | * | |
4439 | * WORK CONSERVING | |
4440 | * | |
4441 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
4442 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
4443 | * tree itself instead of relying on other CPUs to bring it work. | |
4444 | * | |
4445 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
4446 | * time. | |
4447 | * | |
4448 | * [XXX more?] | |
4449 | * | |
4450 | * | |
4451 | * CGROUPS | |
4452 | * | |
4453 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
4454 | * | |
4455 | * s_k,i | |
4456 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
4457 | * S_k | |
4458 | * | |
4459 | * Where | |
4460 | * | |
4461 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4462 | * | |
4463 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4464 | * | |
4465 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4466 | * property. | |
4467 | * | |
4468 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4469 | * rewrite all of this once again.] | |
4470 | */ | |
bf0f6f24 | 4471 | |
ed387b78 HS |
4472 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4473 | ||
ddcdf6e7 | 4474 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4475 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4476 | #define LBF_DST_PINNED 0x04 |
4477 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4478 | |
4479 | struct lb_env { | |
4480 | struct sched_domain *sd; | |
4481 | ||
ddcdf6e7 | 4482 | struct rq *src_rq; |
85c1e7da | 4483 | int src_cpu; |
ddcdf6e7 PZ |
4484 | |
4485 | int dst_cpu; | |
4486 | struct rq *dst_rq; | |
4487 | ||
88b8dac0 SV |
4488 | struct cpumask *dst_grpmask; |
4489 | int new_dst_cpu; | |
ddcdf6e7 | 4490 | enum cpu_idle_type idle; |
bd939f45 | 4491 | long imbalance; |
b9403130 MW |
4492 | /* The set of CPUs under consideration for load-balancing */ |
4493 | struct cpumask *cpus; | |
4494 | ||
ddcdf6e7 | 4495 | unsigned int flags; |
367456c7 PZ |
4496 | |
4497 | unsigned int loop; | |
4498 | unsigned int loop_break; | |
4499 | unsigned int loop_max; | |
ddcdf6e7 PZ |
4500 | }; |
4501 | ||
1e3c88bd | 4502 | /* |
ddcdf6e7 | 4503 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
4504 | * Both runqueues must be locked. |
4505 | */ | |
ddcdf6e7 | 4506 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 4507 | { |
ddcdf6e7 PZ |
4508 | deactivate_task(env->src_rq, p, 0); |
4509 | set_task_cpu(p, env->dst_cpu); | |
4510 | activate_task(env->dst_rq, p, 0); | |
4511 | check_preempt_curr(env->dst_rq, p, 0); | |
6fe6b2d6 RR |
4512 | #ifdef CONFIG_NUMA_BALANCING |
4513 | if (p->numa_preferred_nid != -1) { | |
4514 | int src_nid = cpu_to_node(env->src_cpu); | |
4515 | int dst_nid = cpu_to_node(env->dst_cpu); | |
4516 | ||
4517 | /* | |
4518 | * If the load balancer has moved the task then limit | |
4519 | * migrations from taking place in the short term in | |
4520 | * case this is a short-lived migration. | |
4521 | */ | |
4522 | if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid) | |
4523 | p->numa_migrate_seq = 0; | |
4524 | } | |
4525 | #endif | |
1e3c88bd PZ |
4526 | } |
4527 | ||
029632fb PZ |
4528 | /* |
4529 | * Is this task likely cache-hot: | |
4530 | */ | |
4531 | static int | |
4532 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
4533 | { | |
4534 | s64 delta; | |
4535 | ||
4536 | if (p->sched_class != &fair_sched_class) | |
4537 | return 0; | |
4538 | ||
4539 | if (unlikely(p->policy == SCHED_IDLE)) | |
4540 | return 0; | |
4541 | ||
4542 | /* | |
4543 | * Buddy candidates are cache hot: | |
4544 | */ | |
4545 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
4546 | (&p->se == cfs_rq_of(&p->se)->next || | |
4547 | &p->se == cfs_rq_of(&p->se)->last)) | |
4548 | return 1; | |
4549 | ||
4550 | if (sysctl_sched_migration_cost == -1) | |
4551 | return 1; | |
4552 | if (sysctl_sched_migration_cost == 0) | |
4553 | return 0; | |
4554 | ||
4555 | delta = now - p->se.exec_start; | |
4556 | ||
4557 | return delta < (s64)sysctl_sched_migration_cost; | |
4558 | } | |
4559 | ||
3a7053b3 MG |
4560 | #ifdef CONFIG_NUMA_BALANCING |
4561 | /* Returns true if the destination node has incurred more faults */ | |
4562 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
4563 | { | |
4564 | int src_nid, dst_nid; | |
4565 | ||
4566 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || | |
4567 | !(env->sd->flags & SD_NUMA)) { | |
4568 | return false; | |
4569 | } | |
4570 | ||
4571 | src_nid = cpu_to_node(env->src_cpu); | |
4572 | dst_nid = cpu_to_node(env->dst_cpu); | |
4573 | ||
4574 | if (src_nid == dst_nid || | |
4575 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4576 | return false; | |
4577 | ||
4578 | if (dst_nid == p->numa_preferred_nid || | |
ac8e895b | 4579 | task_faults(p, dst_nid) > task_faults(p, src_nid)) |
3a7053b3 MG |
4580 | return true; |
4581 | ||
4582 | return false; | |
4583 | } | |
7a0f3083 MG |
4584 | |
4585 | ||
4586 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
4587 | { | |
4588 | int src_nid, dst_nid; | |
4589 | ||
4590 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
4591 | return false; | |
4592 | ||
4593 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) | |
4594 | return false; | |
4595 | ||
4596 | src_nid = cpu_to_node(env->src_cpu); | |
4597 | dst_nid = cpu_to_node(env->dst_cpu); | |
4598 | ||
4599 | if (src_nid == dst_nid || | |
4600 | p->numa_migrate_seq >= sysctl_numa_balancing_settle_count) | |
4601 | return false; | |
4602 | ||
ac8e895b | 4603 | if (task_faults(p, dst_nid) < task_faults(p, src_nid)) |
7a0f3083 MG |
4604 | return true; |
4605 | ||
4606 | return false; | |
4607 | } | |
4608 | ||
3a7053b3 MG |
4609 | #else |
4610 | static inline bool migrate_improves_locality(struct task_struct *p, | |
4611 | struct lb_env *env) | |
4612 | { | |
4613 | return false; | |
4614 | } | |
7a0f3083 MG |
4615 | |
4616 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
4617 | struct lb_env *env) | |
4618 | { | |
4619 | return false; | |
4620 | } | |
3a7053b3 MG |
4621 | #endif |
4622 | ||
1e3c88bd PZ |
4623 | /* |
4624 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
4625 | */ | |
4626 | static | |
8e45cb54 | 4627 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
4628 | { |
4629 | int tsk_cache_hot = 0; | |
4630 | /* | |
4631 | * We do not migrate tasks that are: | |
d3198084 | 4632 | * 1) throttled_lb_pair, or |
1e3c88bd | 4633 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
4634 | * 3) running (obviously), or |
4635 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 4636 | */ |
d3198084 JK |
4637 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
4638 | return 0; | |
4639 | ||
ddcdf6e7 | 4640 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 4641 | int cpu; |
88b8dac0 | 4642 | |
41acab88 | 4643 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 4644 | |
6263322c PZ |
4645 | env->flags |= LBF_SOME_PINNED; |
4646 | ||
88b8dac0 SV |
4647 | /* |
4648 | * Remember if this task can be migrated to any other cpu in | |
4649 | * our sched_group. We may want to revisit it if we couldn't | |
4650 | * meet load balance goals by pulling other tasks on src_cpu. | |
4651 | * | |
4652 | * Also avoid computing new_dst_cpu if we have already computed | |
4653 | * one in current iteration. | |
4654 | */ | |
6263322c | 4655 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
4656 | return 0; |
4657 | ||
e02e60c1 JK |
4658 | /* Prevent to re-select dst_cpu via env's cpus */ |
4659 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
4660 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 4661 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
4662 | env->new_dst_cpu = cpu; |
4663 | break; | |
4664 | } | |
88b8dac0 | 4665 | } |
e02e60c1 | 4666 | |
1e3c88bd PZ |
4667 | return 0; |
4668 | } | |
88b8dac0 SV |
4669 | |
4670 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 4671 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 4672 | |
ddcdf6e7 | 4673 | if (task_running(env->src_rq, p)) { |
41acab88 | 4674 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
4675 | return 0; |
4676 | } | |
4677 | ||
4678 | /* | |
4679 | * Aggressive migration if: | |
3a7053b3 MG |
4680 | * 1) destination numa is preferred |
4681 | * 2) task is cache cold, or | |
4682 | * 3) too many balance attempts have failed. | |
1e3c88bd | 4683 | */ |
78becc27 | 4684 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
7a0f3083 MG |
4685 | if (!tsk_cache_hot) |
4686 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
4687 | |
4688 | if (migrate_improves_locality(p, env)) { | |
4689 | #ifdef CONFIG_SCHEDSTATS | |
4690 | if (tsk_cache_hot) { | |
4691 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
4692 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
4693 | } | |
4694 | #endif | |
4695 | return 1; | |
4696 | } | |
4697 | ||
1e3c88bd | 4698 | if (!tsk_cache_hot || |
8e45cb54 | 4699 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 4700 | |
1e3c88bd | 4701 | if (tsk_cache_hot) { |
8e45cb54 | 4702 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 4703 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 4704 | } |
4e2dcb73 | 4705 | |
1e3c88bd PZ |
4706 | return 1; |
4707 | } | |
4708 | ||
4e2dcb73 ZH |
4709 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
4710 | return 0; | |
1e3c88bd PZ |
4711 | } |
4712 | ||
897c395f PZ |
4713 | /* |
4714 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
4715 | * part of active balancing operations within "domain". | |
4716 | * Returns 1 if successful and 0 otherwise. | |
4717 | * | |
4718 | * Called with both runqueues locked. | |
4719 | */ | |
8e45cb54 | 4720 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
4721 | { |
4722 | struct task_struct *p, *n; | |
897c395f | 4723 | |
367456c7 | 4724 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
4725 | if (!can_migrate_task(p, env)) |
4726 | continue; | |
897c395f | 4727 | |
367456c7 PZ |
4728 | move_task(p, env); |
4729 | /* | |
4730 | * Right now, this is only the second place move_task() | |
4731 | * is called, so we can safely collect move_task() | |
4732 | * stats here rather than inside move_task(). | |
4733 | */ | |
4734 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
4735 | return 1; | |
897c395f | 4736 | } |
897c395f PZ |
4737 | return 0; |
4738 | } | |
4739 | ||
eb95308e PZ |
4740 | static const unsigned int sched_nr_migrate_break = 32; |
4741 | ||
5d6523eb | 4742 | /* |
bd939f45 | 4743 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
4744 | * this_rq, as part of a balancing operation within domain "sd". |
4745 | * Returns 1 if successful and 0 otherwise. | |
4746 | * | |
4747 | * Called with both runqueues locked. | |
4748 | */ | |
4749 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 4750 | { |
5d6523eb PZ |
4751 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
4752 | struct task_struct *p; | |
367456c7 PZ |
4753 | unsigned long load; |
4754 | int pulled = 0; | |
1e3c88bd | 4755 | |
bd939f45 | 4756 | if (env->imbalance <= 0) |
5d6523eb | 4757 | return 0; |
1e3c88bd | 4758 | |
5d6523eb PZ |
4759 | while (!list_empty(tasks)) { |
4760 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 4761 | |
367456c7 PZ |
4762 | env->loop++; |
4763 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 4764 | if (env->loop > env->loop_max) |
367456c7 | 4765 | break; |
5d6523eb PZ |
4766 | |
4767 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 4768 | if (env->loop > env->loop_break) { |
eb95308e | 4769 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 4770 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 4771 | break; |
a195f004 | 4772 | } |
1e3c88bd | 4773 | |
d3198084 | 4774 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
4775 | goto next; |
4776 | ||
4777 | load = task_h_load(p); | |
5d6523eb | 4778 | |
eb95308e | 4779 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
4780 | goto next; |
4781 | ||
bd939f45 | 4782 | if ((load / 2) > env->imbalance) |
367456c7 | 4783 | goto next; |
1e3c88bd | 4784 | |
ddcdf6e7 | 4785 | move_task(p, env); |
ee00e66f | 4786 | pulled++; |
bd939f45 | 4787 | env->imbalance -= load; |
1e3c88bd PZ |
4788 | |
4789 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
4790 | /* |
4791 | * NEWIDLE balancing is a source of latency, so preemptible | |
4792 | * kernels will stop after the first task is pulled to minimize | |
4793 | * the critical section. | |
4794 | */ | |
5d6523eb | 4795 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 4796 | break; |
1e3c88bd PZ |
4797 | #endif |
4798 | ||
ee00e66f PZ |
4799 | /* |
4800 | * We only want to steal up to the prescribed amount of | |
4801 | * weighted load. | |
4802 | */ | |
bd939f45 | 4803 | if (env->imbalance <= 0) |
ee00e66f | 4804 | break; |
367456c7 PZ |
4805 | |
4806 | continue; | |
4807 | next: | |
5d6523eb | 4808 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 4809 | } |
5d6523eb | 4810 | |
1e3c88bd | 4811 | /* |
ddcdf6e7 PZ |
4812 | * Right now, this is one of only two places move_task() is called, |
4813 | * so we can safely collect move_task() stats here rather than | |
4814 | * inside move_task(). | |
1e3c88bd | 4815 | */ |
8e45cb54 | 4816 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 4817 | |
5d6523eb | 4818 | return pulled; |
1e3c88bd PZ |
4819 | } |
4820 | ||
230059de | 4821 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
4822 | /* |
4823 | * update tg->load_weight by folding this cpu's load_avg | |
4824 | */ | |
48a16753 | 4825 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 4826 | { |
48a16753 PT |
4827 | struct sched_entity *se = tg->se[cpu]; |
4828 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 4829 | |
48a16753 PT |
4830 | /* throttled entities do not contribute to load */ |
4831 | if (throttled_hierarchy(cfs_rq)) | |
4832 | return; | |
9e3081ca | 4833 | |
aff3e498 | 4834 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 4835 | |
82958366 PT |
4836 | if (se) { |
4837 | update_entity_load_avg(se, 1); | |
4838 | /* | |
4839 | * We pivot on our runnable average having decayed to zero for | |
4840 | * list removal. This generally implies that all our children | |
4841 | * have also been removed (modulo rounding error or bandwidth | |
4842 | * control); however, such cases are rare and we can fix these | |
4843 | * at enqueue. | |
4844 | * | |
4845 | * TODO: fix up out-of-order children on enqueue. | |
4846 | */ | |
4847 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
4848 | list_del_leaf_cfs_rq(cfs_rq); | |
4849 | } else { | |
48a16753 | 4850 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
4851 | update_rq_runnable_avg(rq, rq->nr_running); |
4852 | } | |
9e3081ca PZ |
4853 | } |
4854 | ||
48a16753 | 4855 | static void update_blocked_averages(int cpu) |
9e3081ca | 4856 | { |
9e3081ca | 4857 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
4858 | struct cfs_rq *cfs_rq; |
4859 | unsigned long flags; | |
9e3081ca | 4860 | |
48a16753 PT |
4861 | raw_spin_lock_irqsave(&rq->lock, flags); |
4862 | update_rq_clock(rq); | |
9763b67f PZ |
4863 | /* |
4864 | * Iterates the task_group tree in a bottom up fashion, see | |
4865 | * list_add_leaf_cfs_rq() for details. | |
4866 | */ | |
64660c86 | 4867 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
4868 | /* |
4869 | * Note: We may want to consider periodically releasing | |
4870 | * rq->lock about these updates so that creating many task | |
4871 | * groups does not result in continually extending hold time. | |
4872 | */ | |
4873 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 4874 | } |
48a16753 PT |
4875 | |
4876 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
4877 | } |
4878 | ||
9763b67f | 4879 | /* |
68520796 | 4880 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
4881 | * This needs to be done in a top-down fashion because the load of a child |
4882 | * group is a fraction of its parents load. | |
4883 | */ | |
68520796 | 4884 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 4885 | { |
68520796 VD |
4886 | struct rq *rq = rq_of(cfs_rq); |
4887 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 4888 | unsigned long now = jiffies; |
68520796 | 4889 | unsigned long load; |
a35b6466 | 4890 | |
68520796 | 4891 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
4892 | return; |
4893 | ||
68520796 VD |
4894 | cfs_rq->h_load_next = NULL; |
4895 | for_each_sched_entity(se) { | |
4896 | cfs_rq = cfs_rq_of(se); | |
4897 | cfs_rq->h_load_next = se; | |
4898 | if (cfs_rq->last_h_load_update == now) | |
4899 | break; | |
4900 | } | |
a35b6466 | 4901 | |
68520796 | 4902 | if (!se) { |
7e3115ef | 4903 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
4904 | cfs_rq->last_h_load_update = now; |
4905 | } | |
4906 | ||
4907 | while ((se = cfs_rq->h_load_next) != NULL) { | |
4908 | load = cfs_rq->h_load; | |
4909 | load = div64_ul(load * se->avg.load_avg_contrib, | |
4910 | cfs_rq->runnable_load_avg + 1); | |
4911 | cfs_rq = group_cfs_rq(se); | |
4912 | cfs_rq->h_load = load; | |
4913 | cfs_rq->last_h_load_update = now; | |
4914 | } | |
9763b67f PZ |
4915 | } |
4916 | ||
367456c7 | 4917 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 4918 | { |
367456c7 | 4919 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 4920 | |
68520796 | 4921 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
4922 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
4923 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
4924 | } |
4925 | #else | |
48a16753 | 4926 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
4927 | { |
4928 | } | |
4929 | ||
367456c7 | 4930 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 4931 | { |
a003a25b | 4932 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 4933 | } |
230059de | 4934 | #endif |
1e3c88bd | 4935 | |
1e3c88bd | 4936 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
4937 | /* |
4938 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
4939 | */ | |
4940 | struct sg_lb_stats { | |
4941 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
4942 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 4943 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 4944 | unsigned long load_per_task; |
3ae11c90 | 4945 | unsigned long group_power; |
147c5fc2 PZ |
4946 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
4947 | unsigned int group_capacity; | |
4948 | unsigned int idle_cpus; | |
4949 | unsigned int group_weight; | |
1e3c88bd | 4950 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 4951 | int group_has_capacity; /* Is there extra capacity in the group? */ |
1e3c88bd PZ |
4952 | }; |
4953 | ||
56cf515b JK |
4954 | /* |
4955 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
4956 | * during load balancing. | |
4957 | */ | |
4958 | struct sd_lb_stats { | |
4959 | struct sched_group *busiest; /* Busiest group in this sd */ | |
4960 | struct sched_group *local; /* Local group in this sd */ | |
4961 | unsigned long total_load; /* Total load of all groups in sd */ | |
4962 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
4963 | unsigned long avg_load; /* Average load across all groups in sd */ | |
4964 | ||
56cf515b | 4965 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 4966 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
4967 | }; |
4968 | ||
147c5fc2 PZ |
4969 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
4970 | { | |
4971 | /* | |
4972 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
4973 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
4974 | * We must however clear busiest_stat::avg_load because | |
4975 | * update_sd_pick_busiest() reads this before assignment. | |
4976 | */ | |
4977 | *sds = (struct sd_lb_stats){ | |
4978 | .busiest = NULL, | |
4979 | .local = NULL, | |
4980 | .total_load = 0UL, | |
4981 | .total_pwr = 0UL, | |
4982 | .busiest_stat = { | |
4983 | .avg_load = 0UL, | |
4984 | }, | |
4985 | }; | |
4986 | } | |
4987 | ||
1e3c88bd PZ |
4988 | /** |
4989 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
4990 | * @sd: The sched_domain whose load_idx is to be obtained. | |
4991 | * @idle: The Idle status of the CPU for whose sd load_icx is obtained. | |
e69f6186 YB |
4992 | * |
4993 | * Return: The load index. | |
1e3c88bd PZ |
4994 | */ |
4995 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
4996 | enum cpu_idle_type idle) | |
4997 | { | |
4998 | int load_idx; | |
4999 | ||
5000 | switch (idle) { | |
5001 | case CPU_NOT_IDLE: | |
5002 | load_idx = sd->busy_idx; | |
5003 | break; | |
5004 | ||
5005 | case CPU_NEWLY_IDLE: | |
5006 | load_idx = sd->newidle_idx; | |
5007 | break; | |
5008 | default: | |
5009 | load_idx = sd->idle_idx; | |
5010 | break; | |
5011 | } | |
5012 | ||
5013 | return load_idx; | |
5014 | } | |
5015 | ||
15f803c9 | 5016 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5017 | { |
1399fa78 | 5018 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
5019 | } |
5020 | ||
5021 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
5022 | { | |
5023 | return default_scale_freq_power(sd, cpu); | |
5024 | } | |
5025 | ||
15f803c9 | 5026 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5027 | { |
669c55e9 | 5028 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
5029 | unsigned long smt_gain = sd->smt_gain; |
5030 | ||
5031 | smt_gain /= weight; | |
5032 | ||
5033 | return smt_gain; | |
5034 | } | |
5035 | ||
5036 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
5037 | { | |
5038 | return default_scale_smt_power(sd, cpu); | |
5039 | } | |
5040 | ||
15f803c9 | 5041 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
5042 | { |
5043 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5044 | u64 total, available, age_stamp, avg; |
1e3c88bd | 5045 | |
b654f7de PZ |
5046 | /* |
5047 | * Since we're reading these variables without serialization make sure | |
5048 | * we read them once before doing sanity checks on them. | |
5049 | */ | |
5050 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5051 | avg = ACCESS_ONCE(rq->rt_avg); | |
5052 | ||
78becc27 | 5053 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 5054 | |
b654f7de | 5055 | if (unlikely(total < avg)) { |
aa483808 VP |
5056 | /* Ensures that power won't end up being negative */ |
5057 | available = 0; | |
5058 | } else { | |
b654f7de | 5059 | available = total - avg; |
aa483808 | 5060 | } |
1e3c88bd | 5061 | |
1399fa78 NR |
5062 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
5063 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 5064 | |
1399fa78 | 5065 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5066 | |
5067 | return div_u64(available, total); | |
5068 | } | |
5069 | ||
5070 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
5071 | { | |
669c55e9 | 5072 | unsigned long weight = sd->span_weight; |
1399fa78 | 5073 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
5074 | struct sched_group *sdg = sd->groups; |
5075 | ||
1e3c88bd PZ |
5076 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
5077 | if (sched_feat(ARCH_POWER)) | |
5078 | power *= arch_scale_smt_power(sd, cpu); | |
5079 | else | |
5080 | power *= default_scale_smt_power(sd, cpu); | |
5081 | ||
1399fa78 | 5082 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5083 | } |
5084 | ||
9c3f75cb | 5085 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
5086 | |
5087 | if (sched_feat(ARCH_POWER)) | |
5088 | power *= arch_scale_freq_power(sd, cpu); | |
5089 | else | |
5090 | power *= default_scale_freq_power(sd, cpu); | |
5091 | ||
1399fa78 | 5092 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 5093 | |
1e3c88bd | 5094 | power *= scale_rt_power(cpu); |
1399fa78 | 5095 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5096 | |
5097 | if (!power) | |
5098 | power = 1; | |
5099 | ||
e51fd5e2 | 5100 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 5101 | sdg->sgp->power = power; |
1e3c88bd PZ |
5102 | } |
5103 | ||
029632fb | 5104 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
5105 | { |
5106 | struct sched_domain *child = sd->child; | |
5107 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 5108 | unsigned long power, power_orig; |
4ec4412e VG |
5109 | unsigned long interval; |
5110 | ||
5111 | interval = msecs_to_jiffies(sd->balance_interval); | |
5112 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
5113 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
5114 | |
5115 | if (!child) { | |
5116 | update_cpu_power(sd, cpu); | |
5117 | return; | |
5118 | } | |
5119 | ||
863bffc8 | 5120 | power_orig = power = 0; |
1e3c88bd | 5121 | |
74a5ce20 PZ |
5122 | if (child->flags & SD_OVERLAP) { |
5123 | /* | |
5124 | * SD_OVERLAP domains cannot assume that child groups | |
5125 | * span the current group. | |
5126 | */ | |
5127 | ||
863bffc8 PZ |
5128 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
5129 | struct sched_group *sg = cpu_rq(cpu)->sd->groups; | |
5130 | ||
5131 | power_orig += sg->sgp->power_orig; | |
5132 | power += sg->sgp->power; | |
5133 | } | |
74a5ce20 PZ |
5134 | } else { |
5135 | /* | |
5136 | * !SD_OVERLAP domains can assume that child groups | |
5137 | * span the current group. | |
5138 | */ | |
5139 | ||
5140 | group = child->groups; | |
5141 | do { | |
863bffc8 | 5142 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
5143 | power += group->sgp->power; |
5144 | group = group->next; | |
5145 | } while (group != child->groups); | |
5146 | } | |
1e3c88bd | 5147 | |
863bffc8 PZ |
5148 | sdg->sgp->power_orig = power_orig; |
5149 | sdg->sgp->power = power; | |
1e3c88bd PZ |
5150 | } |
5151 | ||
9d5efe05 SV |
5152 | /* |
5153 | * Try and fix up capacity for tiny siblings, this is needed when | |
5154 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
5155 | * which on its own isn't powerful enough. | |
5156 | * | |
5157 | * See update_sd_pick_busiest() and check_asym_packing(). | |
5158 | */ | |
5159 | static inline int | |
5160 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
5161 | { | |
5162 | /* | |
1399fa78 | 5163 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 5164 | */ |
a6c75f2f | 5165 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
5166 | return 0; |
5167 | ||
5168 | /* | |
5169 | * If ~90% of the cpu_power is still there, we're good. | |
5170 | */ | |
9c3f75cb | 5171 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
5172 | return 1; |
5173 | ||
5174 | return 0; | |
5175 | } | |
5176 | ||
30ce5dab PZ |
5177 | /* |
5178 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
5179 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
5180 | * | |
5181 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
5182 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
5183 | * Something like: | |
5184 | * | |
5185 | * { 0 1 2 3 } { 4 5 6 7 } | |
5186 | * * * * * | |
5187 | * | |
5188 | * If we were to balance group-wise we'd place two tasks in the first group and | |
5189 | * two tasks in the second group. Clearly this is undesired as it will overload | |
5190 | * cpu 3 and leave one of the cpus in the second group unused. | |
5191 | * | |
5192 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
5193 | * by noticing the lower domain failed to reach balance and had difficulty |
5194 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
5195 | * |
5196 | * When this is so detected; this group becomes a candidate for busiest; see | |
5197 | * update_sd_pick_busiest(). And calculcate_imbalance() and | |
6263322c | 5198 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
5199 | * to create an effective group imbalance. |
5200 | * | |
5201 | * This is a somewhat tricky proposition since the next run might not find the | |
5202 | * group imbalance and decide the groups need to be balanced again. A most | |
5203 | * subtle and fragile situation. | |
5204 | */ | |
5205 | ||
6263322c | 5206 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 5207 | { |
6263322c | 5208 | return group->sgp->imbalance; |
30ce5dab PZ |
5209 | } |
5210 | ||
b37d9316 PZ |
5211 | /* |
5212 | * Compute the group capacity. | |
5213 | * | |
c61037e9 PZ |
5214 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
5215 | * first dividing out the smt factor and computing the actual number of cores | |
5216 | * and limit power unit capacity with that. | |
b37d9316 PZ |
5217 | */ |
5218 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
5219 | { | |
c61037e9 PZ |
5220 | unsigned int capacity, smt, cpus; |
5221 | unsigned int power, power_orig; | |
5222 | ||
5223 | power = group->sgp->power; | |
5224 | power_orig = group->sgp->power_orig; | |
5225 | cpus = group->group_weight; | |
b37d9316 | 5226 | |
c61037e9 PZ |
5227 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
5228 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
5229 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 5230 | |
c61037e9 | 5231 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
5232 | if (!capacity) |
5233 | capacity = fix_small_capacity(env->sd, group); | |
5234 | ||
5235 | return capacity; | |
5236 | } | |
5237 | ||
1e3c88bd PZ |
5238 | /** |
5239 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 5240 | * @env: The load balancing environment. |
1e3c88bd | 5241 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 5242 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 5243 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
5244 | * @sgs: variable to hold the statistics for this group. |
5245 | */ | |
bd939f45 PZ |
5246 | static inline void update_sg_lb_stats(struct lb_env *env, |
5247 | struct sched_group *group, int load_idx, | |
23f0d209 | 5248 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 5249 | { |
30ce5dab PZ |
5250 | unsigned long nr_running; |
5251 | unsigned long load; | |
bd939f45 | 5252 | int i; |
1e3c88bd | 5253 | |
b72ff13c PZ |
5254 | memset(sgs, 0, sizeof(*sgs)); |
5255 | ||
b9403130 | 5256 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
5257 | struct rq *rq = cpu_rq(i); |
5258 | ||
e44bc5c5 PZ |
5259 | nr_running = rq->nr_running; |
5260 | ||
1e3c88bd | 5261 | /* Bias balancing toward cpus of our domain */ |
6263322c | 5262 | if (local_group) |
04f733b4 | 5263 | load = target_load(i, load_idx); |
6263322c | 5264 | else |
1e3c88bd | 5265 | load = source_load(i, load_idx); |
1e3c88bd PZ |
5266 | |
5267 | sgs->group_load += load; | |
e44bc5c5 | 5268 | sgs->sum_nr_running += nr_running; |
1e3c88bd | 5269 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
5270 | if (idle_cpu(i)) |
5271 | sgs->idle_cpus++; | |
1e3c88bd PZ |
5272 | } |
5273 | ||
1e3c88bd | 5274 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
5275 | sgs->group_power = group->sgp->power; |
5276 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 5277 | |
dd5feea1 | 5278 | if (sgs->sum_nr_running) |
38d0f770 | 5279 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5280 | |
aae6d3dd | 5281 | sgs->group_weight = group->group_weight; |
fab47622 | 5282 | |
b37d9316 PZ |
5283 | sgs->group_imb = sg_imbalanced(group); |
5284 | sgs->group_capacity = sg_capacity(env, group); | |
5285 | ||
fab47622 NR |
5286 | if (sgs->group_capacity > sgs->sum_nr_running) |
5287 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5288 | } |
5289 | ||
532cb4c4 MN |
5290 | /** |
5291 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5292 | * @env: The load balancing environment. |
532cb4c4 MN |
5293 | * @sds: sched_domain statistics |
5294 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5295 | * @sgs: sched_group statistics |
532cb4c4 MN |
5296 | * |
5297 | * Determine if @sg is a busier group than the previously selected | |
5298 | * busiest group. | |
e69f6186 YB |
5299 | * |
5300 | * Return: %true if @sg is a busier group than the previously selected | |
5301 | * busiest group. %false otherwise. | |
532cb4c4 | 5302 | */ |
bd939f45 | 5303 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5304 | struct sd_lb_stats *sds, |
5305 | struct sched_group *sg, | |
bd939f45 | 5306 | struct sg_lb_stats *sgs) |
532cb4c4 | 5307 | { |
56cf515b | 5308 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5309 | return false; |
5310 | ||
5311 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5312 | return true; | |
5313 | ||
5314 | if (sgs->group_imb) | |
5315 | return true; | |
5316 | ||
5317 | /* | |
5318 | * ASYM_PACKING needs to move all the work to the lowest | |
5319 | * numbered CPUs in the group, therefore mark all groups | |
5320 | * higher than ourself as busy. | |
5321 | */ | |
bd939f45 PZ |
5322 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5323 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5324 | if (!sds->busiest) |
5325 | return true; | |
5326 | ||
5327 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5328 | return true; | |
5329 | } | |
5330 | ||
5331 | return false; | |
5332 | } | |
5333 | ||
1e3c88bd | 5334 | /** |
461819ac | 5335 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5336 | * @env: The load balancing environment. |
1e3c88bd PZ |
5337 | * @balance: Should we balance. |
5338 | * @sds: variable to hold the statistics for this sched_domain. | |
5339 | */ | |
bd939f45 | 5340 | static inline void update_sd_lb_stats(struct lb_env *env, |
23f0d209 | 5341 | struct sd_lb_stats *sds) |
1e3c88bd | 5342 | { |
bd939f45 PZ |
5343 | struct sched_domain *child = env->sd->child; |
5344 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5345 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5346 | int load_idx, prefer_sibling = 0; |
5347 | ||
5348 | if (child && child->flags & SD_PREFER_SIBLING) | |
5349 | prefer_sibling = 1; | |
5350 | ||
bd939f45 | 5351 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5352 | |
5353 | do { | |
56cf515b | 5354 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5355 | int local_group; |
5356 | ||
bd939f45 | 5357 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5358 | if (local_group) { |
5359 | sds->local = sg; | |
5360 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5361 | |
5362 | if (env->idle != CPU_NEWLY_IDLE || | |
5363 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5364 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5365 | } |
1e3c88bd | 5366 | |
56cf515b | 5367 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5368 | |
b72ff13c PZ |
5369 | if (local_group) |
5370 | goto next_group; | |
5371 | ||
1e3c88bd PZ |
5372 | /* |
5373 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 5374 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
5375 | * and move all the excess tasks away. We lower the capacity |
5376 | * of a group only if the local group has the capacity to fit | |
5377 | * these excess tasks, i.e. nr_running < group_capacity. The | |
5378 | * extra check prevents the case where you always pull from the | |
5379 | * heaviest group when it is already under-utilized (possible | |
5380 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 5381 | */ |
b72ff13c PZ |
5382 | if (prefer_sibling && sds->local && |
5383 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 5384 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 5385 | |
b72ff13c | 5386 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 5387 | sds->busiest = sg; |
56cf515b | 5388 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
5389 | } |
5390 | ||
b72ff13c PZ |
5391 | next_group: |
5392 | /* Now, start updating sd_lb_stats */ | |
5393 | sds->total_load += sgs->group_load; | |
5394 | sds->total_pwr += sgs->group_power; | |
5395 | ||
532cb4c4 | 5396 | sg = sg->next; |
bd939f45 | 5397 | } while (sg != env->sd->groups); |
532cb4c4 MN |
5398 | } |
5399 | ||
532cb4c4 MN |
5400 | /** |
5401 | * check_asym_packing - Check to see if the group is packed into the | |
5402 | * sched doman. | |
5403 | * | |
5404 | * This is primarily intended to used at the sibling level. Some | |
5405 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
5406 | * case of POWER7, it can move to lower SMT modes only when higher | |
5407 | * threads are idle. When in lower SMT modes, the threads will | |
5408 | * perform better since they share less core resources. Hence when we | |
5409 | * have idle threads, we want them to be the higher ones. | |
5410 | * | |
5411 | * This packing function is run on idle threads. It checks to see if | |
5412 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
5413 | * CPU number than the packing function is being run on. Here we are | |
5414 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
5415 | * number. | |
5416 | * | |
e69f6186 | 5417 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
5418 | * this CPU. The amount of the imbalance is returned in *imbalance. |
5419 | * | |
cd96891d | 5420 | * @env: The load balancing environment. |
532cb4c4 | 5421 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 5422 | */ |
bd939f45 | 5423 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
5424 | { |
5425 | int busiest_cpu; | |
5426 | ||
bd939f45 | 5427 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
5428 | return 0; |
5429 | ||
5430 | if (!sds->busiest) | |
5431 | return 0; | |
5432 | ||
5433 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 5434 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
5435 | return 0; |
5436 | ||
bd939f45 | 5437 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
5438 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
5439 | SCHED_POWER_SCALE); | |
bd939f45 | 5440 | |
532cb4c4 | 5441 | return 1; |
1e3c88bd PZ |
5442 | } |
5443 | ||
5444 | /** | |
5445 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
5446 | * amongst the groups of a sched_domain, during | |
5447 | * load balancing. | |
cd96891d | 5448 | * @env: The load balancing environment. |
1e3c88bd | 5449 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5450 | */ |
bd939f45 PZ |
5451 | static inline |
5452 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
5453 | { |
5454 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
5455 | unsigned int imbn = 2; | |
dd5feea1 | 5456 | unsigned long scaled_busy_load_per_task; |
56cf515b | 5457 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 5458 | |
56cf515b JK |
5459 | local = &sds->local_stat; |
5460 | busiest = &sds->busiest_stat; | |
1e3c88bd | 5461 | |
56cf515b JK |
5462 | if (!local->sum_nr_running) |
5463 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
5464 | else if (busiest->load_per_task > local->load_per_task) | |
5465 | imbn = 1; | |
dd5feea1 | 5466 | |
56cf515b JK |
5467 | scaled_busy_load_per_task = |
5468 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5469 | busiest->group_power; |
56cf515b | 5470 | |
3029ede3 VD |
5471 | if (busiest->avg_load + scaled_busy_load_per_task >= |
5472 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 5473 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5474 | return; |
5475 | } | |
5476 | ||
5477 | /* | |
5478 | * OK, we don't have enough imbalance to justify moving tasks, | |
5479 | * however we may be able to increase total CPU power used by | |
5480 | * moving them. | |
5481 | */ | |
5482 | ||
3ae11c90 | 5483 | pwr_now += busiest->group_power * |
56cf515b | 5484 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 5485 | pwr_now += local->group_power * |
56cf515b | 5486 | min(local->load_per_task, local->avg_load); |
1399fa78 | 5487 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5488 | |
5489 | /* Amount of load we'd subtract */ | |
56cf515b | 5490 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 5491 | busiest->group_power; |
56cf515b | 5492 | if (busiest->avg_load > tmp) { |
3ae11c90 | 5493 | pwr_move += busiest->group_power * |
56cf515b JK |
5494 | min(busiest->load_per_task, |
5495 | busiest->avg_load - tmp); | |
5496 | } | |
1e3c88bd PZ |
5497 | |
5498 | /* Amount of load we'd add */ | |
3ae11c90 | 5499 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 5500 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
5501 | tmp = (busiest->avg_load * busiest->group_power) / |
5502 | local->group_power; | |
56cf515b JK |
5503 | } else { |
5504 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 5505 | local->group_power; |
56cf515b | 5506 | } |
3ae11c90 PZ |
5507 | pwr_move += local->group_power * |
5508 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 5509 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
5510 | |
5511 | /* Move if we gain throughput */ | |
5512 | if (pwr_move > pwr_now) | |
56cf515b | 5513 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
5514 | } |
5515 | ||
5516 | /** | |
5517 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
5518 | * groups of a given sched_domain during load balance. | |
bd939f45 | 5519 | * @env: load balance environment |
1e3c88bd | 5520 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 5521 | */ |
bd939f45 | 5522 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5523 | { |
dd5feea1 | 5524 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
5525 | struct sg_lb_stats *local, *busiest; |
5526 | ||
5527 | local = &sds->local_stat; | |
56cf515b | 5528 | busiest = &sds->busiest_stat; |
dd5feea1 | 5529 | |
56cf515b | 5530 | if (busiest->group_imb) { |
30ce5dab PZ |
5531 | /* |
5532 | * In the group_imb case we cannot rely on group-wide averages | |
5533 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
5534 | */ | |
56cf515b JK |
5535 | busiest->load_per_task = |
5536 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
5537 | } |
5538 | ||
1e3c88bd PZ |
5539 | /* |
5540 | * In the presence of smp nice balancing, certain scenarios can have | |
5541 | * max load less than avg load(as we skip the groups at or below | |
5542 | * its cpu_power, while calculating max_load..) | |
5543 | */ | |
b1885550 VD |
5544 | if (busiest->avg_load <= sds->avg_load || |
5545 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
5546 | env->imbalance = 0; |
5547 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
5548 | } |
5549 | ||
56cf515b | 5550 | if (!busiest->group_imb) { |
dd5feea1 SS |
5551 | /* |
5552 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
5553 | * Except of course for the group_imb case, since then we might |
5554 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 5555 | */ |
56cf515b JK |
5556 | load_above_capacity = |
5557 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 5558 | |
1399fa78 | 5559 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 5560 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
5561 | } |
5562 | ||
5563 | /* | |
5564 | * We're trying to get all the cpus to the average_load, so we don't | |
5565 | * want to push ourselves above the average load, nor do we wish to | |
5566 | * reduce the max loaded cpu below the average load. At the same time, | |
5567 | * we also don't want to reduce the group load below the group capacity | |
5568 | * (so that we can implement power-savings policies etc). Thus we look | |
5569 | * for the minimum possible imbalance. | |
dd5feea1 | 5570 | */ |
30ce5dab | 5571 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
5572 | |
5573 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 5574 | env->imbalance = min( |
3ae11c90 PZ |
5575 | max_pull * busiest->group_power, |
5576 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 5577 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
5578 | |
5579 | /* | |
5580 | * if *imbalance is less than the average load per runnable task | |
25985edc | 5581 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
5582 | * a think about bumping its value to force at least one task to be |
5583 | * moved | |
5584 | */ | |
56cf515b | 5585 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 5586 | return fix_small_imbalance(env, sds); |
1e3c88bd | 5587 | } |
fab47622 | 5588 | |
1e3c88bd PZ |
5589 | /******* find_busiest_group() helpers end here *********************/ |
5590 | ||
5591 | /** | |
5592 | * find_busiest_group - Returns the busiest group within the sched_domain | |
5593 | * if there is an imbalance. If there isn't an imbalance, and | |
5594 | * the user has opted for power-savings, it returns a group whose | |
5595 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
5596 | * such a group exists. | |
5597 | * | |
5598 | * Also calculates the amount of weighted load which should be moved | |
5599 | * to restore balance. | |
5600 | * | |
cd96891d | 5601 | * @env: The load balancing environment. |
1e3c88bd | 5602 | * |
e69f6186 | 5603 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
5604 | * - If no imbalance and user has opted for power-savings balance, |
5605 | * return the least loaded group whose CPUs can be | |
5606 | * put to idle by rebalancing its tasks onto our group. | |
5607 | */ | |
56cf515b | 5608 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 5609 | { |
56cf515b | 5610 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
5611 | struct sd_lb_stats sds; |
5612 | ||
147c5fc2 | 5613 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
5614 | |
5615 | /* | |
5616 | * Compute the various statistics relavent for load balancing at | |
5617 | * this level. | |
5618 | */ | |
23f0d209 | 5619 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
5620 | local = &sds.local_stat; |
5621 | busiest = &sds.busiest_stat; | |
1e3c88bd | 5622 | |
bd939f45 PZ |
5623 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
5624 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
5625 | return sds.busiest; |
5626 | ||
cc57aa8f | 5627 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 5628 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
5629 | goto out_balanced; |
5630 | ||
1399fa78 | 5631 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 5632 | |
866ab43e PZ |
5633 | /* |
5634 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 5635 | * work because they assume all things are equal, which typically |
866ab43e PZ |
5636 | * isn't true due to cpus_allowed constraints and the like. |
5637 | */ | |
56cf515b | 5638 | if (busiest->group_imb) |
866ab43e PZ |
5639 | goto force_balance; |
5640 | ||
cc57aa8f | 5641 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
5642 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
5643 | !busiest->group_has_capacity) | |
fab47622 NR |
5644 | goto force_balance; |
5645 | ||
cc57aa8f PZ |
5646 | /* |
5647 | * If the local group is more busy than the selected busiest group | |
5648 | * don't try and pull any tasks. | |
5649 | */ | |
56cf515b | 5650 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
5651 | goto out_balanced; |
5652 | ||
cc57aa8f PZ |
5653 | /* |
5654 | * Don't pull any tasks if this group is already above the domain | |
5655 | * average load. | |
5656 | */ | |
56cf515b | 5657 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
5658 | goto out_balanced; |
5659 | ||
bd939f45 | 5660 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
5661 | /* |
5662 | * This cpu is idle. If the busiest group load doesn't | |
5663 | * have more tasks than the number of available cpu's and | |
5664 | * there is no imbalance between this and busiest group | |
5665 | * wrt to idle cpu's, it is balanced. | |
5666 | */ | |
56cf515b JK |
5667 | if ((local->idle_cpus < busiest->idle_cpus) && |
5668 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 5669 | goto out_balanced; |
c186fafe PZ |
5670 | } else { |
5671 | /* | |
5672 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
5673 | * imbalance_pct to be conservative. | |
5674 | */ | |
56cf515b JK |
5675 | if (100 * busiest->avg_load <= |
5676 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 5677 | goto out_balanced; |
aae6d3dd | 5678 | } |
1e3c88bd | 5679 | |
fab47622 | 5680 | force_balance: |
1e3c88bd | 5681 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 5682 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
5683 | return sds.busiest; |
5684 | ||
5685 | out_balanced: | |
bd939f45 | 5686 | env->imbalance = 0; |
1e3c88bd PZ |
5687 | return NULL; |
5688 | } | |
5689 | ||
5690 | /* | |
5691 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
5692 | */ | |
bd939f45 | 5693 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 5694 | struct sched_group *group) |
1e3c88bd PZ |
5695 | { |
5696 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 5697 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
5698 | int i; |
5699 | ||
6906a408 | 5700 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd | 5701 | unsigned long power = power_of(i); |
1399fa78 NR |
5702 | unsigned long capacity = DIV_ROUND_CLOSEST(power, |
5703 | SCHED_POWER_SCALE); | |
1e3c88bd PZ |
5704 | unsigned long wl; |
5705 | ||
9d5efe05 | 5706 | if (!capacity) |
bd939f45 | 5707 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 5708 | |
1e3c88bd | 5709 | rq = cpu_rq(i); |
6e40f5bb | 5710 | wl = weighted_cpuload(i); |
1e3c88bd | 5711 | |
6e40f5bb TG |
5712 | /* |
5713 | * When comparing with imbalance, use weighted_cpuload() | |
5714 | * which is not scaled with the cpu power. | |
5715 | */ | |
bd939f45 | 5716 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
5717 | continue; |
5718 | ||
6e40f5bb TG |
5719 | /* |
5720 | * For the load comparisons with the other cpu's, consider | |
5721 | * the weighted_cpuload() scaled with the cpu power, so that | |
5722 | * the load can be moved away from the cpu that is potentially | |
5723 | * running at a lower capacity. | |
95a79b80 JK |
5724 | * |
5725 | * Thus we're looking for max(wl_i / power_i), crosswise | |
5726 | * multiplication to rid ourselves of the division works out | |
5727 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
5728 | * previous maximum. | |
6e40f5bb | 5729 | */ |
95a79b80 JK |
5730 | if (wl * busiest_power > busiest_load * power) { |
5731 | busiest_load = wl; | |
5732 | busiest_power = power; | |
1e3c88bd PZ |
5733 | busiest = rq; |
5734 | } | |
5735 | } | |
5736 | ||
5737 | return busiest; | |
5738 | } | |
5739 | ||
5740 | /* | |
5741 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
5742 | * so long as it is large enough. | |
5743 | */ | |
5744 | #define MAX_PINNED_INTERVAL 512 | |
5745 | ||
5746 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 5747 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 5748 | |
bd939f45 | 5749 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 5750 | { |
bd939f45 PZ |
5751 | struct sched_domain *sd = env->sd; |
5752 | ||
5753 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
5754 | |
5755 | /* | |
5756 | * ASYM_PACKING needs to force migrate tasks from busy but | |
5757 | * higher numbered CPUs in order to pack all tasks in the | |
5758 | * lowest numbered CPUs. | |
5759 | */ | |
bd939f45 | 5760 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 5761 | return 1; |
1af3ed3d PZ |
5762 | } |
5763 | ||
5764 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
5765 | } | |
5766 | ||
969c7921 TH |
5767 | static int active_load_balance_cpu_stop(void *data); |
5768 | ||
23f0d209 JK |
5769 | static int should_we_balance(struct lb_env *env) |
5770 | { | |
5771 | struct sched_group *sg = env->sd->groups; | |
5772 | struct cpumask *sg_cpus, *sg_mask; | |
5773 | int cpu, balance_cpu = -1; | |
5774 | ||
5775 | /* | |
5776 | * In the newly idle case, we will allow all the cpu's | |
5777 | * to do the newly idle load balance. | |
5778 | */ | |
5779 | if (env->idle == CPU_NEWLY_IDLE) | |
5780 | return 1; | |
5781 | ||
5782 | sg_cpus = sched_group_cpus(sg); | |
5783 | sg_mask = sched_group_mask(sg); | |
5784 | /* Try to find first idle cpu */ | |
5785 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
5786 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
5787 | continue; | |
5788 | ||
5789 | balance_cpu = cpu; | |
5790 | break; | |
5791 | } | |
5792 | ||
5793 | if (balance_cpu == -1) | |
5794 | balance_cpu = group_balance_cpu(sg); | |
5795 | ||
5796 | /* | |
5797 | * First idle cpu or the first cpu(busiest) in this sched group | |
5798 | * is eligible for doing load balancing at this and above domains. | |
5799 | */ | |
b0cff9d8 | 5800 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
5801 | } |
5802 | ||
1e3c88bd PZ |
5803 | /* |
5804 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
5805 | * tasks if there is an imbalance. | |
5806 | */ | |
5807 | static int load_balance(int this_cpu, struct rq *this_rq, | |
5808 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 5809 | int *continue_balancing) |
1e3c88bd | 5810 | { |
88b8dac0 | 5811 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 5812 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 5813 | struct sched_group *group; |
1e3c88bd PZ |
5814 | struct rq *busiest; |
5815 | unsigned long flags; | |
e6252c3e | 5816 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 5817 | |
8e45cb54 PZ |
5818 | struct lb_env env = { |
5819 | .sd = sd, | |
ddcdf6e7 PZ |
5820 | .dst_cpu = this_cpu, |
5821 | .dst_rq = this_rq, | |
88b8dac0 | 5822 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 5823 | .idle = idle, |
eb95308e | 5824 | .loop_break = sched_nr_migrate_break, |
b9403130 | 5825 | .cpus = cpus, |
8e45cb54 PZ |
5826 | }; |
5827 | ||
cfc03118 JK |
5828 | /* |
5829 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
5830 | * other cpus in our group | |
5831 | */ | |
e02e60c1 | 5832 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 5833 | env.dst_grpmask = NULL; |
cfc03118 | 5834 | |
1e3c88bd PZ |
5835 | cpumask_copy(cpus, cpu_active_mask); |
5836 | ||
1e3c88bd PZ |
5837 | schedstat_inc(sd, lb_count[idle]); |
5838 | ||
5839 | redo: | |
23f0d209 JK |
5840 | if (!should_we_balance(&env)) { |
5841 | *continue_balancing = 0; | |
1e3c88bd | 5842 | goto out_balanced; |
23f0d209 | 5843 | } |
1e3c88bd | 5844 | |
23f0d209 | 5845 | group = find_busiest_group(&env); |
1e3c88bd PZ |
5846 | if (!group) { |
5847 | schedstat_inc(sd, lb_nobusyg[idle]); | |
5848 | goto out_balanced; | |
5849 | } | |
5850 | ||
b9403130 | 5851 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
5852 | if (!busiest) { |
5853 | schedstat_inc(sd, lb_nobusyq[idle]); | |
5854 | goto out_balanced; | |
5855 | } | |
5856 | ||
78feefc5 | 5857 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 5858 | |
bd939f45 | 5859 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
5860 | |
5861 | ld_moved = 0; | |
5862 | if (busiest->nr_running > 1) { | |
5863 | /* | |
5864 | * Attempt to move tasks. If find_busiest_group has found | |
5865 | * an imbalance but busiest->nr_running <= 1, the group is | |
5866 | * still unbalanced. ld_moved simply stays zero, so it is | |
5867 | * correctly treated as an imbalance. | |
5868 | */ | |
8e45cb54 | 5869 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
5870 | env.src_cpu = busiest->cpu; |
5871 | env.src_rq = busiest; | |
5872 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 5873 | |
5d6523eb | 5874 | more_balance: |
1e3c88bd | 5875 | local_irq_save(flags); |
78feefc5 | 5876 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
5877 | |
5878 | /* | |
5879 | * cur_ld_moved - load moved in current iteration | |
5880 | * ld_moved - cumulative load moved across iterations | |
5881 | */ | |
5882 | cur_ld_moved = move_tasks(&env); | |
5883 | ld_moved += cur_ld_moved; | |
78feefc5 | 5884 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
5885 | local_irq_restore(flags); |
5886 | ||
5887 | /* | |
5888 | * some other cpu did the load balance for us. | |
5889 | */ | |
88b8dac0 SV |
5890 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
5891 | resched_cpu(env.dst_cpu); | |
5892 | ||
f1cd0858 JK |
5893 | if (env.flags & LBF_NEED_BREAK) { |
5894 | env.flags &= ~LBF_NEED_BREAK; | |
5895 | goto more_balance; | |
5896 | } | |
5897 | ||
88b8dac0 SV |
5898 | /* |
5899 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
5900 | * us and move them to an alternate dst_cpu in our sched_group | |
5901 | * where they can run. The upper limit on how many times we | |
5902 | * iterate on same src_cpu is dependent on number of cpus in our | |
5903 | * sched_group. | |
5904 | * | |
5905 | * This changes load balance semantics a bit on who can move | |
5906 | * load to a given_cpu. In addition to the given_cpu itself | |
5907 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
5908 | * nohz-idle), we now have balance_cpu in a position to move | |
5909 | * load to given_cpu. In rare situations, this may cause | |
5910 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
5911 | * _independently_ and at _same_ time to move some load to | |
5912 | * given_cpu) causing exceess load to be moved to given_cpu. | |
5913 | * This however should not happen so much in practice and | |
5914 | * moreover subsequent load balance cycles should correct the | |
5915 | * excess load moved. | |
5916 | */ | |
6263322c | 5917 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 5918 | |
7aff2e3a VD |
5919 | /* Prevent to re-select dst_cpu via env's cpus */ |
5920 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
5921 | ||
78feefc5 | 5922 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 5923 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 5924 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
5925 | env.loop = 0; |
5926 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 5927 | |
88b8dac0 SV |
5928 | /* |
5929 | * Go back to "more_balance" rather than "redo" since we | |
5930 | * need to continue with same src_cpu. | |
5931 | */ | |
5932 | goto more_balance; | |
5933 | } | |
1e3c88bd | 5934 | |
6263322c PZ |
5935 | /* |
5936 | * We failed to reach balance because of affinity. | |
5937 | */ | |
5938 | if (sd_parent) { | |
5939 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
5940 | ||
5941 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
5942 | *group_imbalance = 1; | |
5943 | } else if (*group_imbalance) | |
5944 | *group_imbalance = 0; | |
5945 | } | |
5946 | ||
1e3c88bd | 5947 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 5948 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 5949 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
5950 | if (!cpumask_empty(cpus)) { |
5951 | env.loop = 0; | |
5952 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 5953 | goto redo; |
bbf18b19 | 5954 | } |
1e3c88bd PZ |
5955 | goto out_balanced; |
5956 | } | |
5957 | } | |
5958 | ||
5959 | if (!ld_moved) { | |
5960 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
5961 | /* |
5962 | * Increment the failure counter only on periodic balance. | |
5963 | * We do not want newidle balance, which can be very | |
5964 | * frequent, pollute the failure counter causing | |
5965 | * excessive cache_hot migrations and active balances. | |
5966 | */ | |
5967 | if (idle != CPU_NEWLY_IDLE) | |
5968 | sd->nr_balance_failed++; | |
1e3c88bd | 5969 | |
bd939f45 | 5970 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
5971 | raw_spin_lock_irqsave(&busiest->lock, flags); |
5972 | ||
969c7921 TH |
5973 | /* don't kick the active_load_balance_cpu_stop, |
5974 | * if the curr task on busiest cpu can't be | |
5975 | * moved to this_cpu | |
1e3c88bd PZ |
5976 | */ |
5977 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 5978 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
5979 | raw_spin_unlock_irqrestore(&busiest->lock, |
5980 | flags); | |
8e45cb54 | 5981 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
5982 | goto out_one_pinned; |
5983 | } | |
5984 | ||
969c7921 TH |
5985 | /* |
5986 | * ->active_balance synchronizes accesses to | |
5987 | * ->active_balance_work. Once set, it's cleared | |
5988 | * only after active load balance is finished. | |
5989 | */ | |
1e3c88bd PZ |
5990 | if (!busiest->active_balance) { |
5991 | busiest->active_balance = 1; | |
5992 | busiest->push_cpu = this_cpu; | |
5993 | active_balance = 1; | |
5994 | } | |
5995 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 5996 | |
bd939f45 | 5997 | if (active_balance) { |
969c7921 TH |
5998 | stop_one_cpu_nowait(cpu_of(busiest), |
5999 | active_load_balance_cpu_stop, busiest, | |
6000 | &busiest->active_balance_work); | |
bd939f45 | 6001 | } |
1e3c88bd PZ |
6002 | |
6003 | /* | |
6004 | * We've kicked active balancing, reset the failure | |
6005 | * counter. | |
6006 | */ | |
6007 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
6008 | } | |
6009 | } else | |
6010 | sd->nr_balance_failed = 0; | |
6011 | ||
6012 | if (likely(!active_balance)) { | |
6013 | /* We were unbalanced, so reset the balancing interval */ | |
6014 | sd->balance_interval = sd->min_interval; | |
6015 | } else { | |
6016 | /* | |
6017 | * If we've begun active balancing, start to back off. This | |
6018 | * case may not be covered by the all_pinned logic if there | |
6019 | * is only 1 task on the busy runqueue (because we don't call | |
6020 | * move_tasks). | |
6021 | */ | |
6022 | if (sd->balance_interval < sd->max_interval) | |
6023 | sd->balance_interval *= 2; | |
6024 | } | |
6025 | ||
1e3c88bd PZ |
6026 | goto out; |
6027 | ||
6028 | out_balanced: | |
6029 | schedstat_inc(sd, lb_balanced[idle]); | |
6030 | ||
6031 | sd->nr_balance_failed = 0; | |
6032 | ||
6033 | out_one_pinned: | |
6034 | /* tune up the balancing interval */ | |
8e45cb54 | 6035 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 6036 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
6037 | (sd->balance_interval < sd->max_interval)) |
6038 | sd->balance_interval *= 2; | |
6039 | ||
46e49b38 | 6040 | ld_moved = 0; |
1e3c88bd | 6041 | out: |
1e3c88bd PZ |
6042 | return ld_moved; |
6043 | } | |
6044 | ||
1e3c88bd PZ |
6045 | /* |
6046 | * idle_balance is called by schedule() if this_cpu is about to become | |
6047 | * idle. Attempts to pull tasks from other CPUs. | |
6048 | */ | |
029632fb | 6049 | void idle_balance(int this_cpu, struct rq *this_rq) |
1e3c88bd PZ |
6050 | { |
6051 | struct sched_domain *sd; | |
6052 | int pulled_task = 0; | |
6053 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 6054 | u64 curr_cost = 0; |
1e3c88bd | 6055 | |
78becc27 | 6056 | this_rq->idle_stamp = rq_clock(this_rq); |
1e3c88bd PZ |
6057 | |
6058 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
6059 | return; | |
6060 | ||
f492e12e PZ |
6061 | /* |
6062 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
6063 | */ | |
6064 | raw_spin_unlock(&this_rq->lock); | |
6065 | ||
48a16753 | 6066 | update_blocked_averages(this_cpu); |
dce840a0 | 6067 | rcu_read_lock(); |
1e3c88bd PZ |
6068 | for_each_domain(this_cpu, sd) { |
6069 | unsigned long interval; | |
23f0d209 | 6070 | int continue_balancing = 1; |
9bd721c5 | 6071 | u64 t0, domain_cost; |
1e3c88bd PZ |
6072 | |
6073 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
6074 | continue; | |
6075 | ||
9bd721c5 JL |
6076 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
6077 | break; | |
6078 | ||
f492e12e | 6079 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
6080 | t0 = sched_clock_cpu(this_cpu); |
6081 | ||
1e3c88bd | 6082 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 6083 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
6084 | sd, CPU_NEWLY_IDLE, |
6085 | &continue_balancing); | |
9bd721c5 JL |
6086 | |
6087 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
6088 | if (domain_cost > sd->max_newidle_lb_cost) | |
6089 | sd->max_newidle_lb_cost = domain_cost; | |
6090 | ||
6091 | curr_cost += domain_cost; | |
f492e12e | 6092 | } |
1e3c88bd PZ |
6093 | |
6094 | interval = msecs_to_jiffies(sd->balance_interval); | |
6095 | if (time_after(next_balance, sd->last_balance + interval)) | |
6096 | next_balance = sd->last_balance + interval; | |
d5ad140b NR |
6097 | if (pulled_task) { |
6098 | this_rq->idle_stamp = 0; | |
1e3c88bd | 6099 | break; |
d5ad140b | 6100 | } |
1e3c88bd | 6101 | } |
dce840a0 | 6102 | rcu_read_unlock(); |
f492e12e PZ |
6103 | |
6104 | raw_spin_lock(&this_rq->lock); | |
6105 | ||
1e3c88bd PZ |
6106 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
6107 | /* | |
6108 | * We are going idle. next_balance may be set based on | |
6109 | * a busy processor. So reset next_balance. | |
6110 | */ | |
6111 | this_rq->next_balance = next_balance; | |
6112 | } | |
9bd721c5 JL |
6113 | |
6114 | if (curr_cost > this_rq->max_idle_balance_cost) | |
6115 | this_rq->max_idle_balance_cost = curr_cost; | |
1e3c88bd PZ |
6116 | } |
6117 | ||
6118 | /* | |
969c7921 TH |
6119 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
6120 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
6121 | * least 1 task to be running on each physical CPU where possible, and | |
6122 | * avoids physical / logical imbalances. | |
1e3c88bd | 6123 | */ |
969c7921 | 6124 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 6125 | { |
969c7921 TH |
6126 | struct rq *busiest_rq = data; |
6127 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 6128 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 6129 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 6130 | struct sched_domain *sd; |
969c7921 TH |
6131 | |
6132 | raw_spin_lock_irq(&busiest_rq->lock); | |
6133 | ||
6134 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
6135 | if (unlikely(busiest_cpu != smp_processor_id() || | |
6136 | !busiest_rq->active_balance)) | |
6137 | goto out_unlock; | |
1e3c88bd PZ |
6138 | |
6139 | /* Is there any task to move? */ | |
6140 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 6141 | goto out_unlock; |
1e3c88bd PZ |
6142 | |
6143 | /* | |
6144 | * This condition is "impossible", if it occurs | |
6145 | * we need to fix it. Originally reported by | |
6146 | * Bjorn Helgaas on a 128-cpu setup. | |
6147 | */ | |
6148 | BUG_ON(busiest_rq == target_rq); | |
6149 | ||
6150 | /* move a task from busiest_rq to target_rq */ | |
6151 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
6152 | |
6153 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 6154 | rcu_read_lock(); |
1e3c88bd PZ |
6155 | for_each_domain(target_cpu, sd) { |
6156 | if ((sd->flags & SD_LOAD_BALANCE) && | |
6157 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
6158 | break; | |
6159 | } | |
6160 | ||
6161 | if (likely(sd)) { | |
8e45cb54 PZ |
6162 | struct lb_env env = { |
6163 | .sd = sd, | |
ddcdf6e7 PZ |
6164 | .dst_cpu = target_cpu, |
6165 | .dst_rq = target_rq, | |
6166 | .src_cpu = busiest_rq->cpu, | |
6167 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
6168 | .idle = CPU_IDLE, |
6169 | }; | |
6170 | ||
1e3c88bd PZ |
6171 | schedstat_inc(sd, alb_count); |
6172 | ||
8e45cb54 | 6173 | if (move_one_task(&env)) |
1e3c88bd PZ |
6174 | schedstat_inc(sd, alb_pushed); |
6175 | else | |
6176 | schedstat_inc(sd, alb_failed); | |
6177 | } | |
dce840a0 | 6178 | rcu_read_unlock(); |
1e3c88bd | 6179 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
6180 | out_unlock: |
6181 | busiest_rq->active_balance = 0; | |
6182 | raw_spin_unlock_irq(&busiest_rq->lock); | |
6183 | return 0; | |
1e3c88bd PZ |
6184 | } |
6185 | ||
3451d024 | 6186 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
6187 | /* |
6188 | * idle load balancing details | |
83cd4fe2 VP |
6189 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
6190 | * needed, they will kick the idle load balancer, which then does idle | |
6191 | * load balancing for all the idle CPUs. | |
6192 | */ | |
1e3c88bd | 6193 | static struct { |
83cd4fe2 | 6194 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 6195 | atomic_t nr_cpus; |
83cd4fe2 VP |
6196 | unsigned long next_balance; /* in jiffy units */ |
6197 | } nohz ____cacheline_aligned; | |
1e3c88bd | 6198 | |
8e7fbcbc | 6199 | static inline int find_new_ilb(int call_cpu) |
1e3c88bd | 6200 | { |
0b005cf5 | 6201 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 6202 | |
786d6dc7 SS |
6203 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
6204 | return ilb; | |
6205 | ||
6206 | return nr_cpu_ids; | |
1e3c88bd | 6207 | } |
1e3c88bd | 6208 | |
83cd4fe2 VP |
6209 | /* |
6210 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
6211 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
6212 | * CPU (if there is one). | |
6213 | */ | |
6214 | static void nohz_balancer_kick(int cpu) | |
6215 | { | |
6216 | int ilb_cpu; | |
6217 | ||
6218 | nohz.next_balance++; | |
6219 | ||
0b005cf5 | 6220 | ilb_cpu = find_new_ilb(cpu); |
83cd4fe2 | 6221 | |
0b005cf5 SS |
6222 | if (ilb_cpu >= nr_cpu_ids) |
6223 | return; | |
83cd4fe2 | 6224 | |
cd490c5b | 6225 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
6226 | return; |
6227 | /* | |
6228 | * Use smp_send_reschedule() instead of resched_cpu(). | |
6229 | * This way we generate a sched IPI on the target cpu which | |
6230 | * is idle. And the softirq performing nohz idle load balance | |
6231 | * will be run before returning from the IPI. | |
6232 | */ | |
6233 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
6234 | return; |
6235 | } | |
6236 | ||
c1cc017c | 6237 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
6238 | { |
6239 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
6240 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
6241 | atomic_dec(&nohz.nr_cpus); | |
6242 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
6243 | } | |
6244 | } | |
6245 | ||
69e1e811 SS |
6246 | static inline void set_cpu_sd_state_busy(void) |
6247 | { | |
6248 | struct sched_domain *sd; | |
69e1e811 | 6249 | |
69e1e811 | 6250 | rcu_read_lock(); |
424c93fe | 6251 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
6252 | |
6253 | if (!sd || !sd->nohz_idle) | |
6254 | goto unlock; | |
6255 | sd->nohz_idle = 0; | |
6256 | ||
6257 | for (; sd; sd = sd->parent) | |
69e1e811 | 6258 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6259 | unlock: |
69e1e811 SS |
6260 | rcu_read_unlock(); |
6261 | } | |
6262 | ||
6263 | void set_cpu_sd_state_idle(void) | |
6264 | { | |
6265 | struct sched_domain *sd; | |
69e1e811 | 6266 | |
69e1e811 | 6267 | rcu_read_lock(); |
424c93fe | 6268 | sd = rcu_dereference_check_sched_domain(this_rq()->sd); |
25f55d9d VG |
6269 | |
6270 | if (!sd || sd->nohz_idle) | |
6271 | goto unlock; | |
6272 | sd->nohz_idle = 1; | |
6273 | ||
6274 | for (; sd; sd = sd->parent) | |
69e1e811 | 6275 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6276 | unlock: |
69e1e811 SS |
6277 | rcu_read_unlock(); |
6278 | } | |
6279 | ||
1e3c88bd | 6280 | /* |
c1cc017c | 6281 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 6282 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 6283 | */ |
c1cc017c | 6284 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 6285 | { |
71325960 SS |
6286 | /* |
6287 | * If this cpu is going down, then nothing needs to be done. | |
6288 | */ | |
6289 | if (!cpu_active(cpu)) | |
6290 | return; | |
6291 | ||
c1cc017c AS |
6292 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
6293 | return; | |
1e3c88bd | 6294 | |
c1cc017c AS |
6295 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
6296 | atomic_inc(&nohz.nr_cpus); | |
6297 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 6298 | } |
71325960 | 6299 | |
0db0628d | 6300 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
6301 | unsigned long action, void *hcpu) |
6302 | { | |
6303 | switch (action & ~CPU_TASKS_FROZEN) { | |
6304 | case CPU_DYING: | |
c1cc017c | 6305 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
6306 | return NOTIFY_OK; |
6307 | default: | |
6308 | return NOTIFY_DONE; | |
6309 | } | |
6310 | } | |
1e3c88bd PZ |
6311 | #endif |
6312 | ||
6313 | static DEFINE_SPINLOCK(balancing); | |
6314 | ||
49c022e6 PZ |
6315 | /* |
6316 | * Scale the max load_balance interval with the number of CPUs in the system. | |
6317 | * This trades load-balance latency on larger machines for less cross talk. | |
6318 | */ | |
029632fb | 6319 | void update_max_interval(void) |
49c022e6 PZ |
6320 | { |
6321 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
6322 | } | |
6323 | ||
1e3c88bd PZ |
6324 | /* |
6325 | * It checks each scheduling domain to see if it is due to be balanced, | |
6326 | * and initiates a balancing operation if so. | |
6327 | * | |
b9b0853a | 6328 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd PZ |
6329 | */ |
6330 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) | |
6331 | { | |
23f0d209 | 6332 | int continue_balancing = 1; |
1e3c88bd PZ |
6333 | struct rq *rq = cpu_rq(cpu); |
6334 | unsigned long interval; | |
04f733b4 | 6335 | struct sched_domain *sd; |
1e3c88bd PZ |
6336 | /* Earliest time when we have to do rebalance again */ |
6337 | unsigned long next_balance = jiffies + 60*HZ; | |
6338 | int update_next_balance = 0; | |
f48627e6 JL |
6339 | int need_serialize, need_decay = 0; |
6340 | u64 max_cost = 0; | |
1e3c88bd | 6341 | |
48a16753 | 6342 | update_blocked_averages(cpu); |
2069dd75 | 6343 | |
dce840a0 | 6344 | rcu_read_lock(); |
1e3c88bd | 6345 | for_each_domain(cpu, sd) { |
f48627e6 JL |
6346 | /* |
6347 | * Decay the newidle max times here because this is a regular | |
6348 | * visit to all the domains. Decay ~1% per second. | |
6349 | */ | |
6350 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
6351 | sd->max_newidle_lb_cost = | |
6352 | (sd->max_newidle_lb_cost * 253) / 256; | |
6353 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
6354 | need_decay = 1; | |
6355 | } | |
6356 | max_cost += sd->max_newidle_lb_cost; | |
6357 | ||
1e3c88bd PZ |
6358 | if (!(sd->flags & SD_LOAD_BALANCE)) |
6359 | continue; | |
6360 | ||
f48627e6 JL |
6361 | /* |
6362 | * Stop the load balance at this level. There is another | |
6363 | * CPU in our sched group which is doing load balancing more | |
6364 | * actively. | |
6365 | */ | |
6366 | if (!continue_balancing) { | |
6367 | if (need_decay) | |
6368 | continue; | |
6369 | break; | |
6370 | } | |
6371 | ||
1e3c88bd PZ |
6372 | interval = sd->balance_interval; |
6373 | if (idle != CPU_IDLE) | |
6374 | interval *= sd->busy_factor; | |
6375 | ||
6376 | /* scale ms to jiffies */ | |
6377 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 6378 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
6379 | |
6380 | need_serialize = sd->flags & SD_SERIALIZE; | |
6381 | ||
6382 | if (need_serialize) { | |
6383 | if (!spin_trylock(&balancing)) | |
6384 | goto out; | |
6385 | } | |
6386 | ||
6387 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 6388 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 6389 | /* |
6263322c | 6390 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
6391 | * env->dst_cpu, so we can't know our idle |
6392 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 6393 | */ |
de5eb2dd | 6394 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
6395 | } |
6396 | sd->last_balance = jiffies; | |
6397 | } | |
6398 | if (need_serialize) | |
6399 | spin_unlock(&balancing); | |
6400 | out: | |
6401 | if (time_after(next_balance, sd->last_balance + interval)) { | |
6402 | next_balance = sd->last_balance + interval; | |
6403 | update_next_balance = 1; | |
6404 | } | |
f48627e6 JL |
6405 | } |
6406 | if (need_decay) { | |
1e3c88bd | 6407 | /* |
f48627e6 JL |
6408 | * Ensure the rq-wide value also decays but keep it at a |
6409 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 6410 | */ |
f48627e6 JL |
6411 | rq->max_idle_balance_cost = |
6412 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 6413 | } |
dce840a0 | 6414 | rcu_read_unlock(); |
1e3c88bd PZ |
6415 | |
6416 | /* | |
6417 | * next_balance will be updated only when there is a need. | |
6418 | * When the cpu is attached to null domain for ex, it will not be | |
6419 | * updated. | |
6420 | */ | |
6421 | if (likely(update_next_balance)) | |
6422 | rq->next_balance = next_balance; | |
6423 | } | |
6424 | ||
3451d024 | 6425 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 6426 | /* |
3451d024 | 6427 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
6428 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
6429 | */ | |
83cd4fe2 VP |
6430 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) |
6431 | { | |
6432 | struct rq *this_rq = cpu_rq(this_cpu); | |
6433 | struct rq *rq; | |
6434 | int balance_cpu; | |
6435 | ||
1c792db7 SS |
6436 | if (idle != CPU_IDLE || |
6437 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
6438 | goto end; | |
83cd4fe2 VP |
6439 | |
6440 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 6441 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
6442 | continue; |
6443 | ||
6444 | /* | |
6445 | * If this cpu gets work to do, stop the load balancing | |
6446 | * work being done for other cpus. Next load | |
6447 | * balancing owner will pick it up. | |
6448 | */ | |
1c792db7 | 6449 | if (need_resched()) |
83cd4fe2 | 6450 | break; |
83cd4fe2 | 6451 | |
5ed4f1d9 VG |
6452 | rq = cpu_rq(balance_cpu); |
6453 | ||
6454 | raw_spin_lock_irq(&rq->lock); | |
6455 | update_rq_clock(rq); | |
6456 | update_idle_cpu_load(rq); | |
6457 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 VP |
6458 | |
6459 | rebalance_domains(balance_cpu, CPU_IDLE); | |
6460 | ||
83cd4fe2 VP |
6461 | if (time_after(this_rq->next_balance, rq->next_balance)) |
6462 | this_rq->next_balance = rq->next_balance; | |
6463 | } | |
6464 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
6465 | end: |
6466 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
6467 | } |
6468 | ||
6469 | /* | |
0b005cf5 SS |
6470 | * Current heuristic for kicking the idle load balancer in the presence |
6471 | * of an idle cpu is the system. | |
6472 | * - This rq has more than one task. | |
6473 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
6474 | * busy cpu's exceeding the group's power. | |
6475 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
6476 | * domain span are idle. | |
83cd4fe2 VP |
6477 | */ |
6478 | static inline int nohz_kick_needed(struct rq *rq, int cpu) | |
6479 | { | |
6480 | unsigned long now = jiffies; | |
0b005cf5 | 6481 | struct sched_domain *sd; |
83cd4fe2 | 6482 | |
1c792db7 | 6483 | if (unlikely(idle_cpu(cpu))) |
83cd4fe2 VP |
6484 | return 0; |
6485 | ||
1c792db7 SS |
6486 | /* |
6487 | * We may be recently in ticked or tickless idle mode. At the first | |
6488 | * busy tick after returning from idle, we will update the busy stats. | |
6489 | */ | |
69e1e811 | 6490 | set_cpu_sd_state_busy(); |
c1cc017c | 6491 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
6492 | |
6493 | /* | |
6494 | * None are in tickless mode and hence no need for NOHZ idle load | |
6495 | * balancing. | |
6496 | */ | |
6497 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
6498 | return 0; | |
1c792db7 SS |
6499 | |
6500 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
6501 | return 0; |
6502 | ||
0b005cf5 SS |
6503 | if (rq->nr_running >= 2) |
6504 | goto need_kick; | |
83cd4fe2 | 6505 | |
067491b7 | 6506 | rcu_read_lock(); |
0b005cf5 SS |
6507 | for_each_domain(cpu, sd) { |
6508 | struct sched_group *sg = sd->groups; | |
6509 | struct sched_group_power *sgp = sg->sgp; | |
6510 | int nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
83cd4fe2 | 6511 | |
0b005cf5 | 6512 | if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) |
067491b7 | 6513 | goto need_kick_unlock; |
0b005cf5 SS |
6514 | |
6515 | if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight | |
6516 | && (cpumask_first_and(nohz.idle_cpus_mask, | |
6517 | sched_domain_span(sd)) < cpu)) | |
067491b7 | 6518 | goto need_kick_unlock; |
0b005cf5 SS |
6519 | |
6520 | if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) | |
6521 | break; | |
83cd4fe2 | 6522 | } |
067491b7 | 6523 | rcu_read_unlock(); |
83cd4fe2 | 6524 | return 0; |
067491b7 PZ |
6525 | |
6526 | need_kick_unlock: | |
6527 | rcu_read_unlock(); | |
0b005cf5 SS |
6528 | need_kick: |
6529 | return 1; | |
83cd4fe2 VP |
6530 | } |
6531 | #else | |
6532 | static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } | |
6533 | #endif | |
6534 | ||
6535 | /* | |
6536 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
6537 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
6538 | */ | |
1e3c88bd PZ |
6539 | static void run_rebalance_domains(struct softirq_action *h) |
6540 | { | |
6541 | int this_cpu = smp_processor_id(); | |
6542 | struct rq *this_rq = cpu_rq(this_cpu); | |
6eb57e0d | 6543 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
6544 | CPU_IDLE : CPU_NOT_IDLE; |
6545 | ||
6546 | rebalance_domains(this_cpu, idle); | |
6547 | ||
1e3c88bd | 6548 | /* |
83cd4fe2 | 6549 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
6550 | * balancing on behalf of the other idle cpus whose ticks are |
6551 | * stopped. | |
6552 | */ | |
83cd4fe2 | 6553 | nohz_idle_balance(this_cpu, idle); |
1e3c88bd PZ |
6554 | } |
6555 | ||
6556 | static inline int on_null_domain(int cpu) | |
6557 | { | |
90a6501f | 6558 | return !rcu_dereference_sched(cpu_rq(cpu)->sd); |
1e3c88bd PZ |
6559 | } |
6560 | ||
6561 | /* | |
6562 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 6563 | */ |
029632fb | 6564 | void trigger_load_balance(struct rq *rq, int cpu) |
1e3c88bd | 6565 | { |
1e3c88bd PZ |
6566 | /* Don't need to rebalance while attached to NULL domain */ |
6567 | if (time_after_eq(jiffies, rq->next_balance) && | |
6568 | likely(!on_null_domain(cpu))) | |
6569 | raise_softirq(SCHED_SOFTIRQ); | |
3451d024 | 6570 | #ifdef CONFIG_NO_HZ_COMMON |
1c792db7 | 6571 | if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) |
83cd4fe2 VP |
6572 | nohz_balancer_kick(cpu); |
6573 | #endif | |
1e3c88bd PZ |
6574 | } |
6575 | ||
0bcdcf28 CE |
6576 | static void rq_online_fair(struct rq *rq) |
6577 | { | |
6578 | update_sysctl(); | |
6579 | } | |
6580 | ||
6581 | static void rq_offline_fair(struct rq *rq) | |
6582 | { | |
6583 | update_sysctl(); | |
a4c96ae3 PB |
6584 | |
6585 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
6586 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
6587 | } |
6588 | ||
55e12e5e | 6589 | #endif /* CONFIG_SMP */ |
e1d1484f | 6590 | |
bf0f6f24 IM |
6591 | /* |
6592 | * scheduler tick hitting a task of our scheduling class: | |
6593 | */ | |
8f4d37ec | 6594 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
6595 | { |
6596 | struct cfs_rq *cfs_rq; | |
6597 | struct sched_entity *se = &curr->se; | |
6598 | ||
6599 | for_each_sched_entity(se) { | |
6600 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 6601 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 6602 | } |
18bf2805 | 6603 | |
10e84b97 | 6604 | if (numabalancing_enabled) |
cbee9f88 | 6605 | task_tick_numa(rq, curr); |
3d59eebc | 6606 | |
18bf2805 | 6607 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
6608 | } |
6609 | ||
6610 | /* | |
cd29fe6f PZ |
6611 | * called on fork with the child task as argument from the parent's context |
6612 | * - child not yet on the tasklist | |
6613 | * - preemption disabled | |
bf0f6f24 | 6614 | */ |
cd29fe6f | 6615 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 6616 | { |
4fc420c9 DN |
6617 | struct cfs_rq *cfs_rq; |
6618 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 6619 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
6620 | struct rq *rq = this_rq(); |
6621 | unsigned long flags; | |
6622 | ||
05fa785c | 6623 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 6624 | |
861d034e PZ |
6625 | update_rq_clock(rq); |
6626 | ||
4fc420c9 DN |
6627 | cfs_rq = task_cfs_rq(current); |
6628 | curr = cfs_rq->curr; | |
6629 | ||
6c9a27f5 DN |
6630 | /* |
6631 | * Not only the cpu but also the task_group of the parent might have | |
6632 | * been changed after parent->se.parent,cfs_rq were copied to | |
6633 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
6634 | * of child point to valid ones. | |
6635 | */ | |
6636 | rcu_read_lock(); | |
6637 | __set_task_cpu(p, this_cpu); | |
6638 | rcu_read_unlock(); | |
bf0f6f24 | 6639 | |
7109c442 | 6640 | update_curr(cfs_rq); |
cd29fe6f | 6641 | |
b5d9d734 MG |
6642 | if (curr) |
6643 | se->vruntime = curr->vruntime; | |
aeb73b04 | 6644 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 6645 | |
cd29fe6f | 6646 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 6647 | /* |
edcb60a3 IM |
6648 | * Upon rescheduling, sched_class::put_prev_task() will place |
6649 | * 'current' within the tree based on its new key value. | |
6650 | */ | |
4d78e7b6 | 6651 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 6652 | resched_task(rq->curr); |
4d78e7b6 | 6653 | } |
bf0f6f24 | 6654 | |
88ec22d3 PZ |
6655 | se->vruntime -= cfs_rq->min_vruntime; |
6656 | ||
05fa785c | 6657 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
6658 | } |
6659 | ||
cb469845 SR |
6660 | /* |
6661 | * Priority of the task has changed. Check to see if we preempt | |
6662 | * the current task. | |
6663 | */ | |
da7a735e PZ |
6664 | static void |
6665 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 6666 | { |
da7a735e PZ |
6667 | if (!p->se.on_rq) |
6668 | return; | |
6669 | ||
cb469845 SR |
6670 | /* |
6671 | * Reschedule if we are currently running on this runqueue and | |
6672 | * our priority decreased, or if we are not currently running on | |
6673 | * this runqueue and our priority is higher than the current's | |
6674 | */ | |
da7a735e | 6675 | if (rq->curr == p) { |
cb469845 SR |
6676 | if (p->prio > oldprio) |
6677 | resched_task(rq->curr); | |
6678 | } else | |
15afe09b | 6679 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6680 | } |
6681 | ||
da7a735e PZ |
6682 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
6683 | { | |
6684 | struct sched_entity *se = &p->se; | |
6685 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6686 | ||
6687 | /* | |
6688 | * Ensure the task's vruntime is normalized, so that when its | |
6689 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
6690 | * do the right thing. | |
6691 | * | |
6692 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
6693 | * have normalized the vruntime, if it was !on_rq, then only when | |
6694 | * the task is sleeping will it still have non-normalized vruntime. | |
6695 | */ | |
6696 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
6697 | /* | |
6698 | * Fix up our vruntime so that the current sleep doesn't | |
6699 | * cause 'unlimited' sleep bonus. | |
6700 | */ | |
6701 | place_entity(cfs_rq, se, 0); | |
6702 | se->vruntime -= cfs_rq->min_vruntime; | |
6703 | } | |
9ee474f5 | 6704 | |
141965c7 | 6705 | #ifdef CONFIG_SMP |
9ee474f5 PT |
6706 | /* |
6707 | * Remove our load from contribution when we leave sched_fair | |
6708 | * and ensure we don't carry in an old decay_count if we | |
6709 | * switch back. | |
6710 | */ | |
87e3c8ae KT |
6711 | if (se->avg.decay_count) { |
6712 | __synchronize_entity_decay(se); | |
6713 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
6714 | } |
6715 | #endif | |
da7a735e PZ |
6716 | } |
6717 | ||
cb469845 SR |
6718 | /* |
6719 | * We switched to the sched_fair class. | |
6720 | */ | |
da7a735e | 6721 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 6722 | { |
da7a735e PZ |
6723 | if (!p->se.on_rq) |
6724 | return; | |
6725 | ||
cb469845 SR |
6726 | /* |
6727 | * We were most likely switched from sched_rt, so | |
6728 | * kick off the schedule if running, otherwise just see | |
6729 | * if we can still preempt the current task. | |
6730 | */ | |
da7a735e | 6731 | if (rq->curr == p) |
cb469845 SR |
6732 | resched_task(rq->curr); |
6733 | else | |
15afe09b | 6734 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
6735 | } |
6736 | ||
83b699ed SV |
6737 | /* Account for a task changing its policy or group. |
6738 | * | |
6739 | * This routine is mostly called to set cfs_rq->curr field when a task | |
6740 | * migrates between groups/classes. | |
6741 | */ | |
6742 | static void set_curr_task_fair(struct rq *rq) | |
6743 | { | |
6744 | struct sched_entity *se = &rq->curr->se; | |
6745 | ||
ec12cb7f PT |
6746 | for_each_sched_entity(se) { |
6747 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6748 | ||
6749 | set_next_entity(cfs_rq, se); | |
6750 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
6751 | account_cfs_rq_runtime(cfs_rq, 0); | |
6752 | } | |
83b699ed SV |
6753 | } |
6754 | ||
029632fb PZ |
6755 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
6756 | { | |
6757 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
6758 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
6759 | #ifndef CONFIG_64BIT | |
6760 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
6761 | #endif | |
141965c7 | 6762 | #ifdef CONFIG_SMP |
9ee474f5 | 6763 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 6764 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 6765 | #endif |
029632fb PZ |
6766 | } |
6767 | ||
810b3817 | 6768 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 6769 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 6770 | { |
aff3e498 | 6771 | struct cfs_rq *cfs_rq; |
b2b5ce02 PZ |
6772 | /* |
6773 | * If the task was not on the rq at the time of this cgroup movement | |
6774 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
6775 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
6776 | * bonus in place_entity()). | |
6777 | * | |
6778 | * If it was on the rq, we've just 'preempted' it, which does convert | |
6779 | * ->vruntime to a relative base. | |
6780 | * | |
6781 | * Make sure both cases convert their relative position when migrating | |
6782 | * to another cgroup's rq. This does somewhat interfere with the | |
6783 | * fair sleeper stuff for the first placement, but who cares. | |
6784 | */ | |
7ceff013 DN |
6785 | /* |
6786 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
6787 | * But there are some cases where it has already been normalized: | |
6788 | * | |
6789 | * - Moving a forked child which is waiting for being woken up by | |
6790 | * wake_up_new_task(). | |
62af3783 DN |
6791 | * - Moving a task which has been woken up by try_to_wake_up() and |
6792 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
6793 | * |
6794 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
6795 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
6796 | */ | |
62af3783 | 6797 | if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
6798 | on_rq = 1; |
6799 | ||
b2b5ce02 PZ |
6800 | if (!on_rq) |
6801 | p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; | |
6802 | set_task_rq(p, task_cpu(p)); | |
aff3e498 PT |
6803 | if (!on_rq) { |
6804 | cfs_rq = cfs_rq_of(&p->se); | |
6805 | p->se.vruntime += cfs_rq->min_vruntime; | |
6806 | #ifdef CONFIG_SMP | |
6807 | /* | |
6808 | * migrate_task_rq_fair() will have removed our previous | |
6809 | * contribution, but we must synchronize for ongoing future | |
6810 | * decay. | |
6811 | */ | |
6812 | p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
6813 | cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib; | |
6814 | #endif | |
6815 | } | |
810b3817 | 6816 | } |
029632fb PZ |
6817 | |
6818 | void free_fair_sched_group(struct task_group *tg) | |
6819 | { | |
6820 | int i; | |
6821 | ||
6822 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6823 | ||
6824 | for_each_possible_cpu(i) { | |
6825 | if (tg->cfs_rq) | |
6826 | kfree(tg->cfs_rq[i]); | |
6827 | if (tg->se) | |
6828 | kfree(tg->se[i]); | |
6829 | } | |
6830 | ||
6831 | kfree(tg->cfs_rq); | |
6832 | kfree(tg->se); | |
6833 | } | |
6834 | ||
6835 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6836 | { | |
6837 | struct cfs_rq *cfs_rq; | |
6838 | struct sched_entity *se; | |
6839 | int i; | |
6840 | ||
6841 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
6842 | if (!tg->cfs_rq) | |
6843 | goto err; | |
6844 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
6845 | if (!tg->se) | |
6846 | goto err; | |
6847 | ||
6848 | tg->shares = NICE_0_LOAD; | |
6849 | ||
6850 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
6851 | ||
6852 | for_each_possible_cpu(i) { | |
6853 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
6854 | GFP_KERNEL, cpu_to_node(i)); | |
6855 | if (!cfs_rq) | |
6856 | goto err; | |
6857 | ||
6858 | se = kzalloc_node(sizeof(struct sched_entity), | |
6859 | GFP_KERNEL, cpu_to_node(i)); | |
6860 | if (!se) | |
6861 | goto err_free_rq; | |
6862 | ||
6863 | init_cfs_rq(cfs_rq); | |
6864 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
6865 | } | |
6866 | ||
6867 | return 1; | |
6868 | ||
6869 | err_free_rq: | |
6870 | kfree(cfs_rq); | |
6871 | err: | |
6872 | return 0; | |
6873 | } | |
6874 | ||
6875 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
6876 | { | |
6877 | struct rq *rq = cpu_rq(cpu); | |
6878 | unsigned long flags; | |
6879 | ||
6880 | /* | |
6881 | * Only empty task groups can be destroyed; so we can speculatively | |
6882 | * check on_list without danger of it being re-added. | |
6883 | */ | |
6884 | if (!tg->cfs_rq[cpu]->on_list) | |
6885 | return; | |
6886 | ||
6887 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6888 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
6889 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6890 | } | |
6891 | ||
6892 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
6893 | struct sched_entity *se, int cpu, | |
6894 | struct sched_entity *parent) | |
6895 | { | |
6896 | struct rq *rq = cpu_rq(cpu); | |
6897 | ||
6898 | cfs_rq->tg = tg; | |
6899 | cfs_rq->rq = rq; | |
029632fb PZ |
6900 | init_cfs_rq_runtime(cfs_rq); |
6901 | ||
6902 | tg->cfs_rq[cpu] = cfs_rq; | |
6903 | tg->se[cpu] = se; | |
6904 | ||
6905 | /* se could be NULL for root_task_group */ | |
6906 | if (!se) | |
6907 | return; | |
6908 | ||
6909 | if (!parent) | |
6910 | se->cfs_rq = &rq->cfs; | |
6911 | else | |
6912 | se->cfs_rq = parent->my_q; | |
6913 | ||
6914 | se->my_q = cfs_rq; | |
6915 | update_load_set(&se->load, 0); | |
6916 | se->parent = parent; | |
6917 | } | |
6918 | ||
6919 | static DEFINE_MUTEX(shares_mutex); | |
6920 | ||
6921 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
6922 | { | |
6923 | int i; | |
6924 | unsigned long flags; | |
6925 | ||
6926 | /* | |
6927 | * We can't change the weight of the root cgroup. | |
6928 | */ | |
6929 | if (!tg->se[0]) | |
6930 | return -EINVAL; | |
6931 | ||
6932 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
6933 | ||
6934 | mutex_lock(&shares_mutex); | |
6935 | if (tg->shares == shares) | |
6936 | goto done; | |
6937 | ||
6938 | tg->shares = shares; | |
6939 | for_each_possible_cpu(i) { | |
6940 | struct rq *rq = cpu_rq(i); | |
6941 | struct sched_entity *se; | |
6942 | ||
6943 | se = tg->se[i]; | |
6944 | /* Propagate contribution to hierarchy */ | |
6945 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
6946 | |
6947 | /* Possible calls to update_curr() need rq clock */ | |
6948 | update_rq_clock(rq); | |
17bc14b7 | 6949 | for_each_sched_entity(se) |
029632fb PZ |
6950 | update_cfs_shares(group_cfs_rq(se)); |
6951 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
6952 | } | |
6953 | ||
6954 | done: | |
6955 | mutex_unlock(&shares_mutex); | |
6956 | return 0; | |
6957 | } | |
6958 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6959 | ||
6960 | void free_fair_sched_group(struct task_group *tg) { } | |
6961 | ||
6962 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
6963 | { | |
6964 | return 1; | |
6965 | } | |
6966 | ||
6967 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
6968 | ||
6969 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
6970 | ||
810b3817 | 6971 | |
6d686f45 | 6972 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
6973 | { |
6974 | struct sched_entity *se = &task->se; | |
0d721cea PW |
6975 | unsigned int rr_interval = 0; |
6976 | ||
6977 | /* | |
6978 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
6979 | * idle runqueue: | |
6980 | */ | |
0d721cea | 6981 | if (rq->cfs.load.weight) |
a59f4e07 | 6982 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
6983 | |
6984 | return rr_interval; | |
6985 | } | |
6986 | ||
bf0f6f24 IM |
6987 | /* |
6988 | * All the scheduling class methods: | |
6989 | */ | |
029632fb | 6990 | const struct sched_class fair_sched_class = { |
5522d5d5 | 6991 | .next = &idle_sched_class, |
bf0f6f24 IM |
6992 | .enqueue_task = enqueue_task_fair, |
6993 | .dequeue_task = dequeue_task_fair, | |
6994 | .yield_task = yield_task_fair, | |
d95f4122 | 6995 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 6996 | |
2e09bf55 | 6997 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
6998 | |
6999 | .pick_next_task = pick_next_task_fair, | |
7000 | .put_prev_task = put_prev_task_fair, | |
7001 | ||
681f3e68 | 7002 | #ifdef CONFIG_SMP |
4ce72a2c | 7003 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 7004 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 7005 | |
0bcdcf28 CE |
7006 | .rq_online = rq_online_fair, |
7007 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
7008 | |
7009 | .task_waking = task_waking_fair, | |
681f3e68 | 7010 | #endif |
bf0f6f24 | 7011 | |
83b699ed | 7012 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 7013 | .task_tick = task_tick_fair, |
cd29fe6f | 7014 | .task_fork = task_fork_fair, |
cb469845 SR |
7015 | |
7016 | .prio_changed = prio_changed_fair, | |
da7a735e | 7017 | .switched_from = switched_from_fair, |
cb469845 | 7018 | .switched_to = switched_to_fair, |
810b3817 | 7019 | |
0d721cea PW |
7020 | .get_rr_interval = get_rr_interval_fair, |
7021 | ||
810b3817 | 7022 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7023 | .task_move_group = task_move_group_fair, |
810b3817 | 7024 | #endif |
bf0f6f24 IM |
7025 | }; |
7026 | ||
7027 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 7028 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 7029 | { |
bf0f6f24 IM |
7030 | struct cfs_rq *cfs_rq; |
7031 | ||
5973e5b9 | 7032 | rcu_read_lock(); |
c3b64f1e | 7033 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 7034 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 7035 | rcu_read_unlock(); |
bf0f6f24 IM |
7036 | } |
7037 | #endif | |
029632fb PZ |
7038 | |
7039 | __init void init_sched_fair_class(void) | |
7040 | { | |
7041 | #ifdef CONFIG_SMP | |
7042 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
7043 | ||
3451d024 | 7044 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 7045 | nohz.next_balance = jiffies; |
029632fb | 7046 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 7047 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
7048 | #endif |
7049 | #endif /* SMP */ | |
7050 | ||
7051 | } |