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 | ||
9dbdb155 | 181 | #define WMULT_CONST (~0U) |
029632fb PZ |
182 | #define WMULT_SHIFT 32 |
183 | ||
9dbdb155 PZ |
184 | static void __update_inv_weight(struct load_weight *lw) |
185 | { | |
186 | unsigned long w; | |
187 | ||
188 | if (likely(lw->inv_weight)) | |
189 | return; | |
190 | ||
191 | w = scale_load_down(lw->weight); | |
192 | ||
193 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
194 | lw->inv_weight = 1; | |
195 | else if (unlikely(!w)) | |
196 | lw->inv_weight = WMULT_CONST; | |
197 | else | |
198 | lw->inv_weight = WMULT_CONST / w; | |
199 | } | |
029632fb PZ |
200 | |
201 | /* | |
9dbdb155 PZ |
202 | * delta_exec * weight / lw.weight |
203 | * OR | |
204 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
205 | * | |
206 | * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case | |
207 | * we're guaranteed shift stays positive because inv_weight is guaranteed to | |
208 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
209 | * | |
210 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
211 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 212 | */ |
9dbdb155 | 213 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 214 | { |
9dbdb155 PZ |
215 | u64 fact = scale_load_down(weight); |
216 | int shift = WMULT_SHIFT; | |
029632fb | 217 | |
9dbdb155 | 218 | __update_inv_weight(lw); |
029632fb | 219 | |
9dbdb155 PZ |
220 | if (unlikely(fact >> 32)) { |
221 | while (fact >> 32) { | |
222 | fact >>= 1; | |
223 | shift--; | |
224 | } | |
029632fb PZ |
225 | } |
226 | ||
9dbdb155 PZ |
227 | /* hint to use a 32x32->64 mul */ |
228 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 229 | |
9dbdb155 PZ |
230 | while (fact >> 32) { |
231 | fact >>= 1; | |
232 | shift--; | |
233 | } | |
029632fb | 234 | |
9dbdb155 | 235 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
236 | } |
237 | ||
238 | ||
239 | const struct sched_class fair_sched_class; | |
a4c2f00f | 240 | |
bf0f6f24 IM |
241 | /************************************************************** |
242 | * CFS operations on generic schedulable entities: | |
243 | */ | |
244 | ||
62160e3f | 245 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 246 | |
62160e3f | 247 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
248 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
249 | { | |
62160e3f | 250 | return cfs_rq->rq; |
bf0f6f24 IM |
251 | } |
252 | ||
62160e3f IM |
253 | /* An entity is a task if it doesn't "own" a runqueue */ |
254 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 255 | |
8f48894f PZ |
256 | static inline struct task_struct *task_of(struct sched_entity *se) |
257 | { | |
258 | #ifdef CONFIG_SCHED_DEBUG | |
259 | WARN_ON_ONCE(!entity_is_task(se)); | |
260 | #endif | |
261 | return container_of(se, struct task_struct, se); | |
262 | } | |
263 | ||
b758149c PZ |
264 | /* Walk up scheduling entities hierarchy */ |
265 | #define for_each_sched_entity(se) \ | |
266 | for (; se; se = se->parent) | |
267 | ||
268 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
269 | { | |
270 | return p->se.cfs_rq; | |
271 | } | |
272 | ||
273 | /* runqueue on which this entity is (to be) queued */ | |
274 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
275 | { | |
276 | return se->cfs_rq; | |
277 | } | |
278 | ||
279 | /* runqueue "owned" by this group */ | |
280 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
281 | { | |
282 | return grp->my_q; | |
283 | } | |
284 | ||
aff3e498 PT |
285 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
286 | int force_update); | |
9ee474f5 | 287 | |
3d4b47b4 PZ |
288 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
289 | { | |
290 | if (!cfs_rq->on_list) { | |
67e86250 PT |
291 | /* |
292 | * Ensure we either appear before our parent (if already | |
293 | * enqueued) or force our parent to appear after us when it is | |
294 | * enqueued. The fact that we always enqueue bottom-up | |
295 | * reduces this to two cases. | |
296 | */ | |
297 | if (cfs_rq->tg->parent && | |
298 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
299 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
300 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
301 | } else { | |
302 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 303 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 304 | } |
3d4b47b4 PZ |
305 | |
306 | cfs_rq->on_list = 1; | |
9ee474f5 | 307 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 308 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
309 | } |
310 | } | |
311 | ||
312 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
313 | { | |
314 | if (cfs_rq->on_list) { | |
315 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
316 | cfs_rq->on_list = 0; | |
317 | } | |
318 | } | |
319 | ||
b758149c PZ |
320 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
321 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
322 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
323 | ||
324 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 325 | static inline struct cfs_rq * |
b758149c PZ |
326 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
327 | { | |
328 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 329 | return se->cfs_rq; |
b758149c | 330 | |
fed14d45 | 331 | return NULL; |
b758149c PZ |
332 | } |
333 | ||
334 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
335 | { | |
336 | return se->parent; | |
337 | } | |
338 | ||
464b7527 PZ |
339 | static void |
340 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
341 | { | |
342 | int se_depth, pse_depth; | |
343 | ||
344 | /* | |
345 | * preemption test can be made between sibling entities who are in the | |
346 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
347 | * both tasks until we find their ancestors who are siblings of common | |
348 | * parent. | |
349 | */ | |
350 | ||
351 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
352 | se_depth = (*se)->depth; |
353 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
354 | |
355 | while (se_depth > pse_depth) { | |
356 | se_depth--; | |
357 | *se = parent_entity(*se); | |
358 | } | |
359 | ||
360 | while (pse_depth > se_depth) { | |
361 | pse_depth--; | |
362 | *pse = parent_entity(*pse); | |
363 | } | |
364 | ||
365 | while (!is_same_group(*se, *pse)) { | |
366 | *se = parent_entity(*se); | |
367 | *pse = parent_entity(*pse); | |
368 | } | |
369 | } | |
370 | ||
8f48894f PZ |
371 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
372 | ||
373 | static inline struct task_struct *task_of(struct sched_entity *se) | |
374 | { | |
375 | return container_of(se, struct task_struct, se); | |
376 | } | |
bf0f6f24 | 377 | |
62160e3f IM |
378 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
379 | { | |
380 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
381 | } |
382 | ||
383 | #define entity_is_task(se) 1 | |
384 | ||
b758149c PZ |
385 | #define for_each_sched_entity(se) \ |
386 | for (; se; se = NULL) | |
bf0f6f24 | 387 | |
b758149c | 388 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 389 | { |
b758149c | 390 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
391 | } |
392 | ||
b758149c PZ |
393 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
394 | { | |
395 | struct task_struct *p = task_of(se); | |
396 | struct rq *rq = task_rq(p); | |
397 | ||
398 | return &rq->cfs; | |
399 | } | |
400 | ||
401 | /* runqueue "owned" by this group */ | |
402 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
403 | { | |
404 | return NULL; | |
405 | } | |
406 | ||
3d4b47b4 PZ |
407 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
408 | { | |
409 | } | |
410 | ||
411 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
412 | { | |
413 | } | |
414 | ||
b758149c PZ |
415 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
416 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
417 | ||
b758149c PZ |
418 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
419 | { | |
420 | return NULL; | |
421 | } | |
422 | ||
464b7527 PZ |
423 | static inline void |
424 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
425 | { | |
426 | } | |
427 | ||
b758149c PZ |
428 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
429 | ||
6c16a6dc | 430 | static __always_inline |
9dbdb155 | 431 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
432 | |
433 | /************************************************************** | |
434 | * Scheduling class tree data structure manipulation methods: | |
435 | */ | |
436 | ||
1bf08230 | 437 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 438 | { |
1bf08230 | 439 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 440 | if (delta > 0) |
1bf08230 | 441 | max_vruntime = vruntime; |
02e0431a | 442 | |
1bf08230 | 443 | return max_vruntime; |
02e0431a PZ |
444 | } |
445 | ||
0702e3eb | 446 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
447 | { |
448 | s64 delta = (s64)(vruntime - min_vruntime); | |
449 | if (delta < 0) | |
450 | min_vruntime = vruntime; | |
451 | ||
452 | return min_vruntime; | |
453 | } | |
454 | ||
54fdc581 FC |
455 | static inline int entity_before(struct sched_entity *a, |
456 | struct sched_entity *b) | |
457 | { | |
458 | return (s64)(a->vruntime - b->vruntime) < 0; | |
459 | } | |
460 | ||
1af5f730 PZ |
461 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
462 | { | |
463 | u64 vruntime = cfs_rq->min_vruntime; | |
464 | ||
465 | if (cfs_rq->curr) | |
466 | vruntime = cfs_rq->curr->vruntime; | |
467 | ||
468 | if (cfs_rq->rb_leftmost) { | |
469 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
470 | struct sched_entity, | |
471 | run_node); | |
472 | ||
e17036da | 473 | if (!cfs_rq->curr) |
1af5f730 PZ |
474 | vruntime = se->vruntime; |
475 | else | |
476 | vruntime = min_vruntime(vruntime, se->vruntime); | |
477 | } | |
478 | ||
1bf08230 | 479 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 480 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
481 | #ifndef CONFIG_64BIT |
482 | smp_wmb(); | |
483 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
484 | #endif | |
1af5f730 PZ |
485 | } |
486 | ||
bf0f6f24 IM |
487 | /* |
488 | * Enqueue an entity into the rb-tree: | |
489 | */ | |
0702e3eb | 490 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
491 | { |
492 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
493 | struct rb_node *parent = NULL; | |
494 | struct sched_entity *entry; | |
bf0f6f24 IM |
495 | int leftmost = 1; |
496 | ||
497 | /* | |
498 | * Find the right place in the rbtree: | |
499 | */ | |
500 | while (*link) { | |
501 | parent = *link; | |
502 | entry = rb_entry(parent, struct sched_entity, run_node); | |
503 | /* | |
504 | * We dont care about collisions. Nodes with | |
505 | * the same key stay together. | |
506 | */ | |
2bd2d6f2 | 507 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
508 | link = &parent->rb_left; |
509 | } else { | |
510 | link = &parent->rb_right; | |
511 | leftmost = 0; | |
512 | } | |
513 | } | |
514 | ||
515 | /* | |
516 | * Maintain a cache of leftmost tree entries (it is frequently | |
517 | * used): | |
518 | */ | |
1af5f730 | 519 | if (leftmost) |
57cb499d | 520 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
521 | |
522 | rb_link_node(&se->run_node, parent, link); | |
523 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
524 | } |
525 | ||
0702e3eb | 526 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 527 | { |
3fe69747 PZ |
528 | if (cfs_rq->rb_leftmost == &se->run_node) { |
529 | struct rb_node *next_node; | |
3fe69747 PZ |
530 | |
531 | next_node = rb_next(&se->run_node); | |
532 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 533 | } |
e9acbff6 | 534 | |
bf0f6f24 | 535 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
536 | } |
537 | ||
029632fb | 538 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 539 | { |
f4b6755f PZ |
540 | struct rb_node *left = cfs_rq->rb_leftmost; |
541 | ||
542 | if (!left) | |
543 | return NULL; | |
544 | ||
545 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
546 | } |
547 | ||
ac53db59 RR |
548 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
549 | { | |
550 | struct rb_node *next = rb_next(&se->run_node); | |
551 | ||
552 | if (!next) | |
553 | return NULL; | |
554 | ||
555 | return rb_entry(next, struct sched_entity, run_node); | |
556 | } | |
557 | ||
558 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 559 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 560 | { |
7eee3e67 | 561 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 562 | |
70eee74b BS |
563 | if (!last) |
564 | return NULL; | |
7eee3e67 IM |
565 | |
566 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
567 | } |
568 | ||
bf0f6f24 IM |
569 | /************************************************************** |
570 | * Scheduling class statistics methods: | |
571 | */ | |
572 | ||
acb4a848 | 573 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 574 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
575 | loff_t *ppos) |
576 | { | |
8d65af78 | 577 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 578 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
579 | |
580 | if (ret || !write) | |
581 | return ret; | |
582 | ||
583 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
584 | sysctl_sched_min_granularity); | |
585 | ||
acb4a848 CE |
586 | #define WRT_SYSCTL(name) \ |
587 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
588 | WRT_SYSCTL(sched_min_granularity); | |
589 | WRT_SYSCTL(sched_latency); | |
590 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
591 | #undef WRT_SYSCTL |
592 | ||
b2be5e96 PZ |
593 | return 0; |
594 | } | |
595 | #endif | |
647e7cac | 596 | |
a7be37ac | 597 | /* |
f9c0b095 | 598 | * delta /= w |
a7be37ac | 599 | */ |
9dbdb155 | 600 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 601 | { |
f9c0b095 | 602 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 603 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
604 | |
605 | return delta; | |
606 | } | |
607 | ||
647e7cac IM |
608 | /* |
609 | * The idea is to set a period in which each task runs once. | |
610 | * | |
532b1858 | 611 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
612 | * this period because otherwise the slices get too small. |
613 | * | |
614 | * p = (nr <= nl) ? l : l*nr/nl | |
615 | */ | |
4d78e7b6 PZ |
616 | static u64 __sched_period(unsigned long nr_running) |
617 | { | |
618 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 619 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
620 | |
621 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 622 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 623 | period *= nr_running; |
4d78e7b6 PZ |
624 | } |
625 | ||
626 | return period; | |
627 | } | |
628 | ||
647e7cac IM |
629 | /* |
630 | * We calculate the wall-time slice from the period by taking a part | |
631 | * proportional to the weight. | |
632 | * | |
f9c0b095 | 633 | * s = p*P[w/rw] |
647e7cac | 634 | */ |
6d0f0ebd | 635 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 636 | { |
0a582440 | 637 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 638 | |
0a582440 | 639 | for_each_sched_entity(se) { |
6272d68c | 640 | struct load_weight *load; |
3104bf03 | 641 | struct load_weight lw; |
6272d68c LM |
642 | |
643 | cfs_rq = cfs_rq_of(se); | |
644 | load = &cfs_rq->load; | |
f9c0b095 | 645 | |
0a582440 | 646 | if (unlikely(!se->on_rq)) { |
3104bf03 | 647 | lw = cfs_rq->load; |
0a582440 MG |
648 | |
649 | update_load_add(&lw, se->load.weight); | |
650 | load = &lw; | |
651 | } | |
9dbdb155 | 652 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
653 | } |
654 | return slice; | |
bf0f6f24 IM |
655 | } |
656 | ||
647e7cac | 657 | /* |
660cc00f | 658 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 659 | * |
f9c0b095 | 660 | * vs = s/w |
647e7cac | 661 | */ |
f9c0b095 | 662 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 663 | { |
f9c0b095 | 664 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
665 | } |
666 | ||
a75cdaa9 | 667 | #ifdef CONFIG_SMP |
fb13c7ee MG |
668 | static unsigned long task_h_load(struct task_struct *p); |
669 | ||
a75cdaa9 AS |
670 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
671 | ||
672 | /* Give new task start runnable values to heavy its load in infant time */ | |
673 | void init_task_runnable_average(struct task_struct *p) | |
674 | { | |
675 | u32 slice; | |
676 | ||
677 | p->se.avg.decay_count = 0; | |
678 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | |
679 | p->se.avg.runnable_avg_sum = slice; | |
680 | p->se.avg.runnable_avg_period = slice; | |
681 | __update_task_entity_contrib(&p->se); | |
682 | } | |
683 | #else | |
684 | void init_task_runnable_average(struct task_struct *p) | |
685 | { | |
686 | } | |
687 | #endif | |
688 | ||
bf0f6f24 | 689 | /* |
9dbdb155 | 690 | * Update the current task's runtime statistics. |
bf0f6f24 | 691 | */ |
b7cc0896 | 692 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 693 | { |
429d43bc | 694 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 695 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 696 | u64 delta_exec; |
bf0f6f24 IM |
697 | |
698 | if (unlikely(!curr)) | |
699 | return; | |
700 | ||
9dbdb155 PZ |
701 | delta_exec = now - curr->exec_start; |
702 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 703 | return; |
bf0f6f24 | 704 | |
8ebc91d9 | 705 | curr->exec_start = now; |
d842de87 | 706 | |
9dbdb155 PZ |
707 | schedstat_set(curr->statistics.exec_max, |
708 | max(delta_exec, curr->statistics.exec_max)); | |
709 | ||
710 | curr->sum_exec_runtime += delta_exec; | |
711 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
712 | ||
713 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
714 | update_min_vruntime(cfs_rq); | |
715 | ||
d842de87 SV |
716 | if (entity_is_task(curr)) { |
717 | struct task_struct *curtask = task_of(curr); | |
718 | ||
f977bb49 | 719 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 720 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 721 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 722 | } |
ec12cb7f PT |
723 | |
724 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
725 | } |
726 | ||
727 | static inline void | |
5870db5b | 728 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 729 | { |
78becc27 | 730 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
731 | } |
732 | ||
bf0f6f24 IM |
733 | /* |
734 | * Task is being enqueued - update stats: | |
735 | */ | |
d2417e5a | 736 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 737 | { |
bf0f6f24 IM |
738 | /* |
739 | * Are we enqueueing a waiting task? (for current tasks | |
740 | * a dequeue/enqueue event is a NOP) | |
741 | */ | |
429d43bc | 742 | if (se != cfs_rq->curr) |
5870db5b | 743 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
744 | } |
745 | ||
bf0f6f24 | 746 | static void |
9ef0a961 | 747 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 748 | { |
41acab88 | 749 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 750 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
751 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
752 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 753 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
754 | #ifdef CONFIG_SCHEDSTATS |
755 | if (entity_is_task(se)) { | |
756 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 757 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
758 | } |
759 | #endif | |
41acab88 | 760 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
761 | } |
762 | ||
763 | static inline void | |
19b6a2e3 | 764 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 765 | { |
bf0f6f24 IM |
766 | /* |
767 | * Mark the end of the wait period if dequeueing a | |
768 | * waiting task: | |
769 | */ | |
429d43bc | 770 | if (se != cfs_rq->curr) |
9ef0a961 | 771 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
772 | } |
773 | ||
774 | /* | |
775 | * We are picking a new current task - update its stats: | |
776 | */ | |
777 | static inline void | |
79303e9e | 778 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
779 | { |
780 | /* | |
781 | * We are starting a new run period: | |
782 | */ | |
78becc27 | 783 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
784 | } |
785 | ||
bf0f6f24 IM |
786 | /************************************************** |
787 | * Scheduling class queueing methods: | |
788 | */ | |
789 | ||
cbee9f88 PZ |
790 | #ifdef CONFIG_NUMA_BALANCING |
791 | /* | |
598f0ec0 MG |
792 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
793 | * calculated based on the tasks virtual memory size and | |
794 | * numa_balancing_scan_size. | |
cbee9f88 | 795 | */ |
598f0ec0 MG |
796 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
797 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
798 | |
799 | /* Portion of address space to scan in MB */ | |
800 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 801 | |
4b96a29b PZ |
802 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
803 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
804 | ||
598f0ec0 MG |
805 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
806 | { | |
807 | unsigned long rss = 0; | |
808 | unsigned long nr_scan_pages; | |
809 | ||
810 | /* | |
811 | * Calculations based on RSS as non-present and empty pages are skipped | |
812 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
813 | * on resident pages | |
814 | */ | |
815 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
816 | rss = get_mm_rss(p->mm); | |
817 | if (!rss) | |
818 | rss = nr_scan_pages; | |
819 | ||
820 | rss = round_up(rss, nr_scan_pages); | |
821 | return rss / nr_scan_pages; | |
822 | } | |
823 | ||
824 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
825 | #define MAX_SCAN_WINDOW 2560 | |
826 | ||
827 | static unsigned int task_scan_min(struct task_struct *p) | |
828 | { | |
829 | unsigned int scan, floor; | |
830 | unsigned int windows = 1; | |
831 | ||
832 | if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW) | |
833 | windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size; | |
834 | floor = 1000 / windows; | |
835 | ||
836 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
837 | return max_t(unsigned int, floor, scan); | |
838 | } | |
839 | ||
840 | static unsigned int task_scan_max(struct task_struct *p) | |
841 | { | |
842 | unsigned int smin = task_scan_min(p); | |
843 | unsigned int smax; | |
844 | ||
845 | /* Watch for min being lower than max due to floor calculations */ | |
846 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
847 | return max(smin, smax); | |
848 | } | |
849 | ||
0ec8aa00 PZ |
850 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
851 | { | |
852 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
853 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
854 | } | |
855 | ||
856 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
857 | { | |
858 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
859 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
860 | } | |
861 | ||
8c8a743c PZ |
862 | struct numa_group { |
863 | atomic_t refcount; | |
864 | ||
865 | spinlock_t lock; /* nr_tasks, tasks */ | |
866 | int nr_tasks; | |
e29cf08b | 867 | pid_t gid; |
8c8a743c PZ |
868 | struct list_head task_list; |
869 | ||
870 | struct rcu_head rcu; | |
20e07dea | 871 | nodemask_t active_nodes; |
989348b5 | 872 | unsigned long total_faults; |
7e2703e6 RR |
873 | /* |
874 | * Faults_cpu is used to decide whether memory should move | |
875 | * towards the CPU. As a consequence, these stats are weighted | |
876 | * more by CPU use than by memory faults. | |
877 | */ | |
50ec8a40 | 878 | unsigned long *faults_cpu; |
989348b5 | 879 | unsigned long faults[0]; |
8c8a743c PZ |
880 | }; |
881 | ||
be1e4e76 RR |
882 | /* Shared or private faults. */ |
883 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
884 | ||
885 | /* Memory and CPU locality */ | |
886 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
887 | ||
888 | /* Averaged statistics, and temporary buffers. */ | |
889 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
890 | ||
e29cf08b MG |
891 | pid_t task_numa_group_id(struct task_struct *p) |
892 | { | |
893 | return p->numa_group ? p->numa_group->gid : 0; | |
894 | } | |
895 | ||
ac8e895b MG |
896 | static inline int task_faults_idx(int nid, int priv) |
897 | { | |
be1e4e76 | 898 | return NR_NUMA_HINT_FAULT_TYPES * nid + priv; |
ac8e895b MG |
899 | } |
900 | ||
901 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
902 | { | |
ff1df896 | 903 | if (!p->numa_faults_memory) |
ac8e895b MG |
904 | return 0; |
905 | ||
ff1df896 RR |
906 | return p->numa_faults_memory[task_faults_idx(nid, 0)] + |
907 | p->numa_faults_memory[task_faults_idx(nid, 1)]; | |
ac8e895b MG |
908 | } |
909 | ||
83e1d2cd MG |
910 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
911 | { | |
912 | if (!p->numa_group) | |
913 | return 0; | |
914 | ||
82897b4f WL |
915 | return p->numa_group->faults[task_faults_idx(nid, 0)] + |
916 | p->numa_group->faults[task_faults_idx(nid, 1)]; | |
83e1d2cd MG |
917 | } |
918 | ||
20e07dea RR |
919 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
920 | { | |
921 | return group->faults_cpu[task_faults_idx(nid, 0)] + | |
922 | group->faults_cpu[task_faults_idx(nid, 1)]; | |
923 | } | |
924 | ||
83e1d2cd MG |
925 | /* |
926 | * These return the fraction of accesses done by a particular task, or | |
927 | * task group, on a particular numa node. The group weight is given a | |
928 | * larger multiplier, in order to group tasks together that are almost | |
929 | * evenly spread out between numa nodes. | |
930 | */ | |
931 | static inline unsigned long task_weight(struct task_struct *p, int nid) | |
932 | { | |
933 | unsigned long total_faults; | |
934 | ||
ff1df896 | 935 | if (!p->numa_faults_memory) |
83e1d2cd MG |
936 | return 0; |
937 | ||
938 | total_faults = p->total_numa_faults; | |
939 | ||
940 | if (!total_faults) | |
941 | return 0; | |
942 | ||
943 | return 1000 * task_faults(p, nid) / total_faults; | |
944 | } | |
945 | ||
946 | static inline unsigned long group_weight(struct task_struct *p, int nid) | |
947 | { | |
989348b5 | 948 | if (!p->numa_group || !p->numa_group->total_faults) |
83e1d2cd MG |
949 | return 0; |
950 | ||
989348b5 | 951 | return 1000 * group_faults(p, nid) / p->numa_group->total_faults; |
83e1d2cd MG |
952 | } |
953 | ||
10f39042 RR |
954 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
955 | int src_nid, int dst_cpu) | |
956 | { | |
957 | struct numa_group *ng = p->numa_group; | |
958 | int dst_nid = cpu_to_node(dst_cpu); | |
959 | int last_cpupid, this_cpupid; | |
960 | ||
961 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
962 | ||
963 | /* | |
964 | * Multi-stage node selection is used in conjunction with a periodic | |
965 | * migration fault to build a temporal task<->page relation. By using | |
966 | * a two-stage filter we remove short/unlikely relations. | |
967 | * | |
968 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
969 | * a task's usage of a particular page (n_p) per total usage of this | |
970 | * page (n_t) (in a given time-span) to a probability. | |
971 | * | |
972 | * Our periodic faults will sample this probability and getting the | |
973 | * same result twice in a row, given these samples are fully | |
974 | * independent, is then given by P(n)^2, provided our sample period | |
975 | * is sufficiently short compared to the usage pattern. | |
976 | * | |
977 | * This quadric squishes small probabilities, making it less likely we | |
978 | * act on an unlikely task<->page relation. | |
979 | */ | |
980 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
981 | if (!cpupid_pid_unset(last_cpupid) && | |
982 | cpupid_to_nid(last_cpupid) != dst_nid) | |
983 | return false; | |
984 | ||
985 | /* Always allow migrate on private faults */ | |
986 | if (cpupid_match_pid(p, last_cpupid)) | |
987 | return true; | |
988 | ||
989 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
990 | if (!ng) | |
991 | return true; | |
992 | ||
993 | /* | |
994 | * Do not migrate if the destination is not a node that | |
995 | * is actively used by this numa group. | |
996 | */ | |
997 | if (!node_isset(dst_nid, ng->active_nodes)) | |
998 | return false; | |
999 | ||
1000 | /* | |
1001 | * Source is a node that is not actively used by this | |
1002 | * numa group, while the destination is. Migrate. | |
1003 | */ | |
1004 | if (!node_isset(src_nid, ng->active_nodes)) | |
1005 | return true; | |
1006 | ||
1007 | /* | |
1008 | * Both source and destination are nodes in active | |
1009 | * use by this numa group. Maximize memory bandwidth | |
1010 | * by migrating from more heavily used groups, to less | |
1011 | * heavily used ones, spreading the load around. | |
1012 | * Use a 1/4 hysteresis to avoid spurious page movement. | |
1013 | */ | |
1014 | return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); | |
1015 | } | |
1016 | ||
e6628d5b | 1017 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
1018 | static unsigned long source_load(int cpu, int type); |
1019 | static unsigned long target_load(int cpu, int type); | |
1020 | static unsigned long power_of(int cpu); | |
1021 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | |
1022 | ||
fb13c7ee | 1023 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1024 | struct numa_stats { |
fb13c7ee | 1025 | unsigned long nr_running; |
58d081b5 | 1026 | unsigned long load; |
fb13c7ee MG |
1027 | |
1028 | /* Total compute capacity of CPUs on a node */ | |
1029 | unsigned long power; | |
1030 | ||
1031 | /* Approximate capacity in terms of runnable tasks on a node */ | |
1032 | unsigned long capacity; | |
1033 | int has_capacity; | |
58d081b5 | 1034 | }; |
e6628d5b | 1035 | |
fb13c7ee MG |
1036 | /* |
1037 | * XXX borrowed from update_sg_lb_stats | |
1038 | */ | |
1039 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1040 | { | |
5eca82a9 | 1041 | int cpu, cpus = 0; |
fb13c7ee MG |
1042 | |
1043 | memset(ns, 0, sizeof(*ns)); | |
1044 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1045 | struct rq *rq = cpu_rq(cpu); | |
1046 | ||
1047 | ns->nr_running += rq->nr_running; | |
1048 | ns->load += weighted_cpuload(cpu); | |
1049 | ns->power += power_of(cpu); | |
5eca82a9 PZ |
1050 | |
1051 | cpus++; | |
fb13c7ee MG |
1052 | } |
1053 | ||
5eca82a9 PZ |
1054 | /* |
1055 | * If we raced with hotplug and there are no CPUs left in our mask | |
1056 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1057 | * not find this node attractive. | |
1058 | * | |
1059 | * We'll either bail at !has_capacity, or we'll detect a huge imbalance | |
1060 | * and bail there. | |
1061 | */ | |
1062 | if (!cpus) | |
1063 | return; | |
1064 | ||
fb13c7ee MG |
1065 | ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power; |
1066 | ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE); | |
1067 | ns->has_capacity = (ns->nr_running < ns->capacity); | |
1068 | } | |
1069 | ||
58d081b5 MG |
1070 | struct task_numa_env { |
1071 | struct task_struct *p; | |
e6628d5b | 1072 | |
58d081b5 MG |
1073 | int src_cpu, src_nid; |
1074 | int dst_cpu, dst_nid; | |
e6628d5b | 1075 | |
58d081b5 | 1076 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1077 | |
40ea2b42 | 1078 | int imbalance_pct; |
fb13c7ee MG |
1079 | |
1080 | struct task_struct *best_task; | |
1081 | long best_imp; | |
58d081b5 MG |
1082 | int best_cpu; |
1083 | }; | |
1084 | ||
fb13c7ee MG |
1085 | static void task_numa_assign(struct task_numa_env *env, |
1086 | struct task_struct *p, long imp) | |
1087 | { | |
1088 | if (env->best_task) | |
1089 | put_task_struct(env->best_task); | |
1090 | if (p) | |
1091 | get_task_struct(p); | |
1092 | ||
1093 | env->best_task = p; | |
1094 | env->best_imp = imp; | |
1095 | env->best_cpu = env->dst_cpu; | |
1096 | } | |
1097 | ||
1098 | /* | |
1099 | * This checks if the overall compute and NUMA accesses of the system would | |
1100 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1101 | * into account that it might be best if task running on the dst_cpu should | |
1102 | * be exchanged with the source task | |
1103 | */ | |
887c290e RR |
1104 | static void task_numa_compare(struct task_numa_env *env, |
1105 | long taskimp, long groupimp) | |
fb13c7ee MG |
1106 | { |
1107 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1108 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1109 | struct task_struct *cur; | |
1110 | long dst_load, src_load; | |
1111 | long load; | |
887c290e | 1112 | long imp = (groupimp > 0) ? groupimp : taskimp; |
fb13c7ee MG |
1113 | |
1114 | rcu_read_lock(); | |
1115 | cur = ACCESS_ONCE(dst_rq->curr); | |
1116 | if (cur->pid == 0) /* idle */ | |
1117 | cur = NULL; | |
1118 | ||
1119 | /* | |
1120 | * "imp" is the fault differential for the source task between the | |
1121 | * source and destination node. Calculate the total differential for | |
1122 | * the source task and potential destination task. The more negative | |
1123 | * the value is, the more rmeote accesses that would be expected to | |
1124 | * be incurred if the tasks were swapped. | |
1125 | */ | |
1126 | if (cur) { | |
1127 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1128 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1129 | goto unlock; | |
1130 | ||
887c290e RR |
1131 | /* |
1132 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1133 | * in any group then look only at task weights. |
887c290e | 1134 | */ |
ca28aa53 | 1135 | if (cur->numa_group == env->p->numa_group) { |
887c290e RR |
1136 | imp = taskimp + task_weight(cur, env->src_nid) - |
1137 | task_weight(cur, env->dst_nid); | |
ca28aa53 RR |
1138 | /* |
1139 | * Add some hysteresis to prevent swapping the | |
1140 | * tasks within a group over tiny differences. | |
1141 | */ | |
1142 | if (cur->numa_group) | |
1143 | imp -= imp/16; | |
887c290e | 1144 | } else { |
ca28aa53 RR |
1145 | /* |
1146 | * Compare the group weights. If a task is all by | |
1147 | * itself (not part of a group), use the task weight | |
1148 | * instead. | |
1149 | */ | |
1150 | if (env->p->numa_group) | |
1151 | imp = groupimp; | |
1152 | else | |
1153 | imp = taskimp; | |
1154 | ||
1155 | if (cur->numa_group) | |
1156 | imp += group_weight(cur, env->src_nid) - | |
1157 | group_weight(cur, env->dst_nid); | |
1158 | else | |
1159 | imp += task_weight(cur, env->src_nid) - | |
1160 | task_weight(cur, env->dst_nid); | |
887c290e | 1161 | } |
fb13c7ee MG |
1162 | } |
1163 | ||
1164 | if (imp < env->best_imp) | |
1165 | goto unlock; | |
1166 | ||
1167 | if (!cur) { | |
1168 | /* Is there capacity at our destination? */ | |
1169 | if (env->src_stats.has_capacity && | |
1170 | !env->dst_stats.has_capacity) | |
1171 | goto unlock; | |
1172 | ||
1173 | goto balance; | |
1174 | } | |
1175 | ||
1176 | /* Balance doesn't matter much if we're running a task per cpu */ | |
1177 | if (src_rq->nr_running == 1 && dst_rq->nr_running == 1) | |
1178 | goto assign; | |
1179 | ||
1180 | /* | |
1181 | * In the overloaded case, try and keep the load balanced. | |
1182 | */ | |
1183 | balance: | |
1184 | dst_load = env->dst_stats.load; | |
1185 | src_load = env->src_stats.load; | |
1186 | ||
1187 | /* XXX missing power terms */ | |
1188 | load = task_h_load(env->p); | |
1189 | dst_load += load; | |
1190 | src_load -= load; | |
1191 | ||
1192 | if (cur) { | |
1193 | load = task_h_load(cur); | |
1194 | dst_load -= load; | |
1195 | src_load += load; | |
1196 | } | |
1197 | ||
1198 | /* make src_load the smaller */ | |
1199 | if (dst_load < src_load) | |
1200 | swap(dst_load, src_load); | |
1201 | ||
1202 | if (src_load * env->imbalance_pct < dst_load * 100) | |
1203 | goto unlock; | |
1204 | ||
1205 | assign: | |
1206 | task_numa_assign(env, cur, imp); | |
1207 | unlock: | |
1208 | rcu_read_unlock(); | |
1209 | } | |
1210 | ||
887c290e RR |
1211 | static void task_numa_find_cpu(struct task_numa_env *env, |
1212 | long taskimp, long groupimp) | |
2c8a50aa MG |
1213 | { |
1214 | int cpu; | |
1215 | ||
1216 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1217 | /* Skip this CPU if the source task cannot migrate */ | |
1218 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1219 | continue; | |
1220 | ||
1221 | env->dst_cpu = cpu; | |
887c290e | 1222 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1223 | } |
1224 | } | |
1225 | ||
58d081b5 MG |
1226 | static int task_numa_migrate(struct task_struct *p) |
1227 | { | |
58d081b5 MG |
1228 | struct task_numa_env env = { |
1229 | .p = p, | |
fb13c7ee | 1230 | |
58d081b5 | 1231 | .src_cpu = task_cpu(p), |
b32e86b4 | 1232 | .src_nid = task_node(p), |
fb13c7ee MG |
1233 | |
1234 | .imbalance_pct = 112, | |
1235 | ||
1236 | .best_task = NULL, | |
1237 | .best_imp = 0, | |
1238 | .best_cpu = -1 | |
58d081b5 MG |
1239 | }; |
1240 | struct sched_domain *sd; | |
887c290e | 1241 | unsigned long taskweight, groupweight; |
2c8a50aa | 1242 | int nid, ret; |
887c290e | 1243 | long taskimp, groupimp; |
e6628d5b | 1244 | |
58d081b5 | 1245 | /* |
fb13c7ee MG |
1246 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1247 | * imbalance and would be the first to start moving tasks about. | |
1248 | * | |
1249 | * And we want to avoid any moving of tasks about, as that would create | |
1250 | * random movement of tasks -- counter the numa conditions we're trying | |
1251 | * to satisfy here. | |
58d081b5 MG |
1252 | */ |
1253 | rcu_read_lock(); | |
fb13c7ee | 1254 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1255 | if (sd) |
1256 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1257 | rcu_read_unlock(); |
1258 | ||
46a73e8a RR |
1259 | /* |
1260 | * Cpusets can break the scheduler domain tree into smaller | |
1261 | * balance domains, some of which do not cross NUMA boundaries. | |
1262 | * Tasks that are "trapped" in such domains cannot be migrated | |
1263 | * elsewhere, so there is no point in (re)trying. | |
1264 | */ | |
1265 | if (unlikely(!sd)) { | |
de1b301a | 1266 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1267 | return -EINVAL; |
1268 | } | |
1269 | ||
887c290e RR |
1270 | taskweight = task_weight(p, env.src_nid); |
1271 | groupweight = group_weight(p, env.src_nid); | |
fb13c7ee | 1272 | update_numa_stats(&env.src_stats, env.src_nid); |
2c8a50aa | 1273 | env.dst_nid = p->numa_preferred_nid; |
887c290e RR |
1274 | taskimp = task_weight(p, env.dst_nid) - taskweight; |
1275 | groupimp = group_weight(p, env.dst_nid) - groupweight; | |
2c8a50aa | 1276 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1277 | |
e1dda8a7 RR |
1278 | /* If the preferred nid has capacity, try to use it. */ |
1279 | if (env.dst_stats.has_capacity) | |
887c290e | 1280 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 RR |
1281 | |
1282 | /* No space available on the preferred nid. Look elsewhere. */ | |
1283 | if (env.best_cpu == -1) { | |
2c8a50aa MG |
1284 | for_each_online_node(nid) { |
1285 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1286 | continue; | |
58d081b5 | 1287 | |
83e1d2cd | 1288 | /* Only consider nodes where both task and groups benefit */ |
887c290e RR |
1289 | taskimp = task_weight(p, nid) - taskweight; |
1290 | groupimp = group_weight(p, nid) - groupweight; | |
1291 | if (taskimp < 0 && groupimp < 0) | |
fb13c7ee MG |
1292 | continue; |
1293 | ||
2c8a50aa MG |
1294 | env.dst_nid = nid; |
1295 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
887c290e | 1296 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1297 | } |
1298 | } | |
1299 | ||
fb13c7ee MG |
1300 | /* No better CPU than the current one was found. */ |
1301 | if (env.best_cpu == -1) | |
1302 | return -EAGAIN; | |
1303 | ||
0ec8aa00 PZ |
1304 | sched_setnuma(p, env.dst_nid); |
1305 | ||
04bb2f94 RR |
1306 | /* |
1307 | * Reset the scan period if the task is being rescheduled on an | |
1308 | * alternative node to recheck if the tasks is now properly placed. | |
1309 | */ | |
1310 | p->numa_scan_period = task_scan_min(p); | |
1311 | ||
fb13c7ee | 1312 | if (env.best_task == NULL) { |
286549dc MG |
1313 | ret = migrate_task_to(p, env.best_cpu); |
1314 | if (ret != 0) | |
1315 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1316 | return ret; |
1317 | } | |
1318 | ||
1319 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1320 | if (ret != 0) |
1321 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1322 | put_task_struct(env.best_task); |
1323 | return ret; | |
e6628d5b MG |
1324 | } |
1325 | ||
6b9a7460 MG |
1326 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1327 | static void numa_migrate_preferred(struct task_struct *p) | |
1328 | { | |
2739d3ee | 1329 | /* This task has no NUMA fault statistics yet */ |
ff1df896 | 1330 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory)) |
6b9a7460 MG |
1331 | return; |
1332 | ||
2739d3ee RR |
1333 | /* Periodically retry migrating the task to the preferred node */ |
1334 | p->numa_migrate_retry = jiffies + HZ; | |
1335 | ||
1336 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1337 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1338 | return; |
1339 | ||
1340 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1341 | task_numa_migrate(p); |
6b9a7460 MG |
1342 | } |
1343 | ||
20e07dea RR |
1344 | /* |
1345 | * Find the nodes on which the workload is actively running. We do this by | |
1346 | * tracking the nodes from which NUMA hinting faults are triggered. This can | |
1347 | * be different from the set of nodes where the workload's memory is currently | |
1348 | * located. | |
1349 | * | |
1350 | * The bitmask is used to make smarter decisions on when to do NUMA page | |
1351 | * migrations, To prevent flip-flopping, and excessive page migrations, nodes | |
1352 | * are added when they cause over 6/16 of the maximum number of faults, but | |
1353 | * only removed when they drop below 3/16. | |
1354 | */ | |
1355 | static void update_numa_active_node_mask(struct numa_group *numa_group) | |
1356 | { | |
1357 | unsigned long faults, max_faults = 0; | |
1358 | int nid; | |
1359 | ||
1360 | for_each_online_node(nid) { | |
1361 | faults = group_faults_cpu(numa_group, nid); | |
1362 | if (faults > max_faults) | |
1363 | max_faults = faults; | |
1364 | } | |
1365 | ||
1366 | for_each_online_node(nid) { | |
1367 | faults = group_faults_cpu(numa_group, nid); | |
1368 | if (!node_isset(nid, numa_group->active_nodes)) { | |
1369 | if (faults > max_faults * 6 / 16) | |
1370 | node_set(nid, numa_group->active_nodes); | |
1371 | } else if (faults < max_faults * 3 / 16) | |
1372 | node_clear(nid, numa_group->active_nodes); | |
1373 | } | |
1374 | } | |
1375 | ||
04bb2f94 RR |
1376 | /* |
1377 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1378 | * increments. The more local the fault statistics are, the higher the scan | |
1379 | * period will be for the next scan window. If local/remote ratio is below | |
1380 | * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the | |
1381 | * scan period will decrease | |
1382 | */ | |
1383 | #define NUMA_PERIOD_SLOTS 10 | |
1384 | #define NUMA_PERIOD_THRESHOLD 3 | |
1385 | ||
1386 | /* | |
1387 | * Increase the scan period (slow down scanning) if the majority of | |
1388 | * our memory is already on our local node, or if the majority of | |
1389 | * the page accesses are shared with other processes. | |
1390 | * Otherwise, decrease the scan period. | |
1391 | */ | |
1392 | static void update_task_scan_period(struct task_struct *p, | |
1393 | unsigned long shared, unsigned long private) | |
1394 | { | |
1395 | unsigned int period_slot; | |
1396 | int ratio; | |
1397 | int diff; | |
1398 | ||
1399 | unsigned long remote = p->numa_faults_locality[0]; | |
1400 | unsigned long local = p->numa_faults_locality[1]; | |
1401 | ||
1402 | /* | |
1403 | * If there were no record hinting faults then either the task is | |
1404 | * completely idle or all activity is areas that are not of interest | |
1405 | * to automatic numa balancing. Scan slower | |
1406 | */ | |
1407 | if (local + shared == 0) { | |
1408 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1409 | p->numa_scan_period << 1); | |
1410 | ||
1411 | p->mm->numa_next_scan = jiffies + | |
1412 | msecs_to_jiffies(p->numa_scan_period); | |
1413 | ||
1414 | return; | |
1415 | } | |
1416 | ||
1417 | /* | |
1418 | * Prepare to scale scan period relative to the current period. | |
1419 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1420 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1421 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1422 | */ | |
1423 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1424 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1425 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1426 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1427 | if (!slot) | |
1428 | slot = 1; | |
1429 | diff = slot * period_slot; | |
1430 | } else { | |
1431 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1432 | ||
1433 | /* | |
1434 | * Scale scan rate increases based on sharing. There is an | |
1435 | * inverse relationship between the degree of sharing and | |
1436 | * the adjustment made to the scanning period. Broadly | |
1437 | * speaking the intent is that there is little point | |
1438 | * scanning faster if shared accesses dominate as it may | |
1439 | * simply bounce migrations uselessly | |
1440 | */ | |
04bb2f94 RR |
1441 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared)); |
1442 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; | |
1443 | } | |
1444 | ||
1445 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1446 | task_scan_min(p), task_scan_max(p)); | |
1447 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1448 | } | |
1449 | ||
7e2703e6 RR |
1450 | /* |
1451 | * Get the fraction of time the task has been running since the last | |
1452 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1453 | * decays those on a 32ms period, which is orders of magnitude off | |
1454 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1455 | * stats only if the task is so new there are no NUMA statistics yet. | |
1456 | */ | |
1457 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1458 | { | |
1459 | u64 runtime, delta, now; | |
1460 | /* Use the start of this time slice to avoid calculations. */ | |
1461 | now = p->se.exec_start; | |
1462 | runtime = p->se.sum_exec_runtime; | |
1463 | ||
1464 | if (p->last_task_numa_placement) { | |
1465 | delta = runtime - p->last_sum_exec_runtime; | |
1466 | *period = now - p->last_task_numa_placement; | |
1467 | } else { | |
1468 | delta = p->se.avg.runnable_avg_sum; | |
1469 | *period = p->se.avg.runnable_avg_period; | |
1470 | } | |
1471 | ||
1472 | p->last_sum_exec_runtime = runtime; | |
1473 | p->last_task_numa_placement = now; | |
1474 | ||
1475 | return delta; | |
1476 | } | |
1477 | ||
cbee9f88 PZ |
1478 | static void task_numa_placement(struct task_struct *p) |
1479 | { | |
83e1d2cd MG |
1480 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1481 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 1482 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
1483 | unsigned long total_faults; |
1484 | u64 runtime, period; | |
7dbd13ed | 1485 | spinlock_t *group_lock = NULL; |
cbee9f88 | 1486 | |
2832bc19 | 1487 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1488 | if (p->numa_scan_seq == seq) |
1489 | return; | |
1490 | p->numa_scan_seq = seq; | |
598f0ec0 | 1491 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1492 | |
7e2703e6 RR |
1493 | total_faults = p->numa_faults_locality[0] + |
1494 | p->numa_faults_locality[1]; | |
1495 | runtime = numa_get_avg_runtime(p, &period); | |
1496 | ||
7dbd13ed MG |
1497 | /* If the task is part of a group prevent parallel updates to group stats */ |
1498 | if (p->numa_group) { | |
1499 | group_lock = &p->numa_group->lock; | |
1500 | spin_lock(group_lock); | |
1501 | } | |
1502 | ||
688b7585 MG |
1503 | /* Find the node with the highest number of faults */ |
1504 | for_each_online_node(nid) { | |
83e1d2cd | 1505 | unsigned long faults = 0, group_faults = 0; |
ac8e895b | 1506 | int priv, i; |
745d6147 | 1507 | |
be1e4e76 | 1508 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 1509 | long diff, f_diff, f_weight; |
8c8a743c | 1510 | |
ac8e895b | 1511 | i = task_faults_idx(nid, priv); |
745d6147 | 1512 | |
ac8e895b | 1513 | /* Decay existing window, copy faults since last scan */ |
35664fd4 | 1514 | diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2; |
ff1df896 RR |
1515 | fault_types[priv] += p->numa_faults_buffer_memory[i]; |
1516 | p->numa_faults_buffer_memory[i] = 0; | |
fb13c7ee | 1517 | |
7e2703e6 RR |
1518 | /* |
1519 | * Normalize the faults_from, so all tasks in a group | |
1520 | * count according to CPU use, instead of by the raw | |
1521 | * number of faults. Tasks with little runtime have | |
1522 | * little over-all impact on throughput, and thus their | |
1523 | * faults are less important. | |
1524 | */ | |
1525 | f_weight = div64_u64(runtime << 16, period + 1); | |
1526 | f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) / | |
1527 | (total_faults + 1); | |
35664fd4 | 1528 | f_diff = f_weight - p->numa_faults_cpu[i] / 2; |
50ec8a40 RR |
1529 | p->numa_faults_buffer_cpu[i] = 0; |
1530 | ||
35664fd4 RR |
1531 | p->numa_faults_memory[i] += diff; |
1532 | p->numa_faults_cpu[i] += f_diff; | |
ff1df896 | 1533 | faults += p->numa_faults_memory[i]; |
83e1d2cd | 1534 | p->total_numa_faults += diff; |
8c8a743c PZ |
1535 | if (p->numa_group) { |
1536 | /* safe because we can only change our own group */ | |
989348b5 | 1537 | p->numa_group->faults[i] += diff; |
50ec8a40 | 1538 | p->numa_group->faults_cpu[i] += f_diff; |
989348b5 MG |
1539 | p->numa_group->total_faults += diff; |
1540 | group_faults += p->numa_group->faults[i]; | |
8c8a743c | 1541 | } |
ac8e895b MG |
1542 | } |
1543 | ||
688b7585 MG |
1544 | if (faults > max_faults) { |
1545 | max_faults = faults; | |
1546 | max_nid = nid; | |
1547 | } | |
83e1d2cd MG |
1548 | |
1549 | if (group_faults > max_group_faults) { | |
1550 | max_group_faults = group_faults; | |
1551 | max_group_nid = nid; | |
1552 | } | |
1553 | } | |
1554 | ||
04bb2f94 RR |
1555 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1556 | ||
7dbd13ed | 1557 | if (p->numa_group) { |
20e07dea | 1558 | update_numa_active_node_mask(p->numa_group); |
7dbd13ed MG |
1559 | /* |
1560 | * If the preferred task and group nids are different, | |
1561 | * iterate over the nodes again to find the best place. | |
1562 | */ | |
1563 | if (max_nid != max_group_nid) { | |
1564 | unsigned long weight, max_weight = 0; | |
1565 | ||
1566 | for_each_online_node(nid) { | |
1567 | weight = task_weight(p, nid) + group_weight(p, nid); | |
1568 | if (weight > max_weight) { | |
1569 | max_weight = weight; | |
1570 | max_nid = nid; | |
1571 | } | |
83e1d2cd MG |
1572 | } |
1573 | } | |
7dbd13ed MG |
1574 | |
1575 | spin_unlock(group_lock); | |
688b7585 MG |
1576 | } |
1577 | ||
6b9a7460 | 1578 | /* Preferred node as the node with the most faults */ |
3a7053b3 | 1579 | if (max_faults && max_nid != p->numa_preferred_nid) { |
e6628d5b | 1580 | /* Update the preferred nid and migrate task if possible */ |
0ec8aa00 | 1581 | sched_setnuma(p, max_nid); |
6b9a7460 | 1582 | numa_migrate_preferred(p); |
3a7053b3 | 1583 | } |
cbee9f88 PZ |
1584 | } |
1585 | ||
8c8a743c PZ |
1586 | static inline int get_numa_group(struct numa_group *grp) |
1587 | { | |
1588 | return atomic_inc_not_zero(&grp->refcount); | |
1589 | } | |
1590 | ||
1591 | static inline void put_numa_group(struct numa_group *grp) | |
1592 | { | |
1593 | if (atomic_dec_and_test(&grp->refcount)) | |
1594 | kfree_rcu(grp, rcu); | |
1595 | } | |
1596 | ||
3e6a9418 MG |
1597 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1598 | int *priv) | |
8c8a743c PZ |
1599 | { |
1600 | struct numa_group *grp, *my_grp; | |
1601 | struct task_struct *tsk; | |
1602 | bool join = false; | |
1603 | int cpu = cpupid_to_cpu(cpupid); | |
1604 | int i; | |
1605 | ||
1606 | if (unlikely(!p->numa_group)) { | |
1607 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 1608 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
1609 | |
1610 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1611 | if (!grp) | |
1612 | return; | |
1613 | ||
1614 | atomic_set(&grp->refcount, 1); | |
1615 | spin_lock_init(&grp->lock); | |
1616 | INIT_LIST_HEAD(&grp->task_list); | |
e29cf08b | 1617 | grp->gid = p->pid; |
50ec8a40 | 1618 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
1619 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
1620 | nr_node_ids; | |
8c8a743c | 1621 | |
20e07dea RR |
1622 | node_set(task_node(current), grp->active_nodes); |
1623 | ||
be1e4e76 | 1624 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
ff1df896 | 1625 | grp->faults[i] = p->numa_faults_memory[i]; |
8c8a743c | 1626 | |
989348b5 | 1627 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 1628 | |
8c8a743c PZ |
1629 | list_add(&p->numa_entry, &grp->task_list); |
1630 | grp->nr_tasks++; | |
1631 | rcu_assign_pointer(p->numa_group, grp); | |
1632 | } | |
1633 | ||
1634 | rcu_read_lock(); | |
1635 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1636 | ||
1637 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 1638 | goto no_join; |
8c8a743c PZ |
1639 | |
1640 | grp = rcu_dereference(tsk->numa_group); | |
1641 | if (!grp) | |
3354781a | 1642 | goto no_join; |
8c8a743c PZ |
1643 | |
1644 | my_grp = p->numa_group; | |
1645 | if (grp == my_grp) | |
3354781a | 1646 | goto no_join; |
8c8a743c PZ |
1647 | |
1648 | /* | |
1649 | * Only join the other group if its bigger; if we're the bigger group, | |
1650 | * the other task will join us. | |
1651 | */ | |
1652 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 1653 | goto no_join; |
8c8a743c PZ |
1654 | |
1655 | /* | |
1656 | * Tie-break on the grp address. | |
1657 | */ | |
1658 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 1659 | goto no_join; |
8c8a743c | 1660 | |
dabe1d99 RR |
1661 | /* Always join threads in the same process. */ |
1662 | if (tsk->mm == current->mm) | |
1663 | join = true; | |
1664 | ||
1665 | /* Simple filter to avoid false positives due to PID collisions */ | |
1666 | if (flags & TNF_SHARED) | |
1667 | join = true; | |
8c8a743c | 1668 | |
3e6a9418 MG |
1669 | /* Update priv based on whether false sharing was detected */ |
1670 | *priv = !join; | |
1671 | ||
dabe1d99 | 1672 | if (join && !get_numa_group(grp)) |
3354781a | 1673 | goto no_join; |
8c8a743c | 1674 | |
8c8a743c PZ |
1675 | rcu_read_unlock(); |
1676 | ||
1677 | if (!join) | |
1678 | return; | |
1679 | ||
989348b5 MG |
1680 | double_lock(&my_grp->lock, &grp->lock); |
1681 | ||
be1e4e76 | 1682 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
ff1df896 RR |
1683 | my_grp->faults[i] -= p->numa_faults_memory[i]; |
1684 | grp->faults[i] += p->numa_faults_memory[i]; | |
8c8a743c | 1685 | } |
989348b5 MG |
1686 | my_grp->total_faults -= p->total_numa_faults; |
1687 | grp->total_faults += p->total_numa_faults; | |
8c8a743c PZ |
1688 | |
1689 | list_move(&p->numa_entry, &grp->task_list); | |
1690 | my_grp->nr_tasks--; | |
1691 | grp->nr_tasks++; | |
1692 | ||
1693 | spin_unlock(&my_grp->lock); | |
1694 | spin_unlock(&grp->lock); | |
1695 | ||
1696 | rcu_assign_pointer(p->numa_group, grp); | |
1697 | ||
1698 | put_numa_group(my_grp); | |
3354781a PZ |
1699 | return; |
1700 | ||
1701 | no_join: | |
1702 | rcu_read_unlock(); | |
1703 | return; | |
8c8a743c PZ |
1704 | } |
1705 | ||
1706 | void task_numa_free(struct task_struct *p) | |
1707 | { | |
1708 | struct numa_group *grp = p->numa_group; | |
1709 | int i; | |
ff1df896 | 1710 | void *numa_faults = p->numa_faults_memory; |
8c8a743c PZ |
1711 | |
1712 | if (grp) { | |
989348b5 | 1713 | spin_lock(&grp->lock); |
be1e4e76 | 1714 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
ff1df896 | 1715 | grp->faults[i] -= p->numa_faults_memory[i]; |
989348b5 | 1716 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 1717 | |
8c8a743c PZ |
1718 | list_del(&p->numa_entry); |
1719 | grp->nr_tasks--; | |
1720 | spin_unlock(&grp->lock); | |
1721 | rcu_assign_pointer(p->numa_group, NULL); | |
1722 | put_numa_group(grp); | |
1723 | } | |
1724 | ||
ff1df896 RR |
1725 | p->numa_faults_memory = NULL; |
1726 | p->numa_faults_buffer_memory = NULL; | |
50ec8a40 RR |
1727 | p->numa_faults_cpu= NULL; |
1728 | p->numa_faults_buffer_cpu = NULL; | |
82727018 | 1729 | kfree(numa_faults); |
8c8a743c PZ |
1730 | } |
1731 | ||
cbee9f88 PZ |
1732 | /* |
1733 | * Got a PROT_NONE fault for a page on @node. | |
1734 | */ | |
58b46da3 | 1735 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
1736 | { |
1737 | struct task_struct *p = current; | |
6688cc05 | 1738 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 1739 | int cpu_node = task_node(current); |
ac8e895b | 1740 | int priv; |
cbee9f88 | 1741 | |
10e84b97 | 1742 | if (!numabalancing_enabled) |
1a687c2e MG |
1743 | return; |
1744 | ||
9ff1d9ff MG |
1745 | /* for example, ksmd faulting in a user's mm */ |
1746 | if (!p->mm) | |
1747 | return; | |
1748 | ||
82727018 RR |
1749 | /* Do not worry about placement if exiting */ |
1750 | if (p->state == TASK_DEAD) | |
1751 | return; | |
1752 | ||
f809ca9a | 1753 | /* Allocate buffer to track faults on a per-node basis */ |
ff1df896 | 1754 | if (unlikely(!p->numa_faults_memory)) { |
be1e4e76 RR |
1755 | int size = sizeof(*p->numa_faults_memory) * |
1756 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; | |
f809ca9a | 1757 | |
be1e4e76 | 1758 | p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
ff1df896 | 1759 | if (!p->numa_faults_memory) |
f809ca9a | 1760 | return; |
745d6147 | 1761 | |
ff1df896 | 1762 | BUG_ON(p->numa_faults_buffer_memory); |
be1e4e76 RR |
1763 | /* |
1764 | * The averaged statistics, shared & private, memory & cpu, | |
1765 | * occupy the first half of the array. The second half of the | |
1766 | * array is for current counters, which are averaged into the | |
1767 | * first set by task_numa_placement. | |
1768 | */ | |
50ec8a40 RR |
1769 | p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids); |
1770 | p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids); | |
1771 | p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids); | |
83e1d2cd | 1772 | p->total_numa_faults = 0; |
04bb2f94 | 1773 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 1774 | } |
cbee9f88 | 1775 | |
8c8a743c PZ |
1776 | /* |
1777 | * First accesses are treated as private, otherwise consider accesses | |
1778 | * to be private if the accessing pid has not changed | |
1779 | */ | |
1780 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
1781 | priv = 1; | |
1782 | } else { | |
1783 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 1784 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 1785 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
1786 | } |
1787 | ||
cbee9f88 | 1788 | task_numa_placement(p); |
f809ca9a | 1789 | |
2739d3ee RR |
1790 | /* |
1791 | * Retry task to preferred node migration periodically, in case it | |
1792 | * case it previously failed, or the scheduler moved us. | |
1793 | */ | |
1794 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
1795 | numa_migrate_preferred(p); |
1796 | ||
b32e86b4 IM |
1797 | if (migrated) |
1798 | p->numa_pages_migrated += pages; | |
1799 | ||
58b46da3 RR |
1800 | p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages; |
1801 | p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages; | |
04bb2f94 | 1802 | p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages; |
cbee9f88 PZ |
1803 | } |
1804 | ||
6e5fb223 PZ |
1805 | static void reset_ptenuma_scan(struct task_struct *p) |
1806 | { | |
1807 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
1808 | p->mm->numa_scan_offset = 0; | |
1809 | } | |
1810 | ||
cbee9f88 PZ |
1811 | /* |
1812 | * The expensive part of numa migration is done from task_work context. | |
1813 | * Triggered from task_tick_numa(). | |
1814 | */ | |
1815 | void task_numa_work(struct callback_head *work) | |
1816 | { | |
1817 | unsigned long migrate, next_scan, now = jiffies; | |
1818 | struct task_struct *p = current; | |
1819 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 1820 | struct vm_area_struct *vma; |
9f40604c | 1821 | unsigned long start, end; |
598f0ec0 | 1822 | unsigned long nr_pte_updates = 0; |
9f40604c | 1823 | long pages; |
cbee9f88 PZ |
1824 | |
1825 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
1826 | ||
1827 | work->next = work; /* protect against double add */ | |
1828 | /* | |
1829 | * Who cares about NUMA placement when they're dying. | |
1830 | * | |
1831 | * NOTE: make sure not to dereference p->mm before this check, | |
1832 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
1833 | * without p->mm even though we still had it when we enqueued this | |
1834 | * work. | |
1835 | */ | |
1836 | if (p->flags & PF_EXITING) | |
1837 | return; | |
1838 | ||
930aa174 | 1839 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
1840 | mm->numa_next_scan = now + |
1841 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
1842 | } |
1843 | ||
cbee9f88 PZ |
1844 | /* |
1845 | * Enforce maximal scan/migration frequency.. | |
1846 | */ | |
1847 | migrate = mm->numa_next_scan; | |
1848 | if (time_before(now, migrate)) | |
1849 | return; | |
1850 | ||
598f0ec0 MG |
1851 | if (p->numa_scan_period == 0) { |
1852 | p->numa_scan_period_max = task_scan_max(p); | |
1853 | p->numa_scan_period = task_scan_min(p); | |
1854 | } | |
cbee9f88 | 1855 | |
fb003b80 | 1856 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
1857 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
1858 | return; | |
1859 | ||
19a78d11 PZ |
1860 | /* |
1861 | * Delay this task enough that another task of this mm will likely win | |
1862 | * the next time around. | |
1863 | */ | |
1864 | p->node_stamp += 2 * TICK_NSEC; | |
1865 | ||
9f40604c MG |
1866 | start = mm->numa_scan_offset; |
1867 | pages = sysctl_numa_balancing_scan_size; | |
1868 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
1869 | if (!pages) | |
1870 | return; | |
cbee9f88 | 1871 | |
6e5fb223 | 1872 | down_read(&mm->mmap_sem); |
9f40604c | 1873 | vma = find_vma(mm, start); |
6e5fb223 PZ |
1874 | if (!vma) { |
1875 | reset_ptenuma_scan(p); | |
9f40604c | 1876 | start = 0; |
6e5fb223 PZ |
1877 | vma = mm->mmap; |
1878 | } | |
9f40604c | 1879 | for (; vma; vma = vma->vm_next) { |
fc314724 | 1880 | if (!vma_migratable(vma) || !vma_policy_mof(p, vma)) |
6e5fb223 PZ |
1881 | continue; |
1882 | ||
4591ce4f MG |
1883 | /* |
1884 | * Shared library pages mapped by multiple processes are not | |
1885 | * migrated as it is expected they are cache replicated. Avoid | |
1886 | * hinting faults in read-only file-backed mappings or the vdso | |
1887 | * as migrating the pages will be of marginal benefit. | |
1888 | */ | |
1889 | if (!vma->vm_mm || | |
1890 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
1891 | continue; | |
1892 | ||
3c67f474 MG |
1893 | /* |
1894 | * Skip inaccessible VMAs to avoid any confusion between | |
1895 | * PROT_NONE and NUMA hinting ptes | |
1896 | */ | |
1897 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
1898 | continue; | |
4591ce4f | 1899 | |
9f40604c MG |
1900 | do { |
1901 | start = max(start, vma->vm_start); | |
1902 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
1903 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
1904 | nr_pte_updates += change_prot_numa(vma, start, end); |
1905 | ||
1906 | /* | |
1907 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
1908 | * at least one PTE is updated so that unused virtual | |
1909 | * address space is quickly skipped. | |
1910 | */ | |
1911 | if (nr_pte_updates) | |
1912 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 1913 | |
9f40604c MG |
1914 | start = end; |
1915 | if (pages <= 0) | |
1916 | goto out; | |
1917 | } while (end != vma->vm_end); | |
cbee9f88 | 1918 | } |
6e5fb223 | 1919 | |
9f40604c | 1920 | out: |
6e5fb223 | 1921 | /* |
c69307d5 PZ |
1922 | * It is possible to reach the end of the VMA list but the last few |
1923 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
1924 | * would find the !migratable VMA on the next scan but not reset the | |
1925 | * scanner to the start so check it now. | |
6e5fb223 PZ |
1926 | */ |
1927 | if (vma) | |
9f40604c | 1928 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
1929 | else |
1930 | reset_ptenuma_scan(p); | |
1931 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
1932 | } |
1933 | ||
1934 | /* | |
1935 | * Drive the periodic memory faults.. | |
1936 | */ | |
1937 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1938 | { | |
1939 | struct callback_head *work = &curr->numa_work; | |
1940 | u64 period, now; | |
1941 | ||
1942 | /* | |
1943 | * We don't care about NUMA placement if we don't have memory. | |
1944 | */ | |
1945 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
1946 | return; | |
1947 | ||
1948 | /* | |
1949 | * Using runtime rather than walltime has the dual advantage that | |
1950 | * we (mostly) drive the selection from busy threads and that the | |
1951 | * task needs to have done some actual work before we bother with | |
1952 | * NUMA placement. | |
1953 | */ | |
1954 | now = curr->se.sum_exec_runtime; | |
1955 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
1956 | ||
1957 | if (now - curr->node_stamp > period) { | |
4b96a29b | 1958 | if (!curr->node_stamp) |
598f0ec0 | 1959 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 1960 | curr->node_stamp += period; |
cbee9f88 PZ |
1961 | |
1962 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
1963 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
1964 | task_work_add(curr, work, true); | |
1965 | } | |
1966 | } | |
1967 | } | |
1968 | #else | |
1969 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
1970 | { | |
1971 | } | |
0ec8aa00 PZ |
1972 | |
1973 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
1974 | { | |
1975 | } | |
1976 | ||
1977 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1978 | { | |
1979 | } | |
cbee9f88 PZ |
1980 | #endif /* CONFIG_NUMA_BALANCING */ |
1981 | ||
30cfdcfc DA |
1982 | static void |
1983 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
1984 | { | |
1985 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 1986 | if (!parent_entity(se)) |
029632fb | 1987 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 1988 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
1989 | if (entity_is_task(se)) { |
1990 | struct rq *rq = rq_of(cfs_rq); | |
1991 | ||
1992 | account_numa_enqueue(rq, task_of(se)); | |
1993 | list_add(&se->group_node, &rq->cfs_tasks); | |
1994 | } | |
367456c7 | 1995 | #endif |
30cfdcfc | 1996 | cfs_rq->nr_running++; |
30cfdcfc DA |
1997 | } |
1998 | ||
1999 | static void | |
2000 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2001 | { | |
2002 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2003 | if (!parent_entity(se)) |
029632fb | 2004 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
0ec8aa00 PZ |
2005 | if (entity_is_task(se)) { |
2006 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2007 | list_del_init(&se->group_node); |
0ec8aa00 | 2008 | } |
30cfdcfc | 2009 | cfs_rq->nr_running--; |
30cfdcfc DA |
2010 | } |
2011 | ||
3ff6dcac YZ |
2012 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2013 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
2014 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
2015 | { | |
2016 | long tg_weight; | |
2017 | ||
2018 | /* | |
2019 | * Use this CPU's actual weight instead of the last load_contribution | |
2020 | * to gain a more accurate current total weight. See | |
2021 | * update_cfs_rq_load_contribution(). | |
2022 | */ | |
bf5b986e | 2023 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 2024 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
2025 | tg_weight += cfs_rq->load.weight; |
2026 | ||
2027 | return tg_weight; | |
2028 | } | |
2029 | ||
6d5ab293 | 2030 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 2031 | { |
cf5f0acf | 2032 | long tg_weight, load, shares; |
3ff6dcac | 2033 | |
cf5f0acf | 2034 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 2035 | load = cfs_rq->load.weight; |
3ff6dcac | 2036 | |
3ff6dcac | 2037 | shares = (tg->shares * load); |
cf5f0acf PZ |
2038 | if (tg_weight) |
2039 | shares /= tg_weight; | |
3ff6dcac YZ |
2040 | |
2041 | if (shares < MIN_SHARES) | |
2042 | shares = MIN_SHARES; | |
2043 | if (shares > tg->shares) | |
2044 | shares = tg->shares; | |
2045 | ||
2046 | return shares; | |
2047 | } | |
3ff6dcac | 2048 | # else /* CONFIG_SMP */ |
6d5ab293 | 2049 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
2050 | { |
2051 | return tg->shares; | |
2052 | } | |
3ff6dcac | 2053 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
2054 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2055 | unsigned long weight) | |
2056 | { | |
19e5eebb PT |
2057 | if (se->on_rq) { |
2058 | /* commit outstanding execution time */ | |
2059 | if (cfs_rq->curr == se) | |
2060 | update_curr(cfs_rq); | |
2069dd75 | 2061 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 2062 | } |
2069dd75 PZ |
2063 | |
2064 | update_load_set(&se->load, weight); | |
2065 | ||
2066 | if (se->on_rq) | |
2067 | account_entity_enqueue(cfs_rq, se); | |
2068 | } | |
2069 | ||
82958366 PT |
2070 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2071 | ||
6d5ab293 | 2072 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2073 | { |
2074 | struct task_group *tg; | |
2075 | struct sched_entity *se; | |
3ff6dcac | 2076 | long shares; |
2069dd75 | 2077 | |
2069dd75 PZ |
2078 | tg = cfs_rq->tg; |
2079 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 2080 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 2081 | return; |
3ff6dcac YZ |
2082 | #ifndef CONFIG_SMP |
2083 | if (likely(se->load.weight == tg->shares)) | |
2084 | return; | |
2085 | #endif | |
6d5ab293 | 2086 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
2087 | |
2088 | reweight_entity(cfs_rq_of(se), se, shares); | |
2089 | } | |
2090 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 2091 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2092 | { |
2093 | } | |
2094 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
2095 | ||
141965c7 | 2096 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
2097 | /* |
2098 | * We choose a half-life close to 1 scheduling period. | |
2099 | * Note: The tables below are dependent on this value. | |
2100 | */ | |
2101 | #define LOAD_AVG_PERIOD 32 | |
2102 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
2103 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
2104 | ||
2105 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
2106 | static const u32 runnable_avg_yN_inv[] = { | |
2107 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
2108 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
2109 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
2110 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
2111 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
2112 | 0x85aac367, 0x82cd8698, | |
2113 | }; | |
2114 | ||
2115 | /* | |
2116 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
2117 | * over-estimates when re-combining. | |
2118 | */ | |
2119 | static const u32 runnable_avg_yN_sum[] = { | |
2120 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
2121 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
2122 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
2123 | }; | |
2124 | ||
9d85f21c PT |
2125 | /* |
2126 | * Approximate: | |
2127 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
2128 | */ | |
2129 | static __always_inline u64 decay_load(u64 val, u64 n) | |
2130 | { | |
5b51f2f8 PT |
2131 | unsigned int local_n; |
2132 | ||
2133 | if (!n) | |
2134 | return val; | |
2135 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
2136 | return 0; | |
2137 | ||
2138 | /* after bounds checking we can collapse to 32-bit */ | |
2139 | local_n = n; | |
2140 | ||
2141 | /* | |
2142 | * As y^PERIOD = 1/2, we can combine | |
2143 | * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD) | |
2144 | * With a look-up table which covers k^n (n<PERIOD) | |
2145 | * | |
2146 | * To achieve constant time decay_load. | |
2147 | */ | |
2148 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2149 | val >>= local_n / LOAD_AVG_PERIOD; | |
2150 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2151 | } |
2152 | ||
5b51f2f8 PT |
2153 | val *= runnable_avg_yN_inv[local_n]; |
2154 | /* We don't use SRR here since we always want to round down. */ | |
2155 | return val >> 32; | |
2156 | } | |
2157 | ||
2158 | /* | |
2159 | * For updates fully spanning n periods, the contribution to runnable | |
2160 | * average will be: \Sum 1024*y^n | |
2161 | * | |
2162 | * We can compute this reasonably efficiently by combining: | |
2163 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2164 | */ | |
2165 | static u32 __compute_runnable_contrib(u64 n) | |
2166 | { | |
2167 | u32 contrib = 0; | |
2168 | ||
2169 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2170 | return runnable_avg_yN_sum[n]; | |
2171 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2172 | return LOAD_AVG_MAX; | |
2173 | ||
2174 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
2175 | do { | |
2176 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
2177 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
2178 | ||
2179 | n -= LOAD_AVG_PERIOD; | |
2180 | } while (n > LOAD_AVG_PERIOD); | |
2181 | ||
2182 | contrib = decay_load(contrib, n); | |
2183 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2184 | } |
2185 | ||
2186 | /* | |
2187 | * We can represent the historical contribution to runnable average as the | |
2188 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2189 | * history into segments of approximately 1ms (1024us); label the segment that | |
2190 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2191 | * | |
2192 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2193 | * p0 p1 p2 | |
2194 | * (now) (~1ms ago) (~2ms ago) | |
2195 | * | |
2196 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2197 | * | |
2198 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2199 | * following representation of historical load: | |
2200 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2201 | * | |
2202 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2203 | * y^32 = 0.5 | |
2204 | * | |
2205 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2206 | * approximately half as much as the contribution to load within the last ms | |
2207 | * (u_0). | |
2208 | * | |
2209 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2210 | * sum again by y is sufficient to update: | |
2211 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2212 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2213 | */ | |
2214 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
2215 | struct sched_avg *sa, | |
2216 | int runnable) | |
2217 | { | |
5b51f2f8 PT |
2218 | u64 delta, periods; |
2219 | u32 runnable_contrib; | |
9d85f21c PT |
2220 | int delta_w, decayed = 0; |
2221 | ||
2222 | delta = now - sa->last_runnable_update; | |
2223 | /* | |
2224 | * This should only happen when time goes backwards, which it | |
2225 | * unfortunately does during sched clock init when we swap over to TSC. | |
2226 | */ | |
2227 | if ((s64)delta < 0) { | |
2228 | sa->last_runnable_update = now; | |
2229 | return 0; | |
2230 | } | |
2231 | ||
2232 | /* | |
2233 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2234 | * approximation of 1us and fast to compute. | |
2235 | */ | |
2236 | delta >>= 10; | |
2237 | if (!delta) | |
2238 | return 0; | |
2239 | sa->last_runnable_update = now; | |
2240 | ||
2241 | /* delta_w is the amount already accumulated against our next period */ | |
2242 | delta_w = sa->runnable_avg_period % 1024; | |
2243 | if (delta + delta_w >= 1024) { | |
2244 | /* period roll-over */ | |
2245 | decayed = 1; | |
2246 | ||
2247 | /* | |
2248 | * Now that we know we're crossing a period boundary, figure | |
2249 | * out how much from delta we need to complete the current | |
2250 | * period and accrue it. | |
2251 | */ | |
2252 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
2253 | if (runnable) |
2254 | sa->runnable_avg_sum += delta_w; | |
2255 | sa->runnable_avg_period += delta_w; | |
2256 | ||
2257 | delta -= delta_w; | |
2258 | ||
2259 | /* Figure out how many additional periods this update spans */ | |
2260 | periods = delta / 1024; | |
2261 | delta %= 1024; | |
2262 | ||
2263 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
2264 | periods + 1); | |
2265 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
2266 | periods + 1); | |
2267 | ||
2268 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
2269 | runnable_contrib = __compute_runnable_contrib(periods); | |
2270 | if (runnable) | |
2271 | sa->runnable_avg_sum += runnable_contrib; | |
2272 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
2273 | } |
2274 | ||
2275 | /* Remainder of delta accrued against u_0` */ | |
2276 | if (runnable) | |
2277 | sa->runnable_avg_sum += delta; | |
2278 | sa->runnable_avg_period += delta; | |
2279 | ||
2280 | return decayed; | |
2281 | } | |
2282 | ||
9ee474f5 | 2283 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 2284 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
2285 | { |
2286 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2287 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
2288 | ||
2289 | decays -= se->avg.decay_count; | |
2290 | if (!decays) | |
aff3e498 | 2291 | return 0; |
9ee474f5 PT |
2292 | |
2293 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
2294 | se->avg.decay_count = 0; | |
aff3e498 PT |
2295 | |
2296 | return decays; | |
9ee474f5 PT |
2297 | } |
2298 | ||
c566e8e9 PT |
2299 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2300 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2301 | int force_update) | |
2302 | { | |
2303 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 2304 | long tg_contrib; |
c566e8e9 PT |
2305 | |
2306 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
2307 | tg_contrib -= cfs_rq->tg_load_contrib; | |
2308 | ||
bf5b986e AS |
2309 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
2310 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
2311 | cfs_rq->tg_load_contrib += tg_contrib; |
2312 | } | |
2313 | } | |
8165e145 | 2314 | |
bb17f655 PT |
2315 | /* |
2316 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
2317 | * representation for computing load contributions. | |
2318 | */ | |
2319 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
2320 | struct cfs_rq *cfs_rq) | |
2321 | { | |
2322 | struct task_group *tg = cfs_rq->tg; | |
2323 | long contrib; | |
2324 | ||
2325 | /* The fraction of a cpu used by this cfs_rq */ | |
85b088e9 | 2326 | contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, |
bb17f655 PT |
2327 | sa->runnable_avg_period + 1); |
2328 | contrib -= cfs_rq->tg_runnable_contrib; | |
2329 | ||
2330 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
2331 | atomic_add(contrib, &tg->runnable_avg); | |
2332 | cfs_rq->tg_runnable_contrib += contrib; | |
2333 | } | |
2334 | } | |
2335 | ||
8165e145 PT |
2336 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
2337 | { | |
2338 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
2339 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
2340 | int runnable_avg; |
2341 | ||
8165e145 PT |
2342 | u64 contrib; |
2343 | ||
2344 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
2345 | se->avg.load_avg_contrib = div_u64(contrib, |
2346 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
2347 | |
2348 | /* | |
2349 | * For group entities we need to compute a correction term in the case | |
2350 | * that they are consuming <1 cpu so that we would contribute the same | |
2351 | * load as a task of equal weight. | |
2352 | * | |
2353 | * Explicitly co-ordinating this measurement would be expensive, but | |
2354 | * fortunately the sum of each cpus contribution forms a usable | |
2355 | * lower-bound on the true value. | |
2356 | * | |
2357 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
2358 | * (and the sum represents true value) or they are disjoint and we are | |
2359 | * understating by the aggregate of their overlap. | |
2360 | * | |
2361 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
2362 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
2363 | * cpus that overlap for this interval and w_i is the interval width. | |
2364 | * | |
2365 | * On a small machine; the first term is well-bounded which bounds the | |
2366 | * total error since w_i is a subset of the period. Whereas on a | |
2367 | * larger machine, while this first term can be larger, if w_i is the | |
2368 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
2369 | * our upper bound of 1-cpu. | |
2370 | */ | |
2371 | runnable_avg = atomic_read(&tg->runnable_avg); | |
2372 | if (runnable_avg < NICE_0_LOAD) { | |
2373 | se->avg.load_avg_contrib *= runnable_avg; | |
2374 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
2375 | } | |
8165e145 | 2376 | } |
c566e8e9 PT |
2377 | #else |
2378 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2379 | int force_update) {} | |
bb17f655 PT |
2380 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2381 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 2382 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
c566e8e9 PT |
2383 | #endif |
2384 | ||
8165e145 PT |
2385 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2386 | { | |
2387 | u32 contrib; | |
2388 | ||
2389 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2390 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2391 | contrib /= (se->avg.runnable_avg_period + 1); | |
2392 | se->avg.load_avg_contrib = scale_load(contrib); | |
2393 | } | |
2394 | ||
2dac754e PT |
2395 | /* Compute the current contribution to load_avg by se, return any delta */ |
2396 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2397 | { | |
2398 | long old_contrib = se->avg.load_avg_contrib; | |
2399 | ||
8165e145 PT |
2400 | if (entity_is_task(se)) { |
2401 | __update_task_entity_contrib(se); | |
2402 | } else { | |
bb17f655 | 2403 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2404 | __update_group_entity_contrib(se); |
2405 | } | |
2dac754e PT |
2406 | |
2407 | return se->avg.load_avg_contrib - old_contrib; | |
2408 | } | |
2409 | ||
9ee474f5 PT |
2410 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2411 | long load_contrib) | |
2412 | { | |
2413 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2414 | cfs_rq->blocked_load_avg -= load_contrib; | |
2415 | else | |
2416 | cfs_rq->blocked_load_avg = 0; | |
2417 | } | |
2418 | ||
f1b17280 PT |
2419 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2420 | ||
9d85f21c | 2421 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2422 | static inline void update_entity_load_avg(struct sched_entity *se, |
2423 | int update_cfs_rq) | |
9d85f21c | 2424 | { |
2dac754e PT |
2425 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2426 | long contrib_delta; | |
f1b17280 | 2427 | u64 now; |
2dac754e | 2428 | |
f1b17280 PT |
2429 | /* |
2430 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2431 | * case they are the parent of a throttled hierarchy. | |
2432 | */ | |
2433 | if (entity_is_task(se)) | |
2434 | now = cfs_rq_clock_task(cfs_rq); | |
2435 | else | |
2436 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2437 | ||
2438 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2439 | return; |
2440 | ||
2441 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2442 | |
2443 | if (!update_cfs_rq) | |
2444 | return; | |
2445 | ||
2dac754e PT |
2446 | if (se->on_rq) |
2447 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2448 | else |
2449 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2450 | } | |
2451 | ||
2452 | /* | |
2453 | * Decay the load contributed by all blocked children and account this so that | |
2454 | * their contribution may appropriately discounted when they wake up. | |
2455 | */ | |
aff3e498 | 2456 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2457 | { |
f1b17280 | 2458 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2459 | u64 decays; |
2460 | ||
2461 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2462 | if (!decays && !force_update) |
9ee474f5 PT |
2463 | return; |
2464 | ||
2509940f AS |
2465 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2466 | unsigned long removed_load; | |
2467 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2468 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2469 | } | |
9ee474f5 | 2470 | |
aff3e498 PT |
2471 | if (decays) { |
2472 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2473 | decays); | |
2474 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2475 | cfs_rq->last_decay = now; | |
2476 | } | |
c566e8e9 PT |
2477 | |
2478 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2479 | } |
18bf2805 BS |
2480 | |
2481 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2482 | { | |
78becc27 | 2483 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
bb17f655 | 2484 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
18bf2805 | 2485 | } |
2dac754e PT |
2486 | |
2487 | /* Add the load generated by se into cfs_rq's child load-average */ | |
2488 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2489 | struct sched_entity *se, |
2490 | int wakeup) | |
2dac754e | 2491 | { |
aff3e498 PT |
2492 | /* |
2493 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2494 | * migration we use a negative decay count to track the remote decays | |
2495 | * accumulated while sleeping. | |
a75cdaa9 AS |
2496 | * |
2497 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2498 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2499 | * constructed load_avg_contrib. | |
aff3e498 PT |
2500 | */ |
2501 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2502 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2503 | if (se->avg.decay_count) { |
2504 | /* | |
2505 | * In a wake-up migration we have to approximate the | |
2506 | * time sleeping. This is because we can't synchronize | |
2507 | * clock_task between the two cpus, and it is not | |
2508 | * guaranteed to be read-safe. Instead, we can | |
2509 | * approximate this using our carried decays, which are | |
2510 | * explicitly atomically readable. | |
2511 | */ | |
2512 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2513 | << 20; | |
2514 | update_entity_load_avg(se, 0); | |
2515 | /* Indicate that we're now synchronized and on-rq */ | |
2516 | se->avg.decay_count = 0; | |
2517 | } | |
9ee474f5 PT |
2518 | wakeup = 0; |
2519 | } else { | |
9390675a | 2520 | __synchronize_entity_decay(se); |
9ee474f5 PT |
2521 | } |
2522 | ||
aff3e498 PT |
2523 | /* migrated tasks did not contribute to our blocked load */ |
2524 | if (wakeup) { | |
9ee474f5 | 2525 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2526 | update_entity_load_avg(se, 0); |
2527 | } | |
9ee474f5 | 2528 | |
2dac754e | 2529 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2530 | /* we force update consideration on load-balancer moves */ |
2531 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2532 | } |
2533 | ||
9ee474f5 PT |
2534 | /* |
2535 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2536 | * transitioning to a blocked state we track its projected decay using | |
2537 | * blocked_load_avg. | |
2538 | */ | |
2dac754e | 2539 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2540 | struct sched_entity *se, |
2541 | int sleep) | |
2dac754e | 2542 | { |
9ee474f5 | 2543 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2544 | /* we force update consideration on load-balancer moves */ |
2545 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2546 | |
2dac754e | 2547 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2548 | if (sleep) { |
2549 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2550 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2551 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2552 | } |
642dbc39 VG |
2553 | |
2554 | /* | |
2555 | * Update the rq's load with the elapsed running time before entering | |
2556 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2557 | * be the only way to update the runnable statistic. | |
2558 | */ | |
2559 | void idle_enter_fair(struct rq *this_rq) | |
2560 | { | |
2561 | update_rq_runnable_avg(this_rq, 1); | |
2562 | } | |
2563 | ||
2564 | /* | |
2565 | * Update the rq's load with the elapsed idle time before a task is | |
2566 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2567 | * be the only way to update the runnable statistic. | |
2568 | */ | |
2569 | void idle_exit_fair(struct rq *this_rq) | |
2570 | { | |
2571 | update_rq_runnable_avg(this_rq, 0); | |
2572 | } | |
2573 | ||
38033c37 PZ |
2574 | #else /* CONFIG_SMP */ |
2575 | ||
9ee474f5 PT |
2576 | static inline void update_entity_load_avg(struct sched_entity *se, |
2577 | int update_cfs_rq) {} | |
18bf2805 | 2578 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2579 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2580 | struct sched_entity *se, |
2581 | int wakeup) {} | |
2dac754e | 2582 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2583 | struct sched_entity *se, |
2584 | int sleep) {} | |
aff3e498 PT |
2585 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2586 | int force_update) {} | |
38033c37 | 2587 | #endif /* CONFIG_SMP */ |
9d85f21c | 2588 | |
2396af69 | 2589 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2590 | { |
bf0f6f24 | 2591 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2592 | struct task_struct *tsk = NULL; |
2593 | ||
2594 | if (entity_is_task(se)) | |
2595 | tsk = task_of(se); | |
2596 | ||
41acab88 | 2597 | if (se->statistics.sleep_start) { |
78becc27 | 2598 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2599 | |
2600 | if ((s64)delta < 0) | |
2601 | delta = 0; | |
2602 | ||
41acab88 LDM |
2603 | if (unlikely(delta > se->statistics.sleep_max)) |
2604 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2605 | |
8c79a045 | 2606 | se->statistics.sleep_start = 0; |
41acab88 | 2607 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2608 | |
768d0c27 | 2609 | if (tsk) { |
e414314c | 2610 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2611 | trace_sched_stat_sleep(tsk, delta); |
2612 | } | |
bf0f6f24 | 2613 | } |
41acab88 | 2614 | if (se->statistics.block_start) { |
78becc27 | 2615 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2616 | |
2617 | if ((s64)delta < 0) | |
2618 | delta = 0; | |
2619 | ||
41acab88 LDM |
2620 | if (unlikely(delta > se->statistics.block_max)) |
2621 | se->statistics.block_max = delta; | |
bf0f6f24 | 2622 | |
8c79a045 | 2623 | se->statistics.block_start = 0; |
41acab88 | 2624 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2625 | |
e414314c | 2626 | if (tsk) { |
8f0dfc34 | 2627 | if (tsk->in_iowait) { |
41acab88 LDM |
2628 | se->statistics.iowait_sum += delta; |
2629 | se->statistics.iowait_count++; | |
768d0c27 | 2630 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2631 | } |
2632 | ||
b781a602 AV |
2633 | trace_sched_stat_blocked(tsk, delta); |
2634 | ||
e414314c PZ |
2635 | /* |
2636 | * Blocking time is in units of nanosecs, so shift by | |
2637 | * 20 to get a milliseconds-range estimation of the | |
2638 | * amount of time that the task spent sleeping: | |
2639 | */ | |
2640 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2641 | profile_hits(SLEEP_PROFILING, | |
2642 | (void *)get_wchan(tsk), | |
2643 | delta >> 20); | |
2644 | } | |
2645 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2646 | } |
bf0f6f24 IM |
2647 | } |
2648 | #endif | |
2649 | } | |
2650 | ||
ddc97297 PZ |
2651 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2652 | { | |
2653 | #ifdef CONFIG_SCHED_DEBUG | |
2654 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2655 | ||
2656 | if (d < 0) | |
2657 | d = -d; | |
2658 | ||
2659 | if (d > 3*sysctl_sched_latency) | |
2660 | schedstat_inc(cfs_rq, nr_spread_over); | |
2661 | #endif | |
2662 | } | |
2663 | ||
aeb73b04 PZ |
2664 | static void |
2665 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2666 | { | |
1af5f730 | 2667 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2668 | |
2cb8600e PZ |
2669 | /* |
2670 | * The 'current' period is already promised to the current tasks, | |
2671 | * however the extra weight of the new task will slow them down a | |
2672 | * little, place the new task so that it fits in the slot that | |
2673 | * stays open at the end. | |
2674 | */ | |
94dfb5e7 | 2675 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2676 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2677 | |
a2e7a7eb | 2678 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2679 | if (!initial) { |
a2e7a7eb | 2680 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2681 | |
a2e7a7eb MG |
2682 | /* |
2683 | * Halve their sleep time's effect, to allow | |
2684 | * for a gentler effect of sleepers: | |
2685 | */ | |
2686 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2687 | thresh >>= 1; | |
51e0304c | 2688 | |
a2e7a7eb | 2689 | vruntime -= thresh; |
aeb73b04 PZ |
2690 | } |
2691 | ||
b5d9d734 | 2692 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 2693 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
2694 | } |
2695 | ||
d3d9dc33 PT |
2696 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2697 | ||
bf0f6f24 | 2698 | static void |
88ec22d3 | 2699 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2700 | { |
88ec22d3 PZ |
2701 | /* |
2702 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 2703 | * through calling update_curr(). |
88ec22d3 | 2704 | */ |
371fd7e7 | 2705 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
2706 | se->vruntime += cfs_rq->min_vruntime; |
2707 | ||
bf0f6f24 | 2708 | /* |
a2a2d680 | 2709 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2710 | */ |
b7cc0896 | 2711 | update_curr(cfs_rq); |
f269ae04 | 2712 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
2713 | account_entity_enqueue(cfs_rq, se); |
2714 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 2715 | |
88ec22d3 | 2716 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 2717 | place_entity(cfs_rq, se, 0); |
2396af69 | 2718 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 2719 | } |
bf0f6f24 | 2720 | |
d2417e5a | 2721 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 2722 | check_spread(cfs_rq, se); |
83b699ed SV |
2723 | if (se != cfs_rq->curr) |
2724 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 2725 | se->on_rq = 1; |
3d4b47b4 | 2726 | |
d3d9dc33 | 2727 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 2728 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
2729 | check_enqueue_throttle(cfs_rq); |
2730 | } | |
bf0f6f24 IM |
2731 | } |
2732 | ||
2c13c919 | 2733 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 2734 | { |
2c13c919 RR |
2735 | for_each_sched_entity(se) { |
2736 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 2737 | if (cfs_rq->last != se) |
2c13c919 | 2738 | break; |
f1044799 PZ |
2739 | |
2740 | cfs_rq->last = NULL; | |
2c13c919 RR |
2741 | } |
2742 | } | |
2002c695 | 2743 | |
2c13c919 RR |
2744 | static void __clear_buddies_next(struct sched_entity *se) |
2745 | { | |
2746 | for_each_sched_entity(se) { | |
2747 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 2748 | if (cfs_rq->next != se) |
2c13c919 | 2749 | break; |
f1044799 PZ |
2750 | |
2751 | cfs_rq->next = NULL; | |
2c13c919 | 2752 | } |
2002c695 PZ |
2753 | } |
2754 | ||
ac53db59 RR |
2755 | static void __clear_buddies_skip(struct sched_entity *se) |
2756 | { | |
2757 | for_each_sched_entity(se) { | |
2758 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 2759 | if (cfs_rq->skip != se) |
ac53db59 | 2760 | break; |
f1044799 PZ |
2761 | |
2762 | cfs_rq->skip = NULL; | |
ac53db59 RR |
2763 | } |
2764 | } | |
2765 | ||
a571bbea PZ |
2766 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2767 | { | |
2c13c919 RR |
2768 | if (cfs_rq->last == se) |
2769 | __clear_buddies_last(se); | |
2770 | ||
2771 | if (cfs_rq->next == se) | |
2772 | __clear_buddies_next(se); | |
ac53db59 RR |
2773 | |
2774 | if (cfs_rq->skip == se) | |
2775 | __clear_buddies_skip(se); | |
a571bbea PZ |
2776 | } |
2777 | ||
6c16a6dc | 2778 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 2779 | |
bf0f6f24 | 2780 | static void |
371fd7e7 | 2781 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 2782 | { |
a2a2d680 DA |
2783 | /* |
2784 | * Update run-time statistics of the 'current'. | |
2785 | */ | |
2786 | update_curr(cfs_rq); | |
17bc14b7 | 2787 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 2788 | |
19b6a2e3 | 2789 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 2790 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 2791 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
2792 | if (entity_is_task(se)) { |
2793 | struct task_struct *tsk = task_of(se); | |
2794 | ||
2795 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 2796 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2797 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 2798 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 2799 | } |
db36cc7d | 2800 | #endif |
67e9fb2a PZ |
2801 | } |
2802 | ||
2002c695 | 2803 | clear_buddies(cfs_rq, se); |
4793241b | 2804 | |
83b699ed | 2805 | if (se != cfs_rq->curr) |
30cfdcfc | 2806 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 2807 | se->on_rq = 0; |
30cfdcfc | 2808 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
2809 | |
2810 | /* | |
2811 | * Normalize the entity after updating the min_vruntime because the | |
2812 | * update can refer to the ->curr item and we need to reflect this | |
2813 | * movement in our normalized position. | |
2814 | */ | |
371fd7e7 | 2815 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 2816 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 2817 | |
d8b4986d PT |
2818 | /* return excess runtime on last dequeue */ |
2819 | return_cfs_rq_runtime(cfs_rq); | |
2820 | ||
1e876231 | 2821 | update_min_vruntime(cfs_rq); |
17bc14b7 | 2822 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
2823 | } |
2824 | ||
2825 | /* | |
2826 | * Preempt the current task with a newly woken task if needed: | |
2827 | */ | |
7c92e54f | 2828 | static void |
2e09bf55 | 2829 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 2830 | { |
11697830 | 2831 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
2832 | struct sched_entity *se; |
2833 | s64 delta; | |
11697830 | 2834 | |
6d0f0ebd | 2835 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 2836 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 2837 | if (delta_exec > ideal_runtime) { |
bf0f6f24 | 2838 | resched_task(rq_of(cfs_rq)->curr); |
a9f3e2b5 MG |
2839 | /* |
2840 | * The current task ran long enough, ensure it doesn't get | |
2841 | * re-elected due to buddy favours. | |
2842 | */ | |
2843 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
2844 | return; |
2845 | } | |
2846 | ||
2847 | /* | |
2848 | * Ensure that a task that missed wakeup preemption by a | |
2849 | * narrow margin doesn't have to wait for a full slice. | |
2850 | * This also mitigates buddy induced latencies under load. | |
2851 | */ | |
f685ceac MG |
2852 | if (delta_exec < sysctl_sched_min_granularity) |
2853 | return; | |
2854 | ||
f4cfb33e WX |
2855 | se = __pick_first_entity(cfs_rq); |
2856 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 2857 | |
f4cfb33e WX |
2858 | if (delta < 0) |
2859 | return; | |
d7d82944 | 2860 | |
f4cfb33e WX |
2861 | if (delta > ideal_runtime) |
2862 | resched_task(rq_of(cfs_rq)->curr); | |
bf0f6f24 IM |
2863 | } |
2864 | ||
83b699ed | 2865 | static void |
8494f412 | 2866 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2867 | { |
83b699ed SV |
2868 | /* 'current' is not kept within the tree. */ |
2869 | if (se->on_rq) { | |
2870 | /* | |
2871 | * Any task has to be enqueued before it get to execute on | |
2872 | * a CPU. So account for the time it spent waiting on the | |
2873 | * runqueue. | |
2874 | */ | |
2875 | update_stats_wait_end(cfs_rq, se); | |
2876 | __dequeue_entity(cfs_rq, se); | |
2877 | } | |
2878 | ||
79303e9e | 2879 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 2880 | cfs_rq->curr = se; |
eba1ed4b IM |
2881 | #ifdef CONFIG_SCHEDSTATS |
2882 | /* | |
2883 | * Track our maximum slice length, if the CPU's load is at | |
2884 | * least twice that of our own weight (i.e. dont track it | |
2885 | * when there are only lesser-weight tasks around): | |
2886 | */ | |
495eca49 | 2887 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 2888 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
2889 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
2890 | } | |
2891 | #endif | |
4a55b450 | 2892 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
2893 | } |
2894 | ||
3f3a4904 PZ |
2895 | static int |
2896 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
2897 | ||
ac53db59 RR |
2898 | /* |
2899 | * Pick the next process, keeping these things in mind, in this order: | |
2900 | * 1) keep things fair between processes/task groups | |
2901 | * 2) pick the "next" process, since someone really wants that to run | |
2902 | * 3) pick the "last" process, for cache locality | |
2903 | * 4) do not run the "skip" process, if something else is available | |
2904 | */ | |
678d5718 PZ |
2905 | static struct sched_entity * |
2906 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 2907 | { |
678d5718 PZ |
2908 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
2909 | struct sched_entity *se; | |
2910 | ||
2911 | /* | |
2912 | * If curr is set we have to see if its left of the leftmost entity | |
2913 | * still in the tree, provided there was anything in the tree at all. | |
2914 | */ | |
2915 | if (!left || (curr && entity_before(curr, left))) | |
2916 | left = curr; | |
2917 | ||
2918 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 2919 | |
ac53db59 RR |
2920 | /* |
2921 | * Avoid running the skip buddy, if running something else can | |
2922 | * be done without getting too unfair. | |
2923 | */ | |
2924 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
2925 | struct sched_entity *second; |
2926 | ||
2927 | if (se == curr) { | |
2928 | second = __pick_first_entity(cfs_rq); | |
2929 | } else { | |
2930 | second = __pick_next_entity(se); | |
2931 | if (!second || (curr && entity_before(curr, second))) | |
2932 | second = curr; | |
2933 | } | |
2934 | ||
ac53db59 RR |
2935 | if (second && wakeup_preempt_entity(second, left) < 1) |
2936 | se = second; | |
2937 | } | |
aa2ac252 | 2938 | |
f685ceac MG |
2939 | /* |
2940 | * Prefer last buddy, try to return the CPU to a preempted task. | |
2941 | */ | |
2942 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
2943 | se = cfs_rq->last; | |
2944 | ||
ac53db59 RR |
2945 | /* |
2946 | * Someone really wants this to run. If it's not unfair, run it. | |
2947 | */ | |
2948 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
2949 | se = cfs_rq->next; | |
2950 | ||
f685ceac | 2951 | clear_buddies(cfs_rq, se); |
4793241b PZ |
2952 | |
2953 | return se; | |
aa2ac252 PZ |
2954 | } |
2955 | ||
678d5718 | 2956 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 2957 | |
ab6cde26 | 2958 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
2959 | { |
2960 | /* | |
2961 | * If still on the runqueue then deactivate_task() | |
2962 | * was not called and update_curr() has to be done: | |
2963 | */ | |
2964 | if (prev->on_rq) | |
b7cc0896 | 2965 | update_curr(cfs_rq); |
bf0f6f24 | 2966 | |
d3d9dc33 PT |
2967 | /* throttle cfs_rqs exceeding runtime */ |
2968 | check_cfs_rq_runtime(cfs_rq); | |
2969 | ||
ddc97297 | 2970 | check_spread(cfs_rq, prev); |
30cfdcfc | 2971 | if (prev->on_rq) { |
5870db5b | 2972 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
2973 | /* Put 'current' back into the tree. */ |
2974 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 2975 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 2976 | update_entity_load_avg(prev, 1); |
30cfdcfc | 2977 | } |
429d43bc | 2978 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
2979 | } |
2980 | ||
8f4d37ec PZ |
2981 | static void |
2982 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 2983 | { |
bf0f6f24 | 2984 | /* |
30cfdcfc | 2985 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 2986 | */ |
30cfdcfc | 2987 | update_curr(cfs_rq); |
bf0f6f24 | 2988 | |
9d85f21c PT |
2989 | /* |
2990 | * Ensure that runnable average is periodically updated. | |
2991 | */ | |
9ee474f5 | 2992 | update_entity_load_avg(curr, 1); |
aff3e498 | 2993 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 2994 | update_cfs_shares(cfs_rq); |
9d85f21c | 2995 | |
8f4d37ec PZ |
2996 | #ifdef CONFIG_SCHED_HRTICK |
2997 | /* | |
2998 | * queued ticks are scheduled to match the slice, so don't bother | |
2999 | * validating it and just reschedule. | |
3000 | */ | |
983ed7a6 HH |
3001 | if (queued) { |
3002 | resched_task(rq_of(cfs_rq)->curr); | |
3003 | return; | |
3004 | } | |
8f4d37ec PZ |
3005 | /* |
3006 | * don't let the period tick interfere with the hrtick preemption | |
3007 | */ | |
3008 | if (!sched_feat(DOUBLE_TICK) && | |
3009 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
3010 | return; | |
3011 | #endif | |
3012 | ||
2c2efaed | 3013 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 3014 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
3015 | } |
3016 | ||
ab84d31e PT |
3017 | |
3018 | /************************************************** | |
3019 | * CFS bandwidth control machinery | |
3020 | */ | |
3021 | ||
3022 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
3023 | |
3024 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 3025 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
3026 | |
3027 | static inline bool cfs_bandwidth_used(void) | |
3028 | { | |
c5905afb | 3029 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
3030 | } |
3031 | ||
1ee14e6c | 3032 | void cfs_bandwidth_usage_inc(void) |
029632fb | 3033 | { |
1ee14e6c BS |
3034 | static_key_slow_inc(&__cfs_bandwidth_used); |
3035 | } | |
3036 | ||
3037 | void cfs_bandwidth_usage_dec(void) | |
3038 | { | |
3039 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
3040 | } |
3041 | #else /* HAVE_JUMP_LABEL */ | |
3042 | static bool cfs_bandwidth_used(void) | |
3043 | { | |
3044 | return true; | |
3045 | } | |
3046 | ||
1ee14e6c BS |
3047 | void cfs_bandwidth_usage_inc(void) {} |
3048 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
3049 | #endif /* HAVE_JUMP_LABEL */ |
3050 | ||
ab84d31e PT |
3051 | /* |
3052 | * default period for cfs group bandwidth. | |
3053 | * default: 0.1s, units: nanoseconds | |
3054 | */ | |
3055 | static inline u64 default_cfs_period(void) | |
3056 | { | |
3057 | return 100000000ULL; | |
3058 | } | |
ec12cb7f PT |
3059 | |
3060 | static inline u64 sched_cfs_bandwidth_slice(void) | |
3061 | { | |
3062 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
3063 | } | |
3064 | ||
a9cf55b2 PT |
3065 | /* |
3066 | * Replenish runtime according to assigned quota and update expiration time. | |
3067 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
3068 | * additional synchronization around rq->lock. | |
3069 | * | |
3070 | * requires cfs_b->lock | |
3071 | */ | |
029632fb | 3072 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
3073 | { |
3074 | u64 now; | |
3075 | ||
3076 | if (cfs_b->quota == RUNTIME_INF) | |
3077 | return; | |
3078 | ||
3079 | now = sched_clock_cpu(smp_processor_id()); | |
3080 | cfs_b->runtime = cfs_b->quota; | |
3081 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
3082 | } | |
3083 | ||
029632fb PZ |
3084 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3085 | { | |
3086 | return &tg->cfs_bandwidth; | |
3087 | } | |
3088 | ||
f1b17280 PT |
3089 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
3090 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
3091 | { | |
3092 | if (unlikely(cfs_rq->throttle_count)) | |
3093 | return cfs_rq->throttled_clock_task; | |
3094 | ||
78becc27 | 3095 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
3096 | } |
3097 | ||
85dac906 PT |
3098 | /* returns 0 on failure to allocate runtime */ |
3099 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
3100 | { |
3101 | struct task_group *tg = cfs_rq->tg; | |
3102 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 3103 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
3104 | |
3105 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
3106 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
3107 | ||
3108 | raw_spin_lock(&cfs_b->lock); | |
3109 | if (cfs_b->quota == RUNTIME_INF) | |
3110 | amount = min_amount; | |
58088ad0 | 3111 | else { |
a9cf55b2 PT |
3112 | /* |
3113 | * If the bandwidth pool has become inactive, then at least one | |
3114 | * period must have elapsed since the last consumption. | |
3115 | * Refresh the global state and ensure bandwidth timer becomes | |
3116 | * active. | |
3117 | */ | |
3118 | if (!cfs_b->timer_active) { | |
3119 | __refill_cfs_bandwidth_runtime(cfs_b); | |
58088ad0 | 3120 | __start_cfs_bandwidth(cfs_b); |
a9cf55b2 | 3121 | } |
58088ad0 PT |
3122 | |
3123 | if (cfs_b->runtime > 0) { | |
3124 | amount = min(cfs_b->runtime, min_amount); | |
3125 | cfs_b->runtime -= amount; | |
3126 | cfs_b->idle = 0; | |
3127 | } | |
ec12cb7f | 3128 | } |
a9cf55b2 | 3129 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
3130 | raw_spin_unlock(&cfs_b->lock); |
3131 | ||
3132 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
3133 | /* |
3134 | * we may have advanced our local expiration to account for allowed | |
3135 | * spread between our sched_clock and the one on which runtime was | |
3136 | * issued. | |
3137 | */ | |
3138 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
3139 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
3140 | |
3141 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
3142 | } |
3143 | ||
a9cf55b2 PT |
3144 | /* |
3145 | * Note: This depends on the synchronization provided by sched_clock and the | |
3146 | * fact that rq->clock snapshots this value. | |
3147 | */ | |
3148 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 3149 | { |
a9cf55b2 | 3150 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
3151 | |
3152 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 3153 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
3154 | return; |
3155 | ||
a9cf55b2 PT |
3156 | if (cfs_rq->runtime_remaining < 0) |
3157 | return; | |
3158 | ||
3159 | /* | |
3160 | * If the local deadline has passed we have to consider the | |
3161 | * possibility that our sched_clock is 'fast' and the global deadline | |
3162 | * has not truly expired. | |
3163 | * | |
3164 | * Fortunately we can check determine whether this the case by checking | |
3165 | * whether the global deadline has advanced. | |
3166 | */ | |
3167 | ||
3168 | if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { | |
3169 | /* extend local deadline, drift is bounded above by 2 ticks */ | |
3170 | cfs_rq->runtime_expires += TICK_NSEC; | |
3171 | } else { | |
3172 | /* global deadline is ahead, expiration has passed */ | |
3173 | cfs_rq->runtime_remaining = 0; | |
3174 | } | |
3175 | } | |
3176 | ||
9dbdb155 | 3177 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
3178 | { |
3179 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3180 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3181 | expire_cfs_rq_runtime(cfs_rq); |
3182 | ||
3183 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3184 | return; |
3185 | ||
85dac906 PT |
3186 | /* |
3187 | * if we're unable to extend our runtime we resched so that the active | |
3188 | * hierarchy can be throttled | |
3189 | */ | |
3190 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
3191 | resched_task(rq_of(cfs_rq)->curr); | |
ec12cb7f PT |
3192 | } |
3193 | ||
6c16a6dc | 3194 | static __always_inline |
9dbdb155 | 3195 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 3196 | { |
56f570e5 | 3197 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3198 | return; |
3199 | ||
3200 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3201 | } | |
3202 | ||
85dac906 PT |
3203 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3204 | { | |
56f570e5 | 3205 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3206 | } |
3207 | ||
64660c86 PT |
3208 | /* check whether cfs_rq, or any parent, is throttled */ |
3209 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3210 | { | |
56f570e5 | 3211 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3212 | } |
3213 | ||
3214 | /* | |
3215 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3216 | * dest_cpu are members of a throttled hierarchy when performing group | |
3217 | * load-balance operations. | |
3218 | */ | |
3219 | static inline int throttled_lb_pair(struct task_group *tg, | |
3220 | int src_cpu, int dest_cpu) | |
3221 | { | |
3222 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3223 | ||
3224 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3225 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3226 | ||
3227 | return throttled_hierarchy(src_cfs_rq) || | |
3228 | throttled_hierarchy(dest_cfs_rq); | |
3229 | } | |
3230 | ||
3231 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3232 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3233 | { | |
3234 | struct rq *rq = data; | |
3235 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3236 | ||
3237 | cfs_rq->throttle_count--; | |
3238 | #ifdef CONFIG_SMP | |
3239 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 3240 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3241 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3242 | cfs_rq->throttled_clock_task; |
64660c86 PT |
3243 | } |
3244 | #endif | |
3245 | ||
3246 | return 0; | |
3247 | } | |
3248 | ||
3249 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3250 | { | |
3251 | struct rq *rq = data; | |
3252 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3253 | ||
82958366 PT |
3254 | /* group is entering throttled state, stop time */ |
3255 | if (!cfs_rq->throttle_count) | |
78becc27 | 3256 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3257 | cfs_rq->throttle_count++; |
3258 | ||
3259 | return 0; | |
3260 | } | |
3261 | ||
d3d9dc33 | 3262 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3263 | { |
3264 | struct rq *rq = rq_of(cfs_rq); | |
3265 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3266 | struct sched_entity *se; | |
3267 | long task_delta, dequeue = 1; | |
3268 | ||
3269 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3270 | ||
f1b17280 | 3271 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3272 | rcu_read_lock(); |
3273 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3274 | rcu_read_unlock(); | |
85dac906 PT |
3275 | |
3276 | task_delta = cfs_rq->h_nr_running; | |
3277 | for_each_sched_entity(se) { | |
3278 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3279 | /* throttled entity or throttle-on-deactivate */ | |
3280 | if (!se->on_rq) | |
3281 | break; | |
3282 | ||
3283 | if (dequeue) | |
3284 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3285 | qcfs_rq->h_nr_running -= task_delta; | |
3286 | ||
3287 | if (qcfs_rq->load.weight) | |
3288 | dequeue = 0; | |
3289 | } | |
3290 | ||
3291 | if (!se) | |
3292 | rq->nr_running -= task_delta; | |
3293 | ||
3294 | cfs_rq->throttled = 1; | |
78becc27 | 3295 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 PT |
3296 | raw_spin_lock(&cfs_b->lock); |
3297 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
f9f9ffc2 BS |
3298 | if (!cfs_b->timer_active) |
3299 | __start_cfs_bandwidth(cfs_b); | |
85dac906 PT |
3300 | raw_spin_unlock(&cfs_b->lock); |
3301 | } | |
3302 | ||
029632fb | 3303 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3304 | { |
3305 | struct rq *rq = rq_of(cfs_rq); | |
3306 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3307 | struct sched_entity *se; | |
3308 | int enqueue = 1; | |
3309 | long task_delta; | |
3310 | ||
22b958d8 | 3311 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3312 | |
3313 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3314 | |
3315 | update_rq_clock(rq); | |
3316 | ||
671fd9da | 3317 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3318 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3319 | list_del_rcu(&cfs_rq->throttled_list); |
3320 | raw_spin_unlock(&cfs_b->lock); | |
3321 | ||
64660c86 PT |
3322 | /* update hierarchical throttle state */ |
3323 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3324 | ||
671fd9da PT |
3325 | if (!cfs_rq->load.weight) |
3326 | return; | |
3327 | ||
3328 | task_delta = cfs_rq->h_nr_running; | |
3329 | for_each_sched_entity(se) { | |
3330 | if (se->on_rq) | |
3331 | enqueue = 0; | |
3332 | ||
3333 | cfs_rq = cfs_rq_of(se); | |
3334 | if (enqueue) | |
3335 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3336 | cfs_rq->h_nr_running += task_delta; | |
3337 | ||
3338 | if (cfs_rq_throttled(cfs_rq)) | |
3339 | break; | |
3340 | } | |
3341 | ||
3342 | if (!se) | |
3343 | rq->nr_running += task_delta; | |
3344 | ||
3345 | /* determine whether we need to wake up potentially idle cpu */ | |
3346 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
3347 | resched_task(rq->curr); | |
3348 | } | |
3349 | ||
3350 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
3351 | u64 remaining, u64 expires) | |
3352 | { | |
3353 | struct cfs_rq *cfs_rq; | |
3354 | u64 runtime = remaining; | |
3355 | ||
3356 | rcu_read_lock(); | |
3357 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
3358 | throttled_list) { | |
3359 | struct rq *rq = rq_of(cfs_rq); | |
3360 | ||
3361 | raw_spin_lock(&rq->lock); | |
3362 | if (!cfs_rq_throttled(cfs_rq)) | |
3363 | goto next; | |
3364 | ||
3365 | runtime = -cfs_rq->runtime_remaining + 1; | |
3366 | if (runtime > remaining) | |
3367 | runtime = remaining; | |
3368 | remaining -= runtime; | |
3369 | ||
3370 | cfs_rq->runtime_remaining += runtime; | |
3371 | cfs_rq->runtime_expires = expires; | |
3372 | ||
3373 | /* we check whether we're throttled above */ | |
3374 | if (cfs_rq->runtime_remaining > 0) | |
3375 | unthrottle_cfs_rq(cfs_rq); | |
3376 | ||
3377 | next: | |
3378 | raw_spin_unlock(&rq->lock); | |
3379 | ||
3380 | if (!remaining) | |
3381 | break; | |
3382 | } | |
3383 | rcu_read_unlock(); | |
3384 | ||
3385 | return remaining; | |
3386 | } | |
3387 | ||
58088ad0 PT |
3388 | /* |
3389 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
3390 | * cfs_rqs as appropriate. If there has been no activity within the last | |
3391 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
3392 | * used to track this state. | |
3393 | */ | |
3394 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3395 | { | |
671fd9da PT |
3396 | u64 runtime, runtime_expires; |
3397 | int idle = 1, throttled; | |
58088ad0 PT |
3398 | |
3399 | raw_spin_lock(&cfs_b->lock); | |
3400 | /* no need to continue the timer with no bandwidth constraint */ | |
3401 | if (cfs_b->quota == RUNTIME_INF) | |
3402 | goto out_unlock; | |
3403 | ||
671fd9da PT |
3404 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3405 | /* idle depends on !throttled (for the case of a large deficit) */ | |
3406 | idle = cfs_b->idle && !throttled; | |
e8da1b18 | 3407 | cfs_b->nr_periods += overrun; |
671fd9da | 3408 | |
a9cf55b2 PT |
3409 | /* if we're going inactive then everything else can be deferred */ |
3410 | if (idle) | |
3411 | goto out_unlock; | |
3412 | ||
927b54fc BS |
3413 | /* |
3414 | * if we have relooped after returning idle once, we need to update our | |
3415 | * status as actually running, so that other cpus doing | |
3416 | * __start_cfs_bandwidth will stop trying to cancel us. | |
3417 | */ | |
3418 | cfs_b->timer_active = 1; | |
3419 | ||
a9cf55b2 PT |
3420 | __refill_cfs_bandwidth_runtime(cfs_b); |
3421 | ||
671fd9da PT |
3422 | if (!throttled) { |
3423 | /* mark as potentially idle for the upcoming period */ | |
3424 | cfs_b->idle = 1; | |
3425 | goto out_unlock; | |
3426 | } | |
3427 | ||
e8da1b18 NR |
3428 | /* account preceding periods in which throttling occurred */ |
3429 | cfs_b->nr_throttled += overrun; | |
3430 | ||
671fd9da PT |
3431 | /* |
3432 | * There are throttled entities so we must first use the new bandwidth | |
3433 | * to unthrottle them before making it generally available. This | |
3434 | * ensures that all existing debts will be paid before a new cfs_rq is | |
3435 | * allowed to run. | |
3436 | */ | |
3437 | runtime = cfs_b->runtime; | |
3438 | runtime_expires = cfs_b->runtime_expires; | |
3439 | cfs_b->runtime = 0; | |
3440 | ||
3441 | /* | |
3442 | * This check is repeated as we are holding onto the new bandwidth | |
3443 | * while we unthrottle. This can potentially race with an unthrottled | |
3444 | * group trying to acquire new bandwidth from the global pool. | |
3445 | */ | |
3446 | while (throttled && runtime > 0) { | |
3447 | raw_spin_unlock(&cfs_b->lock); | |
3448 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3449 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3450 | runtime_expires); | |
3451 | raw_spin_lock(&cfs_b->lock); | |
3452 | ||
3453 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
3454 | } | |
58088ad0 | 3455 | |
671fd9da PT |
3456 | /* return (any) remaining runtime */ |
3457 | cfs_b->runtime = runtime; | |
3458 | /* | |
3459 | * While we are ensured activity in the period following an | |
3460 | * unthrottle, this also covers the case in which the new bandwidth is | |
3461 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3462 | * timer to remain active while there are any throttled entities.) | |
3463 | */ | |
3464 | cfs_b->idle = 0; | |
58088ad0 PT |
3465 | out_unlock: |
3466 | if (idle) | |
3467 | cfs_b->timer_active = 0; | |
3468 | raw_spin_unlock(&cfs_b->lock); | |
3469 | ||
3470 | return idle; | |
3471 | } | |
d3d9dc33 | 3472 | |
d8b4986d PT |
3473 | /* a cfs_rq won't donate quota below this amount */ |
3474 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3475 | /* minimum remaining period time to redistribute slack quota */ | |
3476 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3477 | /* how long we wait to gather additional slack before distributing */ | |
3478 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3479 | ||
db06e78c BS |
3480 | /* |
3481 | * Are we near the end of the current quota period? | |
3482 | * | |
3483 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
3484 | * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of | |
3485 | * migrate_hrtimers, base is never cleared, so we are fine. | |
3486 | */ | |
d8b4986d PT |
3487 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
3488 | { | |
3489 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3490 | u64 remaining; | |
3491 | ||
3492 | /* if the call-back is running a quota refresh is already occurring */ | |
3493 | if (hrtimer_callback_running(refresh_timer)) | |
3494 | return 1; | |
3495 | ||
3496 | /* is a quota refresh about to occur? */ | |
3497 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3498 | if (remaining < min_expire) | |
3499 | return 1; | |
3500 | ||
3501 | return 0; | |
3502 | } | |
3503 | ||
3504 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3505 | { | |
3506 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3507 | ||
3508 | /* if there's a quota refresh soon don't bother with slack */ | |
3509 | if (runtime_refresh_within(cfs_b, min_left)) | |
3510 | return; | |
3511 | ||
3512 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3513 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3514 | } | |
3515 | ||
3516 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3517 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3518 | { | |
3519 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3520 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3521 | ||
3522 | if (slack_runtime <= 0) | |
3523 | return; | |
3524 | ||
3525 | raw_spin_lock(&cfs_b->lock); | |
3526 | if (cfs_b->quota != RUNTIME_INF && | |
3527 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3528 | cfs_b->runtime += slack_runtime; | |
3529 | ||
3530 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3531 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3532 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3533 | start_cfs_slack_bandwidth(cfs_b); | |
3534 | } | |
3535 | raw_spin_unlock(&cfs_b->lock); | |
3536 | ||
3537 | /* even if it's not valid for return we don't want to try again */ | |
3538 | cfs_rq->runtime_remaining -= slack_runtime; | |
3539 | } | |
3540 | ||
3541 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3542 | { | |
56f570e5 PT |
3543 | if (!cfs_bandwidth_used()) |
3544 | return; | |
3545 | ||
fccfdc6f | 3546 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3547 | return; |
3548 | ||
3549 | __return_cfs_rq_runtime(cfs_rq); | |
3550 | } | |
3551 | ||
3552 | /* | |
3553 | * This is done with a timer (instead of inline with bandwidth return) since | |
3554 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3555 | */ | |
3556 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3557 | { | |
3558 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3559 | u64 expires; | |
3560 | ||
3561 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
3562 | raw_spin_lock(&cfs_b->lock); |
3563 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
3564 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 3565 | return; |
db06e78c | 3566 | } |
d8b4986d | 3567 | |
d8b4986d PT |
3568 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { |
3569 | runtime = cfs_b->runtime; | |
3570 | cfs_b->runtime = 0; | |
3571 | } | |
3572 | expires = cfs_b->runtime_expires; | |
3573 | raw_spin_unlock(&cfs_b->lock); | |
3574 | ||
3575 | if (!runtime) | |
3576 | return; | |
3577 | ||
3578 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3579 | ||
3580 | raw_spin_lock(&cfs_b->lock); | |
3581 | if (expires == cfs_b->runtime_expires) | |
3582 | cfs_b->runtime = runtime; | |
3583 | raw_spin_unlock(&cfs_b->lock); | |
3584 | } | |
3585 | ||
d3d9dc33 PT |
3586 | /* |
3587 | * When a group wakes up we want to make sure that its quota is not already | |
3588 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3589 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3590 | */ | |
3591 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3592 | { | |
56f570e5 PT |
3593 | if (!cfs_bandwidth_used()) |
3594 | return; | |
3595 | ||
d3d9dc33 PT |
3596 | /* an active group must be handled by the update_curr()->put() path */ |
3597 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3598 | return; | |
3599 | ||
3600 | /* ensure the group is not already throttled */ | |
3601 | if (cfs_rq_throttled(cfs_rq)) | |
3602 | return; | |
3603 | ||
3604 | /* update runtime allocation */ | |
3605 | account_cfs_rq_runtime(cfs_rq, 0); | |
3606 | if (cfs_rq->runtime_remaining <= 0) | |
3607 | throttle_cfs_rq(cfs_rq); | |
3608 | } | |
3609 | ||
3610 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
678d5718 | 3611 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 3612 | { |
56f570e5 | 3613 | if (!cfs_bandwidth_used()) |
678d5718 | 3614 | return false; |
56f570e5 | 3615 | |
d3d9dc33 | 3616 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 3617 | return false; |
d3d9dc33 PT |
3618 | |
3619 | /* | |
3620 | * it's possible for a throttled entity to be forced into a running | |
3621 | * state (e.g. set_curr_task), in this case we're finished. | |
3622 | */ | |
3623 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 3624 | return true; |
d3d9dc33 PT |
3625 | |
3626 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 3627 | return true; |
d3d9dc33 | 3628 | } |
029632fb | 3629 | |
029632fb PZ |
3630 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3631 | { | |
3632 | struct cfs_bandwidth *cfs_b = | |
3633 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3634 | do_sched_cfs_slack_timer(cfs_b); | |
3635 | ||
3636 | return HRTIMER_NORESTART; | |
3637 | } | |
3638 | ||
3639 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3640 | { | |
3641 | struct cfs_bandwidth *cfs_b = | |
3642 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3643 | ktime_t now; | |
3644 | int overrun; | |
3645 | int idle = 0; | |
3646 | ||
3647 | for (;;) { | |
3648 | now = hrtimer_cb_get_time(timer); | |
3649 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3650 | ||
3651 | if (!overrun) | |
3652 | break; | |
3653 | ||
3654 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3655 | } | |
3656 | ||
3657 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3658 | } | |
3659 | ||
3660 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3661 | { | |
3662 | raw_spin_lock_init(&cfs_b->lock); | |
3663 | cfs_b->runtime = 0; | |
3664 | cfs_b->quota = RUNTIME_INF; | |
3665 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3666 | ||
3667 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3668 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3669 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3670 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3671 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3672 | } | |
3673 | ||
3674 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3675 | { | |
3676 | cfs_rq->runtime_enabled = 0; | |
3677 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3678 | } | |
3679 | ||
3680 | /* requires cfs_b->lock, may release to reprogram timer */ | |
3681 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3682 | { | |
3683 | /* | |
3684 | * The timer may be active because we're trying to set a new bandwidth | |
3685 | * period or because we're racing with the tear-down path | |
3686 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3687 | * terminates). In either case we ensure that it's re-programmed | |
3688 | */ | |
927b54fc BS |
3689 | while (unlikely(hrtimer_active(&cfs_b->period_timer)) && |
3690 | hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { | |
3691 | /* bounce the lock to allow do_sched_cfs_period_timer to run */ | |
029632fb | 3692 | raw_spin_unlock(&cfs_b->lock); |
927b54fc | 3693 | cpu_relax(); |
029632fb PZ |
3694 | raw_spin_lock(&cfs_b->lock); |
3695 | /* if someone else restarted the timer then we're done */ | |
3696 | if (cfs_b->timer_active) | |
3697 | return; | |
3698 | } | |
3699 | ||
3700 | cfs_b->timer_active = 1; | |
3701 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
3702 | } | |
3703 | ||
3704 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3705 | { | |
3706 | hrtimer_cancel(&cfs_b->period_timer); | |
3707 | hrtimer_cancel(&cfs_b->slack_timer); | |
3708 | } | |
3709 | ||
38dc3348 | 3710 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
3711 | { |
3712 | struct cfs_rq *cfs_rq; | |
3713 | ||
3714 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
3715 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3716 | ||
3717 | if (!cfs_rq->runtime_enabled) | |
3718 | continue; | |
3719 | ||
3720 | /* | |
3721 | * clock_task is not advancing so we just need to make sure | |
3722 | * there's some valid quota amount | |
3723 | */ | |
3724 | cfs_rq->runtime_remaining = cfs_b->quota; | |
3725 | if (cfs_rq_throttled(cfs_rq)) | |
3726 | unthrottle_cfs_rq(cfs_rq); | |
3727 | } | |
3728 | } | |
3729 | ||
3730 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
3731 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3732 | { | |
78becc27 | 3733 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
3734 | } |
3735 | ||
9dbdb155 | 3736 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 3737 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 3738 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
6c16a6dc | 3739 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
3740 | |
3741 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
3742 | { | |
3743 | return 0; | |
3744 | } | |
64660c86 PT |
3745 | |
3746 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3747 | { | |
3748 | return 0; | |
3749 | } | |
3750 | ||
3751 | static inline int throttled_lb_pair(struct task_group *tg, | |
3752 | int src_cpu, int dest_cpu) | |
3753 | { | |
3754 | return 0; | |
3755 | } | |
029632fb PZ |
3756 | |
3757 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
3758 | ||
3759 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
3760 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
3761 | #endif |
3762 | ||
029632fb PZ |
3763 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3764 | { | |
3765 | return NULL; | |
3766 | } | |
3767 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
a4c96ae3 | 3768 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
3769 | |
3770 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
3771 | ||
bf0f6f24 IM |
3772 | /************************************************** |
3773 | * CFS operations on tasks: | |
3774 | */ | |
3775 | ||
8f4d37ec PZ |
3776 | #ifdef CONFIG_SCHED_HRTICK |
3777 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3778 | { | |
8f4d37ec PZ |
3779 | struct sched_entity *se = &p->se; |
3780 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3781 | ||
3782 | WARN_ON(task_rq(p) != rq); | |
3783 | ||
b39e66ea | 3784 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
3785 | u64 slice = sched_slice(cfs_rq, se); |
3786 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
3787 | s64 delta = slice - ran; | |
3788 | ||
3789 | if (delta < 0) { | |
3790 | if (rq->curr == p) | |
3791 | resched_task(p); | |
3792 | return; | |
3793 | } | |
3794 | ||
3795 | /* | |
3796 | * Don't schedule slices shorter than 10000ns, that just | |
3797 | * doesn't make sense. Rely on vruntime for fairness. | |
3798 | */ | |
31656519 | 3799 | if (rq->curr != p) |
157124c1 | 3800 | delta = max_t(s64, 10000LL, delta); |
8f4d37ec | 3801 | |
31656519 | 3802 | hrtick_start(rq, delta); |
8f4d37ec PZ |
3803 | } |
3804 | } | |
a4c2f00f PZ |
3805 | |
3806 | /* | |
3807 | * called from enqueue/dequeue and updates the hrtick when the | |
3808 | * current task is from our class and nr_running is low enough | |
3809 | * to matter. | |
3810 | */ | |
3811 | static void hrtick_update(struct rq *rq) | |
3812 | { | |
3813 | struct task_struct *curr = rq->curr; | |
3814 | ||
b39e66ea | 3815 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
3816 | return; |
3817 | ||
3818 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
3819 | hrtick_start_fair(rq, curr); | |
3820 | } | |
55e12e5e | 3821 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
3822 | static inline void |
3823 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
3824 | { | |
3825 | } | |
a4c2f00f PZ |
3826 | |
3827 | static inline void hrtick_update(struct rq *rq) | |
3828 | { | |
3829 | } | |
8f4d37ec PZ |
3830 | #endif |
3831 | ||
bf0f6f24 IM |
3832 | /* |
3833 | * The enqueue_task method is called before nr_running is | |
3834 | * increased. Here we update the fair scheduling stats and | |
3835 | * then put the task into the rbtree: | |
3836 | */ | |
ea87bb78 | 3837 | static void |
371fd7e7 | 3838 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3839 | { |
3840 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3841 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
3842 | |
3843 | for_each_sched_entity(se) { | |
62fb1851 | 3844 | if (se->on_rq) |
bf0f6f24 IM |
3845 | break; |
3846 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 3847 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
3848 | |
3849 | /* | |
3850 | * end evaluation on encountering a throttled cfs_rq | |
3851 | * | |
3852 | * note: in the case of encountering a throttled cfs_rq we will | |
3853 | * post the final h_nr_running increment below. | |
3854 | */ | |
3855 | if (cfs_rq_throttled(cfs_rq)) | |
3856 | break; | |
953bfcd1 | 3857 | cfs_rq->h_nr_running++; |
85dac906 | 3858 | |
88ec22d3 | 3859 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 3860 | } |
8f4d37ec | 3861 | |
2069dd75 | 3862 | for_each_sched_entity(se) { |
0f317143 | 3863 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3864 | cfs_rq->h_nr_running++; |
2069dd75 | 3865 | |
85dac906 PT |
3866 | if (cfs_rq_throttled(cfs_rq)) |
3867 | break; | |
3868 | ||
17bc14b7 | 3869 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3870 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3871 | } |
3872 | ||
18bf2805 BS |
3873 | if (!se) { |
3874 | update_rq_runnable_avg(rq, rq->nr_running); | |
85dac906 | 3875 | inc_nr_running(rq); |
18bf2805 | 3876 | } |
a4c2f00f | 3877 | hrtick_update(rq); |
bf0f6f24 IM |
3878 | } |
3879 | ||
2f36825b VP |
3880 | static void set_next_buddy(struct sched_entity *se); |
3881 | ||
bf0f6f24 IM |
3882 | /* |
3883 | * The dequeue_task method is called before nr_running is | |
3884 | * decreased. We remove the task from the rbtree and | |
3885 | * update the fair scheduling stats: | |
3886 | */ | |
371fd7e7 | 3887 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
3888 | { |
3889 | struct cfs_rq *cfs_rq; | |
62fb1851 | 3890 | struct sched_entity *se = &p->se; |
2f36825b | 3891 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
3892 | |
3893 | for_each_sched_entity(se) { | |
3894 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 3895 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
3896 | |
3897 | /* | |
3898 | * end evaluation on encountering a throttled cfs_rq | |
3899 | * | |
3900 | * note: in the case of encountering a throttled cfs_rq we will | |
3901 | * post the final h_nr_running decrement below. | |
3902 | */ | |
3903 | if (cfs_rq_throttled(cfs_rq)) | |
3904 | break; | |
953bfcd1 | 3905 | cfs_rq->h_nr_running--; |
2069dd75 | 3906 | |
bf0f6f24 | 3907 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
3908 | if (cfs_rq->load.weight) { |
3909 | /* | |
3910 | * Bias pick_next to pick a task from this cfs_rq, as | |
3911 | * p is sleeping when it is within its sched_slice. | |
3912 | */ | |
3913 | if (task_sleep && parent_entity(se)) | |
3914 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
3915 | |
3916 | /* avoid re-evaluating load for this entity */ | |
3917 | se = parent_entity(se); | |
bf0f6f24 | 3918 | break; |
2f36825b | 3919 | } |
371fd7e7 | 3920 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 3921 | } |
8f4d37ec | 3922 | |
2069dd75 | 3923 | for_each_sched_entity(se) { |
0f317143 | 3924 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 3925 | cfs_rq->h_nr_running--; |
2069dd75 | 3926 | |
85dac906 PT |
3927 | if (cfs_rq_throttled(cfs_rq)) |
3928 | break; | |
3929 | ||
17bc14b7 | 3930 | update_cfs_shares(cfs_rq); |
9ee474f5 | 3931 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
3932 | } |
3933 | ||
18bf2805 | 3934 | if (!se) { |
85dac906 | 3935 | dec_nr_running(rq); |
18bf2805 BS |
3936 | update_rq_runnable_avg(rq, 1); |
3937 | } | |
a4c2f00f | 3938 | hrtick_update(rq); |
bf0f6f24 IM |
3939 | } |
3940 | ||
e7693a36 | 3941 | #ifdef CONFIG_SMP |
029632fb PZ |
3942 | /* Used instead of source_load when we know the type == 0 */ |
3943 | static unsigned long weighted_cpuload(const int cpu) | |
3944 | { | |
b92486cb | 3945 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
3946 | } |
3947 | ||
3948 | /* | |
3949 | * Return a low guess at the load of a migration-source cpu weighted | |
3950 | * according to the scheduling class and "nice" value. | |
3951 | * | |
3952 | * We want to under-estimate the load of migration sources, to | |
3953 | * balance conservatively. | |
3954 | */ | |
3955 | static unsigned long source_load(int cpu, int type) | |
3956 | { | |
3957 | struct rq *rq = cpu_rq(cpu); | |
3958 | unsigned long total = weighted_cpuload(cpu); | |
3959 | ||
3960 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3961 | return total; | |
3962 | ||
3963 | return min(rq->cpu_load[type-1], total); | |
3964 | } | |
3965 | ||
3966 | /* | |
3967 | * Return a high guess at the load of a migration-target cpu weighted | |
3968 | * according to the scheduling class and "nice" value. | |
3969 | */ | |
3970 | static unsigned long target_load(int cpu, int type) | |
3971 | { | |
3972 | struct rq *rq = cpu_rq(cpu); | |
3973 | unsigned long total = weighted_cpuload(cpu); | |
3974 | ||
3975 | if (type == 0 || !sched_feat(LB_BIAS)) | |
3976 | return total; | |
3977 | ||
3978 | return max(rq->cpu_load[type-1], total); | |
3979 | } | |
3980 | ||
3981 | static unsigned long power_of(int cpu) | |
3982 | { | |
3983 | return cpu_rq(cpu)->cpu_power; | |
3984 | } | |
3985 | ||
3986 | static unsigned long cpu_avg_load_per_task(int cpu) | |
3987 | { | |
3988 | struct rq *rq = cpu_rq(cpu); | |
3989 | unsigned long nr_running = ACCESS_ONCE(rq->nr_running); | |
b92486cb | 3990 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
3991 | |
3992 | if (nr_running) | |
b92486cb | 3993 | return load_avg / nr_running; |
029632fb PZ |
3994 | |
3995 | return 0; | |
3996 | } | |
3997 | ||
62470419 MW |
3998 | static void record_wakee(struct task_struct *p) |
3999 | { | |
4000 | /* | |
4001 | * Rough decay (wiping) for cost saving, don't worry | |
4002 | * about the boundary, really active task won't care | |
4003 | * about the loss. | |
4004 | */ | |
4005 | if (jiffies > current->wakee_flip_decay_ts + HZ) { | |
4006 | current->wakee_flips = 0; | |
4007 | current->wakee_flip_decay_ts = jiffies; | |
4008 | } | |
4009 | ||
4010 | if (current->last_wakee != p) { | |
4011 | current->last_wakee = p; | |
4012 | current->wakee_flips++; | |
4013 | } | |
4014 | } | |
098fb9db | 4015 | |
74f8e4b2 | 4016 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
4017 | { |
4018 | struct sched_entity *se = &p->se; | |
4019 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
4020 | u64 min_vruntime; |
4021 | ||
4022 | #ifndef CONFIG_64BIT | |
4023 | u64 min_vruntime_copy; | |
88ec22d3 | 4024 | |
3fe1698b PZ |
4025 | do { |
4026 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
4027 | smp_rmb(); | |
4028 | min_vruntime = cfs_rq->min_vruntime; | |
4029 | } while (min_vruntime != min_vruntime_copy); | |
4030 | #else | |
4031 | min_vruntime = cfs_rq->min_vruntime; | |
4032 | #endif | |
88ec22d3 | 4033 | |
3fe1698b | 4034 | se->vruntime -= min_vruntime; |
62470419 | 4035 | record_wakee(p); |
88ec22d3 PZ |
4036 | } |
4037 | ||
bb3469ac | 4038 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
4039 | /* |
4040 | * effective_load() calculates the load change as seen from the root_task_group | |
4041 | * | |
4042 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
4043 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
4044 | * can calculate the shift in shares. | |
cf5f0acf PZ |
4045 | * |
4046 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
4047 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
4048 | * total group weight. | |
4049 | * | |
4050 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
4051 | * distribution (s_i) using: | |
4052 | * | |
4053 | * s_i = rw_i / \Sum rw_j (1) | |
4054 | * | |
4055 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
4056 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
4057 | * shares distribution (s_i): | |
4058 | * | |
4059 | * rw_i = { 2, 4, 1, 0 } | |
4060 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
4061 | * | |
4062 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
4063 | * task used to run on and the CPU the waker is running on), we need to | |
4064 | * compute the effect of waking a task on either CPU and, in case of a sync | |
4065 | * wakeup, compute the effect of the current task going to sleep. | |
4066 | * | |
4067 | * So for a change of @wl to the local @cpu with an overall group weight change | |
4068 | * of @wl we can compute the new shares distribution (s'_i) using: | |
4069 | * | |
4070 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
4071 | * | |
4072 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
4073 | * differences in waking a task to CPU 0. The additional task changes the | |
4074 | * weight and shares distributions like: | |
4075 | * | |
4076 | * rw'_i = { 3, 4, 1, 0 } | |
4077 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
4078 | * | |
4079 | * We can then compute the difference in effective weight by using: | |
4080 | * | |
4081 | * dw_i = S * (s'_i - s_i) (3) | |
4082 | * | |
4083 | * Where 'S' is the group weight as seen by its parent. | |
4084 | * | |
4085 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
4086 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
4087 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 4088 | */ |
2069dd75 | 4089 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 4090 | { |
4be9daaa | 4091 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 4092 | |
9722c2da | 4093 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
4094 | return wl; |
4095 | ||
4be9daaa | 4096 | for_each_sched_entity(se) { |
cf5f0acf | 4097 | long w, W; |
4be9daaa | 4098 | |
977dda7c | 4099 | tg = se->my_q->tg; |
bb3469ac | 4100 | |
cf5f0acf PZ |
4101 | /* |
4102 | * W = @wg + \Sum rw_j | |
4103 | */ | |
4104 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 4105 | |
cf5f0acf PZ |
4106 | /* |
4107 | * w = rw_i + @wl | |
4108 | */ | |
4109 | w = se->my_q->load.weight + wl; | |
940959e9 | 4110 | |
cf5f0acf PZ |
4111 | /* |
4112 | * wl = S * s'_i; see (2) | |
4113 | */ | |
4114 | if (W > 0 && w < W) | |
4115 | wl = (w * tg->shares) / W; | |
977dda7c PT |
4116 | else |
4117 | wl = tg->shares; | |
940959e9 | 4118 | |
cf5f0acf PZ |
4119 | /* |
4120 | * Per the above, wl is the new se->load.weight value; since | |
4121 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
4122 | * calc_cfs_shares(). | |
4123 | */ | |
977dda7c PT |
4124 | if (wl < MIN_SHARES) |
4125 | wl = MIN_SHARES; | |
cf5f0acf PZ |
4126 | |
4127 | /* | |
4128 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
4129 | */ | |
977dda7c | 4130 | wl -= se->load.weight; |
cf5f0acf PZ |
4131 | |
4132 | /* | |
4133 | * Recursively apply this logic to all parent groups to compute | |
4134 | * the final effective load change on the root group. Since | |
4135 | * only the @tg group gets extra weight, all parent groups can | |
4136 | * only redistribute existing shares. @wl is the shift in shares | |
4137 | * resulting from this level per the above. | |
4138 | */ | |
4be9daaa | 4139 | wg = 0; |
4be9daaa | 4140 | } |
bb3469ac | 4141 | |
4be9daaa | 4142 | return wl; |
bb3469ac PZ |
4143 | } |
4144 | #else | |
4be9daaa | 4145 | |
58d081b5 | 4146 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 4147 | { |
83378269 | 4148 | return wl; |
bb3469ac | 4149 | } |
4be9daaa | 4150 | |
bb3469ac PZ |
4151 | #endif |
4152 | ||
62470419 MW |
4153 | static int wake_wide(struct task_struct *p) |
4154 | { | |
7d9ffa89 | 4155 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
4156 | |
4157 | /* | |
4158 | * Yeah, it's the switching-frequency, could means many wakee or | |
4159 | * rapidly switch, use factor here will just help to automatically | |
4160 | * adjust the loose-degree, so bigger node will lead to more pull. | |
4161 | */ | |
4162 | if (p->wakee_flips > factor) { | |
4163 | /* | |
4164 | * wakee is somewhat hot, it needs certain amount of cpu | |
4165 | * resource, so if waker is far more hot, prefer to leave | |
4166 | * it alone. | |
4167 | */ | |
4168 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
4169 | return 1; | |
4170 | } | |
4171 | ||
4172 | return 0; | |
4173 | } | |
4174 | ||
c88d5910 | 4175 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 4176 | { |
e37b6a7b | 4177 | s64 this_load, load; |
c88d5910 | 4178 | int idx, this_cpu, prev_cpu; |
098fb9db | 4179 | unsigned long tl_per_task; |
c88d5910 | 4180 | struct task_group *tg; |
83378269 | 4181 | unsigned long weight; |
b3137bc8 | 4182 | int balanced; |
098fb9db | 4183 | |
62470419 MW |
4184 | /* |
4185 | * If we wake multiple tasks be careful to not bounce | |
4186 | * ourselves around too much. | |
4187 | */ | |
4188 | if (wake_wide(p)) | |
4189 | return 0; | |
4190 | ||
c88d5910 PZ |
4191 | idx = sd->wake_idx; |
4192 | this_cpu = smp_processor_id(); | |
4193 | prev_cpu = task_cpu(p); | |
4194 | load = source_load(prev_cpu, idx); | |
4195 | this_load = target_load(this_cpu, idx); | |
098fb9db | 4196 | |
b3137bc8 MG |
4197 | /* |
4198 | * If sync wakeup then subtract the (maximum possible) | |
4199 | * effect of the currently running task from the load | |
4200 | * of the current CPU: | |
4201 | */ | |
83378269 PZ |
4202 | if (sync) { |
4203 | tg = task_group(current); | |
4204 | weight = current->se.load.weight; | |
4205 | ||
c88d5910 | 4206 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
4207 | load += effective_load(tg, prev_cpu, 0, -weight); |
4208 | } | |
b3137bc8 | 4209 | |
83378269 PZ |
4210 | tg = task_group(p); |
4211 | weight = p->se.load.weight; | |
b3137bc8 | 4212 | |
71a29aa7 PZ |
4213 | /* |
4214 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
4215 | * due to the sync cause above having dropped this_load to 0, we'll |
4216 | * always have an imbalance, but there's really nothing you can do | |
4217 | * about that, so that's good too. | |
71a29aa7 PZ |
4218 | * |
4219 | * Otherwise check if either cpus are near enough in load to allow this | |
4220 | * task to be woken on this_cpu. | |
4221 | */ | |
e37b6a7b PT |
4222 | if (this_load > 0) { |
4223 | s64 this_eff_load, prev_eff_load; | |
e51fd5e2 PZ |
4224 | |
4225 | this_eff_load = 100; | |
4226 | this_eff_load *= power_of(prev_cpu); | |
4227 | this_eff_load *= this_load + | |
4228 | effective_load(tg, this_cpu, weight, weight); | |
4229 | ||
4230 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | |
4231 | prev_eff_load *= power_of(this_cpu); | |
4232 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | |
4233 | ||
4234 | balanced = this_eff_load <= prev_eff_load; | |
4235 | } else | |
4236 | balanced = true; | |
b3137bc8 | 4237 | |
098fb9db | 4238 | /* |
4ae7d5ce IM |
4239 | * If the currently running task will sleep within |
4240 | * a reasonable amount of time then attract this newly | |
4241 | * woken task: | |
098fb9db | 4242 | */ |
2fb7635c PZ |
4243 | if (sync && balanced) |
4244 | return 1; | |
098fb9db | 4245 | |
41acab88 | 4246 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db IM |
4247 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
4248 | ||
c88d5910 PZ |
4249 | if (balanced || |
4250 | (this_load <= load && | |
4251 | this_load + target_load(prev_cpu, idx) <= tl_per_task)) { | |
098fb9db IM |
4252 | /* |
4253 | * This domain has SD_WAKE_AFFINE and | |
4254 | * p is cache cold in this domain, and | |
4255 | * there is no bad imbalance. | |
4256 | */ | |
c88d5910 | 4257 | schedstat_inc(sd, ttwu_move_affine); |
41acab88 | 4258 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
098fb9db IM |
4259 | |
4260 | return 1; | |
4261 | } | |
4262 | return 0; | |
4263 | } | |
4264 | ||
aaee1203 PZ |
4265 | /* |
4266 | * find_idlest_group finds and returns the least busy CPU group within the | |
4267 | * domain. | |
4268 | */ | |
4269 | static struct sched_group * | |
78e7ed53 | 4270 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 4271 | int this_cpu, int sd_flag) |
e7693a36 | 4272 | { |
b3bd3de6 | 4273 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 4274 | unsigned long min_load = ULONG_MAX, this_load = 0; |
c44f2a02 | 4275 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4276 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 4277 | |
c44f2a02 VG |
4278 | if (sd_flag & SD_BALANCE_WAKE) |
4279 | load_idx = sd->wake_idx; | |
4280 | ||
aaee1203 PZ |
4281 | do { |
4282 | unsigned long load, avg_load; | |
4283 | int local_group; | |
4284 | int i; | |
e7693a36 | 4285 | |
aaee1203 PZ |
4286 | /* Skip over this group if it has no CPUs allowed */ |
4287 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 4288 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
4289 | continue; |
4290 | ||
4291 | local_group = cpumask_test_cpu(this_cpu, | |
4292 | sched_group_cpus(group)); | |
4293 | ||
4294 | /* Tally up the load of all CPUs in the group */ | |
4295 | avg_load = 0; | |
4296 | ||
4297 | for_each_cpu(i, sched_group_cpus(group)) { | |
4298 | /* Bias balancing toward cpus of our domain */ | |
4299 | if (local_group) | |
4300 | load = source_load(i, load_idx); | |
4301 | else | |
4302 | load = target_load(i, load_idx); | |
4303 | ||
4304 | avg_load += load; | |
4305 | } | |
4306 | ||
4307 | /* Adjust by relative CPU power of the group */ | |
9c3f75cb | 4308 | avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; |
aaee1203 PZ |
4309 | |
4310 | if (local_group) { | |
4311 | this_load = avg_load; | |
aaee1203 PZ |
4312 | } else if (avg_load < min_load) { |
4313 | min_load = avg_load; | |
4314 | idlest = group; | |
4315 | } | |
4316 | } while (group = group->next, group != sd->groups); | |
4317 | ||
4318 | if (!idlest || 100*this_load < imbalance*min_load) | |
4319 | return NULL; | |
4320 | return idlest; | |
4321 | } | |
4322 | ||
4323 | /* | |
4324 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
4325 | */ | |
4326 | static int | |
4327 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
4328 | { | |
4329 | unsigned long load, min_load = ULONG_MAX; | |
4330 | int idlest = -1; | |
4331 | int i; | |
4332 | ||
4333 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 4334 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
aaee1203 PZ |
4335 | load = weighted_cpuload(i); |
4336 | ||
4337 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
4338 | min_load = load; | |
4339 | idlest = i; | |
e7693a36 GH |
4340 | } |
4341 | } | |
4342 | ||
aaee1203 PZ |
4343 | return idlest; |
4344 | } | |
e7693a36 | 4345 | |
a50bde51 PZ |
4346 | /* |
4347 | * Try and locate an idle CPU in the sched_domain. | |
4348 | */ | |
99bd5e2f | 4349 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 4350 | { |
99bd5e2f | 4351 | struct sched_domain *sd; |
37407ea7 | 4352 | struct sched_group *sg; |
e0a79f52 | 4353 | int i = task_cpu(p); |
a50bde51 | 4354 | |
e0a79f52 MG |
4355 | if (idle_cpu(target)) |
4356 | return target; | |
99bd5e2f SS |
4357 | |
4358 | /* | |
e0a79f52 | 4359 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 4360 | */ |
e0a79f52 MG |
4361 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4362 | return i; | |
a50bde51 PZ |
4363 | |
4364 | /* | |
37407ea7 | 4365 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 4366 | */ |
518cd623 | 4367 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 4368 | for_each_lower_domain(sd) { |
37407ea7 LT |
4369 | sg = sd->groups; |
4370 | do { | |
4371 | if (!cpumask_intersects(sched_group_cpus(sg), | |
4372 | tsk_cpus_allowed(p))) | |
4373 | goto next; | |
4374 | ||
4375 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 4376 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
4377 | goto next; |
4378 | } | |
970e1789 | 4379 | |
37407ea7 LT |
4380 | target = cpumask_first_and(sched_group_cpus(sg), |
4381 | tsk_cpus_allowed(p)); | |
4382 | goto done; | |
4383 | next: | |
4384 | sg = sg->next; | |
4385 | } while (sg != sd->groups); | |
4386 | } | |
4387 | done: | |
a50bde51 PZ |
4388 | return target; |
4389 | } | |
4390 | ||
aaee1203 PZ |
4391 | /* |
4392 | * sched_balance_self: balance the current task (running on cpu) in domains | |
4393 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
4394 | * SD_BALANCE_EXEC. | |
4395 | * | |
4396 | * Balance, ie. select the least loaded group. | |
4397 | * | |
4398 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
4399 | * | |
4400 | * preempt must be disabled. | |
4401 | */ | |
0017d735 | 4402 | static int |
ac66f547 | 4403 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 4404 | { |
29cd8bae | 4405 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 4406 | int cpu = smp_processor_id(); |
c88d5910 | 4407 | int new_cpu = cpu; |
99bd5e2f | 4408 | int want_affine = 0; |
5158f4e4 | 4409 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 4410 | |
29baa747 | 4411 | if (p->nr_cpus_allowed == 1) |
76854c7e MG |
4412 | return prev_cpu; |
4413 | ||
0763a660 | 4414 | if (sd_flag & SD_BALANCE_WAKE) { |
fa17b507 | 4415 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
c88d5910 PZ |
4416 | want_affine = 1; |
4417 | new_cpu = prev_cpu; | |
4418 | } | |
aaee1203 | 4419 | |
dce840a0 | 4420 | rcu_read_lock(); |
aaee1203 | 4421 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4422 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4423 | continue; | |
4424 | ||
fe3bcfe1 | 4425 | /* |
99bd5e2f SS |
4426 | * If both cpu and prev_cpu are part of this domain, |
4427 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4428 | */ |
99bd5e2f SS |
4429 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4430 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4431 | affine_sd = tmp; | |
29cd8bae | 4432 | break; |
f03542a7 | 4433 | } |
29cd8bae | 4434 | |
f03542a7 | 4435 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4436 | sd = tmp; |
4437 | } | |
4438 | ||
8b911acd | 4439 | if (affine_sd) { |
f03542a7 | 4440 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
dce840a0 PZ |
4441 | prev_cpu = cpu; |
4442 | ||
4443 | new_cpu = select_idle_sibling(p, prev_cpu); | |
4444 | goto unlock; | |
8b911acd | 4445 | } |
e7693a36 | 4446 | |
aaee1203 PZ |
4447 | while (sd) { |
4448 | struct sched_group *group; | |
c88d5910 | 4449 | int weight; |
098fb9db | 4450 | |
0763a660 | 4451 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4452 | sd = sd->child; |
4453 | continue; | |
4454 | } | |
098fb9db | 4455 | |
c44f2a02 | 4456 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
4457 | if (!group) { |
4458 | sd = sd->child; | |
4459 | continue; | |
4460 | } | |
4ae7d5ce | 4461 | |
d7c33c49 | 4462 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4463 | if (new_cpu == -1 || new_cpu == cpu) { |
4464 | /* Now try balancing at a lower domain level of cpu */ | |
4465 | sd = sd->child; | |
4466 | continue; | |
e7693a36 | 4467 | } |
aaee1203 PZ |
4468 | |
4469 | /* Now try balancing at a lower domain level of new_cpu */ | |
4470 | cpu = new_cpu; | |
669c55e9 | 4471 | weight = sd->span_weight; |
aaee1203 PZ |
4472 | sd = NULL; |
4473 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4474 | if (weight <= tmp->span_weight) |
aaee1203 | 4475 | break; |
0763a660 | 4476 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4477 | sd = tmp; |
4478 | } | |
4479 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4480 | } |
dce840a0 PZ |
4481 | unlock: |
4482 | rcu_read_unlock(); | |
e7693a36 | 4483 | |
c88d5910 | 4484 | return new_cpu; |
e7693a36 | 4485 | } |
0a74bef8 PT |
4486 | |
4487 | /* | |
4488 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4489 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4490 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4491 | * other assumptions, including the state of rq->lock, should be made. | |
4492 | */ | |
4493 | static void | |
4494 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4495 | { | |
aff3e498 PT |
4496 | struct sched_entity *se = &p->se; |
4497 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4498 | ||
4499 | /* | |
4500 | * Load tracking: accumulate removed load so that it can be processed | |
4501 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4502 | * to blocked load iff they have a positive decay-count. It can never | |
4503 | * be negative here since on-rq tasks have decay-count == 0. | |
4504 | */ | |
4505 | if (se->avg.decay_count) { | |
4506 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4507 | atomic_long_add(se->avg.load_avg_contrib, |
4508 | &cfs_rq->removed_load); | |
aff3e498 | 4509 | } |
0a74bef8 | 4510 | } |
e7693a36 GH |
4511 | #endif /* CONFIG_SMP */ |
4512 | ||
e52fb7c0 PZ |
4513 | static unsigned long |
4514 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4515 | { |
4516 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4517 | ||
4518 | /* | |
e52fb7c0 PZ |
4519 | * Since its curr running now, convert the gran from real-time |
4520 | * to virtual-time in his units. | |
13814d42 MG |
4521 | * |
4522 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4523 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4524 | * the resulting gran will be larger, therefore penalizing the | |
4525 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4526 | * be smaller, again penalizing the lighter task. | |
4527 | * | |
4528 | * This is especially important for buddies when the leftmost | |
4529 | * task is higher priority than the buddy. | |
0bbd3336 | 4530 | */ |
f4ad9bd2 | 4531 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4532 | } |
4533 | ||
464b7527 PZ |
4534 | /* |
4535 | * Should 'se' preempt 'curr'. | |
4536 | * | |
4537 | * |s1 | |
4538 | * |s2 | |
4539 | * |s3 | |
4540 | * g | |
4541 | * |<--->|c | |
4542 | * | |
4543 | * w(c, s1) = -1 | |
4544 | * w(c, s2) = 0 | |
4545 | * w(c, s3) = 1 | |
4546 | * | |
4547 | */ | |
4548 | static int | |
4549 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4550 | { | |
4551 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4552 | ||
4553 | if (vdiff <= 0) | |
4554 | return -1; | |
4555 | ||
e52fb7c0 | 4556 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4557 | if (vdiff > gran) |
4558 | return 1; | |
4559 | ||
4560 | return 0; | |
4561 | } | |
4562 | ||
02479099 PZ |
4563 | static void set_last_buddy(struct sched_entity *se) |
4564 | { | |
69c80f3e VP |
4565 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4566 | return; | |
4567 | ||
4568 | for_each_sched_entity(se) | |
4569 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4570 | } |
4571 | ||
4572 | static void set_next_buddy(struct sched_entity *se) | |
4573 | { | |
69c80f3e VP |
4574 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4575 | return; | |
4576 | ||
4577 | for_each_sched_entity(se) | |
4578 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4579 | } |
4580 | ||
ac53db59 RR |
4581 | static void set_skip_buddy(struct sched_entity *se) |
4582 | { | |
69c80f3e VP |
4583 | for_each_sched_entity(se) |
4584 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4585 | } |
4586 | ||
bf0f6f24 IM |
4587 | /* |
4588 | * Preempt the current task with a newly woken task if needed: | |
4589 | */ | |
5a9b86f6 | 4590 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4591 | { |
4592 | struct task_struct *curr = rq->curr; | |
8651a86c | 4593 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4594 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4595 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4596 | int next_buddy_marked = 0; |
bf0f6f24 | 4597 | |
4ae7d5ce IM |
4598 | if (unlikely(se == pse)) |
4599 | return; | |
4600 | ||
5238cdd3 | 4601 | /* |
ddcdf6e7 | 4602 | * This is possible from callers such as move_task(), in which we |
5238cdd3 PT |
4603 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4604 | * lead to a throttle). This both saves work and prevents false | |
4605 | * next-buddy nomination below. | |
4606 | */ | |
4607 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4608 | return; | |
4609 | ||
2f36825b | 4610 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4611 | set_next_buddy(pse); |
2f36825b VP |
4612 | next_buddy_marked = 1; |
4613 | } | |
57fdc26d | 4614 | |
aec0a514 BR |
4615 | /* |
4616 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4617 | * wake up path. | |
5238cdd3 PT |
4618 | * |
4619 | * Note: this also catches the edge-case of curr being in a throttled | |
4620 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4621 | * enqueue of curr) will have resulted in resched being set. This | |
4622 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4623 | * below. | |
aec0a514 BR |
4624 | */ |
4625 | if (test_tsk_need_resched(curr)) | |
4626 | return; | |
4627 | ||
a2f5c9ab DH |
4628 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4629 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4630 | likely(p->policy != SCHED_IDLE)) | |
4631 | goto preempt; | |
4632 | ||
91c234b4 | 4633 | /* |
a2f5c9ab DH |
4634 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4635 | * is driven by the tick): | |
91c234b4 | 4636 | */ |
8ed92e51 | 4637 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4638 | return; |
bf0f6f24 | 4639 | |
464b7527 | 4640 | find_matching_se(&se, &pse); |
9bbd7374 | 4641 | update_curr(cfs_rq_of(se)); |
002f128b | 4642 | BUG_ON(!pse); |
2f36825b VP |
4643 | if (wakeup_preempt_entity(se, pse) == 1) { |
4644 | /* | |
4645 | * Bias pick_next to pick the sched entity that is | |
4646 | * triggering this preemption. | |
4647 | */ | |
4648 | if (!next_buddy_marked) | |
4649 | set_next_buddy(pse); | |
3a7e73a2 | 4650 | goto preempt; |
2f36825b | 4651 | } |
464b7527 | 4652 | |
3a7e73a2 | 4653 | return; |
a65ac745 | 4654 | |
3a7e73a2 PZ |
4655 | preempt: |
4656 | resched_task(curr); | |
4657 | /* | |
4658 | * Only set the backward buddy when the current task is still | |
4659 | * on the rq. This can happen when a wakeup gets interleaved | |
4660 | * with schedule on the ->pre_schedule() or idle_balance() | |
4661 | * point, either of which can * drop the rq lock. | |
4662 | * | |
4663 | * Also, during early boot the idle thread is in the fair class, | |
4664 | * for obvious reasons its a bad idea to schedule back to it. | |
4665 | */ | |
4666 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
4667 | return; | |
4668 | ||
4669 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
4670 | set_last_buddy(se); | |
bf0f6f24 IM |
4671 | } |
4672 | ||
606dba2e PZ |
4673 | static struct task_struct * |
4674 | pick_next_task_fair(struct rq *rq, struct task_struct *prev) | |
bf0f6f24 IM |
4675 | { |
4676 | struct cfs_rq *cfs_rq = &rq->cfs; | |
4677 | struct sched_entity *se; | |
678d5718 PZ |
4678 | struct task_struct *p; |
4679 | ||
38033c37 | 4680 | again: __maybe_unused |
678d5718 PZ |
4681 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4682 | if (!cfs_rq->nr_running) | |
38033c37 | 4683 | goto idle; |
678d5718 PZ |
4684 | |
4685 | if (!prev || prev->sched_class != &fair_sched_class) | |
4686 | goto simple; | |
4687 | ||
4688 | /* | |
4689 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
4690 | * likely that a next task is from the same cgroup as the current. | |
4691 | * | |
4692 | * Therefore attempt to avoid putting and setting the entire cgroup | |
4693 | * hierarchy, only change the part that actually changes. | |
4694 | */ | |
4695 | ||
4696 | do { | |
4697 | struct sched_entity *curr = cfs_rq->curr; | |
4698 | ||
4699 | /* | |
4700 | * Since we got here without doing put_prev_entity() we also | |
4701 | * have to consider cfs_rq->curr. If it is still a runnable | |
4702 | * entity, update_curr() will update its vruntime, otherwise | |
4703 | * forget we've ever seen it. | |
4704 | */ | |
4705 | if (curr && curr->on_rq) | |
4706 | update_curr(cfs_rq); | |
4707 | else | |
4708 | curr = NULL; | |
4709 | ||
4710 | /* | |
4711 | * This call to check_cfs_rq_runtime() will do the throttle and | |
4712 | * dequeue its entity in the parent(s). Therefore the 'simple' | |
4713 | * nr_running test will indeed be correct. | |
4714 | */ | |
4715 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
4716 | goto simple; | |
4717 | ||
4718 | se = pick_next_entity(cfs_rq, curr); | |
4719 | cfs_rq = group_cfs_rq(se); | |
4720 | } while (cfs_rq); | |
4721 | ||
4722 | p = task_of(se); | |
4723 | ||
4724 | /* | |
4725 | * Since we haven't yet done put_prev_entity and if the selected task | |
4726 | * is a different task than we started out with, try and touch the | |
4727 | * least amount of cfs_rqs. | |
4728 | */ | |
4729 | if (prev != p) { | |
4730 | struct sched_entity *pse = &prev->se; | |
4731 | ||
4732 | while (!(cfs_rq = is_same_group(se, pse))) { | |
4733 | int se_depth = se->depth; | |
4734 | int pse_depth = pse->depth; | |
4735 | ||
4736 | if (se_depth <= pse_depth) { | |
4737 | put_prev_entity(cfs_rq_of(pse), pse); | |
4738 | pse = parent_entity(pse); | |
4739 | } | |
4740 | if (se_depth >= pse_depth) { | |
4741 | set_next_entity(cfs_rq_of(se), se); | |
4742 | se = parent_entity(se); | |
4743 | } | |
4744 | } | |
4745 | ||
4746 | put_prev_entity(cfs_rq, pse); | |
4747 | set_next_entity(cfs_rq, se); | |
4748 | } | |
4749 | ||
4750 | if (hrtick_enabled(rq)) | |
4751 | hrtick_start_fair(rq, p); | |
4752 | ||
4753 | return p; | |
4754 | simple: | |
4755 | cfs_rq = &rq->cfs; | |
4756 | #endif | |
bf0f6f24 | 4757 | |
36ace27e | 4758 | if (!cfs_rq->nr_running) |
38033c37 | 4759 | goto idle; |
bf0f6f24 | 4760 | |
606dba2e PZ |
4761 | if (prev) |
4762 | prev->sched_class->put_prev_task(rq, prev); | |
4763 | ||
bf0f6f24 | 4764 | do { |
678d5718 | 4765 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 4766 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
4767 | cfs_rq = group_cfs_rq(se); |
4768 | } while (cfs_rq); | |
4769 | ||
8f4d37ec | 4770 | p = task_of(se); |
678d5718 | 4771 | |
b39e66ea MG |
4772 | if (hrtick_enabled(rq)) |
4773 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
4774 | |
4775 | return p; | |
38033c37 PZ |
4776 | |
4777 | idle: | |
4778 | #ifdef CONFIG_SMP | |
4779 | idle_enter_fair(rq); | |
4780 | /* | |
4781 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
4782 | * measure the duration of idle_balance() as idle time. | |
4783 | */ | |
4784 | rq->idle_stamp = rq_clock(rq); | |
4785 | if (idle_balance(rq)) { /* drops rq->lock */ | |
4786 | rq->idle_stamp = 0; | |
4787 | goto again; | |
4788 | } | |
4789 | #endif | |
4790 | ||
4791 | return NULL; | |
bf0f6f24 IM |
4792 | } |
4793 | ||
4794 | /* | |
4795 | * Account for a descheduled task: | |
4796 | */ | |
31ee529c | 4797 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
4798 | { |
4799 | struct sched_entity *se = &prev->se; | |
4800 | struct cfs_rq *cfs_rq; | |
4801 | ||
4802 | for_each_sched_entity(se) { | |
4803 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 4804 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
4805 | } |
4806 | } | |
4807 | ||
ac53db59 RR |
4808 | /* |
4809 | * sched_yield() is very simple | |
4810 | * | |
4811 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
4812 | */ | |
4813 | static void yield_task_fair(struct rq *rq) | |
4814 | { | |
4815 | struct task_struct *curr = rq->curr; | |
4816 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
4817 | struct sched_entity *se = &curr->se; | |
4818 | ||
4819 | /* | |
4820 | * Are we the only task in the tree? | |
4821 | */ | |
4822 | if (unlikely(rq->nr_running == 1)) | |
4823 | return; | |
4824 | ||
4825 | clear_buddies(cfs_rq, se); | |
4826 | ||
4827 | if (curr->policy != SCHED_BATCH) { | |
4828 | update_rq_clock(rq); | |
4829 | /* | |
4830 | * Update run-time statistics of the 'current'. | |
4831 | */ | |
4832 | update_curr(cfs_rq); | |
916671c0 MG |
4833 | /* |
4834 | * Tell update_rq_clock() that we've just updated, | |
4835 | * so we don't do microscopic update in schedule() | |
4836 | * and double the fastpath cost. | |
4837 | */ | |
4838 | rq->skip_clock_update = 1; | |
ac53db59 RR |
4839 | } |
4840 | ||
4841 | set_skip_buddy(se); | |
4842 | } | |
4843 | ||
d95f4122 MG |
4844 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
4845 | { | |
4846 | struct sched_entity *se = &p->se; | |
4847 | ||
5238cdd3 PT |
4848 | /* throttled hierarchies are not runnable */ |
4849 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
4850 | return false; |
4851 | ||
4852 | /* Tell the scheduler that we'd really like pse to run next. */ | |
4853 | set_next_buddy(se); | |
4854 | ||
d95f4122 MG |
4855 | yield_task_fair(rq); |
4856 | ||
4857 | return true; | |
4858 | } | |
4859 | ||
681f3e68 | 4860 | #ifdef CONFIG_SMP |
bf0f6f24 | 4861 | /************************************************** |
e9c84cb8 PZ |
4862 | * Fair scheduling class load-balancing methods. |
4863 | * | |
4864 | * BASICS | |
4865 | * | |
4866 | * The purpose of load-balancing is to achieve the same basic fairness the | |
4867 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
4868 | * time to each task. This is expressed in the following equation: | |
4869 | * | |
4870 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
4871 | * | |
4872 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
4873 | * W_i,0 is defined as: | |
4874 | * | |
4875 | * W_i,0 = \Sum_j w_i,j (2) | |
4876 | * | |
4877 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
4878 | * is derived from the nice value as per prio_to_weight[]. | |
4879 | * | |
4880 | * The weight average is an exponential decay average of the instantaneous | |
4881 | * weight: | |
4882 | * | |
4883 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
4884 | * | |
4885 | * P_i is the cpu power (or compute capacity) of cpu i, typically it is the | |
4886 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | |
4887 | * can also include other factors [XXX]. | |
4888 | * | |
4889 | * To achieve this balance we define a measure of imbalance which follows | |
4890 | * directly from (1): | |
4891 | * | |
4892 | * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4) | |
4893 | * | |
4894 | * We them move tasks around to minimize the imbalance. In the continuous | |
4895 | * function space it is obvious this converges, in the discrete case we get | |
4896 | * a few fun cases generally called infeasible weight scenarios. | |
4897 | * | |
4898 | * [XXX expand on: | |
4899 | * - infeasible weights; | |
4900 | * - local vs global optima in the discrete case. ] | |
4901 | * | |
4902 | * | |
4903 | * SCHED DOMAINS | |
4904 | * | |
4905 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
4906 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
4907 | * topology where each level pairs two lower groups (or better). This results | |
4908 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
4909 | * tree to only the first of the previous level and we decrease the frequency | |
4910 | * of load-balance at each level inv. proportional to the number of cpus in | |
4911 | * the groups. | |
4912 | * | |
4913 | * This yields: | |
4914 | * | |
4915 | * log_2 n 1 n | |
4916 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
4917 | * i = 0 2^i 2^i | |
4918 | * `- size of each group | |
4919 | * | | `- number of cpus doing load-balance | |
4920 | * | `- freq | |
4921 | * `- sum over all levels | |
4922 | * | |
4923 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
4924 | * this makes (5) the runtime complexity of the balancer. | |
4925 | * | |
4926 | * An important property here is that each CPU is still (indirectly) connected | |
4927 | * to every other cpu in at most O(log n) steps: | |
4928 | * | |
4929 | * The adjacency matrix of the resulting graph is given by: | |
4930 | * | |
4931 | * log_2 n | |
4932 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
4933 | * k = 0 | |
4934 | * | |
4935 | * And you'll find that: | |
4936 | * | |
4937 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
4938 | * | |
4939 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
4940 | * The task movement gives a factor of O(m), giving a convergence complexity | |
4941 | * of: | |
4942 | * | |
4943 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
4944 | * | |
4945 | * | |
4946 | * WORK CONSERVING | |
4947 | * | |
4948 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
4949 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
4950 | * tree itself instead of relying on other CPUs to bring it work. | |
4951 | * | |
4952 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
4953 | * time. | |
4954 | * | |
4955 | * [XXX more?] | |
4956 | * | |
4957 | * | |
4958 | * CGROUPS | |
4959 | * | |
4960 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
4961 | * | |
4962 | * s_k,i | |
4963 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
4964 | * S_k | |
4965 | * | |
4966 | * Where | |
4967 | * | |
4968 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
4969 | * | |
4970 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
4971 | * | |
4972 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
4973 | * property. | |
4974 | * | |
4975 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
4976 | * rewrite all of this once again.] | |
4977 | */ | |
bf0f6f24 | 4978 | |
ed387b78 HS |
4979 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
4980 | ||
0ec8aa00 PZ |
4981 | enum fbq_type { regular, remote, all }; |
4982 | ||
ddcdf6e7 | 4983 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 4984 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
4985 | #define LBF_DST_PINNED 0x04 |
4986 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
4987 | |
4988 | struct lb_env { | |
4989 | struct sched_domain *sd; | |
4990 | ||
ddcdf6e7 | 4991 | struct rq *src_rq; |
85c1e7da | 4992 | int src_cpu; |
ddcdf6e7 PZ |
4993 | |
4994 | int dst_cpu; | |
4995 | struct rq *dst_rq; | |
4996 | ||
88b8dac0 SV |
4997 | struct cpumask *dst_grpmask; |
4998 | int new_dst_cpu; | |
ddcdf6e7 | 4999 | enum cpu_idle_type idle; |
bd939f45 | 5000 | long imbalance; |
b9403130 MW |
5001 | /* The set of CPUs under consideration for load-balancing */ |
5002 | struct cpumask *cpus; | |
5003 | ||
ddcdf6e7 | 5004 | unsigned int flags; |
367456c7 PZ |
5005 | |
5006 | unsigned int loop; | |
5007 | unsigned int loop_break; | |
5008 | unsigned int loop_max; | |
0ec8aa00 PZ |
5009 | |
5010 | enum fbq_type fbq_type; | |
ddcdf6e7 PZ |
5011 | }; |
5012 | ||
1e3c88bd | 5013 | /* |
ddcdf6e7 | 5014 | * move_task - move a task from one runqueue to another runqueue. |
1e3c88bd PZ |
5015 | * Both runqueues must be locked. |
5016 | */ | |
ddcdf6e7 | 5017 | static void move_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 5018 | { |
ddcdf6e7 PZ |
5019 | deactivate_task(env->src_rq, p, 0); |
5020 | set_task_cpu(p, env->dst_cpu); | |
5021 | activate_task(env->dst_rq, p, 0); | |
5022 | check_preempt_curr(env->dst_rq, p, 0); | |
1e3c88bd PZ |
5023 | } |
5024 | ||
029632fb PZ |
5025 | /* |
5026 | * Is this task likely cache-hot: | |
5027 | */ | |
5028 | static int | |
5029 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) | |
5030 | { | |
5031 | s64 delta; | |
5032 | ||
5033 | if (p->sched_class != &fair_sched_class) | |
5034 | return 0; | |
5035 | ||
5036 | if (unlikely(p->policy == SCHED_IDLE)) | |
5037 | return 0; | |
5038 | ||
5039 | /* | |
5040 | * Buddy candidates are cache hot: | |
5041 | */ | |
5042 | if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && | |
5043 | (&p->se == cfs_rq_of(&p->se)->next || | |
5044 | &p->se == cfs_rq_of(&p->se)->last)) | |
5045 | return 1; | |
5046 | ||
5047 | if (sysctl_sched_migration_cost == -1) | |
5048 | return 1; | |
5049 | if (sysctl_sched_migration_cost == 0) | |
5050 | return 0; | |
5051 | ||
5052 | delta = now - p->se.exec_start; | |
5053 | ||
5054 | return delta < (s64)sysctl_sched_migration_cost; | |
5055 | } | |
5056 | ||
3a7053b3 MG |
5057 | #ifdef CONFIG_NUMA_BALANCING |
5058 | /* Returns true if the destination node has incurred more faults */ | |
5059 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
5060 | { | |
5061 | int src_nid, dst_nid; | |
5062 | ||
ff1df896 | 5063 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory || |
3a7053b3 MG |
5064 | !(env->sd->flags & SD_NUMA)) { |
5065 | return false; | |
5066 | } | |
5067 | ||
5068 | src_nid = cpu_to_node(env->src_cpu); | |
5069 | dst_nid = cpu_to_node(env->dst_cpu); | |
5070 | ||
83e1d2cd | 5071 | if (src_nid == dst_nid) |
3a7053b3 MG |
5072 | return false; |
5073 | ||
83e1d2cd MG |
5074 | /* Always encourage migration to the preferred node. */ |
5075 | if (dst_nid == p->numa_preferred_nid) | |
5076 | return true; | |
5077 | ||
887c290e RR |
5078 | /* If both task and group weight improve, this move is a winner. */ |
5079 | if (task_weight(p, dst_nid) > task_weight(p, src_nid) && | |
5080 | group_weight(p, dst_nid) > group_weight(p, src_nid)) | |
3a7053b3 MG |
5081 | return true; |
5082 | ||
5083 | return false; | |
5084 | } | |
7a0f3083 MG |
5085 | |
5086 | ||
5087 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
5088 | { | |
5089 | int src_nid, dst_nid; | |
5090 | ||
5091 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
5092 | return false; | |
5093 | ||
ff1df896 | 5094 | if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA)) |
7a0f3083 MG |
5095 | return false; |
5096 | ||
5097 | src_nid = cpu_to_node(env->src_cpu); | |
5098 | dst_nid = cpu_to_node(env->dst_cpu); | |
5099 | ||
83e1d2cd | 5100 | if (src_nid == dst_nid) |
7a0f3083 MG |
5101 | return false; |
5102 | ||
83e1d2cd MG |
5103 | /* Migrating away from the preferred node is always bad. */ |
5104 | if (src_nid == p->numa_preferred_nid) | |
5105 | return true; | |
5106 | ||
887c290e RR |
5107 | /* If either task or group weight get worse, don't do it. */ |
5108 | if (task_weight(p, dst_nid) < task_weight(p, src_nid) || | |
5109 | group_weight(p, dst_nid) < group_weight(p, src_nid)) | |
7a0f3083 MG |
5110 | return true; |
5111 | ||
5112 | return false; | |
5113 | } | |
5114 | ||
3a7053b3 MG |
5115 | #else |
5116 | static inline bool migrate_improves_locality(struct task_struct *p, | |
5117 | struct lb_env *env) | |
5118 | { | |
5119 | return false; | |
5120 | } | |
7a0f3083 MG |
5121 | |
5122 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
5123 | struct lb_env *env) | |
5124 | { | |
5125 | return false; | |
5126 | } | |
3a7053b3 MG |
5127 | #endif |
5128 | ||
1e3c88bd PZ |
5129 | /* |
5130 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
5131 | */ | |
5132 | static | |
8e45cb54 | 5133 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
5134 | { |
5135 | int tsk_cache_hot = 0; | |
5136 | /* | |
5137 | * We do not migrate tasks that are: | |
d3198084 | 5138 | * 1) throttled_lb_pair, or |
1e3c88bd | 5139 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
5140 | * 3) running (obviously), or |
5141 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 5142 | */ |
d3198084 JK |
5143 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
5144 | return 0; | |
5145 | ||
ddcdf6e7 | 5146 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 5147 | int cpu; |
88b8dac0 | 5148 | |
41acab88 | 5149 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 5150 | |
6263322c PZ |
5151 | env->flags |= LBF_SOME_PINNED; |
5152 | ||
88b8dac0 SV |
5153 | /* |
5154 | * Remember if this task can be migrated to any other cpu in | |
5155 | * our sched_group. We may want to revisit it if we couldn't | |
5156 | * meet load balance goals by pulling other tasks on src_cpu. | |
5157 | * | |
5158 | * Also avoid computing new_dst_cpu if we have already computed | |
5159 | * one in current iteration. | |
5160 | */ | |
6263322c | 5161 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
5162 | return 0; |
5163 | ||
e02e60c1 JK |
5164 | /* Prevent to re-select dst_cpu via env's cpus */ |
5165 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
5166 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 5167 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
5168 | env->new_dst_cpu = cpu; |
5169 | break; | |
5170 | } | |
88b8dac0 | 5171 | } |
e02e60c1 | 5172 | |
1e3c88bd PZ |
5173 | return 0; |
5174 | } | |
88b8dac0 SV |
5175 | |
5176 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 5177 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 5178 | |
ddcdf6e7 | 5179 | if (task_running(env->src_rq, p)) { |
41acab88 | 5180 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
5181 | return 0; |
5182 | } | |
5183 | ||
5184 | /* | |
5185 | * Aggressive migration if: | |
3a7053b3 MG |
5186 | * 1) destination numa is preferred |
5187 | * 2) task is cache cold, or | |
5188 | * 3) too many balance attempts have failed. | |
1e3c88bd | 5189 | */ |
78becc27 | 5190 | tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd); |
7a0f3083 MG |
5191 | if (!tsk_cache_hot) |
5192 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 MG |
5193 | |
5194 | if (migrate_improves_locality(p, env)) { | |
5195 | #ifdef CONFIG_SCHEDSTATS | |
5196 | if (tsk_cache_hot) { | |
5197 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
5198 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
5199 | } | |
5200 | #endif | |
5201 | return 1; | |
5202 | } | |
5203 | ||
1e3c88bd | 5204 | if (!tsk_cache_hot || |
8e45cb54 | 5205 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
4e2dcb73 | 5206 | |
1e3c88bd | 5207 | if (tsk_cache_hot) { |
8e45cb54 | 5208 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
41acab88 | 5209 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
1e3c88bd | 5210 | } |
4e2dcb73 | 5211 | |
1e3c88bd PZ |
5212 | return 1; |
5213 | } | |
5214 | ||
4e2dcb73 ZH |
5215 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
5216 | return 0; | |
1e3c88bd PZ |
5217 | } |
5218 | ||
897c395f PZ |
5219 | /* |
5220 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
5221 | * part of active balancing operations within "domain". | |
5222 | * Returns 1 if successful and 0 otherwise. | |
5223 | * | |
5224 | * Called with both runqueues locked. | |
5225 | */ | |
8e45cb54 | 5226 | static int move_one_task(struct lb_env *env) |
897c395f PZ |
5227 | { |
5228 | struct task_struct *p, *n; | |
897c395f | 5229 | |
367456c7 | 5230 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
5231 | if (!can_migrate_task(p, env)) |
5232 | continue; | |
897c395f | 5233 | |
367456c7 PZ |
5234 | move_task(p, env); |
5235 | /* | |
5236 | * Right now, this is only the second place move_task() | |
5237 | * is called, so we can safely collect move_task() | |
5238 | * stats here rather than inside move_task(). | |
5239 | */ | |
5240 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
5241 | return 1; | |
897c395f | 5242 | } |
897c395f PZ |
5243 | return 0; |
5244 | } | |
5245 | ||
eb95308e PZ |
5246 | static const unsigned int sched_nr_migrate_break = 32; |
5247 | ||
5d6523eb | 5248 | /* |
bd939f45 | 5249 | * move_tasks tries to move up to imbalance weighted load from busiest to |
5d6523eb PZ |
5250 | * this_rq, as part of a balancing operation within domain "sd". |
5251 | * Returns 1 if successful and 0 otherwise. | |
5252 | * | |
5253 | * Called with both runqueues locked. | |
5254 | */ | |
5255 | static int move_tasks(struct lb_env *env) | |
1e3c88bd | 5256 | { |
5d6523eb PZ |
5257 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
5258 | struct task_struct *p; | |
367456c7 PZ |
5259 | unsigned long load; |
5260 | int pulled = 0; | |
1e3c88bd | 5261 | |
bd939f45 | 5262 | if (env->imbalance <= 0) |
5d6523eb | 5263 | return 0; |
1e3c88bd | 5264 | |
5d6523eb PZ |
5265 | while (!list_empty(tasks)) { |
5266 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 5267 | |
367456c7 PZ |
5268 | env->loop++; |
5269 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 5270 | if (env->loop > env->loop_max) |
367456c7 | 5271 | break; |
5d6523eb PZ |
5272 | |
5273 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 5274 | if (env->loop > env->loop_break) { |
eb95308e | 5275 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 5276 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 5277 | break; |
a195f004 | 5278 | } |
1e3c88bd | 5279 | |
d3198084 | 5280 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
5281 | goto next; |
5282 | ||
5283 | load = task_h_load(p); | |
5d6523eb | 5284 | |
eb95308e | 5285 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
5286 | goto next; |
5287 | ||
bd939f45 | 5288 | if ((load / 2) > env->imbalance) |
367456c7 | 5289 | goto next; |
1e3c88bd | 5290 | |
ddcdf6e7 | 5291 | move_task(p, env); |
ee00e66f | 5292 | pulled++; |
bd939f45 | 5293 | env->imbalance -= load; |
1e3c88bd PZ |
5294 | |
5295 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
5296 | /* |
5297 | * NEWIDLE balancing is a source of latency, so preemptible | |
5298 | * kernels will stop after the first task is pulled to minimize | |
5299 | * the critical section. | |
5300 | */ | |
5d6523eb | 5301 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 5302 | break; |
1e3c88bd PZ |
5303 | #endif |
5304 | ||
ee00e66f PZ |
5305 | /* |
5306 | * We only want to steal up to the prescribed amount of | |
5307 | * weighted load. | |
5308 | */ | |
bd939f45 | 5309 | if (env->imbalance <= 0) |
ee00e66f | 5310 | break; |
367456c7 PZ |
5311 | |
5312 | continue; | |
5313 | next: | |
5d6523eb | 5314 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 5315 | } |
5d6523eb | 5316 | |
1e3c88bd | 5317 | /* |
ddcdf6e7 PZ |
5318 | * Right now, this is one of only two places move_task() is called, |
5319 | * so we can safely collect move_task() stats here rather than | |
5320 | * inside move_task(). | |
1e3c88bd | 5321 | */ |
8e45cb54 | 5322 | schedstat_add(env->sd, lb_gained[env->idle], pulled); |
1e3c88bd | 5323 | |
5d6523eb | 5324 | return pulled; |
1e3c88bd PZ |
5325 | } |
5326 | ||
230059de | 5327 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
5328 | /* |
5329 | * update tg->load_weight by folding this cpu's load_avg | |
5330 | */ | |
48a16753 | 5331 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 5332 | { |
48a16753 PT |
5333 | struct sched_entity *se = tg->se[cpu]; |
5334 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 5335 | |
48a16753 PT |
5336 | /* throttled entities do not contribute to load */ |
5337 | if (throttled_hierarchy(cfs_rq)) | |
5338 | return; | |
9e3081ca | 5339 | |
aff3e498 | 5340 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 5341 | |
82958366 PT |
5342 | if (se) { |
5343 | update_entity_load_avg(se, 1); | |
5344 | /* | |
5345 | * We pivot on our runnable average having decayed to zero for | |
5346 | * list removal. This generally implies that all our children | |
5347 | * have also been removed (modulo rounding error or bandwidth | |
5348 | * control); however, such cases are rare and we can fix these | |
5349 | * at enqueue. | |
5350 | * | |
5351 | * TODO: fix up out-of-order children on enqueue. | |
5352 | */ | |
5353 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
5354 | list_del_leaf_cfs_rq(cfs_rq); | |
5355 | } else { | |
48a16753 | 5356 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
5357 | update_rq_runnable_avg(rq, rq->nr_running); |
5358 | } | |
9e3081ca PZ |
5359 | } |
5360 | ||
48a16753 | 5361 | static void update_blocked_averages(int cpu) |
9e3081ca | 5362 | { |
9e3081ca | 5363 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
5364 | struct cfs_rq *cfs_rq; |
5365 | unsigned long flags; | |
9e3081ca | 5366 | |
48a16753 PT |
5367 | raw_spin_lock_irqsave(&rq->lock, flags); |
5368 | update_rq_clock(rq); | |
9763b67f PZ |
5369 | /* |
5370 | * Iterates the task_group tree in a bottom up fashion, see | |
5371 | * list_add_leaf_cfs_rq() for details. | |
5372 | */ | |
64660c86 | 5373 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
5374 | /* |
5375 | * Note: We may want to consider periodically releasing | |
5376 | * rq->lock about these updates so that creating many task | |
5377 | * groups does not result in continually extending hold time. | |
5378 | */ | |
5379 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 5380 | } |
48a16753 PT |
5381 | |
5382 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
5383 | } |
5384 | ||
9763b67f | 5385 | /* |
68520796 | 5386 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
5387 | * This needs to be done in a top-down fashion because the load of a child |
5388 | * group is a fraction of its parents load. | |
5389 | */ | |
68520796 | 5390 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 5391 | { |
68520796 VD |
5392 | struct rq *rq = rq_of(cfs_rq); |
5393 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 5394 | unsigned long now = jiffies; |
68520796 | 5395 | unsigned long load; |
a35b6466 | 5396 | |
68520796 | 5397 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
5398 | return; |
5399 | ||
68520796 VD |
5400 | cfs_rq->h_load_next = NULL; |
5401 | for_each_sched_entity(se) { | |
5402 | cfs_rq = cfs_rq_of(se); | |
5403 | cfs_rq->h_load_next = se; | |
5404 | if (cfs_rq->last_h_load_update == now) | |
5405 | break; | |
5406 | } | |
a35b6466 | 5407 | |
68520796 | 5408 | if (!se) { |
7e3115ef | 5409 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
5410 | cfs_rq->last_h_load_update = now; |
5411 | } | |
5412 | ||
5413 | while ((se = cfs_rq->h_load_next) != NULL) { | |
5414 | load = cfs_rq->h_load; | |
5415 | load = div64_ul(load * se->avg.load_avg_contrib, | |
5416 | cfs_rq->runnable_load_avg + 1); | |
5417 | cfs_rq = group_cfs_rq(se); | |
5418 | cfs_rq->h_load = load; | |
5419 | cfs_rq->last_h_load_update = now; | |
5420 | } | |
9763b67f PZ |
5421 | } |
5422 | ||
367456c7 | 5423 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 5424 | { |
367456c7 | 5425 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 5426 | |
68520796 | 5427 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
5428 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
5429 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
5430 | } |
5431 | #else | |
48a16753 | 5432 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
5433 | { |
5434 | } | |
5435 | ||
367456c7 | 5436 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 5437 | { |
a003a25b | 5438 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 5439 | } |
230059de | 5440 | #endif |
1e3c88bd | 5441 | |
1e3c88bd | 5442 | /********** Helpers for find_busiest_group ************************/ |
1e3c88bd PZ |
5443 | /* |
5444 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
5445 | */ | |
5446 | struct sg_lb_stats { | |
5447 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
5448 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 5449 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 5450 | unsigned long load_per_task; |
3ae11c90 | 5451 | unsigned long group_power; |
147c5fc2 PZ |
5452 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
5453 | unsigned int group_capacity; | |
5454 | unsigned int idle_cpus; | |
5455 | unsigned int group_weight; | |
1e3c88bd | 5456 | int group_imb; /* Is there an imbalance in the group ? */ |
fab47622 | 5457 | int group_has_capacity; /* Is there extra capacity in the group? */ |
0ec8aa00 PZ |
5458 | #ifdef CONFIG_NUMA_BALANCING |
5459 | unsigned int nr_numa_running; | |
5460 | unsigned int nr_preferred_running; | |
5461 | #endif | |
1e3c88bd PZ |
5462 | }; |
5463 | ||
56cf515b JK |
5464 | /* |
5465 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
5466 | * during load balancing. | |
5467 | */ | |
5468 | struct sd_lb_stats { | |
5469 | struct sched_group *busiest; /* Busiest group in this sd */ | |
5470 | struct sched_group *local; /* Local group in this sd */ | |
5471 | unsigned long total_load; /* Total load of all groups in sd */ | |
5472 | unsigned long total_pwr; /* Total power of all groups in sd */ | |
5473 | unsigned long avg_load; /* Average load across all groups in sd */ | |
5474 | ||
56cf515b | 5475 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 5476 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
5477 | }; |
5478 | ||
147c5fc2 PZ |
5479 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
5480 | { | |
5481 | /* | |
5482 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
5483 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
5484 | * We must however clear busiest_stat::avg_load because | |
5485 | * update_sd_pick_busiest() reads this before assignment. | |
5486 | */ | |
5487 | *sds = (struct sd_lb_stats){ | |
5488 | .busiest = NULL, | |
5489 | .local = NULL, | |
5490 | .total_load = 0UL, | |
5491 | .total_pwr = 0UL, | |
5492 | .busiest_stat = { | |
5493 | .avg_load = 0UL, | |
5494 | }, | |
5495 | }; | |
5496 | } | |
5497 | ||
1e3c88bd PZ |
5498 | /** |
5499 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
5500 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 5501 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
5502 | * |
5503 | * Return: The load index. | |
1e3c88bd PZ |
5504 | */ |
5505 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
5506 | enum cpu_idle_type idle) | |
5507 | { | |
5508 | int load_idx; | |
5509 | ||
5510 | switch (idle) { | |
5511 | case CPU_NOT_IDLE: | |
5512 | load_idx = sd->busy_idx; | |
5513 | break; | |
5514 | ||
5515 | case CPU_NEWLY_IDLE: | |
5516 | load_idx = sd->newidle_idx; | |
5517 | break; | |
5518 | default: | |
5519 | load_idx = sd->idle_idx; | |
5520 | break; | |
5521 | } | |
5522 | ||
5523 | return load_idx; | |
5524 | } | |
5525 | ||
15f803c9 | 5526 | static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5527 | { |
1399fa78 | 5528 | return SCHED_POWER_SCALE; |
1e3c88bd PZ |
5529 | } |
5530 | ||
5531 | unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
5532 | { | |
5533 | return default_scale_freq_power(sd, cpu); | |
5534 | } | |
5535 | ||
15f803c9 | 5536 | static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) |
1e3c88bd | 5537 | { |
669c55e9 | 5538 | unsigned long weight = sd->span_weight; |
1e3c88bd PZ |
5539 | unsigned long smt_gain = sd->smt_gain; |
5540 | ||
5541 | smt_gain /= weight; | |
5542 | ||
5543 | return smt_gain; | |
5544 | } | |
5545 | ||
5546 | unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) | |
5547 | { | |
5548 | return default_scale_smt_power(sd, cpu); | |
5549 | } | |
5550 | ||
15f803c9 | 5551 | static unsigned long scale_rt_power(int cpu) |
1e3c88bd PZ |
5552 | { |
5553 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5554 | u64 total, available, age_stamp, avg; |
1e3c88bd | 5555 | |
b654f7de PZ |
5556 | /* |
5557 | * Since we're reading these variables without serialization make sure | |
5558 | * we read them once before doing sanity checks on them. | |
5559 | */ | |
5560 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5561 | avg = ACCESS_ONCE(rq->rt_avg); | |
5562 | ||
78becc27 | 5563 | total = sched_avg_period() + (rq_clock(rq) - age_stamp); |
aa483808 | 5564 | |
b654f7de | 5565 | if (unlikely(total < avg)) { |
aa483808 VP |
5566 | /* Ensures that power won't end up being negative */ |
5567 | available = 0; | |
5568 | } else { | |
b654f7de | 5569 | available = total - avg; |
aa483808 | 5570 | } |
1e3c88bd | 5571 | |
1399fa78 NR |
5572 | if (unlikely((s64)total < SCHED_POWER_SCALE)) |
5573 | total = SCHED_POWER_SCALE; | |
1e3c88bd | 5574 | |
1399fa78 | 5575 | total >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5576 | |
5577 | return div_u64(available, total); | |
5578 | } | |
5579 | ||
5580 | static void update_cpu_power(struct sched_domain *sd, int cpu) | |
5581 | { | |
669c55e9 | 5582 | unsigned long weight = sd->span_weight; |
1399fa78 | 5583 | unsigned long power = SCHED_POWER_SCALE; |
1e3c88bd PZ |
5584 | struct sched_group *sdg = sd->groups; |
5585 | ||
1e3c88bd PZ |
5586 | if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { |
5587 | if (sched_feat(ARCH_POWER)) | |
5588 | power *= arch_scale_smt_power(sd, cpu); | |
5589 | else | |
5590 | power *= default_scale_smt_power(sd, cpu); | |
5591 | ||
1399fa78 | 5592 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5593 | } |
5594 | ||
9c3f75cb | 5595 | sdg->sgp->power_orig = power; |
9d5efe05 SV |
5596 | |
5597 | if (sched_feat(ARCH_POWER)) | |
5598 | power *= arch_scale_freq_power(sd, cpu); | |
5599 | else | |
5600 | power *= default_scale_freq_power(sd, cpu); | |
5601 | ||
1399fa78 | 5602 | power >>= SCHED_POWER_SHIFT; |
9d5efe05 | 5603 | |
1e3c88bd | 5604 | power *= scale_rt_power(cpu); |
1399fa78 | 5605 | power >>= SCHED_POWER_SHIFT; |
1e3c88bd PZ |
5606 | |
5607 | if (!power) | |
5608 | power = 1; | |
5609 | ||
e51fd5e2 | 5610 | cpu_rq(cpu)->cpu_power = power; |
9c3f75cb | 5611 | sdg->sgp->power = power; |
1e3c88bd PZ |
5612 | } |
5613 | ||
029632fb | 5614 | void update_group_power(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
5615 | { |
5616 | struct sched_domain *child = sd->child; | |
5617 | struct sched_group *group, *sdg = sd->groups; | |
863bffc8 | 5618 | unsigned long power, power_orig; |
4ec4412e VG |
5619 | unsigned long interval; |
5620 | ||
5621 | interval = msecs_to_jiffies(sd->balance_interval); | |
5622 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
5623 | sdg->sgp->next_update = jiffies + interval; | |
1e3c88bd PZ |
5624 | |
5625 | if (!child) { | |
5626 | update_cpu_power(sd, cpu); | |
5627 | return; | |
5628 | } | |
5629 | ||
863bffc8 | 5630 | power_orig = power = 0; |
1e3c88bd | 5631 | |
74a5ce20 PZ |
5632 | if (child->flags & SD_OVERLAP) { |
5633 | /* | |
5634 | * SD_OVERLAP domains cannot assume that child groups | |
5635 | * span the current group. | |
5636 | */ | |
5637 | ||
863bffc8 | 5638 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
9abf24d4 SD |
5639 | struct sched_group_power *sgp; |
5640 | struct rq *rq = cpu_rq(cpu); | |
863bffc8 | 5641 | |
9abf24d4 SD |
5642 | /* |
5643 | * build_sched_domains() -> init_sched_groups_power() | |
5644 | * gets here before we've attached the domains to the | |
5645 | * runqueues. | |
5646 | * | |
5647 | * Use power_of(), which is set irrespective of domains | |
5648 | * in update_cpu_power(). | |
5649 | * | |
5650 | * This avoids power/power_orig from being 0 and | |
5651 | * causing divide-by-zero issues on boot. | |
5652 | * | |
5653 | * Runtime updates will correct power_orig. | |
5654 | */ | |
5655 | if (unlikely(!rq->sd)) { | |
5656 | power_orig += power_of(cpu); | |
5657 | power += power_of(cpu); | |
5658 | continue; | |
5659 | } | |
863bffc8 | 5660 | |
9abf24d4 SD |
5661 | sgp = rq->sd->groups->sgp; |
5662 | power_orig += sgp->power_orig; | |
5663 | power += sgp->power; | |
863bffc8 | 5664 | } |
74a5ce20 PZ |
5665 | } else { |
5666 | /* | |
5667 | * !SD_OVERLAP domains can assume that child groups | |
5668 | * span the current group. | |
5669 | */ | |
5670 | ||
5671 | group = child->groups; | |
5672 | do { | |
863bffc8 | 5673 | power_orig += group->sgp->power_orig; |
74a5ce20 PZ |
5674 | power += group->sgp->power; |
5675 | group = group->next; | |
5676 | } while (group != child->groups); | |
5677 | } | |
1e3c88bd | 5678 | |
863bffc8 PZ |
5679 | sdg->sgp->power_orig = power_orig; |
5680 | sdg->sgp->power = power; | |
1e3c88bd PZ |
5681 | } |
5682 | ||
9d5efe05 SV |
5683 | /* |
5684 | * Try and fix up capacity for tiny siblings, this is needed when | |
5685 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
5686 | * which on its own isn't powerful enough. | |
5687 | * | |
5688 | * See update_sd_pick_busiest() and check_asym_packing(). | |
5689 | */ | |
5690 | static inline int | |
5691 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
5692 | { | |
5693 | /* | |
1399fa78 | 5694 | * Only siblings can have significantly less than SCHED_POWER_SCALE |
9d5efe05 | 5695 | */ |
a6c75f2f | 5696 | if (!(sd->flags & SD_SHARE_CPUPOWER)) |
9d5efe05 SV |
5697 | return 0; |
5698 | ||
5699 | /* | |
5700 | * If ~90% of the cpu_power is still there, we're good. | |
5701 | */ | |
9c3f75cb | 5702 | if (group->sgp->power * 32 > group->sgp->power_orig * 29) |
9d5efe05 SV |
5703 | return 1; |
5704 | ||
5705 | return 0; | |
5706 | } | |
5707 | ||
30ce5dab PZ |
5708 | /* |
5709 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
5710 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
5711 | * | |
5712 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
5713 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
5714 | * Something like: | |
5715 | * | |
5716 | * { 0 1 2 3 } { 4 5 6 7 } | |
5717 | * * * * * | |
5718 | * | |
5719 | * If we were to balance group-wise we'd place two tasks in the first group and | |
5720 | * two tasks in the second group. Clearly this is undesired as it will overload | |
5721 | * cpu 3 and leave one of the cpus in the second group unused. | |
5722 | * | |
5723 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
5724 | * by noticing the lower domain failed to reach balance and had difficulty |
5725 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
5726 | * |
5727 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 5728 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 5729 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
5730 | * to create an effective group imbalance. |
5731 | * | |
5732 | * This is a somewhat tricky proposition since the next run might not find the | |
5733 | * group imbalance and decide the groups need to be balanced again. A most | |
5734 | * subtle and fragile situation. | |
5735 | */ | |
5736 | ||
6263322c | 5737 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 5738 | { |
6263322c | 5739 | return group->sgp->imbalance; |
30ce5dab PZ |
5740 | } |
5741 | ||
b37d9316 PZ |
5742 | /* |
5743 | * Compute the group capacity. | |
5744 | * | |
c61037e9 PZ |
5745 | * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by |
5746 | * first dividing out the smt factor and computing the actual number of cores | |
5747 | * and limit power unit capacity with that. | |
b37d9316 PZ |
5748 | */ |
5749 | static inline int sg_capacity(struct lb_env *env, struct sched_group *group) | |
5750 | { | |
c61037e9 PZ |
5751 | unsigned int capacity, smt, cpus; |
5752 | unsigned int power, power_orig; | |
5753 | ||
5754 | power = group->sgp->power; | |
5755 | power_orig = group->sgp->power_orig; | |
5756 | cpus = group->group_weight; | |
b37d9316 | 5757 | |
c61037e9 PZ |
5758 | /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */ |
5759 | smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig); | |
5760 | capacity = cpus / smt; /* cores */ | |
b37d9316 | 5761 | |
c61037e9 | 5762 | capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE)); |
b37d9316 PZ |
5763 | if (!capacity) |
5764 | capacity = fix_small_capacity(env->sd, group); | |
5765 | ||
5766 | return capacity; | |
5767 | } | |
5768 | ||
1e3c88bd PZ |
5769 | /** |
5770 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 5771 | * @env: The load balancing environment. |
1e3c88bd | 5772 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 5773 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 5774 | * @local_group: Does group contain this_cpu. |
1e3c88bd PZ |
5775 | * @sgs: variable to hold the statistics for this group. |
5776 | */ | |
bd939f45 PZ |
5777 | static inline void update_sg_lb_stats(struct lb_env *env, |
5778 | struct sched_group *group, int load_idx, | |
23f0d209 | 5779 | int local_group, struct sg_lb_stats *sgs) |
1e3c88bd | 5780 | { |
30ce5dab | 5781 | unsigned long load; |
bd939f45 | 5782 | int i; |
1e3c88bd | 5783 | |
b72ff13c PZ |
5784 | memset(sgs, 0, sizeof(*sgs)); |
5785 | ||
b9403130 | 5786 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
5787 | struct rq *rq = cpu_rq(i); |
5788 | ||
1e3c88bd | 5789 | /* Bias balancing toward cpus of our domain */ |
6263322c | 5790 | if (local_group) |
04f733b4 | 5791 | load = target_load(i, load_idx); |
6263322c | 5792 | else |
1e3c88bd | 5793 | load = source_load(i, load_idx); |
1e3c88bd PZ |
5794 | |
5795 | sgs->group_load += load; | |
380c9077 | 5796 | sgs->sum_nr_running += rq->nr_running; |
0ec8aa00 PZ |
5797 | #ifdef CONFIG_NUMA_BALANCING |
5798 | sgs->nr_numa_running += rq->nr_numa_running; | |
5799 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
5800 | #endif | |
1e3c88bd | 5801 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
5802 | if (idle_cpu(i)) |
5803 | sgs->idle_cpus++; | |
1e3c88bd PZ |
5804 | } |
5805 | ||
1e3c88bd | 5806 | /* Adjust by relative CPU power of the group */ |
3ae11c90 PZ |
5807 | sgs->group_power = group->sgp->power; |
5808 | sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power; | |
1e3c88bd | 5809 | |
dd5feea1 | 5810 | if (sgs->sum_nr_running) |
38d0f770 | 5811 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 5812 | |
aae6d3dd | 5813 | sgs->group_weight = group->group_weight; |
fab47622 | 5814 | |
b37d9316 PZ |
5815 | sgs->group_imb = sg_imbalanced(group); |
5816 | sgs->group_capacity = sg_capacity(env, group); | |
5817 | ||
fab47622 NR |
5818 | if (sgs->group_capacity > sgs->sum_nr_running) |
5819 | sgs->group_has_capacity = 1; | |
1e3c88bd PZ |
5820 | } |
5821 | ||
532cb4c4 MN |
5822 | /** |
5823 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 5824 | * @env: The load balancing environment. |
532cb4c4 MN |
5825 | * @sds: sched_domain statistics |
5826 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 5827 | * @sgs: sched_group statistics |
532cb4c4 MN |
5828 | * |
5829 | * Determine if @sg is a busier group than the previously selected | |
5830 | * busiest group. | |
e69f6186 YB |
5831 | * |
5832 | * Return: %true if @sg is a busier group than the previously selected | |
5833 | * busiest group. %false otherwise. | |
532cb4c4 | 5834 | */ |
bd939f45 | 5835 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
5836 | struct sd_lb_stats *sds, |
5837 | struct sched_group *sg, | |
bd939f45 | 5838 | struct sg_lb_stats *sgs) |
532cb4c4 | 5839 | { |
56cf515b | 5840 | if (sgs->avg_load <= sds->busiest_stat.avg_load) |
532cb4c4 MN |
5841 | return false; |
5842 | ||
5843 | if (sgs->sum_nr_running > sgs->group_capacity) | |
5844 | return true; | |
5845 | ||
5846 | if (sgs->group_imb) | |
5847 | return true; | |
5848 | ||
5849 | /* | |
5850 | * ASYM_PACKING needs to move all the work to the lowest | |
5851 | * numbered CPUs in the group, therefore mark all groups | |
5852 | * higher than ourself as busy. | |
5853 | */ | |
bd939f45 PZ |
5854 | if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && |
5855 | env->dst_cpu < group_first_cpu(sg)) { | |
532cb4c4 MN |
5856 | if (!sds->busiest) |
5857 | return true; | |
5858 | ||
5859 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
5860 | return true; | |
5861 | } | |
5862 | ||
5863 | return false; | |
5864 | } | |
5865 | ||
0ec8aa00 PZ |
5866 | #ifdef CONFIG_NUMA_BALANCING |
5867 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5868 | { | |
5869 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
5870 | return regular; | |
5871 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
5872 | return remote; | |
5873 | return all; | |
5874 | } | |
5875 | ||
5876 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5877 | { | |
5878 | if (rq->nr_running > rq->nr_numa_running) | |
5879 | return regular; | |
5880 | if (rq->nr_running > rq->nr_preferred_running) | |
5881 | return remote; | |
5882 | return all; | |
5883 | } | |
5884 | #else | |
5885 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
5886 | { | |
5887 | return all; | |
5888 | } | |
5889 | ||
5890 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
5891 | { | |
5892 | return regular; | |
5893 | } | |
5894 | #endif /* CONFIG_NUMA_BALANCING */ | |
5895 | ||
1e3c88bd | 5896 | /** |
461819ac | 5897 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 5898 | * @env: The load balancing environment. |
1e3c88bd PZ |
5899 | * @sds: variable to hold the statistics for this sched_domain. |
5900 | */ | |
0ec8aa00 | 5901 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 5902 | { |
bd939f45 PZ |
5903 | struct sched_domain *child = env->sd->child; |
5904 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 5905 | struct sg_lb_stats tmp_sgs; |
1e3c88bd PZ |
5906 | int load_idx, prefer_sibling = 0; |
5907 | ||
5908 | if (child && child->flags & SD_PREFER_SIBLING) | |
5909 | prefer_sibling = 1; | |
5910 | ||
bd939f45 | 5911 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
5912 | |
5913 | do { | |
56cf515b | 5914 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
5915 | int local_group; |
5916 | ||
bd939f45 | 5917 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
5918 | if (local_group) { |
5919 | sds->local = sg; | |
5920 | sgs = &sds->local_stat; | |
b72ff13c PZ |
5921 | |
5922 | if (env->idle != CPU_NEWLY_IDLE || | |
5923 | time_after_eq(jiffies, sg->sgp->next_update)) | |
5924 | update_group_power(env->sd, env->dst_cpu); | |
56cf515b | 5925 | } |
1e3c88bd | 5926 | |
56cf515b | 5927 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs); |
1e3c88bd | 5928 | |
b72ff13c PZ |
5929 | if (local_group) |
5930 | goto next_group; | |
5931 | ||
1e3c88bd PZ |
5932 | /* |
5933 | * In case the child domain prefers tasks go to siblings | |
532cb4c4 | 5934 | * first, lower the sg capacity to one so that we'll try |
75dd321d NR |
5935 | * and move all the excess tasks away. We lower the capacity |
5936 | * of a group only if the local group has the capacity to fit | |
5937 | * these excess tasks, i.e. nr_running < group_capacity. The | |
5938 | * extra check prevents the case where you always pull from the | |
5939 | * heaviest group when it is already under-utilized (possible | |
5940 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 5941 | */ |
b72ff13c PZ |
5942 | if (prefer_sibling && sds->local && |
5943 | sds->local_stat.group_has_capacity) | |
147c5fc2 | 5944 | sgs->group_capacity = min(sgs->group_capacity, 1U); |
1e3c88bd | 5945 | |
b72ff13c | 5946 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 5947 | sds->busiest = sg; |
56cf515b | 5948 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
5949 | } |
5950 | ||
b72ff13c PZ |
5951 | next_group: |
5952 | /* Now, start updating sd_lb_stats */ | |
5953 | sds->total_load += sgs->group_load; | |
5954 | sds->total_pwr += sgs->group_power; | |
5955 | ||
532cb4c4 | 5956 | sg = sg->next; |
bd939f45 | 5957 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
5958 | |
5959 | if (env->sd->flags & SD_NUMA) | |
5960 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
532cb4c4 MN |
5961 | } |
5962 | ||
532cb4c4 MN |
5963 | /** |
5964 | * check_asym_packing - Check to see if the group is packed into the | |
5965 | * sched doman. | |
5966 | * | |
5967 | * This is primarily intended to used at the sibling level. Some | |
5968 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
5969 | * case of POWER7, it can move to lower SMT modes only when higher | |
5970 | * threads are idle. When in lower SMT modes, the threads will | |
5971 | * perform better since they share less core resources. Hence when we | |
5972 | * have idle threads, we want them to be the higher ones. | |
5973 | * | |
5974 | * This packing function is run on idle threads. It checks to see if | |
5975 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
5976 | * CPU number than the packing function is being run on. Here we are | |
5977 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
5978 | * number. | |
5979 | * | |
e69f6186 | 5980 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
5981 | * this CPU. The amount of the imbalance is returned in *imbalance. |
5982 | * | |
cd96891d | 5983 | * @env: The load balancing environment. |
532cb4c4 | 5984 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 5985 | */ |
bd939f45 | 5986 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
5987 | { |
5988 | int busiest_cpu; | |
5989 | ||
bd939f45 | 5990 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
5991 | return 0; |
5992 | ||
5993 | if (!sds->busiest) | |
5994 | return 0; | |
5995 | ||
5996 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 5997 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
5998 | return 0; |
5999 | ||
bd939f45 | 6000 | env->imbalance = DIV_ROUND_CLOSEST( |
3ae11c90 PZ |
6001 | sds->busiest_stat.avg_load * sds->busiest_stat.group_power, |
6002 | SCHED_POWER_SCALE); | |
bd939f45 | 6003 | |
532cb4c4 | 6004 | return 1; |
1e3c88bd PZ |
6005 | } |
6006 | ||
6007 | /** | |
6008 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
6009 | * amongst the groups of a sched_domain, during | |
6010 | * load balancing. | |
cd96891d | 6011 | * @env: The load balancing environment. |
1e3c88bd | 6012 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6013 | */ |
bd939f45 PZ |
6014 | static inline |
6015 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd PZ |
6016 | { |
6017 | unsigned long tmp, pwr_now = 0, pwr_move = 0; | |
6018 | unsigned int imbn = 2; | |
dd5feea1 | 6019 | unsigned long scaled_busy_load_per_task; |
56cf515b | 6020 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 6021 | |
56cf515b JK |
6022 | local = &sds->local_stat; |
6023 | busiest = &sds->busiest_stat; | |
1e3c88bd | 6024 | |
56cf515b JK |
6025 | if (!local->sum_nr_running) |
6026 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
6027 | else if (busiest->load_per_task > local->load_per_task) | |
6028 | imbn = 1; | |
dd5feea1 | 6029 | |
56cf515b JK |
6030 | scaled_busy_load_per_task = |
6031 | (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 6032 | busiest->group_power; |
56cf515b | 6033 | |
3029ede3 VD |
6034 | if (busiest->avg_load + scaled_busy_load_per_task >= |
6035 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 6036 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6037 | return; |
6038 | } | |
6039 | ||
6040 | /* | |
6041 | * OK, we don't have enough imbalance to justify moving tasks, | |
6042 | * however we may be able to increase total CPU power used by | |
6043 | * moving them. | |
6044 | */ | |
6045 | ||
3ae11c90 | 6046 | pwr_now += busiest->group_power * |
56cf515b | 6047 | min(busiest->load_per_task, busiest->avg_load); |
3ae11c90 | 6048 | pwr_now += local->group_power * |
56cf515b | 6049 | min(local->load_per_task, local->avg_load); |
1399fa78 | 6050 | pwr_now /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
6051 | |
6052 | /* Amount of load we'd subtract */ | |
56cf515b | 6053 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / |
3ae11c90 | 6054 | busiest->group_power; |
56cf515b | 6055 | if (busiest->avg_load > tmp) { |
3ae11c90 | 6056 | pwr_move += busiest->group_power * |
56cf515b JK |
6057 | min(busiest->load_per_task, |
6058 | busiest->avg_load - tmp); | |
6059 | } | |
1e3c88bd PZ |
6060 | |
6061 | /* Amount of load we'd add */ | |
3ae11c90 | 6062 | if (busiest->avg_load * busiest->group_power < |
56cf515b | 6063 | busiest->load_per_task * SCHED_POWER_SCALE) { |
3ae11c90 PZ |
6064 | tmp = (busiest->avg_load * busiest->group_power) / |
6065 | local->group_power; | |
56cf515b JK |
6066 | } else { |
6067 | tmp = (busiest->load_per_task * SCHED_POWER_SCALE) / | |
3ae11c90 | 6068 | local->group_power; |
56cf515b | 6069 | } |
3ae11c90 PZ |
6070 | pwr_move += local->group_power * |
6071 | min(local->load_per_task, local->avg_load + tmp); | |
1399fa78 | 6072 | pwr_move /= SCHED_POWER_SCALE; |
1e3c88bd PZ |
6073 | |
6074 | /* Move if we gain throughput */ | |
6075 | if (pwr_move > pwr_now) | |
56cf515b | 6076 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6077 | } |
6078 | ||
6079 | /** | |
6080 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
6081 | * groups of a given sched_domain during load balance. | |
bd939f45 | 6082 | * @env: load balance environment |
1e3c88bd | 6083 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6084 | */ |
bd939f45 | 6085 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6086 | { |
dd5feea1 | 6087 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
6088 | struct sg_lb_stats *local, *busiest; |
6089 | ||
6090 | local = &sds->local_stat; | |
56cf515b | 6091 | busiest = &sds->busiest_stat; |
dd5feea1 | 6092 | |
56cf515b | 6093 | if (busiest->group_imb) { |
30ce5dab PZ |
6094 | /* |
6095 | * In the group_imb case we cannot rely on group-wide averages | |
6096 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
6097 | */ | |
56cf515b JK |
6098 | busiest->load_per_task = |
6099 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
6100 | } |
6101 | ||
1e3c88bd PZ |
6102 | /* |
6103 | * In the presence of smp nice balancing, certain scenarios can have | |
6104 | * max load less than avg load(as we skip the groups at or below | |
6105 | * its cpu_power, while calculating max_load..) | |
6106 | */ | |
b1885550 VD |
6107 | if (busiest->avg_load <= sds->avg_load || |
6108 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
6109 | env->imbalance = 0; |
6110 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
6111 | } |
6112 | ||
56cf515b | 6113 | if (!busiest->group_imb) { |
dd5feea1 SS |
6114 | /* |
6115 | * Don't want to pull so many tasks that a group would go idle. | |
30ce5dab PZ |
6116 | * Except of course for the group_imb case, since then we might |
6117 | * have to drop below capacity to reach cpu-load equilibrium. | |
dd5feea1 | 6118 | */ |
56cf515b JK |
6119 | load_above_capacity = |
6120 | (busiest->sum_nr_running - busiest->group_capacity); | |
dd5feea1 | 6121 | |
1399fa78 | 6122 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); |
3ae11c90 | 6123 | load_above_capacity /= busiest->group_power; |
dd5feea1 SS |
6124 | } |
6125 | ||
6126 | /* | |
6127 | * We're trying to get all the cpus to the average_load, so we don't | |
6128 | * want to push ourselves above the average load, nor do we wish to | |
6129 | * reduce the max loaded cpu below the average load. At the same time, | |
6130 | * we also don't want to reduce the group load below the group capacity | |
6131 | * (so that we can implement power-savings policies etc). Thus we look | |
6132 | * for the minimum possible imbalance. | |
dd5feea1 | 6133 | */ |
30ce5dab | 6134 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
6135 | |
6136 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 6137 | env->imbalance = min( |
3ae11c90 PZ |
6138 | max_pull * busiest->group_power, |
6139 | (sds->avg_load - local->avg_load) * local->group_power | |
56cf515b | 6140 | ) / SCHED_POWER_SCALE; |
1e3c88bd PZ |
6141 | |
6142 | /* | |
6143 | * if *imbalance is less than the average load per runnable task | |
25985edc | 6144 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
6145 | * a think about bumping its value to force at least one task to be |
6146 | * moved | |
6147 | */ | |
56cf515b | 6148 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 6149 | return fix_small_imbalance(env, sds); |
1e3c88bd | 6150 | } |
fab47622 | 6151 | |
1e3c88bd PZ |
6152 | /******* find_busiest_group() helpers end here *********************/ |
6153 | ||
6154 | /** | |
6155 | * find_busiest_group - Returns the busiest group within the sched_domain | |
6156 | * if there is an imbalance. If there isn't an imbalance, and | |
6157 | * the user has opted for power-savings, it returns a group whose | |
6158 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
6159 | * such a group exists. | |
6160 | * | |
6161 | * Also calculates the amount of weighted load which should be moved | |
6162 | * to restore balance. | |
6163 | * | |
cd96891d | 6164 | * @env: The load balancing environment. |
1e3c88bd | 6165 | * |
e69f6186 | 6166 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
6167 | * - If no imbalance and user has opted for power-savings balance, |
6168 | * return the least loaded group whose CPUs can be | |
6169 | * put to idle by rebalancing its tasks onto our group. | |
6170 | */ | |
56cf515b | 6171 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 6172 | { |
56cf515b | 6173 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
6174 | struct sd_lb_stats sds; |
6175 | ||
147c5fc2 | 6176 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
6177 | |
6178 | /* | |
6179 | * Compute the various statistics relavent for load balancing at | |
6180 | * this level. | |
6181 | */ | |
23f0d209 | 6182 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
6183 | local = &sds.local_stat; |
6184 | busiest = &sds.busiest_stat; | |
1e3c88bd | 6185 | |
bd939f45 PZ |
6186 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
6187 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
6188 | return sds.busiest; |
6189 | ||
cc57aa8f | 6190 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 6191 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
6192 | goto out_balanced; |
6193 | ||
1399fa78 | 6194 | sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; |
b0432d8f | 6195 | |
866ab43e PZ |
6196 | /* |
6197 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 6198 | * work because they assume all things are equal, which typically |
866ab43e PZ |
6199 | * isn't true due to cpus_allowed constraints and the like. |
6200 | */ | |
56cf515b | 6201 | if (busiest->group_imb) |
866ab43e PZ |
6202 | goto force_balance; |
6203 | ||
cc57aa8f | 6204 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
56cf515b JK |
6205 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity && |
6206 | !busiest->group_has_capacity) | |
fab47622 NR |
6207 | goto force_balance; |
6208 | ||
cc57aa8f PZ |
6209 | /* |
6210 | * If the local group is more busy than the selected busiest group | |
6211 | * don't try and pull any tasks. | |
6212 | */ | |
56cf515b | 6213 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
6214 | goto out_balanced; |
6215 | ||
cc57aa8f PZ |
6216 | /* |
6217 | * Don't pull any tasks if this group is already above the domain | |
6218 | * average load. | |
6219 | */ | |
56cf515b | 6220 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
6221 | goto out_balanced; |
6222 | ||
bd939f45 | 6223 | if (env->idle == CPU_IDLE) { |
aae6d3dd SS |
6224 | /* |
6225 | * This cpu is idle. If the busiest group load doesn't | |
6226 | * have more tasks than the number of available cpu's and | |
6227 | * there is no imbalance between this and busiest group | |
6228 | * wrt to idle cpu's, it is balanced. | |
6229 | */ | |
56cf515b JK |
6230 | if ((local->idle_cpus < busiest->idle_cpus) && |
6231 | busiest->sum_nr_running <= busiest->group_weight) | |
aae6d3dd | 6232 | goto out_balanced; |
c186fafe PZ |
6233 | } else { |
6234 | /* | |
6235 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
6236 | * imbalance_pct to be conservative. | |
6237 | */ | |
56cf515b JK |
6238 | if (100 * busiest->avg_load <= |
6239 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 6240 | goto out_balanced; |
aae6d3dd | 6241 | } |
1e3c88bd | 6242 | |
fab47622 | 6243 | force_balance: |
1e3c88bd | 6244 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 6245 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
6246 | return sds.busiest; |
6247 | ||
6248 | out_balanced: | |
bd939f45 | 6249 | env->imbalance = 0; |
1e3c88bd PZ |
6250 | return NULL; |
6251 | } | |
6252 | ||
6253 | /* | |
6254 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
6255 | */ | |
bd939f45 | 6256 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 6257 | struct sched_group *group) |
1e3c88bd PZ |
6258 | { |
6259 | struct rq *busiest = NULL, *rq; | |
95a79b80 | 6260 | unsigned long busiest_load = 0, busiest_power = 1; |
1e3c88bd PZ |
6261 | int i; |
6262 | ||
6906a408 | 6263 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
0ec8aa00 PZ |
6264 | unsigned long power, capacity, wl; |
6265 | enum fbq_type rt; | |
6266 | ||
6267 | rq = cpu_rq(i); | |
6268 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 6269 | |
0ec8aa00 PZ |
6270 | /* |
6271 | * We classify groups/runqueues into three groups: | |
6272 | * - regular: there are !numa tasks | |
6273 | * - remote: there are numa tasks that run on the 'wrong' node | |
6274 | * - all: there is no distinction | |
6275 | * | |
6276 | * In order to avoid migrating ideally placed numa tasks, | |
6277 | * ignore those when there's better options. | |
6278 | * | |
6279 | * If we ignore the actual busiest queue to migrate another | |
6280 | * task, the next balance pass can still reduce the busiest | |
6281 | * queue by moving tasks around inside the node. | |
6282 | * | |
6283 | * If we cannot move enough load due to this classification | |
6284 | * the next pass will adjust the group classification and | |
6285 | * allow migration of more tasks. | |
6286 | * | |
6287 | * Both cases only affect the total convergence complexity. | |
6288 | */ | |
6289 | if (rt > env->fbq_type) | |
6290 | continue; | |
6291 | ||
6292 | power = power_of(i); | |
6293 | capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); | |
9d5efe05 | 6294 | if (!capacity) |
bd939f45 | 6295 | capacity = fix_small_capacity(env->sd, group); |
9d5efe05 | 6296 | |
6e40f5bb | 6297 | wl = weighted_cpuload(i); |
1e3c88bd | 6298 | |
6e40f5bb TG |
6299 | /* |
6300 | * When comparing with imbalance, use weighted_cpuload() | |
6301 | * which is not scaled with the cpu power. | |
6302 | */ | |
bd939f45 | 6303 | if (capacity && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
6304 | continue; |
6305 | ||
6e40f5bb TG |
6306 | /* |
6307 | * For the load comparisons with the other cpu's, consider | |
6308 | * the weighted_cpuload() scaled with the cpu power, so that | |
6309 | * the load can be moved away from the cpu that is potentially | |
6310 | * running at a lower capacity. | |
95a79b80 JK |
6311 | * |
6312 | * Thus we're looking for max(wl_i / power_i), crosswise | |
6313 | * multiplication to rid ourselves of the division works out | |
6314 | * to: wl_i * power_j > wl_j * power_i; where j is our | |
6315 | * previous maximum. | |
6e40f5bb | 6316 | */ |
95a79b80 JK |
6317 | if (wl * busiest_power > busiest_load * power) { |
6318 | busiest_load = wl; | |
6319 | busiest_power = power; | |
1e3c88bd PZ |
6320 | busiest = rq; |
6321 | } | |
6322 | } | |
6323 | ||
6324 | return busiest; | |
6325 | } | |
6326 | ||
6327 | /* | |
6328 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
6329 | * so long as it is large enough. | |
6330 | */ | |
6331 | #define MAX_PINNED_INTERVAL 512 | |
6332 | ||
6333 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 6334 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 6335 | |
bd939f45 | 6336 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 6337 | { |
bd939f45 PZ |
6338 | struct sched_domain *sd = env->sd; |
6339 | ||
6340 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
6341 | |
6342 | /* | |
6343 | * ASYM_PACKING needs to force migrate tasks from busy but | |
6344 | * higher numbered CPUs in order to pack all tasks in the | |
6345 | * lowest numbered CPUs. | |
6346 | */ | |
bd939f45 | 6347 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 6348 | return 1; |
1af3ed3d PZ |
6349 | } |
6350 | ||
6351 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
6352 | } | |
6353 | ||
969c7921 TH |
6354 | static int active_load_balance_cpu_stop(void *data); |
6355 | ||
23f0d209 JK |
6356 | static int should_we_balance(struct lb_env *env) |
6357 | { | |
6358 | struct sched_group *sg = env->sd->groups; | |
6359 | struct cpumask *sg_cpus, *sg_mask; | |
6360 | int cpu, balance_cpu = -1; | |
6361 | ||
6362 | /* | |
6363 | * In the newly idle case, we will allow all the cpu's | |
6364 | * to do the newly idle load balance. | |
6365 | */ | |
6366 | if (env->idle == CPU_NEWLY_IDLE) | |
6367 | return 1; | |
6368 | ||
6369 | sg_cpus = sched_group_cpus(sg); | |
6370 | sg_mask = sched_group_mask(sg); | |
6371 | /* Try to find first idle cpu */ | |
6372 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
6373 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
6374 | continue; | |
6375 | ||
6376 | balance_cpu = cpu; | |
6377 | break; | |
6378 | } | |
6379 | ||
6380 | if (balance_cpu == -1) | |
6381 | balance_cpu = group_balance_cpu(sg); | |
6382 | ||
6383 | /* | |
6384 | * First idle cpu or the first cpu(busiest) in this sched group | |
6385 | * is eligible for doing load balancing at this and above domains. | |
6386 | */ | |
b0cff9d8 | 6387 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
6388 | } |
6389 | ||
1e3c88bd PZ |
6390 | /* |
6391 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
6392 | * tasks if there is an imbalance. | |
6393 | */ | |
6394 | static int load_balance(int this_cpu, struct rq *this_rq, | |
6395 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 6396 | int *continue_balancing) |
1e3c88bd | 6397 | { |
88b8dac0 | 6398 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 6399 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 6400 | struct sched_group *group; |
1e3c88bd PZ |
6401 | struct rq *busiest; |
6402 | unsigned long flags; | |
e6252c3e | 6403 | struct cpumask *cpus = __get_cpu_var(load_balance_mask); |
1e3c88bd | 6404 | |
8e45cb54 PZ |
6405 | struct lb_env env = { |
6406 | .sd = sd, | |
ddcdf6e7 PZ |
6407 | .dst_cpu = this_cpu, |
6408 | .dst_rq = this_rq, | |
88b8dac0 | 6409 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 6410 | .idle = idle, |
eb95308e | 6411 | .loop_break = sched_nr_migrate_break, |
b9403130 | 6412 | .cpus = cpus, |
0ec8aa00 | 6413 | .fbq_type = all, |
8e45cb54 PZ |
6414 | }; |
6415 | ||
cfc03118 JK |
6416 | /* |
6417 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
6418 | * other cpus in our group | |
6419 | */ | |
e02e60c1 | 6420 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 6421 | env.dst_grpmask = NULL; |
cfc03118 | 6422 | |
1e3c88bd PZ |
6423 | cpumask_copy(cpus, cpu_active_mask); |
6424 | ||
1e3c88bd PZ |
6425 | schedstat_inc(sd, lb_count[idle]); |
6426 | ||
6427 | redo: | |
23f0d209 JK |
6428 | if (!should_we_balance(&env)) { |
6429 | *continue_balancing = 0; | |
1e3c88bd | 6430 | goto out_balanced; |
23f0d209 | 6431 | } |
1e3c88bd | 6432 | |
23f0d209 | 6433 | group = find_busiest_group(&env); |
1e3c88bd PZ |
6434 | if (!group) { |
6435 | schedstat_inc(sd, lb_nobusyg[idle]); | |
6436 | goto out_balanced; | |
6437 | } | |
6438 | ||
b9403130 | 6439 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
6440 | if (!busiest) { |
6441 | schedstat_inc(sd, lb_nobusyq[idle]); | |
6442 | goto out_balanced; | |
6443 | } | |
6444 | ||
78feefc5 | 6445 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 6446 | |
bd939f45 | 6447 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
6448 | |
6449 | ld_moved = 0; | |
6450 | if (busiest->nr_running > 1) { | |
6451 | /* | |
6452 | * Attempt to move tasks. If find_busiest_group has found | |
6453 | * an imbalance but busiest->nr_running <= 1, the group is | |
6454 | * still unbalanced. ld_moved simply stays zero, so it is | |
6455 | * correctly treated as an imbalance. | |
6456 | */ | |
8e45cb54 | 6457 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
6458 | env.src_cpu = busiest->cpu; |
6459 | env.src_rq = busiest; | |
6460 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 6461 | |
5d6523eb | 6462 | more_balance: |
1e3c88bd | 6463 | local_irq_save(flags); |
78feefc5 | 6464 | double_rq_lock(env.dst_rq, busiest); |
88b8dac0 SV |
6465 | |
6466 | /* | |
6467 | * cur_ld_moved - load moved in current iteration | |
6468 | * ld_moved - cumulative load moved across iterations | |
6469 | */ | |
6470 | cur_ld_moved = move_tasks(&env); | |
6471 | ld_moved += cur_ld_moved; | |
78feefc5 | 6472 | double_rq_unlock(env.dst_rq, busiest); |
1e3c88bd PZ |
6473 | local_irq_restore(flags); |
6474 | ||
6475 | /* | |
6476 | * some other cpu did the load balance for us. | |
6477 | */ | |
88b8dac0 SV |
6478 | if (cur_ld_moved && env.dst_cpu != smp_processor_id()) |
6479 | resched_cpu(env.dst_cpu); | |
6480 | ||
f1cd0858 JK |
6481 | if (env.flags & LBF_NEED_BREAK) { |
6482 | env.flags &= ~LBF_NEED_BREAK; | |
6483 | goto more_balance; | |
6484 | } | |
6485 | ||
88b8dac0 SV |
6486 | /* |
6487 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
6488 | * us and move them to an alternate dst_cpu in our sched_group | |
6489 | * where they can run. The upper limit on how many times we | |
6490 | * iterate on same src_cpu is dependent on number of cpus in our | |
6491 | * sched_group. | |
6492 | * | |
6493 | * This changes load balance semantics a bit on who can move | |
6494 | * load to a given_cpu. In addition to the given_cpu itself | |
6495 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
6496 | * nohz-idle), we now have balance_cpu in a position to move | |
6497 | * load to given_cpu. In rare situations, this may cause | |
6498 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
6499 | * _independently_ and at _same_ time to move some load to | |
6500 | * given_cpu) causing exceess load to be moved to given_cpu. | |
6501 | * This however should not happen so much in practice and | |
6502 | * moreover subsequent load balance cycles should correct the | |
6503 | * excess load moved. | |
6504 | */ | |
6263322c | 6505 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 6506 | |
7aff2e3a VD |
6507 | /* Prevent to re-select dst_cpu via env's cpus */ |
6508 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
6509 | ||
78feefc5 | 6510 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 6511 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 6512 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
6513 | env.loop = 0; |
6514 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 6515 | |
88b8dac0 SV |
6516 | /* |
6517 | * Go back to "more_balance" rather than "redo" since we | |
6518 | * need to continue with same src_cpu. | |
6519 | */ | |
6520 | goto more_balance; | |
6521 | } | |
1e3c88bd | 6522 | |
6263322c PZ |
6523 | /* |
6524 | * We failed to reach balance because of affinity. | |
6525 | */ | |
6526 | if (sd_parent) { | |
6527 | int *group_imbalance = &sd_parent->groups->sgp->imbalance; | |
6528 | ||
6529 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) { | |
6530 | *group_imbalance = 1; | |
6531 | } else if (*group_imbalance) | |
6532 | *group_imbalance = 0; | |
6533 | } | |
6534 | ||
1e3c88bd | 6535 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 6536 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 6537 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
6538 | if (!cpumask_empty(cpus)) { |
6539 | env.loop = 0; | |
6540 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 6541 | goto redo; |
bbf18b19 | 6542 | } |
1e3c88bd PZ |
6543 | goto out_balanced; |
6544 | } | |
6545 | } | |
6546 | ||
6547 | if (!ld_moved) { | |
6548 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
6549 | /* |
6550 | * Increment the failure counter only on periodic balance. | |
6551 | * We do not want newidle balance, which can be very | |
6552 | * frequent, pollute the failure counter causing | |
6553 | * excessive cache_hot migrations and active balances. | |
6554 | */ | |
6555 | if (idle != CPU_NEWLY_IDLE) | |
6556 | sd->nr_balance_failed++; | |
1e3c88bd | 6557 | |
bd939f45 | 6558 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
6559 | raw_spin_lock_irqsave(&busiest->lock, flags); |
6560 | ||
969c7921 TH |
6561 | /* don't kick the active_load_balance_cpu_stop, |
6562 | * if the curr task on busiest cpu can't be | |
6563 | * moved to this_cpu | |
1e3c88bd PZ |
6564 | */ |
6565 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 6566 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
6567 | raw_spin_unlock_irqrestore(&busiest->lock, |
6568 | flags); | |
8e45cb54 | 6569 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
6570 | goto out_one_pinned; |
6571 | } | |
6572 | ||
969c7921 TH |
6573 | /* |
6574 | * ->active_balance synchronizes accesses to | |
6575 | * ->active_balance_work. Once set, it's cleared | |
6576 | * only after active load balance is finished. | |
6577 | */ | |
1e3c88bd PZ |
6578 | if (!busiest->active_balance) { |
6579 | busiest->active_balance = 1; | |
6580 | busiest->push_cpu = this_cpu; | |
6581 | active_balance = 1; | |
6582 | } | |
6583 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 6584 | |
bd939f45 | 6585 | if (active_balance) { |
969c7921 TH |
6586 | stop_one_cpu_nowait(cpu_of(busiest), |
6587 | active_load_balance_cpu_stop, busiest, | |
6588 | &busiest->active_balance_work); | |
bd939f45 | 6589 | } |
1e3c88bd PZ |
6590 | |
6591 | /* | |
6592 | * We've kicked active balancing, reset the failure | |
6593 | * counter. | |
6594 | */ | |
6595 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
6596 | } | |
6597 | } else | |
6598 | sd->nr_balance_failed = 0; | |
6599 | ||
6600 | if (likely(!active_balance)) { | |
6601 | /* We were unbalanced, so reset the balancing interval */ | |
6602 | sd->balance_interval = sd->min_interval; | |
6603 | } else { | |
6604 | /* | |
6605 | * If we've begun active balancing, start to back off. This | |
6606 | * case may not be covered by the all_pinned logic if there | |
6607 | * is only 1 task on the busy runqueue (because we don't call | |
6608 | * move_tasks). | |
6609 | */ | |
6610 | if (sd->balance_interval < sd->max_interval) | |
6611 | sd->balance_interval *= 2; | |
6612 | } | |
6613 | ||
1e3c88bd PZ |
6614 | goto out; |
6615 | ||
6616 | out_balanced: | |
6617 | schedstat_inc(sd, lb_balanced[idle]); | |
6618 | ||
6619 | sd->nr_balance_failed = 0; | |
6620 | ||
6621 | out_one_pinned: | |
6622 | /* tune up the balancing interval */ | |
8e45cb54 | 6623 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 6624 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
6625 | (sd->balance_interval < sd->max_interval)) |
6626 | sd->balance_interval *= 2; | |
6627 | ||
46e49b38 | 6628 | ld_moved = 0; |
1e3c88bd | 6629 | out: |
1e3c88bd PZ |
6630 | return ld_moved; |
6631 | } | |
6632 | ||
1e3c88bd PZ |
6633 | /* |
6634 | * idle_balance is called by schedule() if this_cpu is about to become | |
6635 | * idle. Attempts to pull tasks from other CPUs. | |
6636 | */ | |
3c4017c1 | 6637 | int idle_balance(struct rq *this_rq) |
1e3c88bd PZ |
6638 | { |
6639 | struct sched_domain *sd; | |
6640 | int pulled_task = 0; | |
6641 | unsigned long next_balance = jiffies + HZ; | |
9bd721c5 | 6642 | u64 curr_cost = 0; |
b4f2ab43 | 6643 | int this_cpu = this_rq->cpu; |
1e3c88bd | 6644 | |
1e3c88bd | 6645 | if (this_rq->avg_idle < sysctl_sched_migration_cost) |
3c4017c1 | 6646 | return 0; |
1e3c88bd | 6647 | |
f492e12e PZ |
6648 | /* |
6649 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
6650 | */ | |
6651 | raw_spin_unlock(&this_rq->lock); | |
6652 | ||
48a16753 | 6653 | update_blocked_averages(this_cpu); |
dce840a0 | 6654 | rcu_read_lock(); |
1e3c88bd PZ |
6655 | for_each_domain(this_cpu, sd) { |
6656 | unsigned long interval; | |
23f0d209 | 6657 | int continue_balancing = 1; |
9bd721c5 | 6658 | u64 t0, domain_cost; |
1e3c88bd PZ |
6659 | |
6660 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
6661 | continue; | |
6662 | ||
9bd721c5 JL |
6663 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) |
6664 | break; | |
6665 | ||
f492e12e | 6666 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
6667 | t0 = sched_clock_cpu(this_cpu); |
6668 | ||
1e3c88bd | 6669 | /* If we've pulled tasks over stop searching: */ |
f492e12e | 6670 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
6671 | sd, CPU_NEWLY_IDLE, |
6672 | &continue_balancing); | |
9bd721c5 JL |
6673 | |
6674 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
6675 | if (domain_cost > sd->max_newidle_lb_cost) | |
6676 | sd->max_newidle_lb_cost = domain_cost; | |
6677 | ||
6678 | curr_cost += domain_cost; | |
f492e12e | 6679 | } |
1e3c88bd PZ |
6680 | |
6681 | interval = msecs_to_jiffies(sd->balance_interval); | |
6682 | if (time_after(next_balance, sd->last_balance + interval)) | |
6683 | next_balance = sd->last_balance + interval; | |
3c4017c1 | 6684 | if (pulled_task) |
1e3c88bd | 6685 | break; |
1e3c88bd | 6686 | } |
dce840a0 | 6687 | rcu_read_unlock(); |
f492e12e PZ |
6688 | |
6689 | raw_spin_lock(&this_rq->lock); | |
6690 | ||
e5fc6611 DL |
6691 | /* |
6692 | * While browsing the domains, we released the rq lock. | |
6693 | * A task could have be enqueued in the meantime | |
6694 | */ | |
6695 | if (this_rq->nr_running && !pulled_task) | |
3c4017c1 | 6696 | return 1; |
e5fc6611 | 6697 | |
1e3c88bd PZ |
6698 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
6699 | /* | |
6700 | * We are going idle. next_balance may be set based on | |
6701 | * a busy processor. So reset next_balance. | |
6702 | */ | |
6703 | this_rq->next_balance = next_balance; | |
6704 | } | |
9bd721c5 JL |
6705 | |
6706 | if (curr_cost > this_rq->max_idle_balance_cost) | |
6707 | this_rq->max_idle_balance_cost = curr_cost; | |
3c4017c1 DL |
6708 | |
6709 | return pulled_task; | |
1e3c88bd PZ |
6710 | } |
6711 | ||
6712 | /* | |
969c7921 TH |
6713 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
6714 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
6715 | * least 1 task to be running on each physical CPU where possible, and | |
6716 | * avoids physical / logical imbalances. | |
1e3c88bd | 6717 | */ |
969c7921 | 6718 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 6719 | { |
969c7921 TH |
6720 | struct rq *busiest_rq = data; |
6721 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 6722 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 6723 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 6724 | struct sched_domain *sd; |
969c7921 TH |
6725 | |
6726 | raw_spin_lock_irq(&busiest_rq->lock); | |
6727 | ||
6728 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
6729 | if (unlikely(busiest_cpu != smp_processor_id() || | |
6730 | !busiest_rq->active_balance)) | |
6731 | goto out_unlock; | |
1e3c88bd PZ |
6732 | |
6733 | /* Is there any task to move? */ | |
6734 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 6735 | goto out_unlock; |
1e3c88bd PZ |
6736 | |
6737 | /* | |
6738 | * This condition is "impossible", if it occurs | |
6739 | * we need to fix it. Originally reported by | |
6740 | * Bjorn Helgaas on a 128-cpu setup. | |
6741 | */ | |
6742 | BUG_ON(busiest_rq == target_rq); | |
6743 | ||
6744 | /* move a task from busiest_rq to target_rq */ | |
6745 | double_lock_balance(busiest_rq, target_rq); | |
1e3c88bd PZ |
6746 | |
6747 | /* Search for an sd spanning us and the target CPU. */ | |
dce840a0 | 6748 | rcu_read_lock(); |
1e3c88bd PZ |
6749 | for_each_domain(target_cpu, sd) { |
6750 | if ((sd->flags & SD_LOAD_BALANCE) && | |
6751 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
6752 | break; | |
6753 | } | |
6754 | ||
6755 | if (likely(sd)) { | |
8e45cb54 PZ |
6756 | struct lb_env env = { |
6757 | .sd = sd, | |
ddcdf6e7 PZ |
6758 | .dst_cpu = target_cpu, |
6759 | .dst_rq = target_rq, | |
6760 | .src_cpu = busiest_rq->cpu, | |
6761 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
6762 | .idle = CPU_IDLE, |
6763 | }; | |
6764 | ||
1e3c88bd PZ |
6765 | schedstat_inc(sd, alb_count); |
6766 | ||
8e45cb54 | 6767 | if (move_one_task(&env)) |
1e3c88bd PZ |
6768 | schedstat_inc(sd, alb_pushed); |
6769 | else | |
6770 | schedstat_inc(sd, alb_failed); | |
6771 | } | |
dce840a0 | 6772 | rcu_read_unlock(); |
1e3c88bd | 6773 | double_unlock_balance(busiest_rq, target_rq); |
969c7921 TH |
6774 | out_unlock: |
6775 | busiest_rq->active_balance = 0; | |
6776 | raw_spin_unlock_irq(&busiest_rq->lock); | |
6777 | return 0; | |
1e3c88bd PZ |
6778 | } |
6779 | ||
3451d024 | 6780 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
6781 | /* |
6782 | * idle load balancing details | |
83cd4fe2 VP |
6783 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
6784 | * needed, they will kick the idle load balancer, which then does idle | |
6785 | * load balancing for all the idle CPUs. | |
6786 | */ | |
1e3c88bd | 6787 | static struct { |
83cd4fe2 | 6788 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 6789 | atomic_t nr_cpus; |
83cd4fe2 VP |
6790 | unsigned long next_balance; /* in jiffy units */ |
6791 | } nohz ____cacheline_aligned; | |
1e3c88bd | 6792 | |
3dd0337d | 6793 | static inline int find_new_ilb(void) |
1e3c88bd | 6794 | { |
0b005cf5 | 6795 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 6796 | |
786d6dc7 SS |
6797 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
6798 | return ilb; | |
6799 | ||
6800 | return nr_cpu_ids; | |
1e3c88bd | 6801 | } |
1e3c88bd | 6802 | |
83cd4fe2 VP |
6803 | /* |
6804 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
6805 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
6806 | * CPU (if there is one). | |
6807 | */ | |
0aeeeeba | 6808 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
6809 | { |
6810 | int ilb_cpu; | |
6811 | ||
6812 | nohz.next_balance++; | |
6813 | ||
3dd0337d | 6814 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 6815 | |
0b005cf5 SS |
6816 | if (ilb_cpu >= nr_cpu_ids) |
6817 | return; | |
83cd4fe2 | 6818 | |
cd490c5b | 6819 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
6820 | return; |
6821 | /* | |
6822 | * Use smp_send_reschedule() instead of resched_cpu(). | |
6823 | * This way we generate a sched IPI on the target cpu which | |
6824 | * is idle. And the softirq performing nohz idle load balance | |
6825 | * will be run before returning from the IPI. | |
6826 | */ | |
6827 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
6828 | return; |
6829 | } | |
6830 | ||
c1cc017c | 6831 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
6832 | { |
6833 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
6834 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
6835 | atomic_dec(&nohz.nr_cpus); | |
6836 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
6837 | } | |
6838 | } | |
6839 | ||
69e1e811 SS |
6840 | static inline void set_cpu_sd_state_busy(void) |
6841 | { | |
6842 | struct sched_domain *sd; | |
37dc6b50 | 6843 | int cpu = smp_processor_id(); |
69e1e811 | 6844 | |
69e1e811 | 6845 | rcu_read_lock(); |
37dc6b50 | 6846 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
6847 | |
6848 | if (!sd || !sd->nohz_idle) | |
6849 | goto unlock; | |
6850 | sd->nohz_idle = 0; | |
6851 | ||
37dc6b50 | 6852 | atomic_inc(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6853 | unlock: |
69e1e811 SS |
6854 | rcu_read_unlock(); |
6855 | } | |
6856 | ||
6857 | void set_cpu_sd_state_idle(void) | |
6858 | { | |
6859 | struct sched_domain *sd; | |
37dc6b50 | 6860 | int cpu = smp_processor_id(); |
69e1e811 | 6861 | |
69e1e811 | 6862 | rcu_read_lock(); |
37dc6b50 | 6863 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
6864 | |
6865 | if (!sd || sd->nohz_idle) | |
6866 | goto unlock; | |
6867 | sd->nohz_idle = 1; | |
6868 | ||
37dc6b50 | 6869 | atomic_dec(&sd->groups->sgp->nr_busy_cpus); |
25f55d9d | 6870 | unlock: |
69e1e811 SS |
6871 | rcu_read_unlock(); |
6872 | } | |
6873 | ||
1e3c88bd | 6874 | /* |
c1cc017c | 6875 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 6876 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 6877 | */ |
c1cc017c | 6878 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 6879 | { |
71325960 SS |
6880 | /* |
6881 | * If this cpu is going down, then nothing needs to be done. | |
6882 | */ | |
6883 | if (!cpu_active(cpu)) | |
6884 | return; | |
6885 | ||
c1cc017c AS |
6886 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
6887 | return; | |
1e3c88bd | 6888 | |
c1cc017c AS |
6889 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
6890 | atomic_inc(&nohz.nr_cpus); | |
6891 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 6892 | } |
71325960 | 6893 | |
0db0628d | 6894 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
6895 | unsigned long action, void *hcpu) |
6896 | { | |
6897 | switch (action & ~CPU_TASKS_FROZEN) { | |
6898 | case CPU_DYING: | |
c1cc017c | 6899 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
6900 | return NOTIFY_OK; |
6901 | default: | |
6902 | return NOTIFY_DONE; | |
6903 | } | |
6904 | } | |
1e3c88bd PZ |
6905 | #endif |
6906 | ||
6907 | static DEFINE_SPINLOCK(balancing); | |
6908 | ||
49c022e6 PZ |
6909 | /* |
6910 | * Scale the max load_balance interval with the number of CPUs in the system. | |
6911 | * This trades load-balance latency on larger machines for less cross talk. | |
6912 | */ | |
029632fb | 6913 | void update_max_interval(void) |
49c022e6 PZ |
6914 | { |
6915 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
6916 | } | |
6917 | ||
1e3c88bd PZ |
6918 | /* |
6919 | * It checks each scheduling domain to see if it is due to be balanced, | |
6920 | * and initiates a balancing operation if so. | |
6921 | * | |
b9b0853a | 6922 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 6923 | */ |
f7ed0a89 | 6924 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 6925 | { |
23f0d209 | 6926 | int continue_balancing = 1; |
f7ed0a89 | 6927 | int cpu = rq->cpu; |
1e3c88bd | 6928 | unsigned long interval; |
04f733b4 | 6929 | struct sched_domain *sd; |
1e3c88bd PZ |
6930 | /* Earliest time when we have to do rebalance again */ |
6931 | unsigned long next_balance = jiffies + 60*HZ; | |
6932 | int update_next_balance = 0; | |
f48627e6 JL |
6933 | int need_serialize, need_decay = 0; |
6934 | u64 max_cost = 0; | |
1e3c88bd | 6935 | |
48a16753 | 6936 | update_blocked_averages(cpu); |
2069dd75 | 6937 | |
dce840a0 | 6938 | rcu_read_lock(); |
1e3c88bd | 6939 | for_each_domain(cpu, sd) { |
f48627e6 JL |
6940 | /* |
6941 | * Decay the newidle max times here because this is a regular | |
6942 | * visit to all the domains. Decay ~1% per second. | |
6943 | */ | |
6944 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
6945 | sd->max_newidle_lb_cost = | |
6946 | (sd->max_newidle_lb_cost * 253) / 256; | |
6947 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
6948 | need_decay = 1; | |
6949 | } | |
6950 | max_cost += sd->max_newidle_lb_cost; | |
6951 | ||
1e3c88bd PZ |
6952 | if (!(sd->flags & SD_LOAD_BALANCE)) |
6953 | continue; | |
6954 | ||
f48627e6 JL |
6955 | /* |
6956 | * Stop the load balance at this level. There is another | |
6957 | * CPU in our sched group which is doing load balancing more | |
6958 | * actively. | |
6959 | */ | |
6960 | if (!continue_balancing) { | |
6961 | if (need_decay) | |
6962 | continue; | |
6963 | break; | |
6964 | } | |
6965 | ||
1e3c88bd PZ |
6966 | interval = sd->balance_interval; |
6967 | if (idle != CPU_IDLE) | |
6968 | interval *= sd->busy_factor; | |
6969 | ||
6970 | /* scale ms to jiffies */ | |
6971 | interval = msecs_to_jiffies(interval); | |
49c022e6 | 6972 | interval = clamp(interval, 1UL, max_load_balance_interval); |
1e3c88bd PZ |
6973 | |
6974 | need_serialize = sd->flags & SD_SERIALIZE; | |
6975 | ||
6976 | if (need_serialize) { | |
6977 | if (!spin_trylock(&balancing)) | |
6978 | goto out; | |
6979 | } | |
6980 | ||
6981 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 6982 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 6983 | /* |
6263322c | 6984 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
6985 | * env->dst_cpu, so we can't know our idle |
6986 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 6987 | */ |
de5eb2dd | 6988 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
6989 | } |
6990 | sd->last_balance = jiffies; | |
6991 | } | |
6992 | if (need_serialize) | |
6993 | spin_unlock(&balancing); | |
6994 | out: | |
6995 | if (time_after(next_balance, sd->last_balance + interval)) { | |
6996 | next_balance = sd->last_balance + interval; | |
6997 | update_next_balance = 1; | |
6998 | } | |
f48627e6 JL |
6999 | } |
7000 | if (need_decay) { | |
1e3c88bd | 7001 | /* |
f48627e6 JL |
7002 | * Ensure the rq-wide value also decays but keep it at a |
7003 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 7004 | */ |
f48627e6 JL |
7005 | rq->max_idle_balance_cost = |
7006 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 7007 | } |
dce840a0 | 7008 | rcu_read_unlock(); |
1e3c88bd PZ |
7009 | |
7010 | /* | |
7011 | * next_balance will be updated only when there is a need. | |
7012 | * When the cpu is attached to null domain for ex, it will not be | |
7013 | * updated. | |
7014 | */ | |
7015 | if (likely(update_next_balance)) | |
7016 | rq->next_balance = next_balance; | |
7017 | } | |
7018 | ||
3451d024 | 7019 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 7020 | /* |
3451d024 | 7021 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
7022 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
7023 | */ | |
208cb16b | 7024 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 7025 | { |
208cb16b | 7026 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
7027 | struct rq *rq; |
7028 | int balance_cpu; | |
7029 | ||
1c792db7 SS |
7030 | if (idle != CPU_IDLE || |
7031 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
7032 | goto end; | |
83cd4fe2 VP |
7033 | |
7034 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 7035 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
7036 | continue; |
7037 | ||
7038 | /* | |
7039 | * If this cpu gets work to do, stop the load balancing | |
7040 | * work being done for other cpus. Next load | |
7041 | * balancing owner will pick it up. | |
7042 | */ | |
1c792db7 | 7043 | if (need_resched()) |
83cd4fe2 | 7044 | break; |
83cd4fe2 | 7045 | |
5ed4f1d9 VG |
7046 | rq = cpu_rq(balance_cpu); |
7047 | ||
7048 | raw_spin_lock_irq(&rq->lock); | |
7049 | update_rq_clock(rq); | |
7050 | update_idle_cpu_load(rq); | |
7051 | raw_spin_unlock_irq(&rq->lock); | |
83cd4fe2 | 7052 | |
f7ed0a89 | 7053 | rebalance_domains(rq, CPU_IDLE); |
83cd4fe2 | 7054 | |
83cd4fe2 VP |
7055 | if (time_after(this_rq->next_balance, rq->next_balance)) |
7056 | this_rq->next_balance = rq->next_balance; | |
7057 | } | |
7058 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
7059 | end: |
7060 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
7061 | } |
7062 | ||
7063 | /* | |
0b005cf5 SS |
7064 | * Current heuristic for kicking the idle load balancer in the presence |
7065 | * of an idle cpu is the system. | |
7066 | * - This rq has more than one task. | |
7067 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
7068 | * busy cpu's exceeding the group's power. | |
7069 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
7070 | * domain span are idle. | |
83cd4fe2 | 7071 | */ |
4a725627 | 7072 | static inline int nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
7073 | { |
7074 | unsigned long now = jiffies; | |
0b005cf5 | 7075 | struct sched_domain *sd; |
37dc6b50 | 7076 | struct sched_group_power *sgp; |
4a725627 | 7077 | int nr_busy, cpu = rq->cpu; |
83cd4fe2 | 7078 | |
4a725627 | 7079 | if (unlikely(rq->idle_balance)) |
83cd4fe2 VP |
7080 | return 0; |
7081 | ||
1c792db7 SS |
7082 | /* |
7083 | * We may be recently in ticked or tickless idle mode. At the first | |
7084 | * busy tick after returning from idle, we will update the busy stats. | |
7085 | */ | |
69e1e811 | 7086 | set_cpu_sd_state_busy(); |
c1cc017c | 7087 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
7088 | |
7089 | /* | |
7090 | * None are in tickless mode and hence no need for NOHZ idle load | |
7091 | * balancing. | |
7092 | */ | |
7093 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
7094 | return 0; | |
1c792db7 SS |
7095 | |
7096 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
7097 | return 0; |
7098 | ||
0b005cf5 SS |
7099 | if (rq->nr_running >= 2) |
7100 | goto need_kick; | |
83cd4fe2 | 7101 | |
067491b7 | 7102 | rcu_read_lock(); |
37dc6b50 | 7103 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
83cd4fe2 | 7104 | |
37dc6b50 PM |
7105 | if (sd) { |
7106 | sgp = sd->groups->sgp; | |
7107 | nr_busy = atomic_read(&sgp->nr_busy_cpus); | |
0b005cf5 | 7108 | |
37dc6b50 | 7109 | if (nr_busy > 1) |
067491b7 | 7110 | goto need_kick_unlock; |
83cd4fe2 | 7111 | } |
37dc6b50 PM |
7112 | |
7113 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); | |
7114 | ||
7115 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, | |
7116 | sched_domain_span(sd)) < cpu)) | |
7117 | goto need_kick_unlock; | |
7118 | ||
067491b7 | 7119 | rcu_read_unlock(); |
83cd4fe2 | 7120 | return 0; |
067491b7 PZ |
7121 | |
7122 | need_kick_unlock: | |
7123 | rcu_read_unlock(); | |
0b005cf5 SS |
7124 | need_kick: |
7125 | return 1; | |
83cd4fe2 VP |
7126 | } |
7127 | #else | |
208cb16b | 7128 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
7129 | #endif |
7130 | ||
7131 | /* | |
7132 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
7133 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
7134 | */ | |
1e3c88bd PZ |
7135 | static void run_rebalance_domains(struct softirq_action *h) |
7136 | { | |
208cb16b | 7137 | struct rq *this_rq = this_rq(); |
6eb57e0d | 7138 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
7139 | CPU_IDLE : CPU_NOT_IDLE; |
7140 | ||
f7ed0a89 | 7141 | rebalance_domains(this_rq, idle); |
1e3c88bd | 7142 | |
1e3c88bd | 7143 | /* |
83cd4fe2 | 7144 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
7145 | * balancing on behalf of the other idle cpus whose ticks are |
7146 | * stopped. | |
7147 | */ | |
208cb16b | 7148 | nohz_idle_balance(this_rq, idle); |
1e3c88bd PZ |
7149 | } |
7150 | ||
63f609b1 | 7151 | static inline int on_null_domain(struct rq *rq) |
1e3c88bd | 7152 | { |
63f609b1 | 7153 | return !rcu_dereference_sched(rq->sd); |
1e3c88bd PZ |
7154 | } |
7155 | ||
7156 | /* | |
7157 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 7158 | */ |
7caff66f | 7159 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 7160 | { |
1e3c88bd | 7161 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
7162 | if (unlikely(on_null_domain(rq))) |
7163 | return; | |
7164 | ||
7165 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 7166 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 7167 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 7168 | if (nohz_kick_needed(rq)) |
0aeeeeba | 7169 | nohz_balancer_kick(); |
83cd4fe2 | 7170 | #endif |
1e3c88bd PZ |
7171 | } |
7172 | ||
0bcdcf28 CE |
7173 | static void rq_online_fair(struct rq *rq) |
7174 | { | |
7175 | update_sysctl(); | |
7176 | } | |
7177 | ||
7178 | static void rq_offline_fair(struct rq *rq) | |
7179 | { | |
7180 | update_sysctl(); | |
a4c96ae3 PB |
7181 | |
7182 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
7183 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
7184 | } |
7185 | ||
55e12e5e | 7186 | #endif /* CONFIG_SMP */ |
e1d1484f | 7187 | |
bf0f6f24 IM |
7188 | /* |
7189 | * scheduler tick hitting a task of our scheduling class: | |
7190 | */ | |
8f4d37ec | 7191 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
7192 | { |
7193 | struct cfs_rq *cfs_rq; | |
7194 | struct sched_entity *se = &curr->se; | |
7195 | ||
7196 | for_each_sched_entity(se) { | |
7197 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 7198 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 7199 | } |
18bf2805 | 7200 | |
10e84b97 | 7201 | if (numabalancing_enabled) |
cbee9f88 | 7202 | task_tick_numa(rq, curr); |
3d59eebc | 7203 | |
18bf2805 | 7204 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
7205 | } |
7206 | ||
7207 | /* | |
cd29fe6f PZ |
7208 | * called on fork with the child task as argument from the parent's context |
7209 | * - child not yet on the tasklist | |
7210 | * - preemption disabled | |
bf0f6f24 | 7211 | */ |
cd29fe6f | 7212 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 7213 | { |
4fc420c9 DN |
7214 | struct cfs_rq *cfs_rq; |
7215 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 7216 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
7217 | struct rq *rq = this_rq(); |
7218 | unsigned long flags; | |
7219 | ||
05fa785c | 7220 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 7221 | |
861d034e PZ |
7222 | update_rq_clock(rq); |
7223 | ||
4fc420c9 DN |
7224 | cfs_rq = task_cfs_rq(current); |
7225 | curr = cfs_rq->curr; | |
7226 | ||
6c9a27f5 DN |
7227 | /* |
7228 | * Not only the cpu but also the task_group of the parent might have | |
7229 | * been changed after parent->se.parent,cfs_rq were copied to | |
7230 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
7231 | * of child point to valid ones. | |
7232 | */ | |
7233 | rcu_read_lock(); | |
7234 | __set_task_cpu(p, this_cpu); | |
7235 | rcu_read_unlock(); | |
bf0f6f24 | 7236 | |
7109c442 | 7237 | update_curr(cfs_rq); |
cd29fe6f | 7238 | |
b5d9d734 MG |
7239 | if (curr) |
7240 | se->vruntime = curr->vruntime; | |
aeb73b04 | 7241 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 7242 | |
cd29fe6f | 7243 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 7244 | /* |
edcb60a3 IM |
7245 | * Upon rescheduling, sched_class::put_prev_task() will place |
7246 | * 'current' within the tree based on its new key value. | |
7247 | */ | |
4d78e7b6 | 7248 | swap(curr->vruntime, se->vruntime); |
aec0a514 | 7249 | resched_task(rq->curr); |
4d78e7b6 | 7250 | } |
bf0f6f24 | 7251 | |
88ec22d3 PZ |
7252 | se->vruntime -= cfs_rq->min_vruntime; |
7253 | ||
05fa785c | 7254 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
7255 | } |
7256 | ||
cb469845 SR |
7257 | /* |
7258 | * Priority of the task has changed. Check to see if we preempt | |
7259 | * the current task. | |
7260 | */ | |
da7a735e PZ |
7261 | static void |
7262 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 7263 | { |
da7a735e PZ |
7264 | if (!p->se.on_rq) |
7265 | return; | |
7266 | ||
cb469845 SR |
7267 | /* |
7268 | * Reschedule if we are currently running on this runqueue and | |
7269 | * our priority decreased, or if we are not currently running on | |
7270 | * this runqueue and our priority is higher than the current's | |
7271 | */ | |
da7a735e | 7272 | if (rq->curr == p) { |
cb469845 SR |
7273 | if (p->prio > oldprio) |
7274 | resched_task(rq->curr); | |
7275 | } else | |
15afe09b | 7276 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7277 | } |
7278 | ||
da7a735e PZ |
7279 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
7280 | { | |
7281 | struct sched_entity *se = &p->se; | |
7282 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7283 | ||
7284 | /* | |
7285 | * Ensure the task's vruntime is normalized, so that when its | |
7286 | * switched back to the fair class the enqueue_entity(.flags=0) will | |
7287 | * do the right thing. | |
7288 | * | |
7289 | * If it was on_rq, then the dequeue_entity(.flags=0) will already | |
7290 | * have normalized the vruntime, if it was !on_rq, then only when | |
7291 | * the task is sleeping will it still have non-normalized vruntime. | |
7292 | */ | |
7293 | if (!se->on_rq && p->state != TASK_RUNNING) { | |
7294 | /* | |
7295 | * Fix up our vruntime so that the current sleep doesn't | |
7296 | * cause 'unlimited' sleep bonus. | |
7297 | */ | |
7298 | place_entity(cfs_rq, se, 0); | |
7299 | se->vruntime -= cfs_rq->min_vruntime; | |
7300 | } | |
9ee474f5 | 7301 | |
141965c7 | 7302 | #ifdef CONFIG_SMP |
9ee474f5 PT |
7303 | /* |
7304 | * Remove our load from contribution when we leave sched_fair | |
7305 | * and ensure we don't carry in an old decay_count if we | |
7306 | * switch back. | |
7307 | */ | |
87e3c8ae KT |
7308 | if (se->avg.decay_count) { |
7309 | __synchronize_entity_decay(se); | |
7310 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
7311 | } |
7312 | #endif | |
da7a735e PZ |
7313 | } |
7314 | ||
cb469845 SR |
7315 | /* |
7316 | * We switched to the sched_fair class. | |
7317 | */ | |
da7a735e | 7318 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 7319 | { |
da7a735e PZ |
7320 | if (!p->se.on_rq) |
7321 | return; | |
7322 | ||
cb469845 SR |
7323 | /* |
7324 | * We were most likely switched from sched_rt, so | |
7325 | * kick off the schedule if running, otherwise just see | |
7326 | * if we can still preempt the current task. | |
7327 | */ | |
da7a735e | 7328 | if (rq->curr == p) |
cb469845 SR |
7329 | resched_task(rq->curr); |
7330 | else | |
15afe09b | 7331 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7332 | } |
7333 | ||
83b699ed SV |
7334 | /* Account for a task changing its policy or group. |
7335 | * | |
7336 | * This routine is mostly called to set cfs_rq->curr field when a task | |
7337 | * migrates between groups/classes. | |
7338 | */ | |
7339 | static void set_curr_task_fair(struct rq *rq) | |
7340 | { | |
7341 | struct sched_entity *se = &rq->curr->se; | |
7342 | ||
ec12cb7f PT |
7343 | for_each_sched_entity(se) { |
7344 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7345 | ||
7346 | set_next_entity(cfs_rq, se); | |
7347 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
7348 | account_cfs_rq_runtime(cfs_rq, 0); | |
7349 | } | |
83b699ed SV |
7350 | } |
7351 | ||
029632fb PZ |
7352 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
7353 | { | |
7354 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
7355 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
7356 | #ifndef CONFIG_64BIT | |
7357 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
7358 | #endif | |
141965c7 | 7359 | #ifdef CONFIG_SMP |
9ee474f5 | 7360 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 7361 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 7362 | #endif |
029632fb PZ |
7363 | } |
7364 | ||
810b3817 | 7365 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7366 | static void task_move_group_fair(struct task_struct *p, int on_rq) |
810b3817 | 7367 | { |
fed14d45 | 7368 | struct sched_entity *se = &p->se; |
aff3e498 | 7369 | struct cfs_rq *cfs_rq; |
fed14d45 | 7370 | |
b2b5ce02 PZ |
7371 | /* |
7372 | * If the task was not on the rq at the time of this cgroup movement | |
7373 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
7374 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
7375 | * bonus in place_entity()). | |
7376 | * | |
7377 | * If it was on the rq, we've just 'preempted' it, which does convert | |
7378 | * ->vruntime to a relative base. | |
7379 | * | |
7380 | * Make sure both cases convert their relative position when migrating | |
7381 | * to another cgroup's rq. This does somewhat interfere with the | |
7382 | * fair sleeper stuff for the first placement, but who cares. | |
7383 | */ | |
7ceff013 DN |
7384 | /* |
7385 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
7386 | * But there are some cases where it has already been normalized: | |
7387 | * | |
7388 | * - Moving a forked child which is waiting for being woken up by | |
7389 | * wake_up_new_task(). | |
62af3783 DN |
7390 | * - Moving a task which has been woken up by try_to_wake_up() and |
7391 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
7392 | * |
7393 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
7394 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
7395 | */ | |
fed14d45 | 7396 | if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING)) |
7ceff013 DN |
7397 | on_rq = 1; |
7398 | ||
b2b5ce02 | 7399 | if (!on_rq) |
fed14d45 | 7400 | se->vruntime -= cfs_rq_of(se)->min_vruntime; |
b2b5ce02 | 7401 | set_task_rq(p, task_cpu(p)); |
fed14d45 | 7402 | se->depth = se->parent ? se->parent->depth + 1 : 0; |
aff3e498 | 7403 | if (!on_rq) { |
fed14d45 PZ |
7404 | cfs_rq = cfs_rq_of(se); |
7405 | se->vruntime += cfs_rq->min_vruntime; | |
aff3e498 PT |
7406 | #ifdef CONFIG_SMP |
7407 | /* | |
7408 | * migrate_task_rq_fair() will have removed our previous | |
7409 | * contribution, but we must synchronize for ongoing future | |
7410 | * decay. | |
7411 | */ | |
fed14d45 PZ |
7412 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); |
7413 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
aff3e498 PT |
7414 | #endif |
7415 | } | |
810b3817 | 7416 | } |
029632fb PZ |
7417 | |
7418 | void free_fair_sched_group(struct task_group *tg) | |
7419 | { | |
7420 | int i; | |
7421 | ||
7422 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7423 | ||
7424 | for_each_possible_cpu(i) { | |
7425 | if (tg->cfs_rq) | |
7426 | kfree(tg->cfs_rq[i]); | |
7427 | if (tg->se) | |
7428 | kfree(tg->se[i]); | |
7429 | } | |
7430 | ||
7431 | kfree(tg->cfs_rq); | |
7432 | kfree(tg->se); | |
7433 | } | |
7434 | ||
7435 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7436 | { | |
7437 | struct cfs_rq *cfs_rq; | |
7438 | struct sched_entity *se; | |
7439 | int i; | |
7440 | ||
7441 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
7442 | if (!tg->cfs_rq) | |
7443 | goto err; | |
7444 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
7445 | if (!tg->se) | |
7446 | goto err; | |
7447 | ||
7448 | tg->shares = NICE_0_LOAD; | |
7449 | ||
7450 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7451 | ||
7452 | for_each_possible_cpu(i) { | |
7453 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
7454 | GFP_KERNEL, cpu_to_node(i)); | |
7455 | if (!cfs_rq) | |
7456 | goto err; | |
7457 | ||
7458 | se = kzalloc_node(sizeof(struct sched_entity), | |
7459 | GFP_KERNEL, cpu_to_node(i)); | |
7460 | if (!se) | |
7461 | goto err_free_rq; | |
7462 | ||
7463 | init_cfs_rq(cfs_rq); | |
7464 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
7465 | } | |
7466 | ||
7467 | return 1; | |
7468 | ||
7469 | err_free_rq: | |
7470 | kfree(cfs_rq); | |
7471 | err: | |
7472 | return 0; | |
7473 | } | |
7474 | ||
7475 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
7476 | { | |
7477 | struct rq *rq = cpu_rq(cpu); | |
7478 | unsigned long flags; | |
7479 | ||
7480 | /* | |
7481 | * Only empty task groups can be destroyed; so we can speculatively | |
7482 | * check on_list without danger of it being re-added. | |
7483 | */ | |
7484 | if (!tg->cfs_rq[cpu]->on_list) | |
7485 | return; | |
7486 | ||
7487 | raw_spin_lock_irqsave(&rq->lock, flags); | |
7488 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
7489 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7490 | } | |
7491 | ||
7492 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
7493 | struct sched_entity *se, int cpu, | |
7494 | struct sched_entity *parent) | |
7495 | { | |
7496 | struct rq *rq = cpu_rq(cpu); | |
7497 | ||
7498 | cfs_rq->tg = tg; | |
7499 | cfs_rq->rq = rq; | |
029632fb PZ |
7500 | init_cfs_rq_runtime(cfs_rq); |
7501 | ||
7502 | tg->cfs_rq[cpu] = cfs_rq; | |
7503 | tg->se[cpu] = se; | |
7504 | ||
7505 | /* se could be NULL for root_task_group */ | |
7506 | if (!se) | |
7507 | return; | |
7508 | ||
fed14d45 | 7509 | if (!parent) { |
029632fb | 7510 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
7511 | se->depth = 0; |
7512 | } else { | |
029632fb | 7513 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
7514 | se->depth = parent->depth + 1; |
7515 | } | |
029632fb PZ |
7516 | |
7517 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
7518 | /* guarantee group entities always have weight */ |
7519 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
7520 | se->parent = parent; |
7521 | } | |
7522 | ||
7523 | static DEFINE_MUTEX(shares_mutex); | |
7524 | ||
7525 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
7526 | { | |
7527 | int i; | |
7528 | unsigned long flags; | |
7529 | ||
7530 | /* | |
7531 | * We can't change the weight of the root cgroup. | |
7532 | */ | |
7533 | if (!tg->se[0]) | |
7534 | return -EINVAL; | |
7535 | ||
7536 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
7537 | ||
7538 | mutex_lock(&shares_mutex); | |
7539 | if (tg->shares == shares) | |
7540 | goto done; | |
7541 | ||
7542 | tg->shares = shares; | |
7543 | for_each_possible_cpu(i) { | |
7544 | struct rq *rq = cpu_rq(i); | |
7545 | struct sched_entity *se; | |
7546 | ||
7547 | se = tg->se[i]; | |
7548 | /* Propagate contribution to hierarchy */ | |
7549 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
7550 | |
7551 | /* Possible calls to update_curr() need rq clock */ | |
7552 | update_rq_clock(rq); | |
17bc14b7 | 7553 | for_each_sched_entity(se) |
029632fb PZ |
7554 | update_cfs_shares(group_cfs_rq(se)); |
7555 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
7556 | } | |
7557 | ||
7558 | done: | |
7559 | mutex_unlock(&shares_mutex); | |
7560 | return 0; | |
7561 | } | |
7562 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
7563 | ||
7564 | void free_fair_sched_group(struct task_group *tg) { } | |
7565 | ||
7566 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7567 | { | |
7568 | return 1; | |
7569 | } | |
7570 | ||
7571 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
7572 | ||
7573 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
7574 | ||
810b3817 | 7575 | |
6d686f45 | 7576 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
7577 | { |
7578 | struct sched_entity *se = &task->se; | |
0d721cea PW |
7579 | unsigned int rr_interval = 0; |
7580 | ||
7581 | /* | |
7582 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
7583 | * idle runqueue: | |
7584 | */ | |
0d721cea | 7585 | if (rq->cfs.load.weight) |
a59f4e07 | 7586 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
7587 | |
7588 | return rr_interval; | |
7589 | } | |
7590 | ||
bf0f6f24 IM |
7591 | /* |
7592 | * All the scheduling class methods: | |
7593 | */ | |
029632fb | 7594 | const struct sched_class fair_sched_class = { |
5522d5d5 | 7595 | .next = &idle_sched_class, |
bf0f6f24 IM |
7596 | .enqueue_task = enqueue_task_fair, |
7597 | .dequeue_task = dequeue_task_fair, | |
7598 | .yield_task = yield_task_fair, | |
d95f4122 | 7599 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 7600 | |
2e09bf55 | 7601 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
7602 | |
7603 | .pick_next_task = pick_next_task_fair, | |
7604 | .put_prev_task = put_prev_task_fair, | |
7605 | ||
681f3e68 | 7606 | #ifdef CONFIG_SMP |
4ce72a2c | 7607 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 7608 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 7609 | |
0bcdcf28 CE |
7610 | .rq_online = rq_online_fair, |
7611 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
7612 | |
7613 | .task_waking = task_waking_fair, | |
681f3e68 | 7614 | #endif |
bf0f6f24 | 7615 | |
83b699ed | 7616 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 7617 | .task_tick = task_tick_fair, |
cd29fe6f | 7618 | .task_fork = task_fork_fair, |
cb469845 SR |
7619 | |
7620 | .prio_changed = prio_changed_fair, | |
da7a735e | 7621 | .switched_from = switched_from_fair, |
cb469845 | 7622 | .switched_to = switched_to_fair, |
810b3817 | 7623 | |
0d721cea PW |
7624 | .get_rr_interval = get_rr_interval_fair, |
7625 | ||
810b3817 | 7626 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 7627 | .task_move_group = task_move_group_fair, |
810b3817 | 7628 | #endif |
bf0f6f24 IM |
7629 | }; |
7630 | ||
7631 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 7632 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 7633 | { |
bf0f6f24 IM |
7634 | struct cfs_rq *cfs_rq; |
7635 | ||
5973e5b9 | 7636 | rcu_read_lock(); |
c3b64f1e | 7637 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 7638 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 7639 | rcu_read_unlock(); |
bf0f6f24 IM |
7640 | } |
7641 | #endif | |
029632fb PZ |
7642 | |
7643 | __init void init_sched_fair_class(void) | |
7644 | { | |
7645 | #ifdef CONFIG_SMP | |
7646 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
7647 | ||
3451d024 | 7648 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 7649 | nohz.next_balance = jiffies; |
029632fb | 7650 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 7651 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
7652 | #endif |
7653 | #endif /* SMP */ | |
7654 | ||
7655 | } |