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> |
83a0a96a | 26 | #include <linux/cpuidle.h> |
029632fb PZ |
27 | #include <linux/slab.h> |
28 | #include <linux/profile.h> | |
29 | #include <linux/interrupt.h> | |
cbee9f88 | 30 | #include <linux/mempolicy.h> |
e14808b4 | 31 | #include <linux/migrate.h> |
cbee9f88 | 32 | #include <linux/task_work.h> |
029632fb PZ |
33 | |
34 | #include <trace/events/sched.h> | |
35 | ||
36 | #include "sched.h" | |
9745512c | 37 | |
bf0f6f24 | 38 | /* |
21805085 | 39 | * Targeted preemption latency for CPU-bound tasks: |
864616ee | 40 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 | 41 | * |
21805085 | 42 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
43 | * 'timeslice length' - timeslices in CFS are of variable length |
44 | * and have no persistent notion like in traditional, time-slice | |
45 | * based scheduling concepts. | |
bf0f6f24 | 46 | * |
d274a4ce IM |
47 | * (to see the precise effective timeslice length of your workload, |
48 | * run vmstat and monitor the context-switches (cs) field) | |
bf0f6f24 | 49 | */ |
21406928 MG |
50 | unsigned int sysctl_sched_latency = 6000000ULL; |
51 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 52 | |
1983a922 CE |
53 | /* |
54 | * The initial- and re-scaling of tunables is configurable | |
55 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
56 | * | |
57 | * Options are: | |
58 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
59 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
60 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
61 | */ | |
62 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | |
63 | = SCHED_TUNABLESCALING_LOG; | |
64 | ||
2bd8e6d4 | 65 | /* |
b2be5e96 | 66 | * Minimal preemption granularity for CPU-bound tasks: |
864616ee | 67 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 68 | */ |
0bf377bb IM |
69 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
70 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
71 | |
72 | /* | |
b2be5e96 PZ |
73 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
74 | */ | |
0bf377bb | 75 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
76 | |
77 | /* | |
2bba22c5 | 78 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 79 | * parent will (try to) run first. |
21805085 | 80 | */ |
2bba22c5 | 81 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 82 | |
bf0f6f24 IM |
83 | /* |
84 | * SCHED_OTHER wake-up granularity. | |
172e082a | 85 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
bf0f6f24 IM |
86 | * |
87 | * This option delays the preemption effects of decoupled workloads | |
88 | * and reduces their over-scheduling. Synchronous workloads will still | |
89 | * have immediate wakeup/sleep latencies. | |
90 | */ | |
172e082a | 91 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
0bcdcf28 | 92 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
bf0f6f24 | 93 | |
da84d961 IM |
94 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
95 | ||
a7a4f8a7 PT |
96 | /* |
97 | * The exponential sliding window over which load is averaged for shares | |
98 | * distribution. | |
99 | * (default: 10msec) | |
100 | */ | |
101 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | |
102 | ||
ec12cb7f PT |
103 | #ifdef CONFIG_CFS_BANDWIDTH |
104 | /* | |
105 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
106 | * each time a cfs_rq requests quota. | |
107 | * | |
108 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
109 | * to consumption or the quota being specified to be smaller than the slice) | |
110 | * we will always only issue the remaining available time. | |
111 | * | |
112 | * default: 5 msec, units: microseconds | |
113 | */ | |
114 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
115 | #endif | |
116 | ||
8527632d PG |
117 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
118 | { | |
119 | lw->weight += inc; | |
120 | lw->inv_weight = 0; | |
121 | } | |
122 | ||
123 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
124 | { | |
125 | lw->weight -= dec; | |
126 | lw->inv_weight = 0; | |
127 | } | |
128 | ||
129 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
130 | { | |
131 | lw->weight = w; | |
132 | lw->inv_weight = 0; | |
133 | } | |
134 | ||
029632fb PZ |
135 | /* |
136 | * Increase the granularity value when there are more CPUs, | |
137 | * because with more CPUs the 'effective latency' as visible | |
138 | * to users decreases. But the relationship is not linear, | |
139 | * so pick a second-best guess by going with the log2 of the | |
140 | * number of CPUs. | |
141 | * | |
142 | * This idea comes from the SD scheduler of Con Kolivas: | |
143 | */ | |
144 | static int get_update_sysctl_factor(void) | |
145 | { | |
146 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | |
147 | unsigned int factor; | |
148 | ||
149 | switch (sysctl_sched_tunable_scaling) { | |
150 | case SCHED_TUNABLESCALING_NONE: | |
151 | factor = 1; | |
152 | break; | |
153 | case SCHED_TUNABLESCALING_LINEAR: | |
154 | factor = cpus; | |
155 | break; | |
156 | case SCHED_TUNABLESCALING_LOG: | |
157 | default: | |
158 | factor = 1 + ilog2(cpus); | |
159 | break; | |
160 | } | |
161 | ||
162 | return factor; | |
163 | } | |
164 | ||
165 | static void update_sysctl(void) | |
166 | { | |
167 | unsigned int factor = get_update_sysctl_factor(); | |
168 | ||
169 | #define SET_SYSCTL(name) \ | |
170 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
171 | SET_SYSCTL(sched_min_granularity); | |
172 | SET_SYSCTL(sched_latency); | |
173 | SET_SYSCTL(sched_wakeup_granularity); | |
174 | #undef SET_SYSCTL | |
175 | } | |
176 | ||
177 | void sched_init_granularity(void) | |
178 | { | |
179 | update_sysctl(); | |
180 | } | |
181 | ||
9dbdb155 | 182 | #define WMULT_CONST (~0U) |
029632fb PZ |
183 | #define WMULT_SHIFT 32 |
184 | ||
9dbdb155 PZ |
185 | static void __update_inv_weight(struct load_weight *lw) |
186 | { | |
187 | unsigned long w; | |
188 | ||
189 | if (likely(lw->inv_weight)) | |
190 | return; | |
191 | ||
192 | w = scale_load_down(lw->weight); | |
193 | ||
194 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
195 | lw->inv_weight = 1; | |
196 | else if (unlikely(!w)) | |
197 | lw->inv_weight = WMULT_CONST; | |
198 | else | |
199 | lw->inv_weight = WMULT_CONST / w; | |
200 | } | |
029632fb PZ |
201 | |
202 | /* | |
9dbdb155 PZ |
203 | * delta_exec * weight / lw.weight |
204 | * OR | |
205 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
206 | * | |
207 | * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case | |
208 | * we're guaranteed shift stays positive because inv_weight is guaranteed to | |
209 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
210 | * | |
211 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
212 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 213 | */ |
9dbdb155 | 214 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 215 | { |
9dbdb155 PZ |
216 | u64 fact = scale_load_down(weight); |
217 | int shift = WMULT_SHIFT; | |
029632fb | 218 | |
9dbdb155 | 219 | __update_inv_weight(lw); |
029632fb | 220 | |
9dbdb155 PZ |
221 | if (unlikely(fact >> 32)) { |
222 | while (fact >> 32) { | |
223 | fact >>= 1; | |
224 | shift--; | |
225 | } | |
029632fb PZ |
226 | } |
227 | ||
9dbdb155 PZ |
228 | /* hint to use a 32x32->64 mul */ |
229 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 230 | |
9dbdb155 PZ |
231 | while (fact >> 32) { |
232 | fact >>= 1; | |
233 | shift--; | |
234 | } | |
029632fb | 235 | |
9dbdb155 | 236 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
237 | } |
238 | ||
239 | ||
240 | const struct sched_class fair_sched_class; | |
a4c2f00f | 241 | |
bf0f6f24 IM |
242 | /************************************************************** |
243 | * CFS operations on generic schedulable entities: | |
244 | */ | |
245 | ||
62160e3f | 246 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 247 | |
62160e3f | 248 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
249 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
250 | { | |
62160e3f | 251 | return cfs_rq->rq; |
bf0f6f24 IM |
252 | } |
253 | ||
62160e3f IM |
254 | /* An entity is a task if it doesn't "own" a runqueue */ |
255 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 256 | |
8f48894f PZ |
257 | static inline struct task_struct *task_of(struct sched_entity *se) |
258 | { | |
259 | #ifdef CONFIG_SCHED_DEBUG | |
260 | WARN_ON_ONCE(!entity_is_task(se)); | |
261 | #endif | |
262 | return container_of(se, struct task_struct, se); | |
263 | } | |
264 | ||
b758149c PZ |
265 | /* Walk up scheduling entities hierarchy */ |
266 | #define for_each_sched_entity(se) \ | |
267 | for (; se; se = se->parent) | |
268 | ||
269 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
270 | { | |
271 | return p->se.cfs_rq; | |
272 | } | |
273 | ||
274 | /* runqueue on which this entity is (to be) queued */ | |
275 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
276 | { | |
277 | return se->cfs_rq; | |
278 | } | |
279 | ||
280 | /* runqueue "owned" by this group */ | |
281 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
282 | { | |
283 | return grp->my_q; | |
284 | } | |
285 | ||
aff3e498 PT |
286 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
287 | int force_update); | |
9ee474f5 | 288 | |
3d4b47b4 PZ |
289 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
290 | { | |
291 | if (!cfs_rq->on_list) { | |
67e86250 PT |
292 | /* |
293 | * Ensure we either appear before our parent (if already | |
294 | * enqueued) or force our parent to appear after us when it is | |
295 | * enqueued. The fact that we always enqueue bottom-up | |
296 | * reduces this to two cases. | |
297 | */ | |
298 | if (cfs_rq->tg->parent && | |
299 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
300 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
302 | } else { | |
303 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 304 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 305 | } |
3d4b47b4 PZ |
306 | |
307 | cfs_rq->on_list = 1; | |
9ee474f5 | 308 | /* We should have no load, but we need to update last_decay. */ |
aff3e498 | 309 | update_cfs_rq_blocked_load(cfs_rq, 0); |
3d4b47b4 PZ |
310 | } |
311 | } | |
312 | ||
313 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
314 | { | |
315 | if (cfs_rq->on_list) { | |
316 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
317 | cfs_rq->on_list = 0; | |
318 | } | |
319 | } | |
320 | ||
b758149c PZ |
321 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
322 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
323 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
324 | ||
325 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 326 | static inline struct cfs_rq * |
b758149c PZ |
327 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
328 | { | |
329 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 330 | return se->cfs_rq; |
b758149c | 331 | |
fed14d45 | 332 | return NULL; |
b758149c PZ |
333 | } |
334 | ||
335 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
336 | { | |
337 | return se->parent; | |
338 | } | |
339 | ||
464b7527 PZ |
340 | static void |
341 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
342 | { | |
343 | int se_depth, pse_depth; | |
344 | ||
345 | /* | |
346 | * preemption test can be made between sibling entities who are in the | |
347 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
348 | * both tasks until we find their ancestors who are siblings of common | |
349 | * parent. | |
350 | */ | |
351 | ||
352 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
353 | se_depth = (*se)->depth; |
354 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
355 | |
356 | while (se_depth > pse_depth) { | |
357 | se_depth--; | |
358 | *se = parent_entity(*se); | |
359 | } | |
360 | ||
361 | while (pse_depth > se_depth) { | |
362 | pse_depth--; | |
363 | *pse = parent_entity(*pse); | |
364 | } | |
365 | ||
366 | while (!is_same_group(*se, *pse)) { | |
367 | *se = parent_entity(*se); | |
368 | *pse = parent_entity(*pse); | |
369 | } | |
370 | } | |
371 | ||
8f48894f PZ |
372 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
373 | ||
374 | static inline struct task_struct *task_of(struct sched_entity *se) | |
375 | { | |
376 | return container_of(se, struct task_struct, se); | |
377 | } | |
bf0f6f24 | 378 | |
62160e3f IM |
379 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
380 | { | |
381 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
382 | } |
383 | ||
384 | #define entity_is_task(se) 1 | |
385 | ||
b758149c PZ |
386 | #define for_each_sched_entity(se) \ |
387 | for (; se; se = NULL) | |
bf0f6f24 | 388 | |
b758149c | 389 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 390 | { |
b758149c | 391 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
392 | } |
393 | ||
b758149c PZ |
394 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
395 | { | |
396 | struct task_struct *p = task_of(se); | |
397 | struct rq *rq = task_rq(p); | |
398 | ||
399 | return &rq->cfs; | |
400 | } | |
401 | ||
402 | /* runqueue "owned" by this group */ | |
403 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
404 | { | |
405 | return NULL; | |
406 | } | |
407 | ||
3d4b47b4 PZ |
408 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
409 | { | |
410 | } | |
411 | ||
412 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
413 | { | |
414 | } | |
415 | ||
b758149c PZ |
416 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
417 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
418 | ||
b758149c PZ |
419 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
420 | { | |
421 | return NULL; | |
422 | } | |
423 | ||
464b7527 PZ |
424 | static inline void |
425 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
426 | { | |
427 | } | |
428 | ||
b758149c PZ |
429 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
430 | ||
6c16a6dc | 431 | static __always_inline |
9dbdb155 | 432 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
433 | |
434 | /************************************************************** | |
435 | * Scheduling class tree data structure manipulation methods: | |
436 | */ | |
437 | ||
1bf08230 | 438 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 439 | { |
1bf08230 | 440 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 441 | if (delta > 0) |
1bf08230 | 442 | max_vruntime = vruntime; |
02e0431a | 443 | |
1bf08230 | 444 | return max_vruntime; |
02e0431a PZ |
445 | } |
446 | ||
0702e3eb | 447 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
448 | { |
449 | s64 delta = (s64)(vruntime - min_vruntime); | |
450 | if (delta < 0) | |
451 | min_vruntime = vruntime; | |
452 | ||
453 | return min_vruntime; | |
454 | } | |
455 | ||
54fdc581 FC |
456 | static inline int entity_before(struct sched_entity *a, |
457 | struct sched_entity *b) | |
458 | { | |
459 | return (s64)(a->vruntime - b->vruntime) < 0; | |
460 | } | |
461 | ||
1af5f730 PZ |
462 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
463 | { | |
464 | u64 vruntime = cfs_rq->min_vruntime; | |
465 | ||
466 | if (cfs_rq->curr) | |
467 | vruntime = cfs_rq->curr->vruntime; | |
468 | ||
469 | if (cfs_rq->rb_leftmost) { | |
470 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
471 | struct sched_entity, | |
472 | run_node); | |
473 | ||
e17036da | 474 | if (!cfs_rq->curr) |
1af5f730 PZ |
475 | vruntime = se->vruntime; |
476 | else | |
477 | vruntime = min_vruntime(vruntime, se->vruntime); | |
478 | } | |
479 | ||
1bf08230 | 480 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 481 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
482 | #ifndef CONFIG_64BIT |
483 | smp_wmb(); | |
484 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
485 | #endif | |
1af5f730 PZ |
486 | } |
487 | ||
bf0f6f24 IM |
488 | /* |
489 | * Enqueue an entity into the rb-tree: | |
490 | */ | |
0702e3eb | 491 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
492 | { |
493 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
494 | struct rb_node *parent = NULL; | |
495 | struct sched_entity *entry; | |
bf0f6f24 IM |
496 | int leftmost = 1; |
497 | ||
498 | /* | |
499 | * Find the right place in the rbtree: | |
500 | */ | |
501 | while (*link) { | |
502 | parent = *link; | |
503 | entry = rb_entry(parent, struct sched_entity, run_node); | |
504 | /* | |
505 | * We dont care about collisions. Nodes with | |
506 | * the same key stay together. | |
507 | */ | |
2bd2d6f2 | 508 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
509 | link = &parent->rb_left; |
510 | } else { | |
511 | link = &parent->rb_right; | |
512 | leftmost = 0; | |
513 | } | |
514 | } | |
515 | ||
516 | /* | |
517 | * Maintain a cache of leftmost tree entries (it is frequently | |
518 | * used): | |
519 | */ | |
1af5f730 | 520 | if (leftmost) |
57cb499d | 521 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
522 | |
523 | rb_link_node(&se->run_node, parent, link); | |
524 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
525 | } |
526 | ||
0702e3eb | 527 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 528 | { |
3fe69747 PZ |
529 | if (cfs_rq->rb_leftmost == &se->run_node) { |
530 | struct rb_node *next_node; | |
3fe69747 PZ |
531 | |
532 | next_node = rb_next(&se->run_node); | |
533 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 534 | } |
e9acbff6 | 535 | |
bf0f6f24 | 536 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
537 | } |
538 | ||
029632fb | 539 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 540 | { |
f4b6755f PZ |
541 | struct rb_node *left = cfs_rq->rb_leftmost; |
542 | ||
543 | if (!left) | |
544 | return NULL; | |
545 | ||
546 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
547 | } |
548 | ||
ac53db59 RR |
549 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
550 | { | |
551 | struct rb_node *next = rb_next(&se->run_node); | |
552 | ||
553 | if (!next) | |
554 | return NULL; | |
555 | ||
556 | return rb_entry(next, struct sched_entity, run_node); | |
557 | } | |
558 | ||
559 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 560 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 561 | { |
7eee3e67 | 562 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 563 | |
70eee74b BS |
564 | if (!last) |
565 | return NULL; | |
7eee3e67 IM |
566 | |
567 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
568 | } |
569 | ||
bf0f6f24 IM |
570 | /************************************************************** |
571 | * Scheduling class statistics methods: | |
572 | */ | |
573 | ||
acb4a848 | 574 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 575 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
576 | loff_t *ppos) |
577 | { | |
8d65af78 | 578 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
acb4a848 | 579 | int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
580 | |
581 | if (ret || !write) | |
582 | return ret; | |
583 | ||
584 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
585 | sysctl_sched_min_granularity); | |
586 | ||
acb4a848 CE |
587 | #define WRT_SYSCTL(name) \ |
588 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
589 | WRT_SYSCTL(sched_min_granularity); | |
590 | WRT_SYSCTL(sched_latency); | |
591 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
592 | #undef WRT_SYSCTL |
593 | ||
b2be5e96 PZ |
594 | return 0; |
595 | } | |
596 | #endif | |
647e7cac | 597 | |
a7be37ac | 598 | /* |
f9c0b095 | 599 | * delta /= w |
a7be37ac | 600 | */ |
9dbdb155 | 601 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 602 | { |
f9c0b095 | 603 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 604 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
605 | |
606 | return delta; | |
607 | } | |
608 | ||
647e7cac IM |
609 | /* |
610 | * The idea is to set a period in which each task runs once. | |
611 | * | |
532b1858 | 612 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
613 | * this period because otherwise the slices get too small. |
614 | * | |
615 | * p = (nr <= nl) ? l : l*nr/nl | |
616 | */ | |
4d78e7b6 PZ |
617 | static u64 __sched_period(unsigned long nr_running) |
618 | { | |
619 | u64 period = sysctl_sched_latency; | |
b2be5e96 | 620 | unsigned long nr_latency = sched_nr_latency; |
4d78e7b6 PZ |
621 | |
622 | if (unlikely(nr_running > nr_latency)) { | |
4bf0b771 | 623 | period = sysctl_sched_min_granularity; |
4d78e7b6 | 624 | period *= nr_running; |
4d78e7b6 PZ |
625 | } |
626 | ||
627 | return period; | |
628 | } | |
629 | ||
647e7cac IM |
630 | /* |
631 | * We calculate the wall-time slice from the period by taking a part | |
632 | * proportional to the weight. | |
633 | * | |
f9c0b095 | 634 | * s = p*P[w/rw] |
647e7cac | 635 | */ |
6d0f0ebd | 636 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 637 | { |
0a582440 | 638 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 639 | |
0a582440 | 640 | for_each_sched_entity(se) { |
6272d68c | 641 | struct load_weight *load; |
3104bf03 | 642 | struct load_weight lw; |
6272d68c LM |
643 | |
644 | cfs_rq = cfs_rq_of(se); | |
645 | load = &cfs_rq->load; | |
f9c0b095 | 646 | |
0a582440 | 647 | if (unlikely(!se->on_rq)) { |
3104bf03 | 648 | lw = cfs_rq->load; |
0a582440 MG |
649 | |
650 | update_load_add(&lw, se->load.weight); | |
651 | load = &lw; | |
652 | } | |
9dbdb155 | 653 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
654 | } |
655 | return slice; | |
bf0f6f24 IM |
656 | } |
657 | ||
647e7cac | 658 | /* |
660cc00f | 659 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 660 | * |
f9c0b095 | 661 | * vs = s/w |
647e7cac | 662 | */ |
f9c0b095 | 663 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 664 | { |
f9c0b095 | 665 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
666 | } |
667 | ||
a75cdaa9 | 668 | #ifdef CONFIG_SMP |
ba7e5a27 | 669 | static int select_idle_sibling(struct task_struct *p, int cpu); |
fb13c7ee MG |
670 | static unsigned long task_h_load(struct task_struct *p); |
671 | ||
a75cdaa9 AS |
672 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
673 | ||
674 | /* Give new task start runnable values to heavy its load in infant time */ | |
675 | void init_task_runnable_average(struct task_struct *p) | |
676 | { | |
677 | u32 slice; | |
678 | ||
a75cdaa9 AS |
679 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; |
680 | p->se.avg.runnable_avg_sum = slice; | |
681 | p->se.avg.runnable_avg_period = slice; | |
682 | __update_task_entity_contrib(&p->se); | |
683 | } | |
684 | #else | |
685 | void init_task_runnable_average(struct task_struct *p) | |
686 | { | |
687 | } | |
688 | #endif | |
689 | ||
bf0f6f24 | 690 | /* |
9dbdb155 | 691 | * Update the current task's runtime statistics. |
bf0f6f24 | 692 | */ |
b7cc0896 | 693 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 694 | { |
429d43bc | 695 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 696 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 697 | u64 delta_exec; |
bf0f6f24 IM |
698 | |
699 | if (unlikely(!curr)) | |
700 | return; | |
701 | ||
9dbdb155 PZ |
702 | delta_exec = now - curr->exec_start; |
703 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 704 | return; |
bf0f6f24 | 705 | |
8ebc91d9 | 706 | curr->exec_start = now; |
d842de87 | 707 | |
9dbdb155 PZ |
708 | schedstat_set(curr->statistics.exec_max, |
709 | max(delta_exec, curr->statistics.exec_max)); | |
710 | ||
711 | curr->sum_exec_runtime += delta_exec; | |
712 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
713 | ||
714 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
715 | update_min_vruntime(cfs_rq); | |
716 | ||
d842de87 SV |
717 | if (entity_is_task(curr)) { |
718 | struct task_struct *curtask = task_of(curr); | |
719 | ||
f977bb49 | 720 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 721 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 722 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 723 | } |
ec12cb7f PT |
724 | |
725 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
726 | } |
727 | ||
6e998916 SG |
728 | static void update_curr_fair(struct rq *rq) |
729 | { | |
730 | update_curr(cfs_rq_of(&rq->curr->se)); | |
731 | } | |
732 | ||
bf0f6f24 | 733 | static inline void |
5870db5b | 734 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 735 | { |
78becc27 | 736 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
bf0f6f24 IM |
737 | } |
738 | ||
bf0f6f24 IM |
739 | /* |
740 | * Task is being enqueued - update stats: | |
741 | */ | |
d2417e5a | 742 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 743 | { |
bf0f6f24 IM |
744 | /* |
745 | * Are we enqueueing a waiting task? (for current tasks | |
746 | * a dequeue/enqueue event is a NOP) | |
747 | */ | |
429d43bc | 748 | if (se != cfs_rq->curr) |
5870db5b | 749 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
750 | } |
751 | ||
bf0f6f24 | 752 | static void |
9ef0a961 | 753 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 754 | { |
41acab88 | 755 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
78becc27 | 756 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
41acab88 LDM |
757 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
758 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | |
78becc27 | 759 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
760 | #ifdef CONFIG_SCHEDSTATS |
761 | if (entity_is_task(se)) { | |
762 | trace_sched_stat_wait(task_of(se), | |
78becc27 | 763 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
768d0c27 PZ |
764 | } |
765 | #endif | |
41acab88 | 766 | schedstat_set(se->statistics.wait_start, 0); |
bf0f6f24 IM |
767 | } |
768 | ||
769 | static inline void | |
19b6a2e3 | 770 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 771 | { |
bf0f6f24 IM |
772 | /* |
773 | * Mark the end of the wait period if dequeueing a | |
774 | * waiting task: | |
775 | */ | |
429d43bc | 776 | if (se != cfs_rq->curr) |
9ef0a961 | 777 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
778 | } |
779 | ||
780 | /* | |
781 | * We are picking a new current task - update its stats: | |
782 | */ | |
783 | static inline void | |
79303e9e | 784 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
785 | { |
786 | /* | |
787 | * We are starting a new run period: | |
788 | */ | |
78becc27 | 789 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
790 | } |
791 | ||
bf0f6f24 IM |
792 | /************************************************** |
793 | * Scheduling class queueing methods: | |
794 | */ | |
795 | ||
cbee9f88 PZ |
796 | #ifdef CONFIG_NUMA_BALANCING |
797 | /* | |
598f0ec0 MG |
798 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
799 | * calculated based on the tasks virtual memory size and | |
800 | * numa_balancing_scan_size. | |
cbee9f88 | 801 | */ |
598f0ec0 MG |
802 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
803 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
804 | |
805 | /* Portion of address space to scan in MB */ | |
806 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 807 | |
4b96a29b PZ |
808 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
809 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
810 | ||
598f0ec0 MG |
811 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
812 | { | |
813 | unsigned long rss = 0; | |
814 | unsigned long nr_scan_pages; | |
815 | ||
816 | /* | |
817 | * Calculations based on RSS as non-present and empty pages are skipped | |
818 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
819 | * on resident pages | |
820 | */ | |
821 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
822 | rss = get_mm_rss(p->mm); | |
823 | if (!rss) | |
824 | rss = nr_scan_pages; | |
825 | ||
826 | rss = round_up(rss, nr_scan_pages); | |
827 | return rss / nr_scan_pages; | |
828 | } | |
829 | ||
830 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
831 | #define MAX_SCAN_WINDOW 2560 | |
832 | ||
833 | static unsigned int task_scan_min(struct task_struct *p) | |
834 | { | |
64192658 | 835 | unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
836 | unsigned int scan, floor; |
837 | unsigned int windows = 1; | |
838 | ||
64192658 KT |
839 | if (scan_size < MAX_SCAN_WINDOW) |
840 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
841 | floor = 1000 / windows; |
842 | ||
843 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
844 | return max_t(unsigned int, floor, scan); | |
845 | } | |
846 | ||
847 | static unsigned int task_scan_max(struct task_struct *p) | |
848 | { | |
849 | unsigned int smin = task_scan_min(p); | |
850 | unsigned int smax; | |
851 | ||
852 | /* Watch for min being lower than max due to floor calculations */ | |
853 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
854 | return max(smin, smax); | |
855 | } | |
856 | ||
0ec8aa00 PZ |
857 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
858 | { | |
859 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
860 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
861 | } | |
862 | ||
863 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
864 | { | |
865 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
866 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
867 | } | |
868 | ||
8c8a743c PZ |
869 | struct numa_group { |
870 | atomic_t refcount; | |
871 | ||
872 | spinlock_t lock; /* nr_tasks, tasks */ | |
873 | int nr_tasks; | |
e29cf08b | 874 | pid_t gid; |
8c8a743c PZ |
875 | |
876 | struct rcu_head rcu; | |
20e07dea | 877 | nodemask_t active_nodes; |
989348b5 | 878 | unsigned long total_faults; |
7e2703e6 RR |
879 | /* |
880 | * Faults_cpu is used to decide whether memory should move | |
881 | * towards the CPU. As a consequence, these stats are weighted | |
882 | * more by CPU use than by memory faults. | |
883 | */ | |
50ec8a40 | 884 | unsigned long *faults_cpu; |
989348b5 | 885 | unsigned long faults[0]; |
8c8a743c PZ |
886 | }; |
887 | ||
be1e4e76 RR |
888 | /* Shared or private faults. */ |
889 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
890 | ||
891 | /* Memory and CPU locality */ | |
892 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
893 | ||
894 | /* Averaged statistics, and temporary buffers. */ | |
895 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
896 | ||
e29cf08b MG |
897 | pid_t task_numa_group_id(struct task_struct *p) |
898 | { | |
899 | return p->numa_group ? p->numa_group->gid : 0; | |
900 | } | |
901 | ||
44dba3d5 IM |
902 | /* |
903 | * The averaged statistics, shared & private, memory & cpu, | |
904 | * occupy the first half of the array. The second half of the | |
905 | * array is for current counters, which are averaged into the | |
906 | * first set by task_numa_placement. | |
907 | */ | |
908 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 909 | { |
44dba3d5 | 910 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
911 | } |
912 | ||
913 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
914 | { | |
44dba3d5 | 915 | if (!p->numa_faults) |
ac8e895b MG |
916 | return 0; |
917 | ||
44dba3d5 IM |
918 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
919 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
920 | } |
921 | ||
83e1d2cd MG |
922 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
923 | { | |
924 | if (!p->numa_group) | |
925 | return 0; | |
926 | ||
44dba3d5 IM |
927 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
928 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
929 | } |
930 | ||
20e07dea RR |
931 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
932 | { | |
44dba3d5 IM |
933 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
934 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
935 | } |
936 | ||
6c6b1193 RR |
937 | /* Handle placement on systems where not all nodes are directly connected. */ |
938 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
939 | int maxdist, bool task) | |
940 | { | |
941 | unsigned long score = 0; | |
942 | int node; | |
943 | ||
944 | /* | |
945 | * All nodes are directly connected, and the same distance | |
946 | * from each other. No need for fancy placement algorithms. | |
947 | */ | |
948 | if (sched_numa_topology_type == NUMA_DIRECT) | |
949 | return 0; | |
950 | ||
951 | /* | |
952 | * This code is called for each node, introducing N^2 complexity, | |
953 | * which should be ok given the number of nodes rarely exceeds 8. | |
954 | */ | |
955 | for_each_online_node(node) { | |
956 | unsigned long faults; | |
957 | int dist = node_distance(nid, node); | |
958 | ||
959 | /* | |
960 | * The furthest away nodes in the system are not interesting | |
961 | * for placement; nid was already counted. | |
962 | */ | |
963 | if (dist == sched_max_numa_distance || node == nid) | |
964 | continue; | |
965 | ||
966 | /* | |
967 | * On systems with a backplane NUMA topology, compare groups | |
968 | * of nodes, and move tasks towards the group with the most | |
969 | * memory accesses. When comparing two nodes at distance | |
970 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
971 | * of each group. Skip other nodes. | |
972 | */ | |
973 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
974 | dist > maxdist) | |
975 | continue; | |
976 | ||
977 | /* Add up the faults from nearby nodes. */ | |
978 | if (task) | |
979 | faults = task_faults(p, node); | |
980 | else | |
981 | faults = group_faults(p, node); | |
982 | ||
983 | /* | |
984 | * On systems with a glueless mesh NUMA topology, there are | |
985 | * no fixed "groups of nodes". Instead, nodes that are not | |
986 | * directly connected bounce traffic through intermediate | |
987 | * nodes; a numa_group can occupy any set of nodes. | |
988 | * The further away a node is, the less the faults count. | |
989 | * This seems to result in good task placement. | |
990 | */ | |
991 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
992 | faults *= (sched_max_numa_distance - dist); | |
993 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
994 | } | |
995 | ||
996 | score += faults; | |
997 | } | |
998 | ||
999 | return score; | |
1000 | } | |
1001 | ||
83e1d2cd MG |
1002 | /* |
1003 | * These return the fraction of accesses done by a particular task, or | |
1004 | * task group, on a particular numa node. The group weight is given a | |
1005 | * larger multiplier, in order to group tasks together that are almost | |
1006 | * evenly spread out between numa nodes. | |
1007 | */ | |
7bd95320 RR |
1008 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1009 | int dist) | |
83e1d2cd | 1010 | { |
7bd95320 | 1011 | unsigned long faults, total_faults; |
83e1d2cd | 1012 | |
44dba3d5 | 1013 | if (!p->numa_faults) |
83e1d2cd MG |
1014 | return 0; |
1015 | ||
1016 | total_faults = p->total_numa_faults; | |
1017 | ||
1018 | if (!total_faults) | |
1019 | return 0; | |
1020 | ||
7bd95320 | 1021 | faults = task_faults(p, nid); |
6c6b1193 RR |
1022 | faults += score_nearby_nodes(p, nid, dist, true); |
1023 | ||
7bd95320 | 1024 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1025 | } |
1026 | ||
7bd95320 RR |
1027 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1028 | int dist) | |
83e1d2cd | 1029 | { |
7bd95320 RR |
1030 | unsigned long faults, total_faults; |
1031 | ||
1032 | if (!p->numa_group) | |
1033 | return 0; | |
1034 | ||
1035 | total_faults = p->numa_group->total_faults; | |
1036 | ||
1037 | if (!total_faults) | |
83e1d2cd MG |
1038 | return 0; |
1039 | ||
7bd95320 | 1040 | faults = group_faults(p, nid); |
6c6b1193 RR |
1041 | faults += score_nearby_nodes(p, nid, dist, false); |
1042 | ||
7bd95320 | 1043 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1044 | } |
1045 | ||
10f39042 RR |
1046 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1047 | int src_nid, int dst_cpu) | |
1048 | { | |
1049 | struct numa_group *ng = p->numa_group; | |
1050 | int dst_nid = cpu_to_node(dst_cpu); | |
1051 | int last_cpupid, this_cpupid; | |
1052 | ||
1053 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
1054 | ||
1055 | /* | |
1056 | * Multi-stage node selection is used in conjunction with a periodic | |
1057 | * migration fault to build a temporal task<->page relation. By using | |
1058 | * a two-stage filter we remove short/unlikely relations. | |
1059 | * | |
1060 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1061 | * a task's usage of a particular page (n_p) per total usage of this | |
1062 | * page (n_t) (in a given time-span) to a probability. | |
1063 | * | |
1064 | * Our periodic faults will sample this probability and getting the | |
1065 | * same result twice in a row, given these samples are fully | |
1066 | * independent, is then given by P(n)^2, provided our sample period | |
1067 | * is sufficiently short compared to the usage pattern. | |
1068 | * | |
1069 | * This quadric squishes small probabilities, making it less likely we | |
1070 | * act on an unlikely task<->page relation. | |
1071 | */ | |
1072 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
1073 | if (!cpupid_pid_unset(last_cpupid) && | |
1074 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1075 | return false; | |
1076 | ||
1077 | /* Always allow migrate on private faults */ | |
1078 | if (cpupid_match_pid(p, last_cpupid)) | |
1079 | return true; | |
1080 | ||
1081 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1082 | if (!ng) | |
1083 | return true; | |
1084 | ||
1085 | /* | |
1086 | * Do not migrate if the destination is not a node that | |
1087 | * is actively used by this numa group. | |
1088 | */ | |
1089 | if (!node_isset(dst_nid, ng->active_nodes)) | |
1090 | return false; | |
1091 | ||
1092 | /* | |
1093 | * Source is a node that is not actively used by this | |
1094 | * numa group, while the destination is. Migrate. | |
1095 | */ | |
1096 | if (!node_isset(src_nid, ng->active_nodes)) | |
1097 | return true; | |
1098 | ||
1099 | /* | |
1100 | * Both source and destination are nodes in active | |
1101 | * use by this numa group. Maximize memory bandwidth | |
1102 | * by migrating from more heavily used groups, to less | |
1103 | * heavily used ones, spreading the load around. | |
1104 | * Use a 1/4 hysteresis to avoid spurious page movement. | |
1105 | */ | |
1106 | return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); | |
1107 | } | |
1108 | ||
e6628d5b | 1109 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
1110 | static unsigned long source_load(int cpu, int type); |
1111 | static unsigned long target_load(int cpu, int type); | |
ced549fa | 1112 | static unsigned long capacity_of(int cpu); |
58d081b5 MG |
1113 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); |
1114 | ||
fb13c7ee | 1115 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1116 | struct numa_stats { |
fb13c7ee | 1117 | unsigned long nr_running; |
58d081b5 | 1118 | unsigned long load; |
fb13c7ee MG |
1119 | |
1120 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1121 | unsigned long compute_capacity; |
fb13c7ee MG |
1122 | |
1123 | /* Approximate capacity in terms of runnable tasks on a node */ | |
5ef20ca1 | 1124 | unsigned long task_capacity; |
1b6a7495 | 1125 | int has_free_capacity; |
58d081b5 | 1126 | }; |
e6628d5b | 1127 | |
fb13c7ee MG |
1128 | /* |
1129 | * XXX borrowed from update_sg_lb_stats | |
1130 | */ | |
1131 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1132 | { | |
83d7f242 RR |
1133 | int smt, cpu, cpus = 0; |
1134 | unsigned long capacity; | |
fb13c7ee MG |
1135 | |
1136 | memset(ns, 0, sizeof(*ns)); | |
1137 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1138 | struct rq *rq = cpu_rq(cpu); | |
1139 | ||
1140 | ns->nr_running += rq->nr_running; | |
1141 | ns->load += weighted_cpuload(cpu); | |
ced549fa | 1142 | ns->compute_capacity += capacity_of(cpu); |
5eca82a9 PZ |
1143 | |
1144 | cpus++; | |
fb13c7ee MG |
1145 | } |
1146 | ||
5eca82a9 PZ |
1147 | /* |
1148 | * If we raced with hotplug and there are no CPUs left in our mask | |
1149 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1150 | * not find this node attractive. | |
1151 | * | |
1b6a7495 NP |
1152 | * We'll either bail at !has_free_capacity, or we'll detect a huge |
1153 | * imbalance and bail there. | |
5eca82a9 PZ |
1154 | */ |
1155 | if (!cpus) | |
1156 | return; | |
1157 | ||
83d7f242 RR |
1158 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ |
1159 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); | |
1160 | capacity = cpus / smt; /* cores */ | |
1161 | ||
1162 | ns->task_capacity = min_t(unsigned, capacity, | |
1163 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); | |
1b6a7495 | 1164 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); |
fb13c7ee MG |
1165 | } |
1166 | ||
58d081b5 MG |
1167 | struct task_numa_env { |
1168 | struct task_struct *p; | |
e6628d5b | 1169 | |
58d081b5 MG |
1170 | int src_cpu, src_nid; |
1171 | int dst_cpu, dst_nid; | |
e6628d5b | 1172 | |
58d081b5 | 1173 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1174 | |
40ea2b42 | 1175 | int imbalance_pct; |
7bd95320 | 1176 | int dist; |
fb13c7ee MG |
1177 | |
1178 | struct task_struct *best_task; | |
1179 | long best_imp; | |
58d081b5 MG |
1180 | int best_cpu; |
1181 | }; | |
1182 | ||
fb13c7ee MG |
1183 | static void task_numa_assign(struct task_numa_env *env, |
1184 | struct task_struct *p, long imp) | |
1185 | { | |
1186 | if (env->best_task) | |
1187 | put_task_struct(env->best_task); | |
1188 | if (p) | |
1189 | get_task_struct(p); | |
1190 | ||
1191 | env->best_task = p; | |
1192 | env->best_imp = imp; | |
1193 | env->best_cpu = env->dst_cpu; | |
1194 | } | |
1195 | ||
28a21745 | 1196 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1197 | struct task_numa_env *env) |
1198 | { | |
28a21745 | 1199 | long src_capacity, dst_capacity; |
095bebf6 RR |
1200 | long orig_src_load; |
1201 | long load_a, load_b; | |
1202 | long moved_load; | |
1203 | long imb; | |
28a21745 RR |
1204 | |
1205 | /* | |
1206 | * The load is corrected for the CPU capacity available on each node. | |
1207 | * | |
1208 | * src_load dst_load | |
1209 | * ------------ vs --------- | |
1210 | * src_capacity dst_capacity | |
1211 | */ | |
1212 | src_capacity = env->src_stats.compute_capacity; | |
1213 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 RR |
1214 | |
1215 | /* We care about the slope of the imbalance, not the direction. */ | |
095bebf6 RR |
1216 | load_a = dst_load; |
1217 | load_b = src_load; | |
1218 | if (load_a < load_b) | |
1219 | swap(load_a, load_b); | |
e63da036 RR |
1220 | |
1221 | /* Is the difference below the threshold? */ | |
095bebf6 RR |
1222 | imb = load_a * src_capacity * 100 - |
1223 | load_b * dst_capacity * env->imbalance_pct; | |
e63da036 RR |
1224 | if (imb <= 0) |
1225 | return false; | |
1226 | ||
1227 | /* | |
1228 | * The imbalance is above the allowed threshold. | |
095bebf6 RR |
1229 | * Allow a move that brings us closer to a balanced situation, |
1230 | * without moving things past the point of balance. | |
e63da036 | 1231 | */ |
28a21745 | 1232 | orig_src_load = env->src_stats.load; |
28a21745 | 1233 | |
095bebf6 RR |
1234 | /* |
1235 | * In a task swap, there will be one load moving from src to dst, | |
1236 | * and another moving back. This is the net sum of both moves. | |
1237 | * A simple task move will always have a positive value. | |
1238 | * Allow the move if it brings the system closer to a balanced | |
1239 | * situation, without crossing over the balance point. | |
1240 | */ | |
1241 | moved_load = orig_src_load - src_load; | |
e63da036 | 1242 | |
095bebf6 RR |
1243 | if (moved_load > 0) |
1244 | /* Moving src -> dst. Did we overshoot balance? */ | |
1245 | return src_load * dst_capacity < dst_load * src_capacity; | |
1246 | else | |
1247 | /* Moving dst -> src. Did we overshoot balance? */ | |
1248 | return dst_load * src_capacity < src_load * dst_capacity; | |
e63da036 RR |
1249 | } |
1250 | ||
fb13c7ee MG |
1251 | /* |
1252 | * This checks if the overall compute and NUMA accesses of the system would | |
1253 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1254 | * into account that it might be best if task running on the dst_cpu should | |
1255 | * be exchanged with the source task | |
1256 | */ | |
887c290e RR |
1257 | static void task_numa_compare(struct task_numa_env *env, |
1258 | long taskimp, long groupimp) | |
fb13c7ee MG |
1259 | { |
1260 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1261 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1262 | struct task_struct *cur; | |
28a21745 | 1263 | long src_load, dst_load; |
fb13c7ee | 1264 | long load; |
1c5d3eb3 | 1265 | long imp = env->p->numa_group ? groupimp : taskimp; |
0132c3e1 | 1266 | long moveimp = imp; |
7bd95320 | 1267 | int dist = env->dist; |
fb13c7ee MG |
1268 | |
1269 | rcu_read_lock(); | |
1effd9f1 KT |
1270 | |
1271 | raw_spin_lock_irq(&dst_rq->lock); | |
1272 | cur = dst_rq->curr; | |
1273 | /* | |
1274 | * No need to move the exiting task, and this ensures that ->curr | |
1275 | * wasn't reaped and thus get_task_struct() in task_numa_assign() | |
1276 | * is safe under RCU read lock. | |
1277 | * Note that rcu_read_lock() itself can't protect from the final | |
1278 | * put_task_struct() after the last schedule(). | |
1279 | */ | |
1280 | if ((cur->flags & PF_EXITING) || is_idle_task(cur)) | |
fb13c7ee | 1281 | cur = NULL; |
1effd9f1 | 1282 | raw_spin_unlock_irq(&dst_rq->lock); |
fb13c7ee | 1283 | |
7af68335 PZ |
1284 | /* |
1285 | * Because we have preemption enabled we can get migrated around and | |
1286 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1287 | */ | |
1288 | if (cur == env->p) | |
1289 | goto unlock; | |
1290 | ||
fb13c7ee MG |
1291 | /* |
1292 | * "imp" is the fault differential for the source task between the | |
1293 | * source and destination node. Calculate the total differential for | |
1294 | * the source task and potential destination task. The more negative | |
1295 | * the value is, the more rmeote accesses that would be expected to | |
1296 | * be incurred if the tasks were swapped. | |
1297 | */ | |
1298 | if (cur) { | |
1299 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1300 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1301 | goto unlock; | |
1302 | ||
887c290e RR |
1303 | /* |
1304 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1305 | * in any group then look only at task weights. |
887c290e | 1306 | */ |
ca28aa53 | 1307 | if (cur->numa_group == env->p->numa_group) { |
7bd95320 RR |
1308 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1309 | task_weight(cur, env->dst_nid, dist); | |
ca28aa53 RR |
1310 | /* |
1311 | * Add some hysteresis to prevent swapping the | |
1312 | * tasks within a group over tiny differences. | |
1313 | */ | |
1314 | if (cur->numa_group) | |
1315 | imp -= imp/16; | |
887c290e | 1316 | } else { |
ca28aa53 RR |
1317 | /* |
1318 | * Compare the group weights. If a task is all by | |
1319 | * itself (not part of a group), use the task weight | |
1320 | * instead. | |
1321 | */ | |
ca28aa53 | 1322 | if (cur->numa_group) |
7bd95320 RR |
1323 | imp += group_weight(cur, env->src_nid, dist) - |
1324 | group_weight(cur, env->dst_nid, dist); | |
ca28aa53 | 1325 | else |
7bd95320 RR |
1326 | imp += task_weight(cur, env->src_nid, dist) - |
1327 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1328 | } |
fb13c7ee MG |
1329 | } |
1330 | ||
0132c3e1 | 1331 | if (imp <= env->best_imp && moveimp <= env->best_imp) |
fb13c7ee MG |
1332 | goto unlock; |
1333 | ||
1334 | if (!cur) { | |
1335 | /* Is there capacity at our destination? */ | |
b932c03c | 1336 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && |
1b6a7495 | 1337 | !env->dst_stats.has_free_capacity) |
fb13c7ee MG |
1338 | goto unlock; |
1339 | ||
1340 | goto balance; | |
1341 | } | |
1342 | ||
1343 | /* Balance doesn't matter much if we're running a task per cpu */ | |
0132c3e1 RR |
1344 | if (imp > env->best_imp && src_rq->nr_running == 1 && |
1345 | dst_rq->nr_running == 1) | |
fb13c7ee MG |
1346 | goto assign; |
1347 | ||
1348 | /* | |
1349 | * In the overloaded case, try and keep the load balanced. | |
1350 | */ | |
1351 | balance: | |
e720fff6 PZ |
1352 | load = task_h_load(env->p); |
1353 | dst_load = env->dst_stats.load + load; | |
1354 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1355 | |
0132c3e1 RR |
1356 | if (moveimp > imp && moveimp > env->best_imp) { |
1357 | /* | |
1358 | * If the improvement from just moving env->p direction is | |
1359 | * better than swapping tasks around, check if a move is | |
1360 | * possible. Store a slightly smaller score than moveimp, | |
1361 | * so an actually idle CPU will win. | |
1362 | */ | |
1363 | if (!load_too_imbalanced(src_load, dst_load, env)) { | |
1364 | imp = moveimp - 1; | |
1365 | cur = NULL; | |
1366 | goto assign; | |
1367 | } | |
1368 | } | |
1369 | ||
1370 | if (imp <= env->best_imp) | |
1371 | goto unlock; | |
1372 | ||
fb13c7ee | 1373 | if (cur) { |
e720fff6 PZ |
1374 | load = task_h_load(cur); |
1375 | dst_load -= load; | |
1376 | src_load += load; | |
fb13c7ee MG |
1377 | } |
1378 | ||
28a21745 | 1379 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1380 | goto unlock; |
1381 | ||
ba7e5a27 RR |
1382 | /* |
1383 | * One idle CPU per node is evaluated for a task numa move. | |
1384 | * Call select_idle_sibling to maybe find a better one. | |
1385 | */ | |
1386 | if (!cur) | |
1387 | env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); | |
1388 | ||
fb13c7ee MG |
1389 | assign: |
1390 | task_numa_assign(env, cur, imp); | |
1391 | unlock: | |
1392 | rcu_read_unlock(); | |
1393 | } | |
1394 | ||
887c290e RR |
1395 | static void task_numa_find_cpu(struct task_numa_env *env, |
1396 | long taskimp, long groupimp) | |
2c8a50aa MG |
1397 | { |
1398 | int cpu; | |
1399 | ||
1400 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1401 | /* Skip this CPU if the source task cannot migrate */ | |
1402 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1403 | continue; | |
1404 | ||
1405 | env->dst_cpu = cpu; | |
887c290e | 1406 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1407 | } |
1408 | } | |
1409 | ||
58d081b5 MG |
1410 | static int task_numa_migrate(struct task_struct *p) |
1411 | { | |
58d081b5 MG |
1412 | struct task_numa_env env = { |
1413 | .p = p, | |
fb13c7ee | 1414 | |
58d081b5 | 1415 | .src_cpu = task_cpu(p), |
b32e86b4 | 1416 | .src_nid = task_node(p), |
fb13c7ee MG |
1417 | |
1418 | .imbalance_pct = 112, | |
1419 | ||
1420 | .best_task = NULL, | |
1421 | .best_imp = 0, | |
1422 | .best_cpu = -1 | |
58d081b5 MG |
1423 | }; |
1424 | struct sched_domain *sd; | |
887c290e | 1425 | unsigned long taskweight, groupweight; |
7bd95320 | 1426 | int nid, ret, dist; |
887c290e | 1427 | long taskimp, groupimp; |
e6628d5b | 1428 | |
58d081b5 | 1429 | /* |
fb13c7ee MG |
1430 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1431 | * imbalance and would be the first to start moving tasks about. | |
1432 | * | |
1433 | * And we want to avoid any moving of tasks about, as that would create | |
1434 | * random movement of tasks -- counter the numa conditions we're trying | |
1435 | * to satisfy here. | |
58d081b5 MG |
1436 | */ |
1437 | rcu_read_lock(); | |
fb13c7ee | 1438 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1439 | if (sd) |
1440 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1441 | rcu_read_unlock(); |
1442 | ||
46a73e8a RR |
1443 | /* |
1444 | * Cpusets can break the scheduler domain tree into smaller | |
1445 | * balance domains, some of which do not cross NUMA boundaries. | |
1446 | * Tasks that are "trapped" in such domains cannot be migrated | |
1447 | * elsewhere, so there is no point in (re)trying. | |
1448 | */ | |
1449 | if (unlikely(!sd)) { | |
de1b301a | 1450 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1451 | return -EINVAL; |
1452 | } | |
1453 | ||
2c8a50aa | 1454 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1455 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1456 | taskweight = task_weight(p, env.src_nid, dist); | |
1457 | groupweight = group_weight(p, env.src_nid, dist); | |
1458 | update_numa_stats(&env.src_stats, env.src_nid); | |
1459 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1460 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1461 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1462 | |
a43455a1 RR |
1463 | /* Try to find a spot on the preferred nid. */ |
1464 | task_numa_find_cpu(&env, taskimp, groupimp); | |
e1dda8a7 | 1465 | |
9de05d48 RR |
1466 | /* |
1467 | * Look at other nodes in these cases: | |
1468 | * - there is no space available on the preferred_nid | |
1469 | * - the task is part of a numa_group that is interleaved across | |
1470 | * multiple NUMA nodes; in order to better consolidate the group, | |
1471 | * we need to check other locations. | |
1472 | */ | |
1473 | if (env.best_cpu == -1 || (p->numa_group && | |
1474 | nodes_weight(p->numa_group->active_nodes) > 1)) { | |
2c8a50aa MG |
1475 | for_each_online_node(nid) { |
1476 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1477 | continue; | |
58d081b5 | 1478 | |
7bd95320 | 1479 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1480 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1481 | dist != env.dist) { | |
1482 | taskweight = task_weight(p, env.src_nid, dist); | |
1483 | groupweight = group_weight(p, env.src_nid, dist); | |
1484 | } | |
7bd95320 | 1485 | |
83e1d2cd | 1486 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1487 | taskimp = task_weight(p, nid, dist) - taskweight; |
1488 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1489 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1490 | continue; |
1491 | ||
7bd95320 | 1492 | env.dist = dist; |
2c8a50aa MG |
1493 | env.dst_nid = nid; |
1494 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
887c290e | 1495 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1496 | } |
1497 | } | |
1498 | ||
68d1b02a RR |
1499 | /* |
1500 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1501 | * and is migrating into one of the workload's active nodes, remember | |
1502 | * this node as the task's preferred numa node, so the workload can | |
1503 | * settle down. | |
1504 | * A task that migrated to a second choice node will be better off | |
1505 | * trying for a better one later. Do not set the preferred node here. | |
1506 | */ | |
db015dae RR |
1507 | if (p->numa_group) { |
1508 | if (env.best_cpu == -1) | |
1509 | nid = env.src_nid; | |
1510 | else | |
1511 | nid = env.dst_nid; | |
1512 | ||
1513 | if (node_isset(nid, p->numa_group->active_nodes)) | |
1514 | sched_setnuma(p, env.dst_nid); | |
1515 | } | |
1516 | ||
1517 | /* No better CPU than the current one was found. */ | |
1518 | if (env.best_cpu == -1) | |
1519 | return -EAGAIN; | |
0ec8aa00 | 1520 | |
04bb2f94 RR |
1521 | /* |
1522 | * Reset the scan period if the task is being rescheduled on an | |
1523 | * alternative node to recheck if the tasks is now properly placed. | |
1524 | */ | |
1525 | p->numa_scan_period = task_scan_min(p); | |
1526 | ||
fb13c7ee | 1527 | if (env.best_task == NULL) { |
286549dc MG |
1528 | ret = migrate_task_to(p, env.best_cpu); |
1529 | if (ret != 0) | |
1530 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1531 | return ret; |
1532 | } | |
1533 | ||
1534 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1535 | if (ret != 0) |
1536 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1537 | put_task_struct(env.best_task); |
1538 | return ret; | |
e6628d5b MG |
1539 | } |
1540 | ||
6b9a7460 MG |
1541 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1542 | static void numa_migrate_preferred(struct task_struct *p) | |
1543 | { | |
5085e2a3 RR |
1544 | unsigned long interval = HZ; |
1545 | ||
2739d3ee | 1546 | /* This task has no NUMA fault statistics yet */ |
44dba3d5 | 1547 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
6b9a7460 MG |
1548 | return; |
1549 | ||
2739d3ee | 1550 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 RR |
1551 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1552 | p->numa_migrate_retry = jiffies + interval; | |
2739d3ee RR |
1553 | |
1554 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1555 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1556 | return; |
1557 | ||
1558 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1559 | task_numa_migrate(p); |
6b9a7460 MG |
1560 | } |
1561 | ||
20e07dea RR |
1562 | /* |
1563 | * Find the nodes on which the workload is actively running. We do this by | |
1564 | * tracking the nodes from which NUMA hinting faults are triggered. This can | |
1565 | * be different from the set of nodes where the workload's memory is currently | |
1566 | * located. | |
1567 | * | |
1568 | * The bitmask is used to make smarter decisions on when to do NUMA page | |
1569 | * migrations, To prevent flip-flopping, and excessive page migrations, nodes | |
1570 | * are added when they cause over 6/16 of the maximum number of faults, but | |
1571 | * only removed when they drop below 3/16. | |
1572 | */ | |
1573 | static void update_numa_active_node_mask(struct numa_group *numa_group) | |
1574 | { | |
1575 | unsigned long faults, max_faults = 0; | |
1576 | int nid; | |
1577 | ||
1578 | for_each_online_node(nid) { | |
1579 | faults = group_faults_cpu(numa_group, nid); | |
1580 | if (faults > max_faults) | |
1581 | max_faults = faults; | |
1582 | } | |
1583 | ||
1584 | for_each_online_node(nid) { | |
1585 | faults = group_faults_cpu(numa_group, nid); | |
1586 | if (!node_isset(nid, numa_group->active_nodes)) { | |
1587 | if (faults > max_faults * 6 / 16) | |
1588 | node_set(nid, numa_group->active_nodes); | |
1589 | } else if (faults < max_faults * 3 / 16) | |
1590 | node_clear(nid, numa_group->active_nodes); | |
1591 | } | |
1592 | } | |
1593 | ||
04bb2f94 RR |
1594 | /* |
1595 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1596 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1597 | * period will be for the next scan window. If local/(local+remote) ratio is |
1598 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1599 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1600 | */ |
1601 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1602 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1603 | |
1604 | /* | |
1605 | * Increase the scan period (slow down scanning) if the majority of | |
1606 | * our memory is already on our local node, or if the majority of | |
1607 | * the page accesses are shared with other processes. | |
1608 | * Otherwise, decrease the scan period. | |
1609 | */ | |
1610 | static void update_task_scan_period(struct task_struct *p, | |
1611 | unsigned long shared, unsigned long private) | |
1612 | { | |
1613 | unsigned int period_slot; | |
1614 | int ratio; | |
1615 | int diff; | |
1616 | ||
1617 | unsigned long remote = p->numa_faults_locality[0]; | |
1618 | unsigned long local = p->numa_faults_locality[1]; | |
1619 | ||
1620 | /* | |
1621 | * If there were no record hinting faults then either the task is | |
1622 | * completely idle or all activity is areas that are not of interest | |
1623 | * to automatic numa balancing. Scan slower | |
1624 | */ | |
1625 | if (local + shared == 0) { | |
1626 | p->numa_scan_period = min(p->numa_scan_period_max, | |
1627 | p->numa_scan_period << 1); | |
1628 | ||
1629 | p->mm->numa_next_scan = jiffies + | |
1630 | msecs_to_jiffies(p->numa_scan_period); | |
1631 | ||
1632 | return; | |
1633 | } | |
1634 | ||
1635 | /* | |
1636 | * Prepare to scale scan period relative to the current period. | |
1637 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1638 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1639 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1640 | */ | |
1641 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1642 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1643 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1644 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1645 | if (!slot) | |
1646 | slot = 1; | |
1647 | diff = slot * period_slot; | |
1648 | } else { | |
1649 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1650 | ||
1651 | /* | |
1652 | * Scale scan rate increases based on sharing. There is an | |
1653 | * inverse relationship between the degree of sharing and | |
1654 | * the adjustment made to the scanning period. Broadly | |
1655 | * speaking the intent is that there is little point | |
1656 | * scanning faster if shared accesses dominate as it may | |
1657 | * simply bounce migrations uselessly | |
1658 | */ | |
2847c90e | 1659 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); |
04bb2f94 RR |
1660 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; |
1661 | } | |
1662 | ||
1663 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1664 | task_scan_min(p), task_scan_max(p)); | |
1665 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1666 | } | |
1667 | ||
7e2703e6 RR |
1668 | /* |
1669 | * Get the fraction of time the task has been running since the last | |
1670 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1671 | * decays those on a 32ms period, which is orders of magnitude off | |
1672 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1673 | * stats only if the task is so new there are no NUMA statistics yet. | |
1674 | */ | |
1675 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1676 | { | |
1677 | u64 runtime, delta, now; | |
1678 | /* Use the start of this time slice to avoid calculations. */ | |
1679 | now = p->se.exec_start; | |
1680 | runtime = p->se.sum_exec_runtime; | |
1681 | ||
1682 | if (p->last_task_numa_placement) { | |
1683 | delta = runtime - p->last_sum_exec_runtime; | |
1684 | *period = now - p->last_task_numa_placement; | |
1685 | } else { | |
1686 | delta = p->se.avg.runnable_avg_sum; | |
1687 | *period = p->se.avg.runnable_avg_period; | |
1688 | } | |
1689 | ||
1690 | p->last_sum_exec_runtime = runtime; | |
1691 | p->last_task_numa_placement = now; | |
1692 | ||
1693 | return delta; | |
1694 | } | |
1695 | ||
54009416 RR |
1696 | /* |
1697 | * Determine the preferred nid for a task in a numa_group. This needs to | |
1698 | * be done in a way that produces consistent results with group_weight, | |
1699 | * otherwise workloads might not converge. | |
1700 | */ | |
1701 | static int preferred_group_nid(struct task_struct *p, int nid) | |
1702 | { | |
1703 | nodemask_t nodes; | |
1704 | int dist; | |
1705 | ||
1706 | /* Direct connections between all NUMA nodes. */ | |
1707 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1708 | return nid; | |
1709 | ||
1710 | /* | |
1711 | * On a system with glueless mesh NUMA topology, group_weight | |
1712 | * scores nodes according to the number of NUMA hinting faults on | |
1713 | * both the node itself, and on nearby nodes. | |
1714 | */ | |
1715 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1716 | unsigned long score, max_score = 0; | |
1717 | int node, max_node = nid; | |
1718 | ||
1719 | dist = sched_max_numa_distance; | |
1720 | ||
1721 | for_each_online_node(node) { | |
1722 | score = group_weight(p, node, dist); | |
1723 | if (score > max_score) { | |
1724 | max_score = score; | |
1725 | max_node = node; | |
1726 | } | |
1727 | } | |
1728 | return max_node; | |
1729 | } | |
1730 | ||
1731 | /* | |
1732 | * Finding the preferred nid in a system with NUMA backplane | |
1733 | * interconnect topology is more involved. The goal is to locate | |
1734 | * tasks from numa_groups near each other in the system, and | |
1735 | * untangle workloads from different sides of the system. This requires | |
1736 | * searching down the hierarchy of node groups, recursively searching | |
1737 | * inside the highest scoring group of nodes. The nodemask tricks | |
1738 | * keep the complexity of the search down. | |
1739 | */ | |
1740 | nodes = node_online_map; | |
1741 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
1742 | unsigned long max_faults = 0; | |
81907478 | 1743 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
1744 | int a, b; |
1745 | ||
1746 | /* Are there nodes at this distance from each other? */ | |
1747 | if (!find_numa_distance(dist)) | |
1748 | continue; | |
1749 | ||
1750 | for_each_node_mask(a, nodes) { | |
1751 | unsigned long faults = 0; | |
1752 | nodemask_t this_group; | |
1753 | nodes_clear(this_group); | |
1754 | ||
1755 | /* Sum group's NUMA faults; includes a==b case. */ | |
1756 | for_each_node_mask(b, nodes) { | |
1757 | if (node_distance(a, b) < dist) { | |
1758 | faults += group_faults(p, b); | |
1759 | node_set(b, this_group); | |
1760 | node_clear(b, nodes); | |
1761 | } | |
1762 | } | |
1763 | ||
1764 | /* Remember the top group. */ | |
1765 | if (faults > max_faults) { | |
1766 | max_faults = faults; | |
1767 | max_group = this_group; | |
1768 | /* | |
1769 | * subtle: at the smallest distance there is | |
1770 | * just one node left in each "group", the | |
1771 | * winner is the preferred nid. | |
1772 | */ | |
1773 | nid = a; | |
1774 | } | |
1775 | } | |
1776 | /* Next round, evaluate the nodes within max_group. */ | |
1777 | nodes = max_group; | |
1778 | } | |
1779 | return nid; | |
1780 | } | |
1781 | ||
cbee9f88 PZ |
1782 | static void task_numa_placement(struct task_struct *p) |
1783 | { | |
83e1d2cd MG |
1784 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1785 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 1786 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
1787 | unsigned long total_faults; |
1788 | u64 runtime, period; | |
7dbd13ed | 1789 | spinlock_t *group_lock = NULL; |
cbee9f88 | 1790 | |
2832bc19 | 1791 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1792 | if (p->numa_scan_seq == seq) |
1793 | return; | |
1794 | p->numa_scan_seq = seq; | |
598f0ec0 | 1795 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1796 | |
7e2703e6 RR |
1797 | total_faults = p->numa_faults_locality[0] + |
1798 | p->numa_faults_locality[1]; | |
1799 | runtime = numa_get_avg_runtime(p, &period); | |
1800 | ||
7dbd13ed MG |
1801 | /* If the task is part of a group prevent parallel updates to group stats */ |
1802 | if (p->numa_group) { | |
1803 | group_lock = &p->numa_group->lock; | |
60e69eed | 1804 | spin_lock_irq(group_lock); |
7dbd13ed MG |
1805 | } |
1806 | ||
688b7585 MG |
1807 | /* Find the node with the highest number of faults */ |
1808 | for_each_online_node(nid) { | |
44dba3d5 IM |
1809 | /* Keep track of the offsets in numa_faults array */ |
1810 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 1811 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 1812 | int priv; |
745d6147 | 1813 | |
be1e4e76 | 1814 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 1815 | long diff, f_diff, f_weight; |
8c8a743c | 1816 | |
44dba3d5 IM |
1817 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
1818 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
1819 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
1820 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 1821 | |
ac8e895b | 1822 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
1823 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
1824 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
1825 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 1826 | |
7e2703e6 RR |
1827 | /* |
1828 | * Normalize the faults_from, so all tasks in a group | |
1829 | * count according to CPU use, instead of by the raw | |
1830 | * number of faults. Tasks with little runtime have | |
1831 | * little over-all impact on throughput, and thus their | |
1832 | * faults are less important. | |
1833 | */ | |
1834 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 1835 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 1836 | (total_faults + 1); |
44dba3d5 IM |
1837 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
1838 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 1839 | |
44dba3d5 IM |
1840 | p->numa_faults[mem_idx] += diff; |
1841 | p->numa_faults[cpu_idx] += f_diff; | |
1842 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 1843 | p->total_numa_faults += diff; |
8c8a743c | 1844 | if (p->numa_group) { |
44dba3d5 IM |
1845 | /* |
1846 | * safe because we can only change our own group | |
1847 | * | |
1848 | * mem_idx represents the offset for a given | |
1849 | * nid and priv in a specific region because it | |
1850 | * is at the beginning of the numa_faults array. | |
1851 | */ | |
1852 | p->numa_group->faults[mem_idx] += diff; | |
1853 | p->numa_group->faults_cpu[mem_idx] += f_diff; | |
989348b5 | 1854 | p->numa_group->total_faults += diff; |
44dba3d5 | 1855 | group_faults += p->numa_group->faults[mem_idx]; |
8c8a743c | 1856 | } |
ac8e895b MG |
1857 | } |
1858 | ||
688b7585 MG |
1859 | if (faults > max_faults) { |
1860 | max_faults = faults; | |
1861 | max_nid = nid; | |
1862 | } | |
83e1d2cd MG |
1863 | |
1864 | if (group_faults > max_group_faults) { | |
1865 | max_group_faults = group_faults; | |
1866 | max_group_nid = nid; | |
1867 | } | |
1868 | } | |
1869 | ||
04bb2f94 RR |
1870 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1871 | ||
7dbd13ed | 1872 | if (p->numa_group) { |
20e07dea | 1873 | update_numa_active_node_mask(p->numa_group); |
60e69eed | 1874 | spin_unlock_irq(group_lock); |
54009416 | 1875 | max_nid = preferred_group_nid(p, max_group_nid); |
688b7585 MG |
1876 | } |
1877 | ||
bb97fc31 RR |
1878 | if (max_faults) { |
1879 | /* Set the new preferred node */ | |
1880 | if (max_nid != p->numa_preferred_nid) | |
1881 | sched_setnuma(p, max_nid); | |
1882 | ||
1883 | if (task_node(p) != p->numa_preferred_nid) | |
1884 | numa_migrate_preferred(p); | |
3a7053b3 | 1885 | } |
cbee9f88 PZ |
1886 | } |
1887 | ||
8c8a743c PZ |
1888 | static inline int get_numa_group(struct numa_group *grp) |
1889 | { | |
1890 | return atomic_inc_not_zero(&grp->refcount); | |
1891 | } | |
1892 | ||
1893 | static inline void put_numa_group(struct numa_group *grp) | |
1894 | { | |
1895 | if (atomic_dec_and_test(&grp->refcount)) | |
1896 | kfree_rcu(grp, rcu); | |
1897 | } | |
1898 | ||
3e6a9418 MG |
1899 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1900 | int *priv) | |
8c8a743c PZ |
1901 | { |
1902 | struct numa_group *grp, *my_grp; | |
1903 | struct task_struct *tsk; | |
1904 | bool join = false; | |
1905 | int cpu = cpupid_to_cpu(cpupid); | |
1906 | int i; | |
1907 | ||
1908 | if (unlikely(!p->numa_group)) { | |
1909 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 1910 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
1911 | |
1912 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1913 | if (!grp) | |
1914 | return; | |
1915 | ||
1916 | atomic_set(&grp->refcount, 1); | |
1917 | spin_lock_init(&grp->lock); | |
e29cf08b | 1918 | grp->gid = p->pid; |
50ec8a40 | 1919 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
1920 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
1921 | nr_node_ids; | |
8c8a743c | 1922 | |
20e07dea RR |
1923 | node_set(task_node(current), grp->active_nodes); |
1924 | ||
be1e4e76 | 1925 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 1926 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 1927 | |
989348b5 | 1928 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 1929 | |
8c8a743c PZ |
1930 | grp->nr_tasks++; |
1931 | rcu_assign_pointer(p->numa_group, grp); | |
1932 | } | |
1933 | ||
1934 | rcu_read_lock(); | |
1935 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | |
1936 | ||
1937 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 1938 | goto no_join; |
8c8a743c PZ |
1939 | |
1940 | grp = rcu_dereference(tsk->numa_group); | |
1941 | if (!grp) | |
3354781a | 1942 | goto no_join; |
8c8a743c PZ |
1943 | |
1944 | my_grp = p->numa_group; | |
1945 | if (grp == my_grp) | |
3354781a | 1946 | goto no_join; |
8c8a743c PZ |
1947 | |
1948 | /* | |
1949 | * Only join the other group if its bigger; if we're the bigger group, | |
1950 | * the other task will join us. | |
1951 | */ | |
1952 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 1953 | goto no_join; |
8c8a743c PZ |
1954 | |
1955 | /* | |
1956 | * Tie-break on the grp address. | |
1957 | */ | |
1958 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 1959 | goto no_join; |
8c8a743c | 1960 | |
dabe1d99 RR |
1961 | /* Always join threads in the same process. */ |
1962 | if (tsk->mm == current->mm) | |
1963 | join = true; | |
1964 | ||
1965 | /* Simple filter to avoid false positives due to PID collisions */ | |
1966 | if (flags & TNF_SHARED) | |
1967 | join = true; | |
8c8a743c | 1968 | |
3e6a9418 MG |
1969 | /* Update priv based on whether false sharing was detected */ |
1970 | *priv = !join; | |
1971 | ||
dabe1d99 | 1972 | if (join && !get_numa_group(grp)) |
3354781a | 1973 | goto no_join; |
8c8a743c | 1974 | |
8c8a743c PZ |
1975 | rcu_read_unlock(); |
1976 | ||
1977 | if (!join) | |
1978 | return; | |
1979 | ||
60e69eed MG |
1980 | BUG_ON(irqs_disabled()); |
1981 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 1982 | |
be1e4e76 | 1983 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
1984 | my_grp->faults[i] -= p->numa_faults[i]; |
1985 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 1986 | } |
989348b5 MG |
1987 | my_grp->total_faults -= p->total_numa_faults; |
1988 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 1989 | |
8c8a743c PZ |
1990 | my_grp->nr_tasks--; |
1991 | grp->nr_tasks++; | |
1992 | ||
1993 | spin_unlock(&my_grp->lock); | |
60e69eed | 1994 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
1995 | |
1996 | rcu_assign_pointer(p->numa_group, grp); | |
1997 | ||
1998 | put_numa_group(my_grp); | |
3354781a PZ |
1999 | return; |
2000 | ||
2001 | no_join: | |
2002 | rcu_read_unlock(); | |
2003 | return; | |
8c8a743c PZ |
2004 | } |
2005 | ||
2006 | void task_numa_free(struct task_struct *p) | |
2007 | { | |
2008 | struct numa_group *grp = p->numa_group; | |
44dba3d5 | 2009 | void *numa_faults = p->numa_faults; |
e9dd685c SR |
2010 | unsigned long flags; |
2011 | int i; | |
8c8a743c PZ |
2012 | |
2013 | if (grp) { | |
e9dd685c | 2014 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2015 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2016 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2017 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2018 | |
8c8a743c | 2019 | grp->nr_tasks--; |
e9dd685c | 2020 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2021 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2022 | put_numa_group(grp); |
2023 | } | |
2024 | ||
44dba3d5 | 2025 | p->numa_faults = NULL; |
82727018 | 2026 | kfree(numa_faults); |
8c8a743c PZ |
2027 | } |
2028 | ||
cbee9f88 PZ |
2029 | /* |
2030 | * Got a PROT_NONE fault for a page on @node. | |
2031 | */ | |
58b46da3 | 2032 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2033 | { |
2034 | struct task_struct *p = current; | |
6688cc05 | 2035 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2036 | int cpu_node = task_node(current); |
792568ec | 2037 | int local = !!(flags & TNF_FAULT_LOCAL); |
ac8e895b | 2038 | int priv; |
cbee9f88 | 2039 | |
10e84b97 | 2040 | if (!numabalancing_enabled) |
1a687c2e MG |
2041 | return; |
2042 | ||
9ff1d9ff MG |
2043 | /* for example, ksmd faulting in a user's mm */ |
2044 | if (!p->mm) | |
2045 | return; | |
2046 | ||
f809ca9a | 2047 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2048 | if (unlikely(!p->numa_faults)) { |
2049 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2050 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2051 | |
44dba3d5 IM |
2052 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2053 | if (!p->numa_faults) | |
f809ca9a | 2054 | return; |
745d6147 | 2055 | |
83e1d2cd | 2056 | p->total_numa_faults = 0; |
04bb2f94 | 2057 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2058 | } |
cbee9f88 | 2059 | |
8c8a743c PZ |
2060 | /* |
2061 | * First accesses are treated as private, otherwise consider accesses | |
2062 | * to be private if the accessing pid has not changed | |
2063 | */ | |
2064 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2065 | priv = 1; | |
2066 | } else { | |
2067 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2068 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2069 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2070 | } |
2071 | ||
792568ec RR |
2072 | /* |
2073 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2074 | * occurs wholly within the set of nodes that the workload is | |
2075 | * actively using should be counted as local. This allows the | |
2076 | * scan rate to slow down when a workload has settled down. | |
2077 | */ | |
2078 | if (!priv && !local && p->numa_group && | |
2079 | node_isset(cpu_node, p->numa_group->active_nodes) && | |
2080 | node_isset(mem_node, p->numa_group->active_nodes)) | |
2081 | local = 1; | |
2082 | ||
cbee9f88 | 2083 | task_numa_placement(p); |
f809ca9a | 2084 | |
2739d3ee RR |
2085 | /* |
2086 | * Retry task to preferred node migration periodically, in case it | |
2087 | * case it previously failed, or the scheduler moved us. | |
2088 | */ | |
2089 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
2090 | numa_migrate_preferred(p); |
2091 | ||
b32e86b4 IM |
2092 | if (migrated) |
2093 | p->numa_pages_migrated += pages; | |
2094 | ||
44dba3d5 IM |
2095 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2096 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2097 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2098 | } |
2099 | ||
6e5fb223 PZ |
2100 | static void reset_ptenuma_scan(struct task_struct *p) |
2101 | { | |
2102 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | |
2103 | p->mm->numa_scan_offset = 0; | |
2104 | } | |
2105 | ||
cbee9f88 PZ |
2106 | /* |
2107 | * The expensive part of numa migration is done from task_work context. | |
2108 | * Triggered from task_tick_numa(). | |
2109 | */ | |
2110 | void task_numa_work(struct callback_head *work) | |
2111 | { | |
2112 | unsigned long migrate, next_scan, now = jiffies; | |
2113 | struct task_struct *p = current; | |
2114 | struct mm_struct *mm = p->mm; | |
6e5fb223 | 2115 | struct vm_area_struct *vma; |
9f40604c | 2116 | unsigned long start, end; |
598f0ec0 | 2117 | unsigned long nr_pte_updates = 0; |
9f40604c | 2118 | long pages; |
cbee9f88 PZ |
2119 | |
2120 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
2121 | ||
2122 | work->next = work; /* protect against double add */ | |
2123 | /* | |
2124 | * Who cares about NUMA placement when they're dying. | |
2125 | * | |
2126 | * NOTE: make sure not to dereference p->mm before this check, | |
2127 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2128 | * without p->mm even though we still had it when we enqueued this | |
2129 | * work. | |
2130 | */ | |
2131 | if (p->flags & PF_EXITING) | |
2132 | return; | |
2133 | ||
930aa174 | 2134 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2135 | mm->numa_next_scan = now + |
2136 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2137 | } |
2138 | ||
cbee9f88 PZ |
2139 | /* |
2140 | * Enforce maximal scan/migration frequency.. | |
2141 | */ | |
2142 | migrate = mm->numa_next_scan; | |
2143 | if (time_before(now, migrate)) | |
2144 | return; | |
2145 | ||
598f0ec0 MG |
2146 | if (p->numa_scan_period == 0) { |
2147 | p->numa_scan_period_max = task_scan_max(p); | |
2148 | p->numa_scan_period = task_scan_min(p); | |
2149 | } | |
cbee9f88 | 2150 | |
fb003b80 | 2151 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2152 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2153 | return; | |
2154 | ||
19a78d11 PZ |
2155 | /* |
2156 | * Delay this task enough that another task of this mm will likely win | |
2157 | * the next time around. | |
2158 | */ | |
2159 | p->node_stamp += 2 * TICK_NSEC; | |
2160 | ||
9f40604c MG |
2161 | start = mm->numa_scan_offset; |
2162 | pages = sysctl_numa_balancing_scan_size; | |
2163 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
2164 | if (!pages) | |
2165 | return; | |
cbee9f88 | 2166 | |
6e5fb223 | 2167 | down_read(&mm->mmap_sem); |
9f40604c | 2168 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2169 | if (!vma) { |
2170 | reset_ptenuma_scan(p); | |
9f40604c | 2171 | start = 0; |
6e5fb223 PZ |
2172 | vma = mm->mmap; |
2173 | } | |
9f40604c | 2174 | for (; vma; vma = vma->vm_next) { |
6b6482bb | 2175 | if (!vma_migratable(vma) || !vma_policy_mof(vma)) |
6e5fb223 PZ |
2176 | continue; |
2177 | ||
4591ce4f MG |
2178 | /* |
2179 | * Shared library pages mapped by multiple processes are not | |
2180 | * migrated as it is expected they are cache replicated. Avoid | |
2181 | * hinting faults in read-only file-backed mappings or the vdso | |
2182 | * as migrating the pages will be of marginal benefit. | |
2183 | */ | |
2184 | if (!vma->vm_mm || | |
2185 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2186 | continue; | |
2187 | ||
3c67f474 MG |
2188 | /* |
2189 | * Skip inaccessible VMAs to avoid any confusion between | |
2190 | * PROT_NONE and NUMA hinting ptes | |
2191 | */ | |
2192 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2193 | continue; | |
4591ce4f | 2194 | |
9f40604c MG |
2195 | do { |
2196 | start = max(start, vma->vm_start); | |
2197 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2198 | end = min(end, vma->vm_end); | |
598f0ec0 MG |
2199 | nr_pte_updates += change_prot_numa(vma, start, end); |
2200 | ||
2201 | /* | |
2202 | * Scan sysctl_numa_balancing_scan_size but ensure that | |
2203 | * at least one PTE is updated so that unused virtual | |
2204 | * address space is quickly skipped. | |
2205 | */ | |
2206 | if (nr_pte_updates) | |
2207 | pages -= (end - start) >> PAGE_SHIFT; | |
6e5fb223 | 2208 | |
9f40604c MG |
2209 | start = end; |
2210 | if (pages <= 0) | |
2211 | goto out; | |
3cf1962c RR |
2212 | |
2213 | cond_resched(); | |
9f40604c | 2214 | } while (end != vma->vm_end); |
cbee9f88 | 2215 | } |
6e5fb223 | 2216 | |
9f40604c | 2217 | out: |
6e5fb223 | 2218 | /* |
c69307d5 PZ |
2219 | * It is possible to reach the end of the VMA list but the last few |
2220 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2221 | * would find the !migratable VMA on the next scan but not reset the | |
2222 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2223 | */ |
2224 | if (vma) | |
9f40604c | 2225 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2226 | else |
2227 | reset_ptenuma_scan(p); | |
2228 | up_read(&mm->mmap_sem); | |
cbee9f88 PZ |
2229 | } |
2230 | ||
2231 | /* | |
2232 | * Drive the periodic memory faults.. | |
2233 | */ | |
2234 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2235 | { | |
2236 | struct callback_head *work = &curr->numa_work; | |
2237 | u64 period, now; | |
2238 | ||
2239 | /* | |
2240 | * We don't care about NUMA placement if we don't have memory. | |
2241 | */ | |
2242 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2243 | return; | |
2244 | ||
2245 | /* | |
2246 | * Using runtime rather than walltime has the dual advantage that | |
2247 | * we (mostly) drive the selection from busy threads and that the | |
2248 | * task needs to have done some actual work before we bother with | |
2249 | * NUMA placement. | |
2250 | */ | |
2251 | now = curr->se.sum_exec_runtime; | |
2252 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2253 | ||
2254 | if (now - curr->node_stamp > period) { | |
4b96a29b | 2255 | if (!curr->node_stamp) |
598f0ec0 | 2256 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 2257 | curr->node_stamp += period; |
cbee9f88 PZ |
2258 | |
2259 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2260 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2261 | task_work_add(curr, work, true); | |
2262 | } | |
2263 | } | |
2264 | } | |
2265 | #else | |
2266 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2267 | { | |
2268 | } | |
0ec8aa00 PZ |
2269 | |
2270 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2271 | { | |
2272 | } | |
2273 | ||
2274 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2275 | { | |
2276 | } | |
cbee9f88 PZ |
2277 | #endif /* CONFIG_NUMA_BALANCING */ |
2278 | ||
30cfdcfc DA |
2279 | static void |
2280 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2281 | { | |
2282 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2283 | if (!parent_entity(se)) |
029632fb | 2284 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2285 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2286 | if (entity_is_task(se)) { |
2287 | struct rq *rq = rq_of(cfs_rq); | |
2288 | ||
2289 | account_numa_enqueue(rq, task_of(se)); | |
2290 | list_add(&se->group_node, &rq->cfs_tasks); | |
2291 | } | |
367456c7 | 2292 | #endif |
30cfdcfc | 2293 | cfs_rq->nr_running++; |
30cfdcfc DA |
2294 | } |
2295 | ||
2296 | static void | |
2297 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2298 | { | |
2299 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2300 | if (!parent_entity(se)) |
029632fb | 2301 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
0ec8aa00 PZ |
2302 | if (entity_is_task(se)) { |
2303 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2304 | list_del_init(&se->group_node); |
0ec8aa00 | 2305 | } |
30cfdcfc | 2306 | cfs_rq->nr_running--; |
30cfdcfc DA |
2307 | } |
2308 | ||
3ff6dcac YZ |
2309 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2310 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
2311 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
2312 | { | |
2313 | long tg_weight; | |
2314 | ||
2315 | /* | |
2316 | * Use this CPU's actual weight instead of the last load_contribution | |
2317 | * to gain a more accurate current total weight. See | |
2318 | * update_cfs_rq_load_contribution(). | |
2319 | */ | |
bf5b986e | 2320 | tg_weight = atomic_long_read(&tg->load_avg); |
82958366 | 2321 | tg_weight -= cfs_rq->tg_load_contrib; |
cf5f0acf PZ |
2322 | tg_weight += cfs_rq->load.weight; |
2323 | ||
2324 | return tg_weight; | |
2325 | } | |
2326 | ||
6d5ab293 | 2327 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 2328 | { |
cf5f0acf | 2329 | long tg_weight, load, shares; |
3ff6dcac | 2330 | |
cf5f0acf | 2331 | tg_weight = calc_tg_weight(tg, cfs_rq); |
6d5ab293 | 2332 | load = cfs_rq->load.weight; |
3ff6dcac | 2333 | |
3ff6dcac | 2334 | shares = (tg->shares * load); |
cf5f0acf PZ |
2335 | if (tg_weight) |
2336 | shares /= tg_weight; | |
3ff6dcac YZ |
2337 | |
2338 | if (shares < MIN_SHARES) | |
2339 | shares = MIN_SHARES; | |
2340 | if (shares > tg->shares) | |
2341 | shares = tg->shares; | |
2342 | ||
2343 | return shares; | |
2344 | } | |
3ff6dcac | 2345 | # else /* CONFIG_SMP */ |
6d5ab293 | 2346 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
2347 | { |
2348 | return tg->shares; | |
2349 | } | |
3ff6dcac | 2350 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
2351 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2352 | unsigned long weight) | |
2353 | { | |
19e5eebb PT |
2354 | if (se->on_rq) { |
2355 | /* commit outstanding execution time */ | |
2356 | if (cfs_rq->curr == se) | |
2357 | update_curr(cfs_rq); | |
2069dd75 | 2358 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 2359 | } |
2069dd75 PZ |
2360 | |
2361 | update_load_set(&se->load, weight); | |
2362 | ||
2363 | if (se->on_rq) | |
2364 | account_entity_enqueue(cfs_rq, se); | |
2365 | } | |
2366 | ||
82958366 PT |
2367 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2368 | ||
6d5ab293 | 2369 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2370 | { |
2371 | struct task_group *tg; | |
2372 | struct sched_entity *se; | |
3ff6dcac | 2373 | long shares; |
2069dd75 | 2374 | |
2069dd75 PZ |
2375 | tg = cfs_rq->tg; |
2376 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 2377 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 2378 | return; |
3ff6dcac YZ |
2379 | #ifndef CONFIG_SMP |
2380 | if (likely(se->load.weight == tg->shares)) | |
2381 | return; | |
2382 | #endif | |
6d5ab293 | 2383 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
2384 | |
2385 | reweight_entity(cfs_rq_of(se), se, shares); | |
2386 | } | |
2387 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 2388 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2389 | { |
2390 | } | |
2391 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
2392 | ||
141965c7 | 2393 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
2394 | /* |
2395 | * We choose a half-life close to 1 scheduling period. | |
2396 | * Note: The tables below are dependent on this value. | |
2397 | */ | |
2398 | #define LOAD_AVG_PERIOD 32 | |
2399 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
2400 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | |
2401 | ||
2402 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | |
2403 | static const u32 runnable_avg_yN_inv[] = { | |
2404 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
2405 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
2406 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
2407 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
2408 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
2409 | 0x85aac367, 0x82cd8698, | |
2410 | }; | |
2411 | ||
2412 | /* | |
2413 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
2414 | * over-estimates when re-combining. | |
2415 | */ | |
2416 | static const u32 runnable_avg_yN_sum[] = { | |
2417 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
2418 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
2419 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
2420 | }; | |
2421 | ||
9d85f21c PT |
2422 | /* |
2423 | * Approximate: | |
2424 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
2425 | */ | |
2426 | static __always_inline u64 decay_load(u64 val, u64 n) | |
2427 | { | |
5b51f2f8 PT |
2428 | unsigned int local_n; |
2429 | ||
2430 | if (!n) | |
2431 | return val; | |
2432 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
2433 | return 0; | |
2434 | ||
2435 | /* after bounds checking we can collapse to 32-bit */ | |
2436 | local_n = n; | |
2437 | ||
2438 | /* | |
2439 | * As y^PERIOD = 1/2, we can combine | |
9c58c79a ZZ |
2440 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) |
2441 | * With a look-up table which covers y^n (n<PERIOD) | |
5b51f2f8 PT |
2442 | * |
2443 | * To achieve constant time decay_load. | |
2444 | */ | |
2445 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2446 | val >>= local_n / LOAD_AVG_PERIOD; | |
2447 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2448 | } |
2449 | ||
5b51f2f8 PT |
2450 | val *= runnable_avg_yN_inv[local_n]; |
2451 | /* We don't use SRR here since we always want to round down. */ | |
2452 | return val >> 32; | |
2453 | } | |
2454 | ||
2455 | /* | |
2456 | * For updates fully spanning n periods, the contribution to runnable | |
2457 | * average will be: \Sum 1024*y^n | |
2458 | * | |
2459 | * We can compute this reasonably efficiently by combining: | |
2460 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2461 | */ | |
2462 | static u32 __compute_runnable_contrib(u64 n) | |
2463 | { | |
2464 | u32 contrib = 0; | |
2465 | ||
2466 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2467 | return runnable_avg_yN_sum[n]; | |
2468 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2469 | return LOAD_AVG_MAX; | |
2470 | ||
2471 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
2472 | do { | |
2473 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
2474 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
2475 | ||
2476 | n -= LOAD_AVG_PERIOD; | |
2477 | } while (n > LOAD_AVG_PERIOD); | |
2478 | ||
2479 | contrib = decay_load(contrib, n); | |
2480 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2481 | } |
2482 | ||
2483 | /* | |
2484 | * We can represent the historical contribution to runnable average as the | |
2485 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2486 | * history into segments of approximately 1ms (1024us); label the segment that | |
2487 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2488 | * | |
2489 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2490 | * p0 p1 p2 | |
2491 | * (now) (~1ms ago) (~2ms ago) | |
2492 | * | |
2493 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2494 | * | |
2495 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2496 | * following representation of historical load: | |
2497 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2498 | * | |
2499 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2500 | * y^32 = 0.5 | |
2501 | * | |
2502 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2503 | * approximately half as much as the contribution to load within the last ms | |
2504 | * (u_0). | |
2505 | * | |
2506 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2507 | * sum again by y is sufficient to update: | |
2508 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2509 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2510 | */ | |
2511 | static __always_inline int __update_entity_runnable_avg(u64 now, | |
2512 | struct sched_avg *sa, | |
2513 | int runnable) | |
2514 | { | |
5b51f2f8 PT |
2515 | u64 delta, periods; |
2516 | u32 runnable_contrib; | |
9d85f21c PT |
2517 | int delta_w, decayed = 0; |
2518 | ||
2519 | delta = now - sa->last_runnable_update; | |
2520 | /* | |
2521 | * This should only happen when time goes backwards, which it | |
2522 | * unfortunately does during sched clock init when we swap over to TSC. | |
2523 | */ | |
2524 | if ((s64)delta < 0) { | |
2525 | sa->last_runnable_update = now; | |
2526 | return 0; | |
2527 | } | |
2528 | ||
2529 | /* | |
2530 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2531 | * approximation of 1us and fast to compute. | |
2532 | */ | |
2533 | delta >>= 10; | |
2534 | if (!delta) | |
2535 | return 0; | |
2536 | sa->last_runnable_update = now; | |
2537 | ||
2538 | /* delta_w is the amount already accumulated against our next period */ | |
2539 | delta_w = sa->runnable_avg_period % 1024; | |
2540 | if (delta + delta_w >= 1024) { | |
2541 | /* period roll-over */ | |
2542 | decayed = 1; | |
2543 | ||
2544 | /* | |
2545 | * Now that we know we're crossing a period boundary, figure | |
2546 | * out how much from delta we need to complete the current | |
2547 | * period and accrue it. | |
2548 | */ | |
2549 | delta_w = 1024 - delta_w; | |
5b51f2f8 PT |
2550 | if (runnable) |
2551 | sa->runnable_avg_sum += delta_w; | |
2552 | sa->runnable_avg_period += delta_w; | |
2553 | ||
2554 | delta -= delta_w; | |
2555 | ||
2556 | /* Figure out how many additional periods this update spans */ | |
2557 | periods = delta / 1024; | |
2558 | delta %= 1024; | |
2559 | ||
2560 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | |
2561 | periods + 1); | |
2562 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | |
2563 | periods + 1); | |
2564 | ||
2565 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
2566 | runnable_contrib = __compute_runnable_contrib(periods); | |
2567 | if (runnable) | |
2568 | sa->runnable_avg_sum += runnable_contrib; | |
2569 | sa->runnable_avg_period += runnable_contrib; | |
9d85f21c PT |
2570 | } |
2571 | ||
2572 | /* Remainder of delta accrued against u_0` */ | |
2573 | if (runnable) | |
2574 | sa->runnable_avg_sum += delta; | |
2575 | sa->runnable_avg_period += delta; | |
2576 | ||
2577 | return decayed; | |
2578 | } | |
2579 | ||
9ee474f5 | 2580 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
aff3e498 | 2581 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
9ee474f5 PT |
2582 | { |
2583 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2584 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | |
2585 | ||
2586 | decays -= se->avg.decay_count; | |
63847600 | 2587 | se->avg.decay_count = 0; |
9ee474f5 | 2588 | if (!decays) |
aff3e498 | 2589 | return 0; |
9ee474f5 PT |
2590 | |
2591 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | |
aff3e498 PT |
2592 | |
2593 | return decays; | |
9ee474f5 PT |
2594 | } |
2595 | ||
c566e8e9 PT |
2596 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2597 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | |
2598 | int force_update) | |
2599 | { | |
2600 | struct task_group *tg = cfs_rq->tg; | |
bf5b986e | 2601 | long tg_contrib; |
c566e8e9 PT |
2602 | |
2603 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | |
2604 | tg_contrib -= cfs_rq->tg_load_contrib; | |
2605 | ||
8236d907 JL |
2606 | if (!tg_contrib) |
2607 | return; | |
2608 | ||
bf5b986e AS |
2609 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
2610 | atomic_long_add(tg_contrib, &tg->load_avg); | |
c566e8e9 PT |
2611 | cfs_rq->tg_load_contrib += tg_contrib; |
2612 | } | |
2613 | } | |
8165e145 | 2614 | |
bb17f655 PT |
2615 | /* |
2616 | * Aggregate cfs_rq runnable averages into an equivalent task_group | |
2617 | * representation for computing load contributions. | |
2618 | */ | |
2619 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | |
2620 | struct cfs_rq *cfs_rq) | |
2621 | { | |
2622 | struct task_group *tg = cfs_rq->tg; | |
2623 | long contrib; | |
2624 | ||
2625 | /* The fraction of a cpu used by this cfs_rq */ | |
85b088e9 | 2626 | contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, |
bb17f655 PT |
2627 | sa->runnable_avg_period + 1); |
2628 | contrib -= cfs_rq->tg_runnable_contrib; | |
2629 | ||
2630 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | |
2631 | atomic_add(contrib, &tg->runnable_avg); | |
2632 | cfs_rq->tg_runnable_contrib += contrib; | |
2633 | } | |
2634 | } | |
2635 | ||
8165e145 PT |
2636 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
2637 | { | |
2638 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
2639 | struct task_group *tg = cfs_rq->tg; | |
bb17f655 PT |
2640 | int runnable_avg; |
2641 | ||
8165e145 PT |
2642 | u64 contrib; |
2643 | ||
2644 | contrib = cfs_rq->tg_load_contrib * tg->shares; | |
bf5b986e AS |
2645 | se->avg.load_avg_contrib = div_u64(contrib, |
2646 | atomic_long_read(&tg->load_avg) + 1); | |
bb17f655 PT |
2647 | |
2648 | /* | |
2649 | * For group entities we need to compute a correction term in the case | |
2650 | * that they are consuming <1 cpu so that we would contribute the same | |
2651 | * load as a task of equal weight. | |
2652 | * | |
2653 | * Explicitly co-ordinating this measurement would be expensive, but | |
2654 | * fortunately the sum of each cpus contribution forms a usable | |
2655 | * lower-bound on the true value. | |
2656 | * | |
2657 | * Consider the aggregate of 2 contributions. Either they are disjoint | |
2658 | * (and the sum represents true value) or they are disjoint and we are | |
2659 | * understating by the aggregate of their overlap. | |
2660 | * | |
2661 | * Extending this to N cpus, for a given overlap, the maximum amount we | |
2662 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | |
2663 | * cpus that overlap for this interval and w_i is the interval width. | |
2664 | * | |
2665 | * On a small machine; the first term is well-bounded which bounds the | |
2666 | * total error since w_i is a subset of the period. Whereas on a | |
2667 | * larger machine, while this first term can be larger, if w_i is the | |
2668 | * of consequential size guaranteed to see n_i*w_i quickly converge to | |
2669 | * our upper bound of 1-cpu. | |
2670 | */ | |
2671 | runnable_avg = atomic_read(&tg->runnable_avg); | |
2672 | if (runnable_avg < NICE_0_LOAD) { | |
2673 | se->avg.load_avg_contrib *= runnable_avg; | |
2674 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | |
2675 | } | |
8165e145 | 2676 | } |
f5f9739d DE |
2677 | |
2678 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | |
2679 | { | |
2680 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); | |
2681 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); | |
2682 | } | |
6e83125c | 2683 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 PT |
2684 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, |
2685 | int force_update) {} | |
bb17f655 PT |
2686 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2687 | struct cfs_rq *cfs_rq) {} | |
8165e145 | 2688 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
f5f9739d | 2689 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
6e83125c | 2690 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 2691 | |
8165e145 PT |
2692 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2693 | { | |
2694 | u32 contrib; | |
2695 | ||
2696 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | |
2697 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | |
2698 | contrib /= (se->avg.runnable_avg_period + 1); | |
2699 | se->avg.load_avg_contrib = scale_load(contrib); | |
2700 | } | |
2701 | ||
2dac754e PT |
2702 | /* Compute the current contribution to load_avg by se, return any delta */ |
2703 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | |
2704 | { | |
2705 | long old_contrib = se->avg.load_avg_contrib; | |
2706 | ||
8165e145 PT |
2707 | if (entity_is_task(se)) { |
2708 | __update_task_entity_contrib(se); | |
2709 | } else { | |
bb17f655 | 2710 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
8165e145 PT |
2711 | __update_group_entity_contrib(se); |
2712 | } | |
2dac754e PT |
2713 | |
2714 | return se->avg.load_avg_contrib - old_contrib; | |
2715 | } | |
2716 | ||
9ee474f5 PT |
2717 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2718 | long load_contrib) | |
2719 | { | |
2720 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | |
2721 | cfs_rq->blocked_load_avg -= load_contrib; | |
2722 | else | |
2723 | cfs_rq->blocked_load_avg = 0; | |
2724 | } | |
2725 | ||
f1b17280 PT |
2726 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2727 | ||
9d85f21c | 2728 | /* Update a sched_entity's runnable average */ |
9ee474f5 PT |
2729 | static inline void update_entity_load_avg(struct sched_entity *se, |
2730 | int update_cfs_rq) | |
9d85f21c | 2731 | { |
2dac754e PT |
2732 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2733 | long contrib_delta; | |
f1b17280 | 2734 | u64 now; |
2dac754e | 2735 | |
f1b17280 PT |
2736 | /* |
2737 | * For a group entity we need to use their owned cfs_rq_clock_task() in | |
2738 | * case they are the parent of a throttled hierarchy. | |
2739 | */ | |
2740 | if (entity_is_task(se)) | |
2741 | now = cfs_rq_clock_task(cfs_rq); | |
2742 | else | |
2743 | now = cfs_rq_clock_task(group_cfs_rq(se)); | |
2744 | ||
2745 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | |
2dac754e PT |
2746 | return; |
2747 | ||
2748 | contrib_delta = __update_entity_load_avg_contrib(se); | |
9ee474f5 PT |
2749 | |
2750 | if (!update_cfs_rq) | |
2751 | return; | |
2752 | ||
2dac754e PT |
2753 | if (se->on_rq) |
2754 | cfs_rq->runnable_load_avg += contrib_delta; | |
9ee474f5 PT |
2755 | else |
2756 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | |
2757 | } | |
2758 | ||
2759 | /* | |
2760 | * Decay the load contributed by all blocked children and account this so that | |
2761 | * their contribution may appropriately discounted when they wake up. | |
2762 | */ | |
aff3e498 | 2763 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
9ee474f5 | 2764 | { |
f1b17280 | 2765 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
9ee474f5 PT |
2766 | u64 decays; |
2767 | ||
2768 | decays = now - cfs_rq->last_decay; | |
aff3e498 | 2769 | if (!decays && !force_update) |
9ee474f5 PT |
2770 | return; |
2771 | ||
2509940f AS |
2772 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2773 | unsigned long removed_load; | |
2774 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | |
aff3e498 PT |
2775 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2776 | } | |
9ee474f5 | 2777 | |
aff3e498 PT |
2778 | if (decays) { |
2779 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | |
2780 | decays); | |
2781 | atomic64_add(decays, &cfs_rq->decay_counter); | |
2782 | cfs_rq->last_decay = now; | |
2783 | } | |
c566e8e9 PT |
2784 | |
2785 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | |
9d85f21c | 2786 | } |
18bf2805 | 2787 | |
2dac754e PT |
2788 | /* Add the load generated by se into cfs_rq's child load-average */ |
2789 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | |
9ee474f5 PT |
2790 | struct sched_entity *se, |
2791 | int wakeup) | |
2dac754e | 2792 | { |
aff3e498 PT |
2793 | /* |
2794 | * We track migrations using entity decay_count <= 0, on a wake-up | |
2795 | * migration we use a negative decay count to track the remote decays | |
2796 | * accumulated while sleeping. | |
a75cdaa9 AS |
2797 | * |
2798 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | |
2799 | * are seen by enqueue_entity_load_avg() as a migration with an already | |
2800 | * constructed load_avg_contrib. | |
aff3e498 PT |
2801 | */ |
2802 | if (unlikely(se->avg.decay_count <= 0)) { | |
78becc27 | 2803 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
aff3e498 PT |
2804 | if (se->avg.decay_count) { |
2805 | /* | |
2806 | * In a wake-up migration we have to approximate the | |
2807 | * time sleeping. This is because we can't synchronize | |
2808 | * clock_task between the two cpus, and it is not | |
2809 | * guaranteed to be read-safe. Instead, we can | |
2810 | * approximate this using our carried decays, which are | |
2811 | * explicitly atomically readable. | |
2812 | */ | |
2813 | se->avg.last_runnable_update -= (-se->avg.decay_count) | |
2814 | << 20; | |
2815 | update_entity_load_avg(se, 0); | |
2816 | /* Indicate that we're now synchronized and on-rq */ | |
2817 | se->avg.decay_count = 0; | |
2818 | } | |
9ee474f5 PT |
2819 | wakeup = 0; |
2820 | } else { | |
9390675a | 2821 | __synchronize_entity_decay(se); |
9ee474f5 PT |
2822 | } |
2823 | ||
aff3e498 PT |
2824 | /* migrated tasks did not contribute to our blocked load */ |
2825 | if (wakeup) { | |
9ee474f5 | 2826 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
aff3e498 PT |
2827 | update_entity_load_avg(se, 0); |
2828 | } | |
9ee474f5 | 2829 | |
2dac754e | 2830 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
aff3e498 PT |
2831 | /* we force update consideration on load-balancer moves */ |
2832 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | |
2dac754e PT |
2833 | } |
2834 | ||
9ee474f5 PT |
2835 | /* |
2836 | * Remove se's load from this cfs_rq child load-average, if the entity is | |
2837 | * transitioning to a blocked state we track its projected decay using | |
2838 | * blocked_load_avg. | |
2839 | */ | |
2dac754e | 2840 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2841 | struct sched_entity *se, |
2842 | int sleep) | |
2dac754e | 2843 | { |
9ee474f5 | 2844 | update_entity_load_avg(se, 1); |
aff3e498 PT |
2845 | /* we force update consideration on load-balancer moves */ |
2846 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | |
9ee474f5 | 2847 | |
2dac754e | 2848 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
9ee474f5 PT |
2849 | if (sleep) { |
2850 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
2851 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | |
2852 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | |
2dac754e | 2853 | } |
642dbc39 VG |
2854 | |
2855 | /* | |
2856 | * Update the rq's load with the elapsed running time before entering | |
2857 | * idle. if the last scheduled task is not a CFS task, idle_enter will | |
2858 | * be the only way to update the runnable statistic. | |
2859 | */ | |
2860 | void idle_enter_fair(struct rq *this_rq) | |
2861 | { | |
2862 | update_rq_runnable_avg(this_rq, 1); | |
2863 | } | |
2864 | ||
2865 | /* | |
2866 | * Update the rq's load with the elapsed idle time before a task is | |
2867 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | |
2868 | * be the only way to update the runnable statistic. | |
2869 | */ | |
2870 | void idle_exit_fair(struct rq *this_rq) | |
2871 | { | |
2872 | update_rq_runnable_avg(this_rq, 0); | |
2873 | } | |
2874 | ||
6e83125c PZ |
2875 | static int idle_balance(struct rq *this_rq); |
2876 | ||
38033c37 PZ |
2877 | #else /* CONFIG_SMP */ |
2878 | ||
9ee474f5 PT |
2879 | static inline void update_entity_load_avg(struct sched_entity *se, |
2880 | int update_cfs_rq) {} | |
18bf2805 | 2881 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2dac754e | 2882 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2883 | struct sched_entity *se, |
2884 | int wakeup) {} | |
2dac754e | 2885 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
9ee474f5 PT |
2886 | struct sched_entity *se, |
2887 | int sleep) {} | |
aff3e498 PT |
2888 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2889 | int force_update) {} | |
6e83125c PZ |
2890 | |
2891 | static inline int idle_balance(struct rq *rq) | |
2892 | { | |
2893 | return 0; | |
2894 | } | |
2895 | ||
38033c37 | 2896 | #endif /* CONFIG_SMP */ |
9d85f21c | 2897 | |
2396af69 | 2898 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2899 | { |
bf0f6f24 | 2900 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2901 | struct task_struct *tsk = NULL; |
2902 | ||
2903 | if (entity_is_task(se)) | |
2904 | tsk = task_of(se); | |
2905 | ||
41acab88 | 2906 | if (se->statistics.sleep_start) { |
78becc27 | 2907 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
2908 | |
2909 | if ((s64)delta < 0) | |
2910 | delta = 0; | |
2911 | ||
41acab88 LDM |
2912 | if (unlikely(delta > se->statistics.sleep_max)) |
2913 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 2914 | |
8c79a045 | 2915 | se->statistics.sleep_start = 0; |
41acab88 | 2916 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 2917 | |
768d0c27 | 2918 | if (tsk) { |
e414314c | 2919 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
2920 | trace_sched_stat_sleep(tsk, delta); |
2921 | } | |
bf0f6f24 | 2922 | } |
41acab88 | 2923 | if (se->statistics.block_start) { |
78becc27 | 2924 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
2925 | |
2926 | if ((s64)delta < 0) | |
2927 | delta = 0; | |
2928 | ||
41acab88 LDM |
2929 | if (unlikely(delta > se->statistics.block_max)) |
2930 | se->statistics.block_max = delta; | |
bf0f6f24 | 2931 | |
8c79a045 | 2932 | se->statistics.block_start = 0; |
41acab88 | 2933 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 2934 | |
e414314c | 2935 | if (tsk) { |
8f0dfc34 | 2936 | if (tsk->in_iowait) { |
41acab88 LDM |
2937 | se->statistics.iowait_sum += delta; |
2938 | se->statistics.iowait_count++; | |
768d0c27 | 2939 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
2940 | } |
2941 | ||
b781a602 AV |
2942 | trace_sched_stat_blocked(tsk, delta); |
2943 | ||
e414314c PZ |
2944 | /* |
2945 | * Blocking time is in units of nanosecs, so shift by | |
2946 | * 20 to get a milliseconds-range estimation of the | |
2947 | * amount of time that the task spent sleeping: | |
2948 | */ | |
2949 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
2950 | profile_hits(SLEEP_PROFILING, | |
2951 | (void *)get_wchan(tsk), | |
2952 | delta >> 20); | |
2953 | } | |
2954 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 2955 | } |
bf0f6f24 IM |
2956 | } |
2957 | #endif | |
2958 | } | |
2959 | ||
ddc97297 PZ |
2960 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2961 | { | |
2962 | #ifdef CONFIG_SCHED_DEBUG | |
2963 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
2964 | ||
2965 | if (d < 0) | |
2966 | d = -d; | |
2967 | ||
2968 | if (d > 3*sysctl_sched_latency) | |
2969 | schedstat_inc(cfs_rq, nr_spread_over); | |
2970 | #endif | |
2971 | } | |
2972 | ||
aeb73b04 PZ |
2973 | static void |
2974 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
2975 | { | |
1af5f730 | 2976 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 2977 | |
2cb8600e PZ |
2978 | /* |
2979 | * The 'current' period is already promised to the current tasks, | |
2980 | * however the extra weight of the new task will slow them down a | |
2981 | * little, place the new task so that it fits in the slot that | |
2982 | * stays open at the end. | |
2983 | */ | |
94dfb5e7 | 2984 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 2985 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 2986 | |
a2e7a7eb | 2987 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 2988 | if (!initial) { |
a2e7a7eb | 2989 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 2990 | |
a2e7a7eb MG |
2991 | /* |
2992 | * Halve their sleep time's effect, to allow | |
2993 | * for a gentler effect of sleepers: | |
2994 | */ | |
2995 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
2996 | thresh >>= 1; | |
51e0304c | 2997 | |
a2e7a7eb | 2998 | vruntime -= thresh; |
aeb73b04 PZ |
2999 | } |
3000 | ||
b5d9d734 | 3001 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3002 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3003 | } |
3004 | ||
d3d9dc33 PT |
3005 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3006 | ||
bf0f6f24 | 3007 | static void |
88ec22d3 | 3008 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3009 | { |
88ec22d3 PZ |
3010 | /* |
3011 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 3012 | * through calling update_curr(). |
88ec22d3 | 3013 | */ |
371fd7e7 | 3014 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
3015 | se->vruntime += cfs_rq->min_vruntime; |
3016 | ||
bf0f6f24 | 3017 | /* |
a2a2d680 | 3018 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3019 | */ |
b7cc0896 | 3020 | update_curr(cfs_rq); |
f269ae04 | 3021 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
17bc14b7 LT |
3022 | account_entity_enqueue(cfs_rq, se); |
3023 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 3024 | |
88ec22d3 | 3025 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 3026 | place_entity(cfs_rq, se, 0); |
2396af69 | 3027 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 3028 | } |
bf0f6f24 | 3029 | |
d2417e5a | 3030 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 3031 | check_spread(cfs_rq, se); |
83b699ed SV |
3032 | if (se != cfs_rq->curr) |
3033 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 3034 | se->on_rq = 1; |
3d4b47b4 | 3035 | |
d3d9dc33 | 3036 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 3037 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
3038 | check_enqueue_throttle(cfs_rq); |
3039 | } | |
bf0f6f24 IM |
3040 | } |
3041 | ||
2c13c919 | 3042 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 3043 | { |
2c13c919 RR |
3044 | for_each_sched_entity(se) { |
3045 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3046 | if (cfs_rq->last != se) |
2c13c919 | 3047 | break; |
f1044799 PZ |
3048 | |
3049 | cfs_rq->last = NULL; | |
2c13c919 RR |
3050 | } |
3051 | } | |
2002c695 | 3052 | |
2c13c919 RR |
3053 | static void __clear_buddies_next(struct sched_entity *se) |
3054 | { | |
3055 | for_each_sched_entity(se) { | |
3056 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3057 | if (cfs_rq->next != se) |
2c13c919 | 3058 | break; |
f1044799 PZ |
3059 | |
3060 | cfs_rq->next = NULL; | |
2c13c919 | 3061 | } |
2002c695 PZ |
3062 | } |
3063 | ||
ac53db59 RR |
3064 | static void __clear_buddies_skip(struct sched_entity *se) |
3065 | { | |
3066 | for_each_sched_entity(se) { | |
3067 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3068 | if (cfs_rq->skip != se) |
ac53db59 | 3069 | break; |
f1044799 PZ |
3070 | |
3071 | cfs_rq->skip = NULL; | |
ac53db59 RR |
3072 | } |
3073 | } | |
3074 | ||
a571bbea PZ |
3075 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3076 | { | |
2c13c919 RR |
3077 | if (cfs_rq->last == se) |
3078 | __clear_buddies_last(se); | |
3079 | ||
3080 | if (cfs_rq->next == se) | |
3081 | __clear_buddies_next(se); | |
ac53db59 RR |
3082 | |
3083 | if (cfs_rq->skip == se) | |
3084 | __clear_buddies_skip(se); | |
a571bbea PZ |
3085 | } |
3086 | ||
6c16a6dc | 3087 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 3088 | |
bf0f6f24 | 3089 | static void |
371fd7e7 | 3090 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3091 | { |
a2a2d680 DA |
3092 | /* |
3093 | * Update run-time statistics of the 'current'. | |
3094 | */ | |
3095 | update_curr(cfs_rq); | |
17bc14b7 | 3096 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
a2a2d680 | 3097 | |
19b6a2e3 | 3098 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 3099 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 3100 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
3101 | if (entity_is_task(se)) { |
3102 | struct task_struct *tsk = task_of(se); | |
3103 | ||
3104 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 3105 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 3106 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 3107 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 3108 | } |
db36cc7d | 3109 | #endif |
67e9fb2a PZ |
3110 | } |
3111 | ||
2002c695 | 3112 | clear_buddies(cfs_rq, se); |
4793241b | 3113 | |
83b699ed | 3114 | if (se != cfs_rq->curr) |
30cfdcfc | 3115 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 3116 | se->on_rq = 0; |
30cfdcfc | 3117 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
3118 | |
3119 | /* | |
3120 | * Normalize the entity after updating the min_vruntime because the | |
3121 | * update can refer to the ->curr item and we need to reflect this | |
3122 | * movement in our normalized position. | |
3123 | */ | |
371fd7e7 | 3124 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 3125 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 3126 | |
d8b4986d PT |
3127 | /* return excess runtime on last dequeue */ |
3128 | return_cfs_rq_runtime(cfs_rq); | |
3129 | ||
1e876231 | 3130 | update_min_vruntime(cfs_rq); |
17bc14b7 | 3131 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
3132 | } |
3133 | ||
3134 | /* | |
3135 | * Preempt the current task with a newly woken task if needed: | |
3136 | */ | |
7c92e54f | 3137 | static void |
2e09bf55 | 3138 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 3139 | { |
11697830 | 3140 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
3141 | struct sched_entity *se; |
3142 | s64 delta; | |
11697830 | 3143 | |
6d0f0ebd | 3144 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 3145 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 3146 | if (delta_exec > ideal_runtime) { |
8875125e | 3147 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
3148 | /* |
3149 | * The current task ran long enough, ensure it doesn't get | |
3150 | * re-elected due to buddy favours. | |
3151 | */ | |
3152 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
3153 | return; |
3154 | } | |
3155 | ||
3156 | /* | |
3157 | * Ensure that a task that missed wakeup preemption by a | |
3158 | * narrow margin doesn't have to wait for a full slice. | |
3159 | * This also mitigates buddy induced latencies under load. | |
3160 | */ | |
f685ceac MG |
3161 | if (delta_exec < sysctl_sched_min_granularity) |
3162 | return; | |
3163 | ||
f4cfb33e WX |
3164 | se = __pick_first_entity(cfs_rq); |
3165 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 3166 | |
f4cfb33e WX |
3167 | if (delta < 0) |
3168 | return; | |
d7d82944 | 3169 | |
f4cfb33e | 3170 | if (delta > ideal_runtime) |
8875125e | 3171 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
3172 | } |
3173 | ||
83b699ed | 3174 | static void |
8494f412 | 3175 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 3176 | { |
83b699ed SV |
3177 | /* 'current' is not kept within the tree. */ |
3178 | if (se->on_rq) { | |
3179 | /* | |
3180 | * Any task has to be enqueued before it get to execute on | |
3181 | * a CPU. So account for the time it spent waiting on the | |
3182 | * runqueue. | |
3183 | */ | |
3184 | update_stats_wait_end(cfs_rq, se); | |
3185 | __dequeue_entity(cfs_rq, se); | |
3186 | } | |
3187 | ||
79303e9e | 3188 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 3189 | cfs_rq->curr = se; |
eba1ed4b IM |
3190 | #ifdef CONFIG_SCHEDSTATS |
3191 | /* | |
3192 | * Track our maximum slice length, if the CPU's load is at | |
3193 | * least twice that of our own weight (i.e. dont track it | |
3194 | * when there are only lesser-weight tasks around): | |
3195 | */ | |
495eca49 | 3196 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 3197 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
3198 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
3199 | } | |
3200 | #endif | |
4a55b450 | 3201 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
3202 | } |
3203 | ||
3f3a4904 PZ |
3204 | static int |
3205 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
3206 | ||
ac53db59 RR |
3207 | /* |
3208 | * Pick the next process, keeping these things in mind, in this order: | |
3209 | * 1) keep things fair between processes/task groups | |
3210 | * 2) pick the "next" process, since someone really wants that to run | |
3211 | * 3) pick the "last" process, for cache locality | |
3212 | * 4) do not run the "skip" process, if something else is available | |
3213 | */ | |
678d5718 PZ |
3214 | static struct sched_entity * |
3215 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 3216 | { |
678d5718 PZ |
3217 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
3218 | struct sched_entity *se; | |
3219 | ||
3220 | /* | |
3221 | * If curr is set we have to see if its left of the leftmost entity | |
3222 | * still in the tree, provided there was anything in the tree at all. | |
3223 | */ | |
3224 | if (!left || (curr && entity_before(curr, left))) | |
3225 | left = curr; | |
3226 | ||
3227 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 3228 | |
ac53db59 RR |
3229 | /* |
3230 | * Avoid running the skip buddy, if running something else can | |
3231 | * be done without getting too unfair. | |
3232 | */ | |
3233 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
3234 | struct sched_entity *second; |
3235 | ||
3236 | if (se == curr) { | |
3237 | second = __pick_first_entity(cfs_rq); | |
3238 | } else { | |
3239 | second = __pick_next_entity(se); | |
3240 | if (!second || (curr && entity_before(curr, second))) | |
3241 | second = curr; | |
3242 | } | |
3243 | ||
ac53db59 RR |
3244 | if (second && wakeup_preempt_entity(second, left) < 1) |
3245 | se = second; | |
3246 | } | |
aa2ac252 | 3247 | |
f685ceac MG |
3248 | /* |
3249 | * Prefer last buddy, try to return the CPU to a preempted task. | |
3250 | */ | |
3251 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
3252 | se = cfs_rq->last; | |
3253 | ||
ac53db59 RR |
3254 | /* |
3255 | * Someone really wants this to run. If it's not unfair, run it. | |
3256 | */ | |
3257 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
3258 | se = cfs_rq->next; | |
3259 | ||
f685ceac | 3260 | clear_buddies(cfs_rq, se); |
4793241b PZ |
3261 | |
3262 | return se; | |
aa2ac252 PZ |
3263 | } |
3264 | ||
678d5718 | 3265 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 3266 | |
ab6cde26 | 3267 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
3268 | { |
3269 | /* | |
3270 | * If still on the runqueue then deactivate_task() | |
3271 | * was not called and update_curr() has to be done: | |
3272 | */ | |
3273 | if (prev->on_rq) | |
b7cc0896 | 3274 | update_curr(cfs_rq); |
bf0f6f24 | 3275 | |
d3d9dc33 PT |
3276 | /* throttle cfs_rqs exceeding runtime */ |
3277 | check_cfs_rq_runtime(cfs_rq); | |
3278 | ||
ddc97297 | 3279 | check_spread(cfs_rq, prev); |
30cfdcfc | 3280 | if (prev->on_rq) { |
5870db5b | 3281 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
3282 | /* Put 'current' back into the tree. */ |
3283 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 3284 | /* in !on_rq case, update occurred at dequeue */ |
9ee474f5 | 3285 | update_entity_load_avg(prev, 1); |
30cfdcfc | 3286 | } |
429d43bc | 3287 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
3288 | } |
3289 | ||
8f4d37ec PZ |
3290 | static void |
3291 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 3292 | { |
bf0f6f24 | 3293 | /* |
30cfdcfc | 3294 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3295 | */ |
30cfdcfc | 3296 | update_curr(cfs_rq); |
bf0f6f24 | 3297 | |
9d85f21c PT |
3298 | /* |
3299 | * Ensure that runnable average is periodically updated. | |
3300 | */ | |
9ee474f5 | 3301 | update_entity_load_avg(curr, 1); |
aff3e498 | 3302 | update_cfs_rq_blocked_load(cfs_rq, 1); |
bf0bd948 | 3303 | update_cfs_shares(cfs_rq); |
9d85f21c | 3304 | |
8f4d37ec PZ |
3305 | #ifdef CONFIG_SCHED_HRTICK |
3306 | /* | |
3307 | * queued ticks are scheduled to match the slice, so don't bother | |
3308 | * validating it and just reschedule. | |
3309 | */ | |
983ed7a6 | 3310 | if (queued) { |
8875125e | 3311 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
3312 | return; |
3313 | } | |
8f4d37ec PZ |
3314 | /* |
3315 | * don't let the period tick interfere with the hrtick preemption | |
3316 | */ | |
3317 | if (!sched_feat(DOUBLE_TICK) && | |
3318 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
3319 | return; | |
3320 | #endif | |
3321 | ||
2c2efaed | 3322 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 3323 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
3324 | } |
3325 | ||
ab84d31e PT |
3326 | |
3327 | /************************************************** | |
3328 | * CFS bandwidth control machinery | |
3329 | */ | |
3330 | ||
3331 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
3332 | |
3333 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 3334 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
3335 | |
3336 | static inline bool cfs_bandwidth_used(void) | |
3337 | { | |
c5905afb | 3338 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
3339 | } |
3340 | ||
1ee14e6c | 3341 | void cfs_bandwidth_usage_inc(void) |
029632fb | 3342 | { |
1ee14e6c BS |
3343 | static_key_slow_inc(&__cfs_bandwidth_used); |
3344 | } | |
3345 | ||
3346 | void cfs_bandwidth_usage_dec(void) | |
3347 | { | |
3348 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
3349 | } |
3350 | #else /* HAVE_JUMP_LABEL */ | |
3351 | static bool cfs_bandwidth_used(void) | |
3352 | { | |
3353 | return true; | |
3354 | } | |
3355 | ||
1ee14e6c BS |
3356 | void cfs_bandwidth_usage_inc(void) {} |
3357 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
3358 | #endif /* HAVE_JUMP_LABEL */ |
3359 | ||
ab84d31e PT |
3360 | /* |
3361 | * default period for cfs group bandwidth. | |
3362 | * default: 0.1s, units: nanoseconds | |
3363 | */ | |
3364 | static inline u64 default_cfs_period(void) | |
3365 | { | |
3366 | return 100000000ULL; | |
3367 | } | |
ec12cb7f PT |
3368 | |
3369 | static inline u64 sched_cfs_bandwidth_slice(void) | |
3370 | { | |
3371 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
3372 | } | |
3373 | ||
a9cf55b2 PT |
3374 | /* |
3375 | * Replenish runtime according to assigned quota and update expiration time. | |
3376 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
3377 | * additional synchronization around rq->lock. | |
3378 | * | |
3379 | * requires cfs_b->lock | |
3380 | */ | |
029632fb | 3381 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
3382 | { |
3383 | u64 now; | |
3384 | ||
3385 | if (cfs_b->quota == RUNTIME_INF) | |
3386 | return; | |
3387 | ||
3388 | now = sched_clock_cpu(smp_processor_id()); | |
3389 | cfs_b->runtime = cfs_b->quota; | |
3390 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
3391 | } | |
3392 | ||
029632fb PZ |
3393 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3394 | { | |
3395 | return &tg->cfs_bandwidth; | |
3396 | } | |
3397 | ||
f1b17280 PT |
3398 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
3399 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
3400 | { | |
3401 | if (unlikely(cfs_rq->throttle_count)) | |
3402 | return cfs_rq->throttled_clock_task; | |
3403 | ||
78becc27 | 3404 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
3405 | } |
3406 | ||
85dac906 PT |
3407 | /* returns 0 on failure to allocate runtime */ |
3408 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
3409 | { |
3410 | struct task_group *tg = cfs_rq->tg; | |
3411 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 3412 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
3413 | |
3414 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
3415 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
3416 | ||
3417 | raw_spin_lock(&cfs_b->lock); | |
3418 | if (cfs_b->quota == RUNTIME_INF) | |
3419 | amount = min_amount; | |
58088ad0 | 3420 | else { |
a9cf55b2 PT |
3421 | /* |
3422 | * If the bandwidth pool has become inactive, then at least one | |
3423 | * period must have elapsed since the last consumption. | |
3424 | * Refresh the global state and ensure bandwidth timer becomes | |
3425 | * active. | |
3426 | */ | |
3427 | if (!cfs_b->timer_active) { | |
3428 | __refill_cfs_bandwidth_runtime(cfs_b); | |
09dc4ab0 | 3429 | __start_cfs_bandwidth(cfs_b, false); |
a9cf55b2 | 3430 | } |
58088ad0 PT |
3431 | |
3432 | if (cfs_b->runtime > 0) { | |
3433 | amount = min(cfs_b->runtime, min_amount); | |
3434 | cfs_b->runtime -= amount; | |
3435 | cfs_b->idle = 0; | |
3436 | } | |
ec12cb7f | 3437 | } |
a9cf55b2 | 3438 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
3439 | raw_spin_unlock(&cfs_b->lock); |
3440 | ||
3441 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
3442 | /* |
3443 | * we may have advanced our local expiration to account for allowed | |
3444 | * spread between our sched_clock and the one on which runtime was | |
3445 | * issued. | |
3446 | */ | |
3447 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
3448 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
3449 | |
3450 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
3451 | } |
3452 | ||
a9cf55b2 PT |
3453 | /* |
3454 | * Note: This depends on the synchronization provided by sched_clock and the | |
3455 | * fact that rq->clock snapshots this value. | |
3456 | */ | |
3457 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 3458 | { |
a9cf55b2 | 3459 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
3460 | |
3461 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 3462 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
3463 | return; |
3464 | ||
a9cf55b2 PT |
3465 | if (cfs_rq->runtime_remaining < 0) |
3466 | return; | |
3467 | ||
3468 | /* | |
3469 | * If the local deadline has passed we have to consider the | |
3470 | * possibility that our sched_clock is 'fast' and the global deadline | |
3471 | * has not truly expired. | |
3472 | * | |
3473 | * Fortunately we can check determine whether this the case by checking | |
51f2176d BS |
3474 | * whether the global deadline has advanced. It is valid to compare |
3475 | * cfs_b->runtime_expires without any locks since we only care about | |
3476 | * exact equality, so a partial write will still work. | |
a9cf55b2 PT |
3477 | */ |
3478 | ||
51f2176d | 3479 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { |
a9cf55b2 PT |
3480 | /* extend local deadline, drift is bounded above by 2 ticks */ |
3481 | cfs_rq->runtime_expires += TICK_NSEC; | |
3482 | } else { | |
3483 | /* global deadline is ahead, expiration has passed */ | |
3484 | cfs_rq->runtime_remaining = 0; | |
3485 | } | |
3486 | } | |
3487 | ||
9dbdb155 | 3488 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
3489 | { |
3490 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3491 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3492 | expire_cfs_rq_runtime(cfs_rq); |
3493 | ||
3494 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3495 | return; |
3496 | ||
85dac906 PT |
3497 | /* |
3498 | * if we're unable to extend our runtime we resched so that the active | |
3499 | * hierarchy can be throttled | |
3500 | */ | |
3501 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 3502 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
3503 | } |
3504 | ||
6c16a6dc | 3505 | static __always_inline |
9dbdb155 | 3506 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 3507 | { |
56f570e5 | 3508 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3509 | return; |
3510 | ||
3511 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3512 | } | |
3513 | ||
85dac906 PT |
3514 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3515 | { | |
56f570e5 | 3516 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3517 | } |
3518 | ||
64660c86 PT |
3519 | /* check whether cfs_rq, or any parent, is throttled */ |
3520 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3521 | { | |
56f570e5 | 3522 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3523 | } |
3524 | ||
3525 | /* | |
3526 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3527 | * dest_cpu are members of a throttled hierarchy when performing group | |
3528 | * load-balance operations. | |
3529 | */ | |
3530 | static inline int throttled_lb_pair(struct task_group *tg, | |
3531 | int src_cpu, int dest_cpu) | |
3532 | { | |
3533 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3534 | ||
3535 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3536 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3537 | ||
3538 | return throttled_hierarchy(src_cfs_rq) || | |
3539 | throttled_hierarchy(dest_cfs_rq); | |
3540 | } | |
3541 | ||
3542 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3543 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3544 | { | |
3545 | struct rq *rq = data; | |
3546 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3547 | ||
3548 | cfs_rq->throttle_count--; | |
3549 | #ifdef CONFIG_SMP | |
3550 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 3551 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3552 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3553 | cfs_rq->throttled_clock_task; |
64660c86 PT |
3554 | } |
3555 | #endif | |
3556 | ||
3557 | return 0; | |
3558 | } | |
3559 | ||
3560 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3561 | { | |
3562 | struct rq *rq = data; | |
3563 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3564 | ||
82958366 PT |
3565 | /* group is entering throttled state, stop time */ |
3566 | if (!cfs_rq->throttle_count) | |
78becc27 | 3567 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3568 | cfs_rq->throttle_count++; |
3569 | ||
3570 | return 0; | |
3571 | } | |
3572 | ||
d3d9dc33 | 3573 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3574 | { |
3575 | struct rq *rq = rq_of(cfs_rq); | |
3576 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3577 | struct sched_entity *se; | |
3578 | long task_delta, dequeue = 1; | |
3579 | ||
3580 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3581 | ||
f1b17280 | 3582 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3583 | rcu_read_lock(); |
3584 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3585 | rcu_read_unlock(); | |
85dac906 PT |
3586 | |
3587 | task_delta = cfs_rq->h_nr_running; | |
3588 | for_each_sched_entity(se) { | |
3589 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3590 | /* throttled entity or throttle-on-deactivate */ | |
3591 | if (!se->on_rq) | |
3592 | break; | |
3593 | ||
3594 | if (dequeue) | |
3595 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3596 | qcfs_rq->h_nr_running -= task_delta; | |
3597 | ||
3598 | if (qcfs_rq->load.weight) | |
3599 | dequeue = 0; | |
3600 | } | |
3601 | ||
3602 | if (!se) | |
72465447 | 3603 | sub_nr_running(rq, task_delta); |
85dac906 PT |
3604 | |
3605 | cfs_rq->throttled = 1; | |
78becc27 | 3606 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 3607 | raw_spin_lock(&cfs_b->lock); |
c06f04c7 BS |
3608 | /* |
3609 | * Add to the _head_ of the list, so that an already-started | |
3610 | * distribute_cfs_runtime will not see us | |
3611 | */ | |
3612 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
f9f9ffc2 | 3613 | if (!cfs_b->timer_active) |
09dc4ab0 | 3614 | __start_cfs_bandwidth(cfs_b, false); |
85dac906 PT |
3615 | raw_spin_unlock(&cfs_b->lock); |
3616 | } | |
3617 | ||
029632fb | 3618 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3619 | { |
3620 | struct rq *rq = rq_of(cfs_rq); | |
3621 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3622 | struct sched_entity *se; | |
3623 | int enqueue = 1; | |
3624 | long task_delta; | |
3625 | ||
22b958d8 | 3626 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3627 | |
3628 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3629 | |
3630 | update_rq_clock(rq); | |
3631 | ||
671fd9da | 3632 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3633 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3634 | list_del_rcu(&cfs_rq->throttled_list); |
3635 | raw_spin_unlock(&cfs_b->lock); | |
3636 | ||
64660c86 PT |
3637 | /* update hierarchical throttle state */ |
3638 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3639 | ||
671fd9da PT |
3640 | if (!cfs_rq->load.weight) |
3641 | return; | |
3642 | ||
3643 | task_delta = cfs_rq->h_nr_running; | |
3644 | for_each_sched_entity(se) { | |
3645 | if (se->on_rq) | |
3646 | enqueue = 0; | |
3647 | ||
3648 | cfs_rq = cfs_rq_of(se); | |
3649 | if (enqueue) | |
3650 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3651 | cfs_rq->h_nr_running += task_delta; | |
3652 | ||
3653 | if (cfs_rq_throttled(cfs_rq)) | |
3654 | break; | |
3655 | } | |
3656 | ||
3657 | if (!se) | |
72465447 | 3658 | add_nr_running(rq, task_delta); |
671fd9da PT |
3659 | |
3660 | /* determine whether we need to wake up potentially idle cpu */ | |
3661 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
8875125e | 3662 | resched_curr(rq); |
671fd9da PT |
3663 | } |
3664 | ||
3665 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
3666 | u64 remaining, u64 expires) | |
3667 | { | |
3668 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
3669 | u64 runtime; |
3670 | u64 starting_runtime = remaining; | |
671fd9da PT |
3671 | |
3672 | rcu_read_lock(); | |
3673 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
3674 | throttled_list) { | |
3675 | struct rq *rq = rq_of(cfs_rq); | |
3676 | ||
3677 | raw_spin_lock(&rq->lock); | |
3678 | if (!cfs_rq_throttled(cfs_rq)) | |
3679 | goto next; | |
3680 | ||
3681 | runtime = -cfs_rq->runtime_remaining + 1; | |
3682 | if (runtime > remaining) | |
3683 | runtime = remaining; | |
3684 | remaining -= runtime; | |
3685 | ||
3686 | cfs_rq->runtime_remaining += runtime; | |
3687 | cfs_rq->runtime_expires = expires; | |
3688 | ||
3689 | /* we check whether we're throttled above */ | |
3690 | if (cfs_rq->runtime_remaining > 0) | |
3691 | unthrottle_cfs_rq(cfs_rq); | |
3692 | ||
3693 | next: | |
3694 | raw_spin_unlock(&rq->lock); | |
3695 | ||
3696 | if (!remaining) | |
3697 | break; | |
3698 | } | |
3699 | rcu_read_unlock(); | |
3700 | ||
c06f04c7 | 3701 | return starting_runtime - remaining; |
671fd9da PT |
3702 | } |
3703 | ||
58088ad0 PT |
3704 | /* |
3705 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
3706 | * cfs_rqs as appropriate. If there has been no activity within the last | |
3707 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
3708 | * used to track this state. | |
3709 | */ | |
3710 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3711 | { | |
671fd9da | 3712 | u64 runtime, runtime_expires; |
51f2176d | 3713 | int throttled; |
58088ad0 | 3714 | |
58088ad0 PT |
3715 | /* no need to continue the timer with no bandwidth constraint */ |
3716 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 3717 | goto out_deactivate; |
58088ad0 | 3718 | |
671fd9da | 3719 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 3720 | cfs_b->nr_periods += overrun; |
671fd9da | 3721 | |
51f2176d BS |
3722 | /* |
3723 | * idle depends on !throttled (for the case of a large deficit), and if | |
3724 | * we're going inactive then everything else can be deferred | |
3725 | */ | |
3726 | if (cfs_b->idle && !throttled) | |
3727 | goto out_deactivate; | |
a9cf55b2 | 3728 | |
927b54fc BS |
3729 | /* |
3730 | * if we have relooped after returning idle once, we need to update our | |
3731 | * status as actually running, so that other cpus doing | |
3732 | * __start_cfs_bandwidth will stop trying to cancel us. | |
3733 | */ | |
3734 | cfs_b->timer_active = 1; | |
3735 | ||
a9cf55b2 PT |
3736 | __refill_cfs_bandwidth_runtime(cfs_b); |
3737 | ||
671fd9da PT |
3738 | if (!throttled) { |
3739 | /* mark as potentially idle for the upcoming period */ | |
3740 | cfs_b->idle = 1; | |
51f2176d | 3741 | return 0; |
671fd9da PT |
3742 | } |
3743 | ||
e8da1b18 NR |
3744 | /* account preceding periods in which throttling occurred */ |
3745 | cfs_b->nr_throttled += overrun; | |
3746 | ||
671fd9da | 3747 | runtime_expires = cfs_b->runtime_expires; |
671fd9da PT |
3748 | |
3749 | /* | |
c06f04c7 BS |
3750 | * This check is repeated as we are holding onto the new bandwidth while |
3751 | * we unthrottle. This can potentially race with an unthrottled group | |
3752 | * trying to acquire new bandwidth from the global pool. This can result | |
3753 | * in us over-using our runtime if it is all used during this loop, but | |
3754 | * only by limited amounts in that extreme case. | |
671fd9da | 3755 | */ |
c06f04c7 BS |
3756 | while (throttled && cfs_b->runtime > 0) { |
3757 | runtime = cfs_b->runtime; | |
671fd9da PT |
3758 | raw_spin_unlock(&cfs_b->lock); |
3759 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3760 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3761 | runtime_expires); | |
3762 | raw_spin_lock(&cfs_b->lock); | |
3763 | ||
3764 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
c06f04c7 BS |
3765 | |
3766 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | |
671fd9da | 3767 | } |
58088ad0 | 3768 | |
671fd9da PT |
3769 | /* |
3770 | * While we are ensured activity in the period following an | |
3771 | * unthrottle, this also covers the case in which the new bandwidth is | |
3772 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3773 | * timer to remain active while there are any throttled entities.) | |
3774 | */ | |
3775 | cfs_b->idle = 0; | |
58088ad0 | 3776 | |
51f2176d BS |
3777 | return 0; |
3778 | ||
3779 | out_deactivate: | |
3780 | cfs_b->timer_active = 0; | |
3781 | return 1; | |
58088ad0 | 3782 | } |
d3d9dc33 | 3783 | |
d8b4986d PT |
3784 | /* a cfs_rq won't donate quota below this amount */ |
3785 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3786 | /* minimum remaining period time to redistribute slack quota */ | |
3787 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3788 | /* how long we wait to gather additional slack before distributing */ | |
3789 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3790 | ||
db06e78c BS |
3791 | /* |
3792 | * Are we near the end of the current quota period? | |
3793 | * | |
3794 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
3795 | * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of | |
3796 | * migrate_hrtimers, base is never cleared, so we are fine. | |
3797 | */ | |
d8b4986d PT |
3798 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
3799 | { | |
3800 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3801 | u64 remaining; | |
3802 | ||
3803 | /* if the call-back is running a quota refresh is already occurring */ | |
3804 | if (hrtimer_callback_running(refresh_timer)) | |
3805 | return 1; | |
3806 | ||
3807 | /* is a quota refresh about to occur? */ | |
3808 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3809 | if (remaining < min_expire) | |
3810 | return 1; | |
3811 | ||
3812 | return 0; | |
3813 | } | |
3814 | ||
3815 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3816 | { | |
3817 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3818 | ||
3819 | /* if there's a quota refresh soon don't bother with slack */ | |
3820 | if (runtime_refresh_within(cfs_b, min_left)) | |
3821 | return; | |
3822 | ||
3823 | start_bandwidth_timer(&cfs_b->slack_timer, | |
3824 | ns_to_ktime(cfs_bandwidth_slack_period)); | |
3825 | } | |
3826 | ||
3827 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3828 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3829 | { | |
3830 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3831 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3832 | ||
3833 | if (slack_runtime <= 0) | |
3834 | return; | |
3835 | ||
3836 | raw_spin_lock(&cfs_b->lock); | |
3837 | if (cfs_b->quota != RUNTIME_INF && | |
3838 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3839 | cfs_b->runtime += slack_runtime; | |
3840 | ||
3841 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3842 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3843 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3844 | start_cfs_slack_bandwidth(cfs_b); | |
3845 | } | |
3846 | raw_spin_unlock(&cfs_b->lock); | |
3847 | ||
3848 | /* even if it's not valid for return we don't want to try again */ | |
3849 | cfs_rq->runtime_remaining -= slack_runtime; | |
3850 | } | |
3851 | ||
3852 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3853 | { | |
56f570e5 PT |
3854 | if (!cfs_bandwidth_used()) |
3855 | return; | |
3856 | ||
fccfdc6f | 3857 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3858 | return; |
3859 | ||
3860 | __return_cfs_rq_runtime(cfs_rq); | |
3861 | } | |
3862 | ||
3863 | /* | |
3864 | * This is done with a timer (instead of inline with bandwidth return) since | |
3865 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3866 | */ | |
3867 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3868 | { | |
3869 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3870 | u64 expires; | |
3871 | ||
3872 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
3873 | raw_spin_lock(&cfs_b->lock); |
3874 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
3875 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 3876 | return; |
db06e78c | 3877 | } |
d8b4986d | 3878 | |
c06f04c7 | 3879 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 3880 | runtime = cfs_b->runtime; |
c06f04c7 | 3881 | |
d8b4986d PT |
3882 | expires = cfs_b->runtime_expires; |
3883 | raw_spin_unlock(&cfs_b->lock); | |
3884 | ||
3885 | if (!runtime) | |
3886 | return; | |
3887 | ||
3888 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3889 | ||
3890 | raw_spin_lock(&cfs_b->lock); | |
3891 | if (expires == cfs_b->runtime_expires) | |
c06f04c7 | 3892 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
d8b4986d PT |
3893 | raw_spin_unlock(&cfs_b->lock); |
3894 | } | |
3895 | ||
d3d9dc33 PT |
3896 | /* |
3897 | * When a group wakes up we want to make sure that its quota is not already | |
3898 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3899 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3900 | */ | |
3901 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3902 | { | |
56f570e5 PT |
3903 | if (!cfs_bandwidth_used()) |
3904 | return; | |
3905 | ||
d3d9dc33 PT |
3906 | /* an active group must be handled by the update_curr()->put() path */ |
3907 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3908 | return; | |
3909 | ||
3910 | /* ensure the group is not already throttled */ | |
3911 | if (cfs_rq_throttled(cfs_rq)) | |
3912 | return; | |
3913 | ||
3914 | /* update runtime allocation */ | |
3915 | account_cfs_rq_runtime(cfs_rq, 0); | |
3916 | if (cfs_rq->runtime_remaining <= 0) | |
3917 | throttle_cfs_rq(cfs_rq); | |
3918 | } | |
3919 | ||
3920 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
678d5718 | 3921 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 3922 | { |
56f570e5 | 3923 | if (!cfs_bandwidth_used()) |
678d5718 | 3924 | return false; |
56f570e5 | 3925 | |
d3d9dc33 | 3926 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 3927 | return false; |
d3d9dc33 PT |
3928 | |
3929 | /* | |
3930 | * it's possible for a throttled entity to be forced into a running | |
3931 | * state (e.g. set_curr_task), in this case we're finished. | |
3932 | */ | |
3933 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 3934 | return true; |
d3d9dc33 PT |
3935 | |
3936 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 3937 | return true; |
d3d9dc33 | 3938 | } |
029632fb | 3939 | |
029632fb PZ |
3940 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3941 | { | |
3942 | struct cfs_bandwidth *cfs_b = | |
3943 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
3944 | do_sched_cfs_slack_timer(cfs_b); | |
3945 | ||
3946 | return HRTIMER_NORESTART; | |
3947 | } | |
3948 | ||
3949 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
3950 | { | |
3951 | struct cfs_bandwidth *cfs_b = | |
3952 | container_of(timer, struct cfs_bandwidth, period_timer); | |
3953 | ktime_t now; | |
3954 | int overrun; | |
3955 | int idle = 0; | |
3956 | ||
51f2176d | 3957 | raw_spin_lock(&cfs_b->lock); |
029632fb PZ |
3958 | for (;;) { |
3959 | now = hrtimer_cb_get_time(timer); | |
3960 | overrun = hrtimer_forward(timer, now, cfs_b->period); | |
3961 | ||
3962 | if (!overrun) | |
3963 | break; | |
3964 | ||
3965 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
3966 | } | |
51f2176d | 3967 | raw_spin_unlock(&cfs_b->lock); |
029632fb PZ |
3968 | |
3969 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
3970 | } | |
3971 | ||
3972 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
3973 | { | |
3974 | raw_spin_lock_init(&cfs_b->lock); | |
3975 | cfs_b->runtime = 0; | |
3976 | cfs_b->quota = RUNTIME_INF; | |
3977 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
3978 | ||
3979 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
3980 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3981 | cfs_b->period_timer.function = sched_cfs_period_timer; | |
3982 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3983 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
3984 | } | |
3985 | ||
3986 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3987 | { | |
3988 | cfs_rq->runtime_enabled = 0; | |
3989 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
3990 | } | |
3991 | ||
3992 | /* requires cfs_b->lock, may release to reprogram timer */ | |
09dc4ab0 | 3993 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force) |
029632fb PZ |
3994 | { |
3995 | /* | |
3996 | * The timer may be active because we're trying to set a new bandwidth | |
3997 | * period or because we're racing with the tear-down path | |
3998 | * (timer_active==0 becomes visible before the hrtimer call-back | |
3999 | * terminates). In either case we ensure that it's re-programmed | |
4000 | */ | |
927b54fc BS |
4001 | while (unlikely(hrtimer_active(&cfs_b->period_timer)) && |
4002 | hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { | |
4003 | /* bounce the lock to allow do_sched_cfs_period_timer to run */ | |
029632fb | 4004 | raw_spin_unlock(&cfs_b->lock); |
927b54fc | 4005 | cpu_relax(); |
029632fb PZ |
4006 | raw_spin_lock(&cfs_b->lock); |
4007 | /* if someone else restarted the timer then we're done */ | |
09dc4ab0 | 4008 | if (!force && cfs_b->timer_active) |
029632fb PZ |
4009 | return; |
4010 | } | |
4011 | ||
4012 | cfs_b->timer_active = 1; | |
4013 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | |
4014 | } | |
4015 | ||
4016 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4017 | { | |
7f1a169b TH |
4018 | /* init_cfs_bandwidth() was not called */ |
4019 | if (!cfs_b->throttled_cfs_rq.next) | |
4020 | return; | |
4021 | ||
029632fb PZ |
4022 | hrtimer_cancel(&cfs_b->period_timer); |
4023 | hrtimer_cancel(&cfs_b->slack_timer); | |
4024 | } | |
4025 | ||
0e59bdae KT |
4026 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
4027 | { | |
4028 | struct cfs_rq *cfs_rq; | |
4029 | ||
4030 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
4031 | struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; | |
4032 | ||
4033 | raw_spin_lock(&cfs_b->lock); | |
4034 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
4035 | raw_spin_unlock(&cfs_b->lock); | |
4036 | } | |
4037 | } | |
4038 | ||
38dc3348 | 4039 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
4040 | { |
4041 | struct cfs_rq *cfs_rq; | |
4042 | ||
4043 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
029632fb PZ |
4044 | if (!cfs_rq->runtime_enabled) |
4045 | continue; | |
4046 | ||
4047 | /* | |
4048 | * clock_task is not advancing so we just need to make sure | |
4049 | * there's some valid quota amount | |
4050 | */ | |
51f2176d | 4051 | cfs_rq->runtime_remaining = 1; |
0e59bdae KT |
4052 | /* |
4053 | * Offline rq is schedulable till cpu is completely disabled | |
4054 | * in take_cpu_down(), so we prevent new cfs throttling here. | |
4055 | */ | |
4056 | cfs_rq->runtime_enabled = 0; | |
4057 | ||
029632fb PZ |
4058 | if (cfs_rq_throttled(cfs_rq)) |
4059 | unthrottle_cfs_rq(cfs_rq); | |
4060 | } | |
4061 | } | |
4062 | ||
4063 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
4064 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
4065 | { | |
78becc27 | 4066 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
4067 | } |
4068 | ||
9dbdb155 | 4069 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 4070 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 4071 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
6c16a6dc | 4072 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
4073 | |
4074 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
4075 | { | |
4076 | return 0; | |
4077 | } | |
64660c86 PT |
4078 | |
4079 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4080 | { | |
4081 | return 0; | |
4082 | } | |
4083 | ||
4084 | static inline int throttled_lb_pair(struct task_group *tg, | |
4085 | int src_cpu, int dest_cpu) | |
4086 | { | |
4087 | return 0; | |
4088 | } | |
029632fb PZ |
4089 | |
4090 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
4091 | ||
4092 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
4093 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
4094 | #endif |
4095 | ||
029632fb PZ |
4096 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4097 | { | |
4098 | return NULL; | |
4099 | } | |
4100 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 4101 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 4102 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
4103 | |
4104 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
4105 | ||
bf0f6f24 IM |
4106 | /************************************************** |
4107 | * CFS operations on tasks: | |
4108 | */ | |
4109 | ||
8f4d37ec PZ |
4110 | #ifdef CONFIG_SCHED_HRTICK |
4111 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4112 | { | |
8f4d37ec PZ |
4113 | struct sched_entity *se = &p->se; |
4114 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4115 | ||
4116 | WARN_ON(task_rq(p) != rq); | |
4117 | ||
b39e66ea | 4118 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
4119 | u64 slice = sched_slice(cfs_rq, se); |
4120 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
4121 | s64 delta = slice - ran; | |
4122 | ||
4123 | if (delta < 0) { | |
4124 | if (rq->curr == p) | |
8875125e | 4125 | resched_curr(rq); |
8f4d37ec PZ |
4126 | return; |
4127 | } | |
31656519 | 4128 | hrtick_start(rq, delta); |
8f4d37ec PZ |
4129 | } |
4130 | } | |
a4c2f00f PZ |
4131 | |
4132 | /* | |
4133 | * called from enqueue/dequeue and updates the hrtick when the | |
4134 | * current task is from our class and nr_running is low enough | |
4135 | * to matter. | |
4136 | */ | |
4137 | static void hrtick_update(struct rq *rq) | |
4138 | { | |
4139 | struct task_struct *curr = rq->curr; | |
4140 | ||
b39e66ea | 4141 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
4142 | return; |
4143 | ||
4144 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
4145 | hrtick_start_fair(rq, curr); | |
4146 | } | |
55e12e5e | 4147 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
4148 | static inline void |
4149 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4150 | { | |
4151 | } | |
a4c2f00f PZ |
4152 | |
4153 | static inline void hrtick_update(struct rq *rq) | |
4154 | { | |
4155 | } | |
8f4d37ec PZ |
4156 | #endif |
4157 | ||
bf0f6f24 IM |
4158 | /* |
4159 | * The enqueue_task method is called before nr_running is | |
4160 | * increased. Here we update the fair scheduling stats and | |
4161 | * then put the task into the rbtree: | |
4162 | */ | |
ea87bb78 | 4163 | static void |
371fd7e7 | 4164 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4165 | { |
4166 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4167 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
4168 | |
4169 | for_each_sched_entity(se) { | |
62fb1851 | 4170 | if (se->on_rq) |
bf0f6f24 IM |
4171 | break; |
4172 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 4173 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
4174 | |
4175 | /* | |
4176 | * end evaluation on encountering a throttled cfs_rq | |
4177 | * | |
4178 | * note: in the case of encountering a throttled cfs_rq we will | |
4179 | * post the final h_nr_running increment below. | |
4180 | */ | |
4181 | if (cfs_rq_throttled(cfs_rq)) | |
4182 | break; | |
953bfcd1 | 4183 | cfs_rq->h_nr_running++; |
85dac906 | 4184 | |
88ec22d3 | 4185 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 4186 | } |
8f4d37ec | 4187 | |
2069dd75 | 4188 | for_each_sched_entity(se) { |
0f317143 | 4189 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4190 | cfs_rq->h_nr_running++; |
2069dd75 | 4191 | |
85dac906 PT |
4192 | if (cfs_rq_throttled(cfs_rq)) |
4193 | break; | |
4194 | ||
17bc14b7 | 4195 | update_cfs_shares(cfs_rq); |
9ee474f5 | 4196 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
4197 | } |
4198 | ||
18bf2805 BS |
4199 | if (!se) { |
4200 | update_rq_runnable_avg(rq, rq->nr_running); | |
72465447 | 4201 | add_nr_running(rq, 1); |
18bf2805 | 4202 | } |
a4c2f00f | 4203 | hrtick_update(rq); |
bf0f6f24 IM |
4204 | } |
4205 | ||
2f36825b VP |
4206 | static void set_next_buddy(struct sched_entity *se); |
4207 | ||
bf0f6f24 IM |
4208 | /* |
4209 | * The dequeue_task method is called before nr_running is | |
4210 | * decreased. We remove the task from the rbtree and | |
4211 | * update the fair scheduling stats: | |
4212 | */ | |
371fd7e7 | 4213 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4214 | { |
4215 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4216 | struct sched_entity *se = &p->se; |
2f36825b | 4217 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
4218 | |
4219 | for_each_sched_entity(se) { | |
4220 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 4221 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
4222 | |
4223 | /* | |
4224 | * end evaluation on encountering a throttled cfs_rq | |
4225 | * | |
4226 | * note: in the case of encountering a throttled cfs_rq we will | |
4227 | * post the final h_nr_running decrement below. | |
4228 | */ | |
4229 | if (cfs_rq_throttled(cfs_rq)) | |
4230 | break; | |
953bfcd1 | 4231 | cfs_rq->h_nr_running--; |
2069dd75 | 4232 | |
bf0f6f24 | 4233 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
4234 | if (cfs_rq->load.weight) { |
4235 | /* | |
4236 | * Bias pick_next to pick a task from this cfs_rq, as | |
4237 | * p is sleeping when it is within its sched_slice. | |
4238 | */ | |
4239 | if (task_sleep && parent_entity(se)) | |
4240 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
4241 | |
4242 | /* avoid re-evaluating load for this entity */ | |
4243 | se = parent_entity(se); | |
bf0f6f24 | 4244 | break; |
2f36825b | 4245 | } |
371fd7e7 | 4246 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 4247 | } |
8f4d37ec | 4248 | |
2069dd75 | 4249 | for_each_sched_entity(se) { |
0f317143 | 4250 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4251 | cfs_rq->h_nr_running--; |
2069dd75 | 4252 | |
85dac906 PT |
4253 | if (cfs_rq_throttled(cfs_rq)) |
4254 | break; | |
4255 | ||
17bc14b7 | 4256 | update_cfs_shares(cfs_rq); |
9ee474f5 | 4257 | update_entity_load_avg(se, 1); |
2069dd75 PZ |
4258 | } |
4259 | ||
18bf2805 | 4260 | if (!se) { |
72465447 | 4261 | sub_nr_running(rq, 1); |
18bf2805 BS |
4262 | update_rq_runnable_avg(rq, 1); |
4263 | } | |
a4c2f00f | 4264 | hrtick_update(rq); |
bf0f6f24 IM |
4265 | } |
4266 | ||
e7693a36 | 4267 | #ifdef CONFIG_SMP |
029632fb PZ |
4268 | /* Used instead of source_load when we know the type == 0 */ |
4269 | static unsigned long weighted_cpuload(const int cpu) | |
4270 | { | |
b92486cb | 4271 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
029632fb PZ |
4272 | } |
4273 | ||
4274 | /* | |
4275 | * Return a low guess at the load of a migration-source cpu weighted | |
4276 | * according to the scheduling class and "nice" value. | |
4277 | * | |
4278 | * We want to under-estimate the load of migration sources, to | |
4279 | * balance conservatively. | |
4280 | */ | |
4281 | static unsigned long source_load(int cpu, int type) | |
4282 | { | |
4283 | struct rq *rq = cpu_rq(cpu); | |
4284 | unsigned long total = weighted_cpuload(cpu); | |
4285 | ||
4286 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4287 | return total; | |
4288 | ||
4289 | return min(rq->cpu_load[type-1], total); | |
4290 | } | |
4291 | ||
4292 | /* | |
4293 | * Return a high guess at the load of a migration-target cpu weighted | |
4294 | * according to the scheduling class and "nice" value. | |
4295 | */ | |
4296 | static unsigned long target_load(int cpu, int type) | |
4297 | { | |
4298 | struct rq *rq = cpu_rq(cpu); | |
4299 | unsigned long total = weighted_cpuload(cpu); | |
4300 | ||
4301 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4302 | return total; | |
4303 | ||
4304 | return max(rq->cpu_load[type-1], total); | |
4305 | } | |
4306 | ||
ced549fa | 4307 | static unsigned long capacity_of(int cpu) |
029632fb | 4308 | { |
ced549fa | 4309 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
4310 | } |
4311 | ||
4312 | static unsigned long cpu_avg_load_per_task(int cpu) | |
4313 | { | |
4314 | struct rq *rq = cpu_rq(cpu); | |
65fdac08 | 4315 | unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running); |
b92486cb | 4316 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
029632fb PZ |
4317 | |
4318 | if (nr_running) | |
b92486cb | 4319 | return load_avg / nr_running; |
029632fb PZ |
4320 | |
4321 | return 0; | |
4322 | } | |
4323 | ||
62470419 MW |
4324 | static void record_wakee(struct task_struct *p) |
4325 | { | |
4326 | /* | |
4327 | * Rough decay (wiping) for cost saving, don't worry | |
4328 | * about the boundary, really active task won't care | |
4329 | * about the loss. | |
4330 | */ | |
2538d960 | 4331 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { |
096aa338 | 4332 | current->wakee_flips >>= 1; |
62470419 MW |
4333 | current->wakee_flip_decay_ts = jiffies; |
4334 | } | |
4335 | ||
4336 | if (current->last_wakee != p) { | |
4337 | current->last_wakee = p; | |
4338 | current->wakee_flips++; | |
4339 | } | |
4340 | } | |
098fb9db | 4341 | |
74f8e4b2 | 4342 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
4343 | { |
4344 | struct sched_entity *se = &p->se; | |
4345 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
4346 | u64 min_vruntime; |
4347 | ||
4348 | #ifndef CONFIG_64BIT | |
4349 | u64 min_vruntime_copy; | |
88ec22d3 | 4350 | |
3fe1698b PZ |
4351 | do { |
4352 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
4353 | smp_rmb(); | |
4354 | min_vruntime = cfs_rq->min_vruntime; | |
4355 | } while (min_vruntime != min_vruntime_copy); | |
4356 | #else | |
4357 | min_vruntime = cfs_rq->min_vruntime; | |
4358 | #endif | |
88ec22d3 | 4359 | |
3fe1698b | 4360 | se->vruntime -= min_vruntime; |
62470419 | 4361 | record_wakee(p); |
88ec22d3 PZ |
4362 | } |
4363 | ||
bb3469ac | 4364 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
4365 | /* |
4366 | * effective_load() calculates the load change as seen from the root_task_group | |
4367 | * | |
4368 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
4369 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
4370 | * can calculate the shift in shares. | |
cf5f0acf PZ |
4371 | * |
4372 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
4373 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
4374 | * total group weight. | |
4375 | * | |
4376 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
4377 | * distribution (s_i) using: | |
4378 | * | |
4379 | * s_i = rw_i / \Sum rw_j (1) | |
4380 | * | |
4381 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
4382 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
4383 | * shares distribution (s_i): | |
4384 | * | |
4385 | * rw_i = { 2, 4, 1, 0 } | |
4386 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
4387 | * | |
4388 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
4389 | * task used to run on and the CPU the waker is running on), we need to | |
4390 | * compute the effect of waking a task on either CPU and, in case of a sync | |
4391 | * wakeup, compute the effect of the current task going to sleep. | |
4392 | * | |
4393 | * So for a change of @wl to the local @cpu with an overall group weight change | |
4394 | * of @wl we can compute the new shares distribution (s'_i) using: | |
4395 | * | |
4396 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
4397 | * | |
4398 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
4399 | * differences in waking a task to CPU 0. The additional task changes the | |
4400 | * weight and shares distributions like: | |
4401 | * | |
4402 | * rw'_i = { 3, 4, 1, 0 } | |
4403 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
4404 | * | |
4405 | * We can then compute the difference in effective weight by using: | |
4406 | * | |
4407 | * dw_i = S * (s'_i - s_i) (3) | |
4408 | * | |
4409 | * Where 'S' is the group weight as seen by its parent. | |
4410 | * | |
4411 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
4412 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
4413 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 4414 | */ |
2069dd75 | 4415 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 4416 | { |
4be9daaa | 4417 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 4418 | |
9722c2da | 4419 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
4420 | return wl; |
4421 | ||
4be9daaa | 4422 | for_each_sched_entity(se) { |
cf5f0acf | 4423 | long w, W; |
4be9daaa | 4424 | |
977dda7c | 4425 | tg = se->my_q->tg; |
bb3469ac | 4426 | |
cf5f0acf PZ |
4427 | /* |
4428 | * W = @wg + \Sum rw_j | |
4429 | */ | |
4430 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 4431 | |
cf5f0acf PZ |
4432 | /* |
4433 | * w = rw_i + @wl | |
4434 | */ | |
4435 | w = se->my_q->load.weight + wl; | |
940959e9 | 4436 | |
cf5f0acf PZ |
4437 | /* |
4438 | * wl = S * s'_i; see (2) | |
4439 | */ | |
4440 | if (W > 0 && w < W) | |
32a8df4e | 4441 | wl = (w * (long)tg->shares) / W; |
977dda7c PT |
4442 | else |
4443 | wl = tg->shares; | |
940959e9 | 4444 | |
cf5f0acf PZ |
4445 | /* |
4446 | * Per the above, wl is the new se->load.weight value; since | |
4447 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
4448 | * calc_cfs_shares(). | |
4449 | */ | |
977dda7c PT |
4450 | if (wl < MIN_SHARES) |
4451 | wl = MIN_SHARES; | |
cf5f0acf PZ |
4452 | |
4453 | /* | |
4454 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
4455 | */ | |
977dda7c | 4456 | wl -= se->load.weight; |
cf5f0acf PZ |
4457 | |
4458 | /* | |
4459 | * Recursively apply this logic to all parent groups to compute | |
4460 | * the final effective load change on the root group. Since | |
4461 | * only the @tg group gets extra weight, all parent groups can | |
4462 | * only redistribute existing shares. @wl is the shift in shares | |
4463 | * resulting from this level per the above. | |
4464 | */ | |
4be9daaa | 4465 | wg = 0; |
4be9daaa | 4466 | } |
bb3469ac | 4467 | |
4be9daaa | 4468 | return wl; |
bb3469ac PZ |
4469 | } |
4470 | #else | |
4be9daaa | 4471 | |
58d081b5 | 4472 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 4473 | { |
83378269 | 4474 | return wl; |
bb3469ac | 4475 | } |
4be9daaa | 4476 | |
bb3469ac PZ |
4477 | #endif |
4478 | ||
62470419 MW |
4479 | static int wake_wide(struct task_struct *p) |
4480 | { | |
7d9ffa89 | 4481 | int factor = this_cpu_read(sd_llc_size); |
62470419 MW |
4482 | |
4483 | /* | |
4484 | * Yeah, it's the switching-frequency, could means many wakee or | |
4485 | * rapidly switch, use factor here will just help to automatically | |
4486 | * adjust the loose-degree, so bigger node will lead to more pull. | |
4487 | */ | |
4488 | if (p->wakee_flips > factor) { | |
4489 | /* | |
4490 | * wakee is somewhat hot, it needs certain amount of cpu | |
4491 | * resource, so if waker is far more hot, prefer to leave | |
4492 | * it alone. | |
4493 | */ | |
4494 | if (current->wakee_flips > (factor * p->wakee_flips)) | |
4495 | return 1; | |
4496 | } | |
4497 | ||
4498 | return 0; | |
4499 | } | |
4500 | ||
c88d5910 | 4501 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 4502 | { |
e37b6a7b | 4503 | s64 this_load, load; |
bd61c98f | 4504 | s64 this_eff_load, prev_eff_load; |
c88d5910 | 4505 | int idx, this_cpu, prev_cpu; |
c88d5910 | 4506 | struct task_group *tg; |
83378269 | 4507 | unsigned long weight; |
b3137bc8 | 4508 | int balanced; |
098fb9db | 4509 | |
62470419 MW |
4510 | /* |
4511 | * If we wake multiple tasks be careful to not bounce | |
4512 | * ourselves around too much. | |
4513 | */ | |
4514 | if (wake_wide(p)) | |
4515 | return 0; | |
4516 | ||
c88d5910 PZ |
4517 | idx = sd->wake_idx; |
4518 | this_cpu = smp_processor_id(); | |
4519 | prev_cpu = task_cpu(p); | |
4520 | load = source_load(prev_cpu, idx); | |
4521 | this_load = target_load(this_cpu, idx); | |
098fb9db | 4522 | |
b3137bc8 MG |
4523 | /* |
4524 | * If sync wakeup then subtract the (maximum possible) | |
4525 | * effect of the currently running task from the load | |
4526 | * of the current CPU: | |
4527 | */ | |
83378269 PZ |
4528 | if (sync) { |
4529 | tg = task_group(current); | |
4530 | weight = current->se.load.weight; | |
4531 | ||
c88d5910 | 4532 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
4533 | load += effective_load(tg, prev_cpu, 0, -weight); |
4534 | } | |
b3137bc8 | 4535 | |
83378269 PZ |
4536 | tg = task_group(p); |
4537 | weight = p->se.load.weight; | |
b3137bc8 | 4538 | |
71a29aa7 PZ |
4539 | /* |
4540 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
4541 | * due to the sync cause above having dropped this_load to 0, we'll |
4542 | * always have an imbalance, but there's really nothing you can do | |
4543 | * about that, so that's good too. | |
71a29aa7 PZ |
4544 | * |
4545 | * Otherwise check if either cpus are near enough in load to allow this | |
4546 | * task to be woken on this_cpu. | |
4547 | */ | |
bd61c98f VG |
4548 | this_eff_load = 100; |
4549 | this_eff_load *= capacity_of(prev_cpu); | |
e51fd5e2 | 4550 | |
bd61c98f VG |
4551 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; |
4552 | prev_eff_load *= capacity_of(this_cpu); | |
e51fd5e2 | 4553 | |
bd61c98f | 4554 | if (this_load > 0) { |
e51fd5e2 PZ |
4555 | this_eff_load *= this_load + |
4556 | effective_load(tg, this_cpu, weight, weight); | |
4557 | ||
e51fd5e2 | 4558 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); |
bd61c98f | 4559 | } |
e51fd5e2 | 4560 | |
bd61c98f | 4561 | balanced = this_eff_load <= prev_eff_load; |
098fb9db | 4562 | |
41acab88 | 4563 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db | 4564 | |
05bfb65f VG |
4565 | if (!balanced) |
4566 | return 0; | |
098fb9db | 4567 | |
05bfb65f VG |
4568 | schedstat_inc(sd, ttwu_move_affine); |
4569 | schedstat_inc(p, se.statistics.nr_wakeups_affine); | |
4570 | ||
4571 | return 1; | |
098fb9db IM |
4572 | } |
4573 | ||
aaee1203 PZ |
4574 | /* |
4575 | * find_idlest_group finds and returns the least busy CPU group within the | |
4576 | * domain. | |
4577 | */ | |
4578 | static struct sched_group * | |
78e7ed53 | 4579 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 4580 | int this_cpu, int sd_flag) |
e7693a36 | 4581 | { |
b3bd3de6 | 4582 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 4583 | unsigned long min_load = ULONG_MAX, this_load = 0; |
c44f2a02 | 4584 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4585 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 4586 | |
c44f2a02 VG |
4587 | if (sd_flag & SD_BALANCE_WAKE) |
4588 | load_idx = sd->wake_idx; | |
4589 | ||
aaee1203 PZ |
4590 | do { |
4591 | unsigned long load, avg_load; | |
4592 | int local_group; | |
4593 | int i; | |
e7693a36 | 4594 | |
aaee1203 PZ |
4595 | /* Skip over this group if it has no CPUs allowed */ |
4596 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 4597 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
4598 | continue; |
4599 | ||
4600 | local_group = cpumask_test_cpu(this_cpu, | |
4601 | sched_group_cpus(group)); | |
4602 | ||
4603 | /* Tally up the load of all CPUs in the group */ | |
4604 | avg_load = 0; | |
4605 | ||
4606 | for_each_cpu(i, sched_group_cpus(group)) { | |
4607 | /* Bias balancing toward cpus of our domain */ | |
4608 | if (local_group) | |
4609 | load = source_load(i, load_idx); | |
4610 | else | |
4611 | load = target_load(i, load_idx); | |
4612 | ||
4613 | avg_load += load; | |
4614 | } | |
4615 | ||
63b2ca30 | 4616 | /* Adjust by relative CPU capacity of the group */ |
ca8ce3d0 | 4617 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; |
aaee1203 PZ |
4618 | |
4619 | if (local_group) { | |
4620 | this_load = avg_load; | |
aaee1203 PZ |
4621 | } else if (avg_load < min_load) { |
4622 | min_load = avg_load; | |
4623 | idlest = group; | |
4624 | } | |
4625 | } while (group = group->next, group != sd->groups); | |
4626 | ||
4627 | if (!idlest || 100*this_load < imbalance*min_load) | |
4628 | return NULL; | |
4629 | return idlest; | |
4630 | } | |
4631 | ||
4632 | /* | |
4633 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
4634 | */ | |
4635 | static int | |
4636 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
4637 | { | |
4638 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
4639 | unsigned int min_exit_latency = UINT_MAX; |
4640 | u64 latest_idle_timestamp = 0; | |
4641 | int least_loaded_cpu = this_cpu; | |
4642 | int shallowest_idle_cpu = -1; | |
aaee1203 PZ |
4643 | int i; |
4644 | ||
4645 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 4646 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
83a0a96a NP |
4647 | if (idle_cpu(i)) { |
4648 | struct rq *rq = cpu_rq(i); | |
4649 | struct cpuidle_state *idle = idle_get_state(rq); | |
4650 | if (idle && idle->exit_latency < min_exit_latency) { | |
4651 | /* | |
4652 | * We give priority to a CPU whose idle state | |
4653 | * has the smallest exit latency irrespective | |
4654 | * of any idle timestamp. | |
4655 | */ | |
4656 | min_exit_latency = idle->exit_latency; | |
4657 | latest_idle_timestamp = rq->idle_stamp; | |
4658 | shallowest_idle_cpu = i; | |
4659 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
4660 | rq->idle_stamp > latest_idle_timestamp) { | |
4661 | /* | |
4662 | * If equal or no active idle state, then | |
4663 | * the most recently idled CPU might have | |
4664 | * a warmer cache. | |
4665 | */ | |
4666 | latest_idle_timestamp = rq->idle_stamp; | |
4667 | shallowest_idle_cpu = i; | |
4668 | } | |
9f96742a | 4669 | } else if (shallowest_idle_cpu == -1) { |
83a0a96a NP |
4670 | load = weighted_cpuload(i); |
4671 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
4672 | min_load = load; | |
4673 | least_loaded_cpu = i; | |
4674 | } | |
e7693a36 GH |
4675 | } |
4676 | } | |
4677 | ||
83a0a96a | 4678 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 4679 | } |
e7693a36 | 4680 | |
a50bde51 PZ |
4681 | /* |
4682 | * Try and locate an idle CPU in the sched_domain. | |
4683 | */ | |
99bd5e2f | 4684 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 4685 | { |
99bd5e2f | 4686 | struct sched_domain *sd; |
37407ea7 | 4687 | struct sched_group *sg; |
e0a79f52 | 4688 | int i = task_cpu(p); |
a50bde51 | 4689 | |
e0a79f52 MG |
4690 | if (idle_cpu(target)) |
4691 | return target; | |
99bd5e2f SS |
4692 | |
4693 | /* | |
e0a79f52 | 4694 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 4695 | */ |
e0a79f52 MG |
4696 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4697 | return i; | |
a50bde51 PZ |
4698 | |
4699 | /* | |
37407ea7 | 4700 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 4701 | */ |
518cd623 | 4702 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 4703 | for_each_lower_domain(sd) { |
37407ea7 LT |
4704 | sg = sd->groups; |
4705 | do { | |
4706 | if (!cpumask_intersects(sched_group_cpus(sg), | |
4707 | tsk_cpus_allowed(p))) | |
4708 | goto next; | |
4709 | ||
4710 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 4711 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
4712 | goto next; |
4713 | } | |
970e1789 | 4714 | |
37407ea7 LT |
4715 | target = cpumask_first_and(sched_group_cpus(sg), |
4716 | tsk_cpus_allowed(p)); | |
4717 | goto done; | |
4718 | next: | |
4719 | sg = sg->next; | |
4720 | } while (sg != sd->groups); | |
4721 | } | |
4722 | done: | |
a50bde51 PZ |
4723 | return target; |
4724 | } | |
4725 | ||
aaee1203 | 4726 | /* |
de91b9cb MR |
4727 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
4728 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
4729 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 4730 | * |
de91b9cb MR |
4731 | * Balances load by selecting the idlest cpu in the idlest group, or under |
4732 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 4733 | * |
de91b9cb | 4734 | * Returns the target cpu number. |
aaee1203 PZ |
4735 | * |
4736 | * preempt must be disabled. | |
4737 | */ | |
0017d735 | 4738 | static int |
ac66f547 | 4739 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 4740 | { |
29cd8bae | 4741 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 4742 | int cpu = smp_processor_id(); |
c88d5910 | 4743 | int new_cpu = cpu; |
99bd5e2f | 4744 | int want_affine = 0; |
5158f4e4 | 4745 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 4746 | |
a8edd075 KT |
4747 | if (sd_flag & SD_BALANCE_WAKE) |
4748 | want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); | |
aaee1203 | 4749 | |
dce840a0 | 4750 | rcu_read_lock(); |
aaee1203 | 4751 | for_each_domain(cpu, tmp) { |
e4f42888 PZ |
4752 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4753 | continue; | |
4754 | ||
fe3bcfe1 | 4755 | /* |
99bd5e2f SS |
4756 | * If both cpu and prev_cpu are part of this domain, |
4757 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 4758 | */ |
99bd5e2f SS |
4759 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4760 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
4761 | affine_sd = tmp; | |
29cd8bae | 4762 | break; |
f03542a7 | 4763 | } |
29cd8bae | 4764 | |
f03542a7 | 4765 | if (tmp->flags & sd_flag) |
29cd8bae PZ |
4766 | sd = tmp; |
4767 | } | |
4768 | ||
8bf21433 RR |
4769 | if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
4770 | prev_cpu = cpu; | |
dce840a0 | 4771 | |
8bf21433 | 4772 | if (sd_flag & SD_BALANCE_WAKE) { |
dce840a0 PZ |
4773 | new_cpu = select_idle_sibling(p, prev_cpu); |
4774 | goto unlock; | |
8b911acd | 4775 | } |
e7693a36 | 4776 | |
aaee1203 PZ |
4777 | while (sd) { |
4778 | struct sched_group *group; | |
c88d5910 | 4779 | int weight; |
098fb9db | 4780 | |
0763a660 | 4781 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
4782 | sd = sd->child; |
4783 | continue; | |
4784 | } | |
098fb9db | 4785 | |
c44f2a02 | 4786 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
4787 | if (!group) { |
4788 | sd = sd->child; | |
4789 | continue; | |
4790 | } | |
4ae7d5ce | 4791 | |
d7c33c49 | 4792 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
4793 | if (new_cpu == -1 || new_cpu == cpu) { |
4794 | /* Now try balancing at a lower domain level of cpu */ | |
4795 | sd = sd->child; | |
4796 | continue; | |
e7693a36 | 4797 | } |
aaee1203 PZ |
4798 | |
4799 | /* Now try balancing at a lower domain level of new_cpu */ | |
4800 | cpu = new_cpu; | |
669c55e9 | 4801 | weight = sd->span_weight; |
aaee1203 PZ |
4802 | sd = NULL; |
4803 | for_each_domain(cpu, tmp) { | |
669c55e9 | 4804 | if (weight <= tmp->span_weight) |
aaee1203 | 4805 | break; |
0763a660 | 4806 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
4807 | sd = tmp; |
4808 | } | |
4809 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 4810 | } |
dce840a0 PZ |
4811 | unlock: |
4812 | rcu_read_unlock(); | |
e7693a36 | 4813 | |
c88d5910 | 4814 | return new_cpu; |
e7693a36 | 4815 | } |
0a74bef8 PT |
4816 | |
4817 | /* | |
4818 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
4819 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
4820 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | |
4821 | * other assumptions, including the state of rq->lock, should be made. | |
4822 | */ | |
4823 | static void | |
4824 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | |
4825 | { | |
aff3e498 PT |
4826 | struct sched_entity *se = &p->se; |
4827 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4828 | ||
4829 | /* | |
4830 | * Load tracking: accumulate removed load so that it can be processed | |
4831 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | |
4832 | * to blocked load iff they have a positive decay-count. It can never | |
4833 | * be negative here since on-rq tasks have decay-count == 0. | |
4834 | */ | |
4835 | if (se->avg.decay_count) { | |
4836 | se->avg.decay_count = -__synchronize_entity_decay(se); | |
2509940f AS |
4837 | atomic_long_add(se->avg.load_avg_contrib, |
4838 | &cfs_rq->removed_load); | |
aff3e498 | 4839 | } |
3944a927 BS |
4840 | |
4841 | /* We have migrated, no longer consider this task hot */ | |
4842 | se->exec_start = 0; | |
0a74bef8 | 4843 | } |
e7693a36 GH |
4844 | #endif /* CONFIG_SMP */ |
4845 | ||
e52fb7c0 PZ |
4846 | static unsigned long |
4847 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
4848 | { |
4849 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
4850 | ||
4851 | /* | |
e52fb7c0 PZ |
4852 | * Since its curr running now, convert the gran from real-time |
4853 | * to virtual-time in his units. | |
13814d42 MG |
4854 | * |
4855 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
4856 | * they get preempted easier. That is, if 'se' < 'curr' then | |
4857 | * the resulting gran will be larger, therefore penalizing the | |
4858 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
4859 | * be smaller, again penalizing the lighter task. | |
4860 | * | |
4861 | * This is especially important for buddies when the leftmost | |
4862 | * task is higher priority than the buddy. | |
0bbd3336 | 4863 | */ |
f4ad9bd2 | 4864 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
4865 | } |
4866 | ||
464b7527 PZ |
4867 | /* |
4868 | * Should 'se' preempt 'curr'. | |
4869 | * | |
4870 | * |s1 | |
4871 | * |s2 | |
4872 | * |s3 | |
4873 | * g | |
4874 | * |<--->|c | |
4875 | * | |
4876 | * w(c, s1) = -1 | |
4877 | * w(c, s2) = 0 | |
4878 | * w(c, s3) = 1 | |
4879 | * | |
4880 | */ | |
4881 | static int | |
4882 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
4883 | { | |
4884 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
4885 | ||
4886 | if (vdiff <= 0) | |
4887 | return -1; | |
4888 | ||
e52fb7c0 | 4889 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
4890 | if (vdiff > gran) |
4891 | return 1; | |
4892 | ||
4893 | return 0; | |
4894 | } | |
4895 | ||
02479099 PZ |
4896 | static void set_last_buddy(struct sched_entity *se) |
4897 | { | |
69c80f3e VP |
4898 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4899 | return; | |
4900 | ||
4901 | for_each_sched_entity(se) | |
4902 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
4903 | } |
4904 | ||
4905 | static void set_next_buddy(struct sched_entity *se) | |
4906 | { | |
69c80f3e VP |
4907 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4908 | return; | |
4909 | ||
4910 | for_each_sched_entity(se) | |
4911 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
4912 | } |
4913 | ||
ac53db59 RR |
4914 | static void set_skip_buddy(struct sched_entity *se) |
4915 | { | |
69c80f3e VP |
4916 | for_each_sched_entity(se) |
4917 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
4918 | } |
4919 | ||
bf0f6f24 IM |
4920 | /* |
4921 | * Preempt the current task with a newly woken task if needed: | |
4922 | */ | |
5a9b86f6 | 4923 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
4924 | { |
4925 | struct task_struct *curr = rq->curr; | |
8651a86c | 4926 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 4927 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 4928 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 4929 | int next_buddy_marked = 0; |
bf0f6f24 | 4930 | |
4ae7d5ce IM |
4931 | if (unlikely(se == pse)) |
4932 | return; | |
4933 | ||
5238cdd3 | 4934 | /* |
163122b7 | 4935 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
4936 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4937 | * lead to a throttle). This both saves work and prevents false | |
4938 | * next-buddy nomination below. | |
4939 | */ | |
4940 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
4941 | return; | |
4942 | ||
2f36825b | 4943 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 4944 | set_next_buddy(pse); |
2f36825b VP |
4945 | next_buddy_marked = 1; |
4946 | } | |
57fdc26d | 4947 | |
aec0a514 BR |
4948 | /* |
4949 | * We can come here with TIF_NEED_RESCHED already set from new task | |
4950 | * wake up path. | |
5238cdd3 PT |
4951 | * |
4952 | * Note: this also catches the edge-case of curr being in a throttled | |
4953 | * group (e.g. via set_curr_task), since update_curr() (in the | |
4954 | * enqueue of curr) will have resulted in resched being set. This | |
4955 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
4956 | * below. | |
aec0a514 BR |
4957 | */ |
4958 | if (test_tsk_need_resched(curr)) | |
4959 | return; | |
4960 | ||
a2f5c9ab DH |
4961 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4962 | if (unlikely(curr->policy == SCHED_IDLE) && | |
4963 | likely(p->policy != SCHED_IDLE)) | |
4964 | goto preempt; | |
4965 | ||
91c234b4 | 4966 | /* |
a2f5c9ab DH |
4967 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4968 | * is driven by the tick): | |
91c234b4 | 4969 | */ |
8ed92e51 | 4970 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 4971 | return; |
bf0f6f24 | 4972 | |
464b7527 | 4973 | find_matching_se(&se, &pse); |
9bbd7374 | 4974 | update_curr(cfs_rq_of(se)); |
002f128b | 4975 | BUG_ON(!pse); |
2f36825b VP |
4976 | if (wakeup_preempt_entity(se, pse) == 1) { |
4977 | /* | |
4978 | * Bias pick_next to pick the sched entity that is | |
4979 | * triggering this preemption. | |
4980 | */ | |
4981 | if (!next_buddy_marked) | |
4982 | set_next_buddy(pse); | |
3a7e73a2 | 4983 | goto preempt; |
2f36825b | 4984 | } |
464b7527 | 4985 | |
3a7e73a2 | 4986 | return; |
a65ac745 | 4987 | |
3a7e73a2 | 4988 | preempt: |
8875125e | 4989 | resched_curr(rq); |
3a7e73a2 PZ |
4990 | /* |
4991 | * Only set the backward buddy when the current task is still | |
4992 | * on the rq. This can happen when a wakeup gets interleaved | |
4993 | * with schedule on the ->pre_schedule() or idle_balance() | |
4994 | * point, either of which can * drop the rq lock. | |
4995 | * | |
4996 | * Also, during early boot the idle thread is in the fair class, | |
4997 | * for obvious reasons its a bad idea to schedule back to it. | |
4998 | */ | |
4999 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
5000 | return; | |
5001 | ||
5002 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
5003 | set_last_buddy(se); | |
bf0f6f24 IM |
5004 | } |
5005 | ||
606dba2e PZ |
5006 | static struct task_struct * |
5007 | pick_next_task_fair(struct rq *rq, struct task_struct *prev) | |
bf0f6f24 IM |
5008 | { |
5009 | struct cfs_rq *cfs_rq = &rq->cfs; | |
5010 | struct sched_entity *se; | |
678d5718 | 5011 | struct task_struct *p; |
37e117c0 | 5012 | int new_tasks; |
678d5718 | 5013 | |
6e83125c | 5014 | again: |
678d5718 PZ |
5015 | #ifdef CONFIG_FAIR_GROUP_SCHED |
5016 | if (!cfs_rq->nr_running) | |
38033c37 | 5017 | goto idle; |
678d5718 | 5018 | |
3f1d2a31 | 5019 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
5020 | goto simple; |
5021 | ||
5022 | /* | |
5023 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
5024 | * likely that a next task is from the same cgroup as the current. | |
5025 | * | |
5026 | * Therefore attempt to avoid putting and setting the entire cgroup | |
5027 | * hierarchy, only change the part that actually changes. | |
5028 | */ | |
5029 | ||
5030 | do { | |
5031 | struct sched_entity *curr = cfs_rq->curr; | |
5032 | ||
5033 | /* | |
5034 | * Since we got here without doing put_prev_entity() we also | |
5035 | * have to consider cfs_rq->curr. If it is still a runnable | |
5036 | * entity, update_curr() will update its vruntime, otherwise | |
5037 | * forget we've ever seen it. | |
5038 | */ | |
5039 | if (curr && curr->on_rq) | |
5040 | update_curr(cfs_rq); | |
5041 | else | |
5042 | curr = NULL; | |
5043 | ||
5044 | /* | |
5045 | * This call to check_cfs_rq_runtime() will do the throttle and | |
5046 | * dequeue its entity in the parent(s). Therefore the 'simple' | |
5047 | * nr_running test will indeed be correct. | |
5048 | */ | |
5049 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
5050 | goto simple; | |
5051 | ||
5052 | se = pick_next_entity(cfs_rq, curr); | |
5053 | cfs_rq = group_cfs_rq(se); | |
5054 | } while (cfs_rq); | |
5055 | ||
5056 | p = task_of(se); | |
5057 | ||
5058 | /* | |
5059 | * Since we haven't yet done put_prev_entity and if the selected task | |
5060 | * is a different task than we started out with, try and touch the | |
5061 | * least amount of cfs_rqs. | |
5062 | */ | |
5063 | if (prev != p) { | |
5064 | struct sched_entity *pse = &prev->se; | |
5065 | ||
5066 | while (!(cfs_rq = is_same_group(se, pse))) { | |
5067 | int se_depth = se->depth; | |
5068 | int pse_depth = pse->depth; | |
5069 | ||
5070 | if (se_depth <= pse_depth) { | |
5071 | put_prev_entity(cfs_rq_of(pse), pse); | |
5072 | pse = parent_entity(pse); | |
5073 | } | |
5074 | if (se_depth >= pse_depth) { | |
5075 | set_next_entity(cfs_rq_of(se), se); | |
5076 | se = parent_entity(se); | |
5077 | } | |
5078 | } | |
5079 | ||
5080 | put_prev_entity(cfs_rq, pse); | |
5081 | set_next_entity(cfs_rq, se); | |
5082 | } | |
5083 | ||
5084 | if (hrtick_enabled(rq)) | |
5085 | hrtick_start_fair(rq, p); | |
5086 | ||
5087 | return p; | |
5088 | simple: | |
5089 | cfs_rq = &rq->cfs; | |
5090 | #endif | |
bf0f6f24 | 5091 | |
36ace27e | 5092 | if (!cfs_rq->nr_running) |
38033c37 | 5093 | goto idle; |
bf0f6f24 | 5094 | |
3f1d2a31 | 5095 | put_prev_task(rq, prev); |
606dba2e | 5096 | |
bf0f6f24 | 5097 | do { |
678d5718 | 5098 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 5099 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
5100 | cfs_rq = group_cfs_rq(se); |
5101 | } while (cfs_rq); | |
5102 | ||
8f4d37ec | 5103 | p = task_of(se); |
678d5718 | 5104 | |
b39e66ea MG |
5105 | if (hrtick_enabled(rq)) |
5106 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
5107 | |
5108 | return p; | |
38033c37 PZ |
5109 | |
5110 | idle: | |
e4aa358b | 5111 | new_tasks = idle_balance(rq); |
37e117c0 PZ |
5112 | /* |
5113 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
5114 | * possible for any higher priority task to appear. In that case we | |
5115 | * must re-start the pick_next_entity() loop. | |
5116 | */ | |
e4aa358b | 5117 | if (new_tasks < 0) |
37e117c0 PZ |
5118 | return RETRY_TASK; |
5119 | ||
e4aa358b | 5120 | if (new_tasks > 0) |
38033c37 | 5121 | goto again; |
38033c37 PZ |
5122 | |
5123 | return NULL; | |
bf0f6f24 IM |
5124 | } |
5125 | ||
5126 | /* | |
5127 | * Account for a descheduled task: | |
5128 | */ | |
31ee529c | 5129 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
5130 | { |
5131 | struct sched_entity *se = &prev->se; | |
5132 | struct cfs_rq *cfs_rq; | |
5133 | ||
5134 | for_each_sched_entity(se) { | |
5135 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 5136 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
5137 | } |
5138 | } | |
5139 | ||
ac53db59 RR |
5140 | /* |
5141 | * sched_yield() is very simple | |
5142 | * | |
5143 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
5144 | */ | |
5145 | static void yield_task_fair(struct rq *rq) | |
5146 | { | |
5147 | struct task_struct *curr = rq->curr; | |
5148 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
5149 | struct sched_entity *se = &curr->se; | |
5150 | ||
5151 | /* | |
5152 | * Are we the only task in the tree? | |
5153 | */ | |
5154 | if (unlikely(rq->nr_running == 1)) | |
5155 | return; | |
5156 | ||
5157 | clear_buddies(cfs_rq, se); | |
5158 | ||
5159 | if (curr->policy != SCHED_BATCH) { | |
5160 | update_rq_clock(rq); | |
5161 | /* | |
5162 | * Update run-time statistics of the 'current'. | |
5163 | */ | |
5164 | update_curr(cfs_rq); | |
916671c0 MG |
5165 | /* |
5166 | * Tell update_rq_clock() that we've just updated, | |
5167 | * so we don't do microscopic update in schedule() | |
5168 | * and double the fastpath cost. | |
5169 | */ | |
9edfbfed | 5170 | rq_clock_skip_update(rq, true); |
ac53db59 RR |
5171 | } |
5172 | ||
5173 | set_skip_buddy(se); | |
5174 | } | |
5175 | ||
d95f4122 MG |
5176 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
5177 | { | |
5178 | struct sched_entity *se = &p->se; | |
5179 | ||
5238cdd3 PT |
5180 | /* throttled hierarchies are not runnable */ |
5181 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
5182 | return false; |
5183 | ||
5184 | /* Tell the scheduler that we'd really like pse to run next. */ | |
5185 | set_next_buddy(se); | |
5186 | ||
d95f4122 MG |
5187 | yield_task_fair(rq); |
5188 | ||
5189 | return true; | |
5190 | } | |
5191 | ||
681f3e68 | 5192 | #ifdef CONFIG_SMP |
bf0f6f24 | 5193 | /************************************************** |
e9c84cb8 PZ |
5194 | * Fair scheduling class load-balancing methods. |
5195 | * | |
5196 | * BASICS | |
5197 | * | |
5198 | * The purpose of load-balancing is to achieve the same basic fairness the | |
5199 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
5200 | * time to each task. This is expressed in the following equation: | |
5201 | * | |
5202 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
5203 | * | |
5204 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
5205 | * W_i,0 is defined as: | |
5206 | * | |
5207 | * W_i,0 = \Sum_j w_i,j (2) | |
5208 | * | |
5209 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
5210 | * is derived from the nice value as per prio_to_weight[]. | |
5211 | * | |
5212 | * The weight average is an exponential decay average of the instantaneous | |
5213 | * weight: | |
5214 | * | |
5215 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
5216 | * | |
ced549fa | 5217 | * C_i is the compute capacity of cpu i, typically it is the |
e9c84cb8 PZ |
5218 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
5219 | * can also include other factors [XXX]. | |
5220 | * | |
5221 | * To achieve this balance we define a measure of imbalance which follows | |
5222 | * directly from (1): | |
5223 | * | |
ced549fa | 5224 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
5225 | * |
5226 | * We them move tasks around to minimize the imbalance. In the continuous | |
5227 | * function space it is obvious this converges, in the discrete case we get | |
5228 | * a few fun cases generally called infeasible weight scenarios. | |
5229 | * | |
5230 | * [XXX expand on: | |
5231 | * - infeasible weights; | |
5232 | * - local vs global optima in the discrete case. ] | |
5233 | * | |
5234 | * | |
5235 | * SCHED DOMAINS | |
5236 | * | |
5237 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
5238 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
5239 | * topology where each level pairs two lower groups (or better). This results | |
5240 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
5241 | * tree to only the first of the previous level and we decrease the frequency | |
5242 | * of load-balance at each level inv. proportional to the number of cpus in | |
5243 | * the groups. | |
5244 | * | |
5245 | * This yields: | |
5246 | * | |
5247 | * log_2 n 1 n | |
5248 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
5249 | * i = 0 2^i 2^i | |
5250 | * `- size of each group | |
5251 | * | | `- number of cpus doing load-balance | |
5252 | * | `- freq | |
5253 | * `- sum over all levels | |
5254 | * | |
5255 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
5256 | * this makes (5) the runtime complexity of the balancer. | |
5257 | * | |
5258 | * An important property here is that each CPU is still (indirectly) connected | |
5259 | * to every other cpu in at most O(log n) steps: | |
5260 | * | |
5261 | * The adjacency matrix of the resulting graph is given by: | |
5262 | * | |
5263 | * log_2 n | |
5264 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
5265 | * k = 0 | |
5266 | * | |
5267 | * And you'll find that: | |
5268 | * | |
5269 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
5270 | * | |
5271 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
5272 | * The task movement gives a factor of O(m), giving a convergence complexity | |
5273 | * of: | |
5274 | * | |
5275 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
5276 | * | |
5277 | * | |
5278 | * WORK CONSERVING | |
5279 | * | |
5280 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
5281 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
5282 | * tree itself instead of relying on other CPUs to bring it work. | |
5283 | * | |
5284 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
5285 | * time. | |
5286 | * | |
5287 | * [XXX more?] | |
5288 | * | |
5289 | * | |
5290 | * CGROUPS | |
5291 | * | |
5292 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
5293 | * | |
5294 | * s_k,i | |
5295 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
5296 | * S_k | |
5297 | * | |
5298 | * Where | |
5299 | * | |
5300 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
5301 | * | |
5302 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
5303 | * | |
5304 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
5305 | * property. | |
5306 | * | |
5307 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
5308 | * rewrite all of this once again.] | |
5309 | */ | |
bf0f6f24 | 5310 | |
ed387b78 HS |
5311 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
5312 | ||
0ec8aa00 PZ |
5313 | enum fbq_type { regular, remote, all }; |
5314 | ||
ddcdf6e7 | 5315 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 5316 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
5317 | #define LBF_DST_PINNED 0x04 |
5318 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
5319 | |
5320 | struct lb_env { | |
5321 | struct sched_domain *sd; | |
5322 | ||
ddcdf6e7 | 5323 | struct rq *src_rq; |
85c1e7da | 5324 | int src_cpu; |
ddcdf6e7 PZ |
5325 | |
5326 | int dst_cpu; | |
5327 | struct rq *dst_rq; | |
5328 | ||
88b8dac0 SV |
5329 | struct cpumask *dst_grpmask; |
5330 | int new_dst_cpu; | |
ddcdf6e7 | 5331 | enum cpu_idle_type idle; |
bd939f45 | 5332 | long imbalance; |
b9403130 MW |
5333 | /* The set of CPUs under consideration for load-balancing */ |
5334 | struct cpumask *cpus; | |
5335 | ||
ddcdf6e7 | 5336 | unsigned int flags; |
367456c7 PZ |
5337 | |
5338 | unsigned int loop; | |
5339 | unsigned int loop_break; | |
5340 | unsigned int loop_max; | |
0ec8aa00 PZ |
5341 | |
5342 | enum fbq_type fbq_type; | |
163122b7 | 5343 | struct list_head tasks; |
ddcdf6e7 PZ |
5344 | }; |
5345 | ||
029632fb PZ |
5346 | /* |
5347 | * Is this task likely cache-hot: | |
5348 | */ | |
5d5e2b1b | 5349 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
5350 | { |
5351 | s64 delta; | |
5352 | ||
e5673f28 KT |
5353 | lockdep_assert_held(&env->src_rq->lock); |
5354 | ||
029632fb PZ |
5355 | if (p->sched_class != &fair_sched_class) |
5356 | return 0; | |
5357 | ||
5358 | if (unlikely(p->policy == SCHED_IDLE)) | |
5359 | return 0; | |
5360 | ||
5361 | /* | |
5362 | * Buddy candidates are cache hot: | |
5363 | */ | |
5d5e2b1b | 5364 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
5365 | (&p->se == cfs_rq_of(&p->se)->next || |
5366 | &p->se == cfs_rq_of(&p->se)->last)) | |
5367 | return 1; | |
5368 | ||
5369 | if (sysctl_sched_migration_cost == -1) | |
5370 | return 1; | |
5371 | if (sysctl_sched_migration_cost == 0) | |
5372 | return 0; | |
5373 | ||
5d5e2b1b | 5374 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
5375 | |
5376 | return delta < (s64)sysctl_sched_migration_cost; | |
5377 | } | |
5378 | ||
3a7053b3 MG |
5379 | #ifdef CONFIG_NUMA_BALANCING |
5380 | /* Returns true if the destination node has incurred more faults */ | |
5381 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | |
5382 | { | |
b1ad065e | 5383 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
3a7053b3 MG |
5384 | int src_nid, dst_nid; |
5385 | ||
44dba3d5 | 5386 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || |
3a7053b3 MG |
5387 | !(env->sd->flags & SD_NUMA)) { |
5388 | return false; | |
5389 | } | |
5390 | ||
5391 | src_nid = cpu_to_node(env->src_cpu); | |
5392 | dst_nid = cpu_to_node(env->dst_cpu); | |
5393 | ||
83e1d2cd | 5394 | if (src_nid == dst_nid) |
3a7053b3 MG |
5395 | return false; |
5396 | ||
b1ad065e RR |
5397 | if (numa_group) { |
5398 | /* Task is already in the group's interleave set. */ | |
5399 | if (node_isset(src_nid, numa_group->active_nodes)) | |
5400 | return false; | |
83e1d2cd | 5401 | |
b1ad065e RR |
5402 | /* Task is moving into the group's interleave set. */ |
5403 | if (node_isset(dst_nid, numa_group->active_nodes)) | |
5404 | return true; | |
83e1d2cd | 5405 | |
b1ad065e RR |
5406 | return group_faults(p, dst_nid) > group_faults(p, src_nid); |
5407 | } | |
5408 | ||
5409 | /* Encourage migration to the preferred node. */ | |
5410 | if (dst_nid == p->numa_preferred_nid) | |
3a7053b3 MG |
5411 | return true; |
5412 | ||
b1ad065e | 5413 | return task_faults(p, dst_nid) > task_faults(p, src_nid); |
3a7053b3 | 5414 | } |
7a0f3083 MG |
5415 | |
5416 | ||
5417 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | |
5418 | { | |
b1ad065e | 5419 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
7a0f3083 MG |
5420 | int src_nid, dst_nid; |
5421 | ||
5422 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | |
5423 | return false; | |
5424 | ||
44dba3d5 | 5425 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
7a0f3083 MG |
5426 | return false; |
5427 | ||
5428 | src_nid = cpu_to_node(env->src_cpu); | |
5429 | dst_nid = cpu_to_node(env->dst_cpu); | |
5430 | ||
83e1d2cd | 5431 | if (src_nid == dst_nid) |
7a0f3083 MG |
5432 | return false; |
5433 | ||
b1ad065e RR |
5434 | if (numa_group) { |
5435 | /* Task is moving within/into the group's interleave set. */ | |
5436 | if (node_isset(dst_nid, numa_group->active_nodes)) | |
5437 | return false; | |
5438 | ||
5439 | /* Task is moving out of the group's interleave set. */ | |
5440 | if (node_isset(src_nid, numa_group->active_nodes)) | |
5441 | return true; | |
5442 | ||
5443 | return group_faults(p, dst_nid) < group_faults(p, src_nid); | |
5444 | } | |
5445 | ||
83e1d2cd MG |
5446 | /* Migrating away from the preferred node is always bad. */ |
5447 | if (src_nid == p->numa_preferred_nid) | |
5448 | return true; | |
5449 | ||
b1ad065e | 5450 | return task_faults(p, dst_nid) < task_faults(p, src_nid); |
7a0f3083 MG |
5451 | } |
5452 | ||
3a7053b3 MG |
5453 | #else |
5454 | static inline bool migrate_improves_locality(struct task_struct *p, | |
5455 | struct lb_env *env) | |
5456 | { | |
5457 | return false; | |
5458 | } | |
7a0f3083 MG |
5459 | |
5460 | static inline bool migrate_degrades_locality(struct task_struct *p, | |
5461 | struct lb_env *env) | |
5462 | { | |
5463 | return false; | |
5464 | } | |
3a7053b3 MG |
5465 | #endif |
5466 | ||
1e3c88bd PZ |
5467 | /* |
5468 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
5469 | */ | |
5470 | static | |
8e45cb54 | 5471 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd PZ |
5472 | { |
5473 | int tsk_cache_hot = 0; | |
e5673f28 KT |
5474 | |
5475 | lockdep_assert_held(&env->src_rq->lock); | |
5476 | ||
1e3c88bd PZ |
5477 | /* |
5478 | * We do not migrate tasks that are: | |
d3198084 | 5479 | * 1) throttled_lb_pair, or |
1e3c88bd | 5480 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
5481 | * 3) running (obviously), or |
5482 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 5483 | */ |
d3198084 JK |
5484 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
5485 | return 0; | |
5486 | ||
ddcdf6e7 | 5487 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 5488 | int cpu; |
88b8dac0 | 5489 | |
41acab88 | 5490 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 5491 | |
6263322c PZ |
5492 | env->flags |= LBF_SOME_PINNED; |
5493 | ||
88b8dac0 SV |
5494 | /* |
5495 | * Remember if this task can be migrated to any other cpu in | |
5496 | * our sched_group. We may want to revisit it if we couldn't | |
5497 | * meet load balance goals by pulling other tasks on src_cpu. | |
5498 | * | |
5499 | * Also avoid computing new_dst_cpu if we have already computed | |
5500 | * one in current iteration. | |
5501 | */ | |
6263322c | 5502 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
5503 | return 0; |
5504 | ||
e02e60c1 JK |
5505 | /* Prevent to re-select dst_cpu via env's cpus */ |
5506 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
5507 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 5508 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
5509 | env->new_dst_cpu = cpu; |
5510 | break; | |
5511 | } | |
88b8dac0 | 5512 | } |
e02e60c1 | 5513 | |
1e3c88bd PZ |
5514 | return 0; |
5515 | } | |
88b8dac0 SV |
5516 | |
5517 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 5518 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 5519 | |
ddcdf6e7 | 5520 | if (task_running(env->src_rq, p)) { |
41acab88 | 5521 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
5522 | return 0; |
5523 | } | |
5524 | ||
5525 | /* | |
5526 | * Aggressive migration if: | |
3a7053b3 MG |
5527 | * 1) destination numa is preferred |
5528 | * 2) task is cache cold, or | |
5529 | * 3) too many balance attempts have failed. | |
1e3c88bd | 5530 | */ |
5d5e2b1b | 5531 | tsk_cache_hot = task_hot(p, env); |
7a0f3083 MG |
5532 | if (!tsk_cache_hot) |
5533 | tsk_cache_hot = migrate_degrades_locality(p, env); | |
3a7053b3 | 5534 | |
7a96c231 KT |
5535 | if (migrate_improves_locality(p, env) || !tsk_cache_hot || |
5536 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { | |
3a7053b3 MG |
5537 | if (tsk_cache_hot) { |
5538 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | |
5539 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
5540 | } | |
1e3c88bd PZ |
5541 | return 1; |
5542 | } | |
5543 | ||
4e2dcb73 ZH |
5544 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
5545 | return 0; | |
1e3c88bd PZ |
5546 | } |
5547 | ||
897c395f | 5548 | /* |
163122b7 KT |
5549 | * detach_task() -- detach the task for the migration specified in env |
5550 | */ | |
5551 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
5552 | { | |
5553 | lockdep_assert_held(&env->src_rq->lock); | |
5554 | ||
5555 | deactivate_task(env->src_rq, p, 0); | |
5556 | p->on_rq = TASK_ON_RQ_MIGRATING; | |
5557 | set_task_cpu(p, env->dst_cpu); | |
5558 | } | |
5559 | ||
897c395f | 5560 | /* |
e5673f28 | 5561 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 5562 | * part of active balancing operations within "domain". |
897c395f | 5563 | * |
e5673f28 | 5564 | * Returns a task if successful and NULL otherwise. |
897c395f | 5565 | */ |
e5673f28 | 5566 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f PZ |
5567 | { |
5568 | struct task_struct *p, *n; | |
897c395f | 5569 | |
e5673f28 KT |
5570 | lockdep_assert_held(&env->src_rq->lock); |
5571 | ||
367456c7 | 5572 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
5573 | if (!can_migrate_task(p, env)) |
5574 | continue; | |
897c395f | 5575 | |
163122b7 | 5576 | detach_task(p, env); |
e5673f28 | 5577 | |
367456c7 | 5578 | /* |
e5673f28 | 5579 | * Right now, this is only the second place where |
163122b7 | 5580 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 5581 | * so we can safely collect stats here rather than |
163122b7 | 5582 | * inside detach_tasks(). |
367456c7 PZ |
5583 | */ |
5584 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
e5673f28 | 5585 | return p; |
897c395f | 5586 | } |
e5673f28 | 5587 | return NULL; |
897c395f PZ |
5588 | } |
5589 | ||
eb95308e PZ |
5590 | static const unsigned int sched_nr_migrate_break = 32; |
5591 | ||
5d6523eb | 5592 | /* |
163122b7 KT |
5593 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
5594 | * busiest_rq, as part of a balancing operation within domain "sd". | |
5d6523eb | 5595 | * |
163122b7 | 5596 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 5597 | */ |
163122b7 | 5598 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 5599 | { |
5d6523eb PZ |
5600 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
5601 | struct task_struct *p; | |
367456c7 | 5602 | unsigned long load; |
163122b7 KT |
5603 | int detached = 0; |
5604 | ||
5605 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 5606 | |
bd939f45 | 5607 | if (env->imbalance <= 0) |
5d6523eb | 5608 | return 0; |
1e3c88bd | 5609 | |
5d6523eb PZ |
5610 | while (!list_empty(tasks)) { |
5611 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
1e3c88bd | 5612 | |
367456c7 PZ |
5613 | env->loop++; |
5614 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 5615 | if (env->loop > env->loop_max) |
367456c7 | 5616 | break; |
5d6523eb PZ |
5617 | |
5618 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 5619 | if (env->loop > env->loop_break) { |
eb95308e | 5620 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 5621 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 5622 | break; |
a195f004 | 5623 | } |
1e3c88bd | 5624 | |
d3198084 | 5625 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
5626 | goto next; |
5627 | ||
5628 | load = task_h_load(p); | |
5d6523eb | 5629 | |
eb95308e | 5630 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
5631 | goto next; |
5632 | ||
bd939f45 | 5633 | if ((load / 2) > env->imbalance) |
367456c7 | 5634 | goto next; |
1e3c88bd | 5635 | |
163122b7 KT |
5636 | detach_task(p, env); |
5637 | list_add(&p->se.group_node, &env->tasks); | |
5638 | ||
5639 | detached++; | |
bd939f45 | 5640 | env->imbalance -= load; |
1e3c88bd PZ |
5641 | |
5642 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
5643 | /* |
5644 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 5645 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
5646 | * the critical section. |
5647 | */ | |
5d6523eb | 5648 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 5649 | break; |
1e3c88bd PZ |
5650 | #endif |
5651 | ||
ee00e66f PZ |
5652 | /* |
5653 | * We only want to steal up to the prescribed amount of | |
5654 | * weighted load. | |
5655 | */ | |
bd939f45 | 5656 | if (env->imbalance <= 0) |
ee00e66f | 5657 | break; |
367456c7 PZ |
5658 | |
5659 | continue; | |
5660 | next: | |
5d6523eb | 5661 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 5662 | } |
5d6523eb | 5663 | |
1e3c88bd | 5664 | /* |
163122b7 KT |
5665 | * Right now, this is one of only two places we collect this stat |
5666 | * so we can safely collect detach_one_task() stats here rather | |
5667 | * than inside detach_one_task(). | |
1e3c88bd | 5668 | */ |
163122b7 | 5669 | schedstat_add(env->sd, lb_gained[env->idle], detached); |
1e3c88bd | 5670 | |
163122b7 KT |
5671 | return detached; |
5672 | } | |
5673 | ||
5674 | /* | |
5675 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
5676 | */ | |
5677 | static void attach_task(struct rq *rq, struct task_struct *p) | |
5678 | { | |
5679 | lockdep_assert_held(&rq->lock); | |
5680 | ||
5681 | BUG_ON(task_rq(p) != rq); | |
5682 | p->on_rq = TASK_ON_RQ_QUEUED; | |
5683 | activate_task(rq, p, 0); | |
5684 | check_preempt_curr(rq, p, 0); | |
5685 | } | |
5686 | ||
5687 | /* | |
5688 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
5689 | * its new rq. | |
5690 | */ | |
5691 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
5692 | { | |
5693 | raw_spin_lock(&rq->lock); | |
5694 | attach_task(rq, p); | |
5695 | raw_spin_unlock(&rq->lock); | |
5696 | } | |
5697 | ||
5698 | /* | |
5699 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
5700 | * new rq. | |
5701 | */ | |
5702 | static void attach_tasks(struct lb_env *env) | |
5703 | { | |
5704 | struct list_head *tasks = &env->tasks; | |
5705 | struct task_struct *p; | |
5706 | ||
5707 | raw_spin_lock(&env->dst_rq->lock); | |
5708 | ||
5709 | while (!list_empty(tasks)) { | |
5710 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
5711 | list_del_init(&p->se.group_node); | |
1e3c88bd | 5712 | |
163122b7 KT |
5713 | attach_task(env->dst_rq, p); |
5714 | } | |
5715 | ||
5716 | raw_spin_unlock(&env->dst_rq->lock); | |
1e3c88bd PZ |
5717 | } |
5718 | ||
230059de | 5719 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9e3081ca PZ |
5720 | /* |
5721 | * update tg->load_weight by folding this cpu's load_avg | |
5722 | */ | |
48a16753 | 5723 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
9e3081ca | 5724 | { |
48a16753 PT |
5725 | struct sched_entity *se = tg->se[cpu]; |
5726 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | |
9e3081ca | 5727 | |
48a16753 PT |
5728 | /* throttled entities do not contribute to load */ |
5729 | if (throttled_hierarchy(cfs_rq)) | |
5730 | return; | |
9e3081ca | 5731 | |
aff3e498 | 5732 | update_cfs_rq_blocked_load(cfs_rq, 1); |
9e3081ca | 5733 | |
82958366 PT |
5734 | if (se) { |
5735 | update_entity_load_avg(se, 1); | |
5736 | /* | |
5737 | * We pivot on our runnable average having decayed to zero for | |
5738 | * list removal. This generally implies that all our children | |
5739 | * have also been removed (modulo rounding error or bandwidth | |
5740 | * control); however, such cases are rare and we can fix these | |
5741 | * at enqueue. | |
5742 | * | |
5743 | * TODO: fix up out-of-order children on enqueue. | |
5744 | */ | |
5745 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | |
5746 | list_del_leaf_cfs_rq(cfs_rq); | |
5747 | } else { | |
48a16753 | 5748 | struct rq *rq = rq_of(cfs_rq); |
82958366 PT |
5749 | update_rq_runnable_avg(rq, rq->nr_running); |
5750 | } | |
9e3081ca PZ |
5751 | } |
5752 | ||
48a16753 | 5753 | static void update_blocked_averages(int cpu) |
9e3081ca | 5754 | { |
9e3081ca | 5755 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
5756 | struct cfs_rq *cfs_rq; |
5757 | unsigned long flags; | |
9e3081ca | 5758 | |
48a16753 PT |
5759 | raw_spin_lock_irqsave(&rq->lock, flags); |
5760 | update_rq_clock(rq); | |
9763b67f PZ |
5761 | /* |
5762 | * Iterates the task_group tree in a bottom up fashion, see | |
5763 | * list_add_leaf_cfs_rq() for details. | |
5764 | */ | |
64660c86 | 5765 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
48a16753 PT |
5766 | /* |
5767 | * Note: We may want to consider periodically releasing | |
5768 | * rq->lock about these updates so that creating many task | |
5769 | * groups does not result in continually extending hold time. | |
5770 | */ | |
5771 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | |
64660c86 | 5772 | } |
48a16753 PT |
5773 | |
5774 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
5775 | } |
5776 | ||
9763b67f | 5777 | /* |
68520796 | 5778 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
5779 | * This needs to be done in a top-down fashion because the load of a child |
5780 | * group is a fraction of its parents load. | |
5781 | */ | |
68520796 | 5782 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 5783 | { |
68520796 VD |
5784 | struct rq *rq = rq_of(cfs_rq); |
5785 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 5786 | unsigned long now = jiffies; |
68520796 | 5787 | unsigned long load; |
a35b6466 | 5788 | |
68520796 | 5789 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
5790 | return; |
5791 | ||
68520796 VD |
5792 | cfs_rq->h_load_next = NULL; |
5793 | for_each_sched_entity(se) { | |
5794 | cfs_rq = cfs_rq_of(se); | |
5795 | cfs_rq->h_load_next = se; | |
5796 | if (cfs_rq->last_h_load_update == now) | |
5797 | break; | |
5798 | } | |
a35b6466 | 5799 | |
68520796 | 5800 | if (!se) { |
7e3115ef | 5801 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
68520796 VD |
5802 | cfs_rq->last_h_load_update = now; |
5803 | } | |
5804 | ||
5805 | while ((se = cfs_rq->h_load_next) != NULL) { | |
5806 | load = cfs_rq->h_load; | |
5807 | load = div64_ul(load * se->avg.load_avg_contrib, | |
5808 | cfs_rq->runnable_load_avg + 1); | |
5809 | cfs_rq = group_cfs_rq(se); | |
5810 | cfs_rq->h_load = load; | |
5811 | cfs_rq->last_h_load_update = now; | |
5812 | } | |
9763b67f PZ |
5813 | } |
5814 | ||
367456c7 | 5815 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 5816 | { |
367456c7 | 5817 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 5818 | |
68520796 | 5819 | update_cfs_rq_h_load(cfs_rq); |
a003a25b AS |
5820 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
5821 | cfs_rq->runnable_load_avg + 1); | |
230059de PZ |
5822 | } |
5823 | #else | |
48a16753 | 5824 | static inline void update_blocked_averages(int cpu) |
9e3081ca PZ |
5825 | { |
5826 | } | |
5827 | ||
367456c7 | 5828 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 5829 | { |
a003a25b | 5830 | return p->se.avg.load_avg_contrib; |
1e3c88bd | 5831 | } |
230059de | 5832 | #endif |
1e3c88bd | 5833 | |
1e3c88bd | 5834 | /********** Helpers for find_busiest_group ************************/ |
caeb178c RR |
5835 | |
5836 | enum group_type { | |
5837 | group_other = 0, | |
5838 | group_imbalanced, | |
5839 | group_overloaded, | |
5840 | }; | |
5841 | ||
1e3c88bd PZ |
5842 | /* |
5843 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
5844 | */ | |
5845 | struct sg_lb_stats { | |
5846 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
5847 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 5848 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 5849 | unsigned long load_per_task; |
63b2ca30 | 5850 | unsigned long group_capacity; |
147c5fc2 | 5851 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
0fedc6c8 | 5852 | unsigned int group_capacity_factor; |
147c5fc2 PZ |
5853 | unsigned int idle_cpus; |
5854 | unsigned int group_weight; | |
caeb178c | 5855 | enum group_type group_type; |
1b6a7495 | 5856 | int group_has_free_capacity; |
0ec8aa00 PZ |
5857 | #ifdef CONFIG_NUMA_BALANCING |
5858 | unsigned int nr_numa_running; | |
5859 | unsigned int nr_preferred_running; | |
5860 | #endif | |
1e3c88bd PZ |
5861 | }; |
5862 | ||
56cf515b JK |
5863 | /* |
5864 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
5865 | * during load balancing. | |
5866 | */ | |
5867 | struct sd_lb_stats { | |
5868 | struct sched_group *busiest; /* Busiest group in this sd */ | |
5869 | struct sched_group *local; /* Local group in this sd */ | |
5870 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 5871 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
5872 | unsigned long avg_load; /* Average load across all groups in sd */ |
5873 | ||
56cf515b | 5874 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 5875 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
5876 | }; |
5877 | ||
147c5fc2 PZ |
5878 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
5879 | { | |
5880 | /* | |
5881 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
5882 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
5883 | * We must however clear busiest_stat::avg_load because | |
5884 | * update_sd_pick_busiest() reads this before assignment. | |
5885 | */ | |
5886 | *sds = (struct sd_lb_stats){ | |
5887 | .busiest = NULL, | |
5888 | .local = NULL, | |
5889 | .total_load = 0UL, | |
63b2ca30 | 5890 | .total_capacity = 0UL, |
147c5fc2 PZ |
5891 | .busiest_stat = { |
5892 | .avg_load = 0UL, | |
caeb178c RR |
5893 | .sum_nr_running = 0, |
5894 | .group_type = group_other, | |
147c5fc2 PZ |
5895 | }, |
5896 | }; | |
5897 | } | |
5898 | ||
1e3c88bd PZ |
5899 | /** |
5900 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
5901 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 5902 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
5903 | * |
5904 | * Return: The load index. | |
1e3c88bd PZ |
5905 | */ |
5906 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
5907 | enum cpu_idle_type idle) | |
5908 | { | |
5909 | int load_idx; | |
5910 | ||
5911 | switch (idle) { | |
5912 | case CPU_NOT_IDLE: | |
5913 | load_idx = sd->busy_idx; | |
5914 | break; | |
5915 | ||
5916 | case CPU_NEWLY_IDLE: | |
5917 | load_idx = sd->newidle_idx; | |
5918 | break; | |
5919 | default: | |
5920 | load_idx = sd->idle_idx; | |
5921 | break; | |
5922 | } | |
5923 | ||
5924 | return load_idx; | |
5925 | } | |
5926 | ||
ced549fa | 5927 | static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 5928 | { |
ca8ce3d0 | 5929 | return SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
5930 | } |
5931 | ||
ca8ce3d0 | 5932 | unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 5933 | { |
ced549fa | 5934 | return default_scale_capacity(sd, cpu); |
1e3c88bd PZ |
5935 | } |
5936 | ||
26bc3c50 | 5937 | static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 5938 | { |
26bc3c50 VG |
5939 | if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) |
5940 | return sd->smt_gain / sd->span_weight; | |
1e3c88bd | 5941 | |
26bc3c50 | 5942 | return SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
5943 | } |
5944 | ||
26bc3c50 | 5945 | unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 5946 | { |
26bc3c50 | 5947 | return default_scale_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
5948 | } |
5949 | ||
ced549fa | 5950 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
5951 | { |
5952 | struct rq *rq = cpu_rq(cpu); | |
b654f7de | 5953 | u64 total, available, age_stamp, avg; |
cadefd3d | 5954 | s64 delta; |
1e3c88bd | 5955 | |
b654f7de PZ |
5956 | /* |
5957 | * Since we're reading these variables without serialization make sure | |
5958 | * we read them once before doing sanity checks on them. | |
5959 | */ | |
5960 | age_stamp = ACCESS_ONCE(rq->age_stamp); | |
5961 | avg = ACCESS_ONCE(rq->rt_avg); | |
cebde6d6 | 5962 | delta = __rq_clock_broken(rq) - age_stamp; |
b654f7de | 5963 | |
cadefd3d PZ |
5964 | if (unlikely(delta < 0)) |
5965 | delta = 0; | |
5966 | ||
5967 | total = sched_avg_period() + delta; | |
aa483808 | 5968 | |
b654f7de | 5969 | if (unlikely(total < avg)) { |
ced549fa | 5970 | /* Ensures that capacity won't end up being negative */ |
aa483808 VP |
5971 | available = 0; |
5972 | } else { | |
b654f7de | 5973 | available = total - avg; |
aa483808 | 5974 | } |
1e3c88bd | 5975 | |
ca8ce3d0 NP |
5976 | if (unlikely((s64)total < SCHED_CAPACITY_SCALE)) |
5977 | total = SCHED_CAPACITY_SCALE; | |
1e3c88bd | 5978 | |
ca8ce3d0 | 5979 | total >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd PZ |
5980 | |
5981 | return div_u64(available, total); | |
5982 | } | |
5983 | ||
ced549fa | 5984 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 5985 | { |
ca8ce3d0 | 5986 | unsigned long capacity = SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
5987 | struct sched_group *sdg = sd->groups; |
5988 | ||
26bc3c50 VG |
5989 | if (sched_feat(ARCH_CAPACITY)) |
5990 | capacity *= arch_scale_cpu_capacity(sd, cpu); | |
5991 | else | |
5992 | capacity *= default_scale_cpu_capacity(sd, cpu); | |
1e3c88bd | 5993 | |
26bc3c50 | 5994 | capacity >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd | 5995 | |
ced549fa | 5996 | sdg->sgc->capacity_orig = capacity; |
9d5efe05 | 5997 | |
5d4dfddd | 5998 | if (sched_feat(ARCH_CAPACITY)) |
ca8ce3d0 | 5999 | capacity *= arch_scale_freq_capacity(sd, cpu); |
9d5efe05 | 6000 | else |
ced549fa | 6001 | capacity *= default_scale_capacity(sd, cpu); |
9d5efe05 | 6002 | |
ca8ce3d0 | 6003 | capacity >>= SCHED_CAPACITY_SHIFT; |
9d5efe05 | 6004 | |
ced549fa | 6005 | capacity *= scale_rt_capacity(cpu); |
ca8ce3d0 | 6006 | capacity >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd | 6007 | |
ced549fa NP |
6008 | if (!capacity) |
6009 | capacity = 1; | |
1e3c88bd | 6010 | |
ced549fa NP |
6011 | cpu_rq(cpu)->cpu_capacity = capacity; |
6012 | sdg->sgc->capacity = capacity; | |
1e3c88bd PZ |
6013 | } |
6014 | ||
63b2ca30 | 6015 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
6016 | { |
6017 | struct sched_domain *child = sd->child; | |
6018 | struct sched_group *group, *sdg = sd->groups; | |
63b2ca30 | 6019 | unsigned long capacity, capacity_orig; |
4ec4412e VG |
6020 | unsigned long interval; |
6021 | ||
6022 | interval = msecs_to_jiffies(sd->balance_interval); | |
6023 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 6024 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
6025 | |
6026 | if (!child) { | |
ced549fa | 6027 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
6028 | return; |
6029 | } | |
6030 | ||
63b2ca30 | 6031 | capacity_orig = capacity = 0; |
1e3c88bd | 6032 | |
74a5ce20 PZ |
6033 | if (child->flags & SD_OVERLAP) { |
6034 | /* | |
6035 | * SD_OVERLAP domains cannot assume that child groups | |
6036 | * span the current group. | |
6037 | */ | |
6038 | ||
863bffc8 | 6039 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
63b2ca30 | 6040 | struct sched_group_capacity *sgc; |
9abf24d4 | 6041 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 6042 | |
9abf24d4 | 6043 | /* |
63b2ca30 | 6044 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
6045 | * gets here before we've attached the domains to the |
6046 | * runqueues. | |
6047 | * | |
ced549fa NP |
6048 | * Use capacity_of(), which is set irrespective of domains |
6049 | * in update_cpu_capacity(). | |
9abf24d4 | 6050 | * |
63b2ca30 | 6051 | * This avoids capacity/capacity_orig from being 0 and |
9abf24d4 SD |
6052 | * causing divide-by-zero issues on boot. |
6053 | * | |
63b2ca30 | 6054 | * Runtime updates will correct capacity_orig. |
9abf24d4 SD |
6055 | */ |
6056 | if (unlikely(!rq->sd)) { | |
ced549fa NP |
6057 | capacity_orig += capacity_of(cpu); |
6058 | capacity += capacity_of(cpu); | |
9abf24d4 SD |
6059 | continue; |
6060 | } | |
863bffc8 | 6061 | |
63b2ca30 NP |
6062 | sgc = rq->sd->groups->sgc; |
6063 | capacity_orig += sgc->capacity_orig; | |
6064 | capacity += sgc->capacity; | |
863bffc8 | 6065 | } |
74a5ce20 PZ |
6066 | } else { |
6067 | /* | |
6068 | * !SD_OVERLAP domains can assume that child groups | |
6069 | * span the current group. | |
6070 | */ | |
6071 | ||
6072 | group = child->groups; | |
6073 | do { | |
63b2ca30 NP |
6074 | capacity_orig += group->sgc->capacity_orig; |
6075 | capacity += group->sgc->capacity; | |
74a5ce20 PZ |
6076 | group = group->next; |
6077 | } while (group != child->groups); | |
6078 | } | |
1e3c88bd | 6079 | |
63b2ca30 NP |
6080 | sdg->sgc->capacity_orig = capacity_orig; |
6081 | sdg->sgc->capacity = capacity; | |
1e3c88bd PZ |
6082 | } |
6083 | ||
9d5efe05 SV |
6084 | /* |
6085 | * Try and fix up capacity for tiny siblings, this is needed when | |
6086 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | |
6087 | * which on its own isn't powerful enough. | |
6088 | * | |
6089 | * See update_sd_pick_busiest() and check_asym_packing(). | |
6090 | */ | |
6091 | static inline int | |
6092 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | |
6093 | { | |
6094 | /* | |
ca8ce3d0 | 6095 | * Only siblings can have significantly less than SCHED_CAPACITY_SCALE |
9d5efe05 | 6096 | */ |
5d4dfddd | 6097 | if (!(sd->flags & SD_SHARE_CPUCAPACITY)) |
9d5efe05 SV |
6098 | return 0; |
6099 | ||
6100 | /* | |
63b2ca30 | 6101 | * If ~90% of the cpu_capacity is still there, we're good. |
9d5efe05 | 6102 | */ |
63b2ca30 | 6103 | if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29) |
9d5efe05 SV |
6104 | return 1; |
6105 | ||
6106 | return 0; | |
6107 | } | |
6108 | ||
30ce5dab PZ |
6109 | /* |
6110 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
6111 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
6112 | * | |
6113 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
6114 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
6115 | * Something like: | |
6116 | * | |
6117 | * { 0 1 2 3 } { 4 5 6 7 } | |
6118 | * * * * * | |
6119 | * | |
6120 | * If we were to balance group-wise we'd place two tasks in the first group and | |
6121 | * two tasks in the second group. Clearly this is undesired as it will overload | |
6122 | * cpu 3 and leave one of the cpus in the second group unused. | |
6123 | * | |
6124 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
6125 | * by noticing the lower domain failed to reach balance and had difficulty |
6126 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
6127 | * |
6128 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 6129 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 6130 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
6131 | * to create an effective group imbalance. |
6132 | * | |
6133 | * This is a somewhat tricky proposition since the next run might not find the | |
6134 | * group imbalance and decide the groups need to be balanced again. A most | |
6135 | * subtle and fragile situation. | |
6136 | */ | |
6137 | ||
6263322c | 6138 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 6139 | { |
63b2ca30 | 6140 | return group->sgc->imbalance; |
30ce5dab PZ |
6141 | } |
6142 | ||
b37d9316 | 6143 | /* |
0fedc6c8 | 6144 | * Compute the group capacity factor. |
b37d9316 | 6145 | * |
ced549fa | 6146 | * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by |
c61037e9 | 6147 | * first dividing out the smt factor and computing the actual number of cores |
63b2ca30 | 6148 | * and limit unit capacity with that. |
b37d9316 | 6149 | */ |
0fedc6c8 | 6150 | static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group) |
b37d9316 | 6151 | { |
0fedc6c8 | 6152 | unsigned int capacity_factor, smt, cpus; |
63b2ca30 | 6153 | unsigned int capacity, capacity_orig; |
c61037e9 | 6154 | |
63b2ca30 NP |
6155 | capacity = group->sgc->capacity; |
6156 | capacity_orig = group->sgc->capacity_orig; | |
c61037e9 | 6157 | cpus = group->group_weight; |
b37d9316 | 6158 | |
63b2ca30 | 6159 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */ |
ca8ce3d0 | 6160 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig); |
0fedc6c8 | 6161 | capacity_factor = cpus / smt; /* cores */ |
b37d9316 | 6162 | |
63b2ca30 | 6163 | capacity_factor = min_t(unsigned, |
ca8ce3d0 | 6164 | capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE)); |
0fedc6c8 NP |
6165 | if (!capacity_factor) |
6166 | capacity_factor = fix_small_capacity(env->sd, group); | |
b37d9316 | 6167 | |
0fedc6c8 | 6168 | return capacity_factor; |
b37d9316 PZ |
6169 | } |
6170 | ||
caeb178c RR |
6171 | static enum group_type |
6172 | group_classify(struct sched_group *group, struct sg_lb_stats *sgs) | |
6173 | { | |
6174 | if (sgs->sum_nr_running > sgs->group_capacity_factor) | |
6175 | return group_overloaded; | |
6176 | ||
6177 | if (sg_imbalanced(group)) | |
6178 | return group_imbalanced; | |
6179 | ||
6180 | return group_other; | |
6181 | } | |
6182 | ||
1e3c88bd PZ |
6183 | /** |
6184 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 6185 | * @env: The load balancing environment. |
1e3c88bd | 6186 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 6187 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 6188 | * @local_group: Does group contain this_cpu. |
1e3c88bd | 6189 | * @sgs: variable to hold the statistics for this group. |
cd3bd4e6 | 6190 | * @overload: Indicate more than one runnable task for any CPU. |
1e3c88bd | 6191 | */ |
bd939f45 PZ |
6192 | static inline void update_sg_lb_stats(struct lb_env *env, |
6193 | struct sched_group *group, int load_idx, | |
4486edd1 TC |
6194 | int local_group, struct sg_lb_stats *sgs, |
6195 | bool *overload) | |
1e3c88bd | 6196 | { |
30ce5dab | 6197 | unsigned long load; |
bd939f45 | 6198 | int i; |
1e3c88bd | 6199 | |
b72ff13c PZ |
6200 | memset(sgs, 0, sizeof(*sgs)); |
6201 | ||
b9403130 | 6202 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
6203 | struct rq *rq = cpu_rq(i); |
6204 | ||
1e3c88bd | 6205 | /* Bias balancing toward cpus of our domain */ |
6263322c | 6206 | if (local_group) |
04f733b4 | 6207 | load = target_load(i, load_idx); |
6263322c | 6208 | else |
1e3c88bd | 6209 | load = source_load(i, load_idx); |
1e3c88bd PZ |
6210 | |
6211 | sgs->group_load += load; | |
65fdac08 | 6212 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 TC |
6213 | |
6214 | if (rq->nr_running > 1) | |
6215 | *overload = true; | |
6216 | ||
0ec8aa00 PZ |
6217 | #ifdef CONFIG_NUMA_BALANCING |
6218 | sgs->nr_numa_running += rq->nr_numa_running; | |
6219 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
6220 | #endif | |
1e3c88bd | 6221 | sgs->sum_weighted_load += weighted_cpuload(i); |
aae6d3dd SS |
6222 | if (idle_cpu(i)) |
6223 | sgs->idle_cpus++; | |
1e3c88bd PZ |
6224 | } |
6225 | ||
63b2ca30 NP |
6226 | /* Adjust by relative CPU capacity of the group */ |
6227 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 6228 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 6229 | |
dd5feea1 | 6230 | if (sgs->sum_nr_running) |
38d0f770 | 6231 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 6232 | |
aae6d3dd | 6233 | sgs->group_weight = group->group_weight; |
0fedc6c8 | 6234 | sgs->group_capacity_factor = sg_capacity_factor(env, group); |
caeb178c | 6235 | sgs->group_type = group_classify(group, sgs); |
b37d9316 | 6236 | |
0fedc6c8 | 6237 | if (sgs->group_capacity_factor > sgs->sum_nr_running) |
1b6a7495 | 6238 | sgs->group_has_free_capacity = 1; |
1e3c88bd PZ |
6239 | } |
6240 | ||
532cb4c4 MN |
6241 | /** |
6242 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 6243 | * @env: The load balancing environment. |
532cb4c4 MN |
6244 | * @sds: sched_domain statistics |
6245 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 6246 | * @sgs: sched_group statistics |
532cb4c4 MN |
6247 | * |
6248 | * Determine if @sg is a busier group than the previously selected | |
6249 | * busiest group. | |
e69f6186 YB |
6250 | * |
6251 | * Return: %true if @sg is a busier group than the previously selected | |
6252 | * busiest group. %false otherwise. | |
532cb4c4 | 6253 | */ |
bd939f45 | 6254 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
6255 | struct sd_lb_stats *sds, |
6256 | struct sched_group *sg, | |
bd939f45 | 6257 | struct sg_lb_stats *sgs) |
532cb4c4 | 6258 | { |
caeb178c | 6259 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 6260 | |
caeb178c | 6261 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
6262 | return true; |
6263 | ||
caeb178c RR |
6264 | if (sgs->group_type < busiest->group_type) |
6265 | return false; | |
6266 | ||
6267 | if (sgs->avg_load <= busiest->avg_load) | |
6268 | return false; | |
6269 | ||
6270 | /* This is the busiest node in its class. */ | |
6271 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
6272 | return true; |
6273 | ||
6274 | /* | |
6275 | * ASYM_PACKING needs to move all the work to the lowest | |
6276 | * numbered CPUs in the group, therefore mark all groups | |
6277 | * higher than ourself as busy. | |
6278 | */ | |
caeb178c | 6279 | if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { |
532cb4c4 MN |
6280 | if (!sds->busiest) |
6281 | return true; | |
6282 | ||
6283 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
6284 | return true; | |
6285 | } | |
6286 | ||
6287 | return false; | |
6288 | } | |
6289 | ||
0ec8aa00 PZ |
6290 | #ifdef CONFIG_NUMA_BALANCING |
6291 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
6292 | { | |
6293 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
6294 | return regular; | |
6295 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
6296 | return remote; | |
6297 | return all; | |
6298 | } | |
6299 | ||
6300 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
6301 | { | |
6302 | if (rq->nr_running > rq->nr_numa_running) | |
6303 | return regular; | |
6304 | if (rq->nr_running > rq->nr_preferred_running) | |
6305 | return remote; | |
6306 | return all; | |
6307 | } | |
6308 | #else | |
6309 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
6310 | { | |
6311 | return all; | |
6312 | } | |
6313 | ||
6314 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
6315 | { | |
6316 | return regular; | |
6317 | } | |
6318 | #endif /* CONFIG_NUMA_BALANCING */ | |
6319 | ||
1e3c88bd | 6320 | /** |
461819ac | 6321 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 6322 | * @env: The load balancing environment. |
1e3c88bd PZ |
6323 | * @sds: variable to hold the statistics for this sched_domain. |
6324 | */ | |
0ec8aa00 | 6325 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6326 | { |
bd939f45 PZ |
6327 | struct sched_domain *child = env->sd->child; |
6328 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 6329 | struct sg_lb_stats tmp_sgs; |
1e3c88bd | 6330 | int load_idx, prefer_sibling = 0; |
4486edd1 | 6331 | bool overload = false; |
1e3c88bd PZ |
6332 | |
6333 | if (child && child->flags & SD_PREFER_SIBLING) | |
6334 | prefer_sibling = 1; | |
6335 | ||
bd939f45 | 6336 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
6337 | |
6338 | do { | |
56cf515b | 6339 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
6340 | int local_group; |
6341 | ||
bd939f45 | 6342 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
6343 | if (local_group) { |
6344 | sds->local = sg; | |
6345 | sgs = &sds->local_stat; | |
b72ff13c PZ |
6346 | |
6347 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
6348 | time_after_eq(jiffies, sg->sgc->next_update)) |
6349 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 6350 | } |
1e3c88bd | 6351 | |
4486edd1 TC |
6352 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, |
6353 | &overload); | |
1e3c88bd | 6354 | |
b72ff13c PZ |
6355 | if (local_group) |
6356 | goto next_group; | |
6357 | ||
1e3c88bd PZ |
6358 | /* |
6359 | * In case the child domain prefers tasks go to siblings | |
0fedc6c8 | 6360 | * first, lower the sg capacity factor to one so that we'll try |
75dd321d NR |
6361 | * and move all the excess tasks away. We lower the capacity |
6362 | * of a group only if the local group has the capacity to fit | |
0fedc6c8 | 6363 | * these excess tasks, i.e. nr_running < group_capacity_factor. The |
75dd321d NR |
6364 | * extra check prevents the case where you always pull from the |
6365 | * heaviest group when it is already under-utilized (possible | |
6366 | * with a large weight task outweighs the tasks on the system). | |
1e3c88bd | 6367 | */ |
b72ff13c | 6368 | if (prefer_sibling && sds->local && |
cb0b9f24 | 6369 | sds->local_stat.group_has_free_capacity) { |
0fedc6c8 | 6370 | sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U); |
cb0b9f24 WL |
6371 | sgs->group_type = group_classify(sg, sgs); |
6372 | } | |
1e3c88bd | 6373 | |
b72ff13c | 6374 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 6375 | sds->busiest = sg; |
56cf515b | 6376 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
6377 | } |
6378 | ||
b72ff13c PZ |
6379 | next_group: |
6380 | /* Now, start updating sd_lb_stats */ | |
6381 | sds->total_load += sgs->group_load; | |
63b2ca30 | 6382 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 6383 | |
532cb4c4 | 6384 | sg = sg->next; |
bd939f45 | 6385 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
6386 | |
6387 | if (env->sd->flags & SD_NUMA) | |
6388 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
6389 | |
6390 | if (!env->sd->parent) { | |
6391 | /* update overload indicator if we are at root domain */ | |
6392 | if (env->dst_rq->rd->overload != overload) | |
6393 | env->dst_rq->rd->overload = overload; | |
6394 | } | |
6395 | ||
532cb4c4 MN |
6396 | } |
6397 | ||
532cb4c4 MN |
6398 | /** |
6399 | * check_asym_packing - Check to see if the group is packed into the | |
6400 | * sched doman. | |
6401 | * | |
6402 | * This is primarily intended to used at the sibling level. Some | |
6403 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
6404 | * case of POWER7, it can move to lower SMT modes only when higher | |
6405 | * threads are idle. When in lower SMT modes, the threads will | |
6406 | * perform better since they share less core resources. Hence when we | |
6407 | * have idle threads, we want them to be the higher ones. | |
6408 | * | |
6409 | * This packing function is run on idle threads. It checks to see if | |
6410 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
6411 | * CPU number than the packing function is being run on. Here we are | |
6412 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
6413 | * number. | |
6414 | * | |
e69f6186 | 6415 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
6416 | * this CPU. The amount of the imbalance is returned in *imbalance. |
6417 | * | |
cd96891d | 6418 | * @env: The load balancing environment. |
532cb4c4 | 6419 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 6420 | */ |
bd939f45 | 6421 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
6422 | { |
6423 | int busiest_cpu; | |
6424 | ||
bd939f45 | 6425 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
6426 | return 0; |
6427 | ||
6428 | if (!sds->busiest) | |
6429 | return 0; | |
6430 | ||
6431 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 6432 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
6433 | return 0; |
6434 | ||
bd939f45 | 6435 | env->imbalance = DIV_ROUND_CLOSEST( |
63b2ca30 | 6436 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, |
ca8ce3d0 | 6437 | SCHED_CAPACITY_SCALE); |
bd939f45 | 6438 | |
532cb4c4 | 6439 | return 1; |
1e3c88bd PZ |
6440 | } |
6441 | ||
6442 | /** | |
6443 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
6444 | * amongst the groups of a sched_domain, during | |
6445 | * load balancing. | |
cd96891d | 6446 | * @env: The load balancing environment. |
1e3c88bd | 6447 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6448 | */ |
bd939f45 PZ |
6449 | static inline |
6450 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 6451 | { |
63b2ca30 | 6452 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 6453 | unsigned int imbn = 2; |
dd5feea1 | 6454 | unsigned long scaled_busy_load_per_task; |
56cf515b | 6455 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 6456 | |
56cf515b JK |
6457 | local = &sds->local_stat; |
6458 | busiest = &sds->busiest_stat; | |
1e3c88bd | 6459 | |
56cf515b JK |
6460 | if (!local->sum_nr_running) |
6461 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
6462 | else if (busiest->load_per_task > local->load_per_task) | |
6463 | imbn = 1; | |
dd5feea1 | 6464 | |
56cf515b | 6465 | scaled_busy_load_per_task = |
ca8ce3d0 | 6466 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 6467 | busiest->group_capacity; |
56cf515b | 6468 | |
3029ede3 VD |
6469 | if (busiest->avg_load + scaled_busy_load_per_task >= |
6470 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 6471 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6472 | return; |
6473 | } | |
6474 | ||
6475 | /* | |
6476 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 6477 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
6478 | * moving them. |
6479 | */ | |
6480 | ||
63b2ca30 | 6481 | capa_now += busiest->group_capacity * |
56cf515b | 6482 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 6483 | capa_now += local->group_capacity * |
56cf515b | 6484 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 6485 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
6486 | |
6487 | /* Amount of load we'd subtract */ | |
a2cd4260 | 6488 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 6489 | capa_move += busiest->group_capacity * |
56cf515b | 6490 | min(busiest->load_per_task, |
a2cd4260 | 6491 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 6492 | } |
1e3c88bd PZ |
6493 | |
6494 | /* Amount of load we'd add */ | |
63b2ca30 | 6495 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 6496 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
6497 | tmp = (busiest->avg_load * busiest->group_capacity) / |
6498 | local->group_capacity; | |
56cf515b | 6499 | } else { |
ca8ce3d0 | 6500 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 6501 | local->group_capacity; |
56cf515b | 6502 | } |
63b2ca30 | 6503 | capa_move += local->group_capacity * |
3ae11c90 | 6504 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 6505 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
6506 | |
6507 | /* Move if we gain throughput */ | |
63b2ca30 | 6508 | if (capa_move > capa_now) |
56cf515b | 6509 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6510 | } |
6511 | ||
6512 | /** | |
6513 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
6514 | * groups of a given sched_domain during load balance. | |
bd939f45 | 6515 | * @env: load balance environment |
1e3c88bd | 6516 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6517 | */ |
bd939f45 | 6518 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6519 | { |
dd5feea1 | 6520 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
6521 | struct sg_lb_stats *local, *busiest; |
6522 | ||
6523 | local = &sds->local_stat; | |
56cf515b | 6524 | busiest = &sds->busiest_stat; |
dd5feea1 | 6525 | |
caeb178c | 6526 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
6527 | /* |
6528 | * In the group_imb case we cannot rely on group-wide averages | |
6529 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
6530 | */ | |
56cf515b JK |
6531 | busiest->load_per_task = |
6532 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
6533 | } |
6534 | ||
1e3c88bd PZ |
6535 | /* |
6536 | * In the presence of smp nice balancing, certain scenarios can have | |
6537 | * max load less than avg load(as we skip the groups at or below | |
ced549fa | 6538 | * its cpu_capacity, while calculating max_load..) |
1e3c88bd | 6539 | */ |
b1885550 VD |
6540 | if (busiest->avg_load <= sds->avg_load || |
6541 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
6542 | env->imbalance = 0; |
6543 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
6544 | } |
6545 | ||
9a5d9ba6 PZ |
6546 | /* |
6547 | * If there aren't any idle cpus, avoid creating some. | |
6548 | */ | |
6549 | if (busiest->group_type == group_overloaded && | |
6550 | local->group_type == group_overloaded) { | |
56cf515b | 6551 | load_above_capacity = |
0fedc6c8 | 6552 | (busiest->sum_nr_running - busiest->group_capacity_factor); |
dd5feea1 | 6553 | |
ca8ce3d0 | 6554 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE); |
63b2ca30 | 6555 | load_above_capacity /= busiest->group_capacity; |
dd5feea1 SS |
6556 | } |
6557 | ||
6558 | /* | |
6559 | * We're trying to get all the cpus to the average_load, so we don't | |
6560 | * want to push ourselves above the average load, nor do we wish to | |
6561 | * reduce the max loaded cpu below the average load. At the same time, | |
6562 | * we also don't want to reduce the group load below the group capacity | |
6563 | * (so that we can implement power-savings policies etc). Thus we look | |
6564 | * for the minimum possible imbalance. | |
dd5feea1 | 6565 | */ |
30ce5dab | 6566 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
6567 | |
6568 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 6569 | env->imbalance = min( |
63b2ca30 NP |
6570 | max_pull * busiest->group_capacity, |
6571 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 6572 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
6573 | |
6574 | /* | |
6575 | * if *imbalance is less than the average load per runnable task | |
25985edc | 6576 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
6577 | * a think about bumping its value to force at least one task to be |
6578 | * moved | |
6579 | */ | |
56cf515b | 6580 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 6581 | return fix_small_imbalance(env, sds); |
1e3c88bd | 6582 | } |
fab47622 | 6583 | |
1e3c88bd PZ |
6584 | /******* find_busiest_group() helpers end here *********************/ |
6585 | ||
6586 | /** | |
6587 | * find_busiest_group - Returns the busiest group within the sched_domain | |
6588 | * if there is an imbalance. If there isn't an imbalance, and | |
6589 | * the user has opted for power-savings, it returns a group whose | |
6590 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
6591 | * such a group exists. | |
6592 | * | |
6593 | * Also calculates the amount of weighted load which should be moved | |
6594 | * to restore balance. | |
6595 | * | |
cd96891d | 6596 | * @env: The load balancing environment. |
1e3c88bd | 6597 | * |
e69f6186 | 6598 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
6599 | * - If no imbalance and user has opted for power-savings balance, |
6600 | * return the least loaded group whose CPUs can be | |
6601 | * put to idle by rebalancing its tasks onto our group. | |
6602 | */ | |
56cf515b | 6603 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 6604 | { |
56cf515b | 6605 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
6606 | struct sd_lb_stats sds; |
6607 | ||
147c5fc2 | 6608 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
6609 | |
6610 | /* | |
6611 | * Compute the various statistics relavent for load balancing at | |
6612 | * this level. | |
6613 | */ | |
23f0d209 | 6614 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
6615 | local = &sds.local_stat; |
6616 | busiest = &sds.busiest_stat; | |
1e3c88bd | 6617 | |
bd939f45 PZ |
6618 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
6619 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
6620 | return sds.busiest; |
6621 | ||
cc57aa8f | 6622 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 6623 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
6624 | goto out_balanced; |
6625 | ||
ca8ce3d0 NP |
6626 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
6627 | / sds.total_capacity; | |
b0432d8f | 6628 | |
866ab43e PZ |
6629 | /* |
6630 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 6631 | * work because they assume all things are equal, which typically |
866ab43e PZ |
6632 | * isn't true due to cpus_allowed constraints and the like. |
6633 | */ | |
caeb178c | 6634 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
6635 | goto force_balance; |
6636 | ||
cc57aa8f | 6637 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
1b6a7495 NP |
6638 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity && |
6639 | !busiest->group_has_free_capacity) | |
fab47622 NR |
6640 | goto force_balance; |
6641 | ||
cc57aa8f | 6642 | /* |
9c58c79a | 6643 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
6644 | * don't try and pull any tasks. |
6645 | */ | |
56cf515b | 6646 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
6647 | goto out_balanced; |
6648 | ||
cc57aa8f PZ |
6649 | /* |
6650 | * Don't pull any tasks if this group is already above the domain | |
6651 | * average load. | |
6652 | */ | |
56cf515b | 6653 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
6654 | goto out_balanced; |
6655 | ||
bd939f45 | 6656 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 6657 | /* |
43f4d666 VG |
6658 | * This cpu is idle. If the busiest group is not overloaded |
6659 | * and there is no imbalance between this and busiest group | |
6660 | * wrt idle cpus, it is balanced. The imbalance becomes | |
6661 | * significant if the diff is greater than 1 otherwise we | |
6662 | * might end up to just move the imbalance on another group | |
aae6d3dd | 6663 | */ |
43f4d666 VG |
6664 | if ((busiest->group_type != group_overloaded) && |
6665 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 6666 | goto out_balanced; |
c186fafe PZ |
6667 | } else { |
6668 | /* | |
6669 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
6670 | * imbalance_pct to be conservative. | |
6671 | */ | |
56cf515b JK |
6672 | if (100 * busiest->avg_load <= |
6673 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 6674 | goto out_balanced; |
aae6d3dd | 6675 | } |
1e3c88bd | 6676 | |
fab47622 | 6677 | force_balance: |
1e3c88bd | 6678 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 6679 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
6680 | return sds.busiest; |
6681 | ||
6682 | out_balanced: | |
bd939f45 | 6683 | env->imbalance = 0; |
1e3c88bd PZ |
6684 | return NULL; |
6685 | } | |
6686 | ||
6687 | /* | |
6688 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
6689 | */ | |
bd939f45 | 6690 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 6691 | struct sched_group *group) |
1e3c88bd PZ |
6692 | { |
6693 | struct rq *busiest = NULL, *rq; | |
ced549fa | 6694 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
6695 | int i; |
6696 | ||
6906a408 | 6697 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
ced549fa | 6698 | unsigned long capacity, capacity_factor, wl; |
0ec8aa00 PZ |
6699 | enum fbq_type rt; |
6700 | ||
6701 | rq = cpu_rq(i); | |
6702 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 6703 | |
0ec8aa00 PZ |
6704 | /* |
6705 | * We classify groups/runqueues into three groups: | |
6706 | * - regular: there are !numa tasks | |
6707 | * - remote: there are numa tasks that run on the 'wrong' node | |
6708 | * - all: there is no distinction | |
6709 | * | |
6710 | * In order to avoid migrating ideally placed numa tasks, | |
6711 | * ignore those when there's better options. | |
6712 | * | |
6713 | * If we ignore the actual busiest queue to migrate another | |
6714 | * task, the next balance pass can still reduce the busiest | |
6715 | * queue by moving tasks around inside the node. | |
6716 | * | |
6717 | * If we cannot move enough load due to this classification | |
6718 | * the next pass will adjust the group classification and | |
6719 | * allow migration of more tasks. | |
6720 | * | |
6721 | * Both cases only affect the total convergence complexity. | |
6722 | */ | |
6723 | if (rt > env->fbq_type) | |
6724 | continue; | |
6725 | ||
ced549fa | 6726 | capacity = capacity_of(i); |
ca8ce3d0 | 6727 | capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE); |
0fedc6c8 NP |
6728 | if (!capacity_factor) |
6729 | capacity_factor = fix_small_capacity(env->sd, group); | |
9d5efe05 | 6730 | |
6e40f5bb | 6731 | wl = weighted_cpuload(i); |
1e3c88bd | 6732 | |
6e40f5bb TG |
6733 | /* |
6734 | * When comparing with imbalance, use weighted_cpuload() | |
ced549fa | 6735 | * which is not scaled with the cpu capacity. |
6e40f5bb | 6736 | */ |
0fedc6c8 | 6737 | if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance) |
1e3c88bd PZ |
6738 | continue; |
6739 | ||
6e40f5bb TG |
6740 | /* |
6741 | * For the load comparisons with the other cpu's, consider | |
ced549fa NP |
6742 | * the weighted_cpuload() scaled with the cpu capacity, so |
6743 | * that the load can be moved away from the cpu that is | |
6744 | * potentially running at a lower capacity. | |
95a79b80 | 6745 | * |
ced549fa | 6746 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
95a79b80 | 6747 | * multiplication to rid ourselves of the division works out |
ced549fa NP |
6748 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
6749 | * our previous maximum. | |
6e40f5bb | 6750 | */ |
ced549fa | 6751 | if (wl * busiest_capacity > busiest_load * capacity) { |
95a79b80 | 6752 | busiest_load = wl; |
ced549fa | 6753 | busiest_capacity = capacity; |
1e3c88bd PZ |
6754 | busiest = rq; |
6755 | } | |
6756 | } | |
6757 | ||
6758 | return busiest; | |
6759 | } | |
6760 | ||
6761 | /* | |
6762 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
6763 | * so long as it is large enough. | |
6764 | */ | |
6765 | #define MAX_PINNED_INTERVAL 512 | |
6766 | ||
6767 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 6768 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 6769 | |
bd939f45 | 6770 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 6771 | { |
bd939f45 PZ |
6772 | struct sched_domain *sd = env->sd; |
6773 | ||
6774 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
6775 | |
6776 | /* | |
6777 | * ASYM_PACKING needs to force migrate tasks from busy but | |
6778 | * higher numbered CPUs in order to pack all tasks in the | |
6779 | * lowest numbered CPUs. | |
6780 | */ | |
bd939f45 | 6781 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 6782 | return 1; |
1af3ed3d PZ |
6783 | } |
6784 | ||
6785 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | |
6786 | } | |
6787 | ||
969c7921 TH |
6788 | static int active_load_balance_cpu_stop(void *data); |
6789 | ||
23f0d209 JK |
6790 | static int should_we_balance(struct lb_env *env) |
6791 | { | |
6792 | struct sched_group *sg = env->sd->groups; | |
6793 | struct cpumask *sg_cpus, *sg_mask; | |
6794 | int cpu, balance_cpu = -1; | |
6795 | ||
6796 | /* | |
6797 | * In the newly idle case, we will allow all the cpu's | |
6798 | * to do the newly idle load balance. | |
6799 | */ | |
6800 | if (env->idle == CPU_NEWLY_IDLE) | |
6801 | return 1; | |
6802 | ||
6803 | sg_cpus = sched_group_cpus(sg); | |
6804 | sg_mask = sched_group_mask(sg); | |
6805 | /* Try to find first idle cpu */ | |
6806 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
6807 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
6808 | continue; | |
6809 | ||
6810 | balance_cpu = cpu; | |
6811 | break; | |
6812 | } | |
6813 | ||
6814 | if (balance_cpu == -1) | |
6815 | balance_cpu = group_balance_cpu(sg); | |
6816 | ||
6817 | /* | |
6818 | * First idle cpu or the first cpu(busiest) in this sched group | |
6819 | * is eligible for doing load balancing at this and above domains. | |
6820 | */ | |
b0cff9d8 | 6821 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
6822 | } |
6823 | ||
1e3c88bd PZ |
6824 | /* |
6825 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
6826 | * tasks if there is an imbalance. | |
6827 | */ | |
6828 | static int load_balance(int this_cpu, struct rq *this_rq, | |
6829 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 6830 | int *continue_balancing) |
1e3c88bd | 6831 | { |
88b8dac0 | 6832 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 6833 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 6834 | struct sched_group *group; |
1e3c88bd PZ |
6835 | struct rq *busiest; |
6836 | unsigned long flags; | |
4ba29684 | 6837 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 6838 | |
8e45cb54 PZ |
6839 | struct lb_env env = { |
6840 | .sd = sd, | |
ddcdf6e7 PZ |
6841 | .dst_cpu = this_cpu, |
6842 | .dst_rq = this_rq, | |
88b8dac0 | 6843 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 6844 | .idle = idle, |
eb95308e | 6845 | .loop_break = sched_nr_migrate_break, |
b9403130 | 6846 | .cpus = cpus, |
0ec8aa00 | 6847 | .fbq_type = all, |
163122b7 | 6848 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
6849 | }; |
6850 | ||
cfc03118 JK |
6851 | /* |
6852 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
6853 | * other cpus in our group | |
6854 | */ | |
e02e60c1 | 6855 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 6856 | env.dst_grpmask = NULL; |
cfc03118 | 6857 | |
1e3c88bd PZ |
6858 | cpumask_copy(cpus, cpu_active_mask); |
6859 | ||
1e3c88bd PZ |
6860 | schedstat_inc(sd, lb_count[idle]); |
6861 | ||
6862 | redo: | |
23f0d209 JK |
6863 | if (!should_we_balance(&env)) { |
6864 | *continue_balancing = 0; | |
1e3c88bd | 6865 | goto out_balanced; |
23f0d209 | 6866 | } |
1e3c88bd | 6867 | |
23f0d209 | 6868 | group = find_busiest_group(&env); |
1e3c88bd PZ |
6869 | if (!group) { |
6870 | schedstat_inc(sd, lb_nobusyg[idle]); | |
6871 | goto out_balanced; | |
6872 | } | |
6873 | ||
b9403130 | 6874 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
6875 | if (!busiest) { |
6876 | schedstat_inc(sd, lb_nobusyq[idle]); | |
6877 | goto out_balanced; | |
6878 | } | |
6879 | ||
78feefc5 | 6880 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 6881 | |
bd939f45 | 6882 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd PZ |
6883 | |
6884 | ld_moved = 0; | |
6885 | if (busiest->nr_running > 1) { | |
6886 | /* | |
6887 | * Attempt to move tasks. If find_busiest_group has found | |
6888 | * an imbalance but busiest->nr_running <= 1, the group is | |
6889 | * still unbalanced. ld_moved simply stays zero, so it is | |
6890 | * correctly treated as an imbalance. | |
6891 | */ | |
8e45cb54 | 6892 | env.flags |= LBF_ALL_PINNED; |
c82513e5 PZ |
6893 | env.src_cpu = busiest->cpu; |
6894 | env.src_rq = busiest; | |
6895 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | |
8e45cb54 | 6896 | |
5d6523eb | 6897 | more_balance: |
163122b7 | 6898 | raw_spin_lock_irqsave(&busiest->lock, flags); |
88b8dac0 SV |
6899 | |
6900 | /* | |
6901 | * cur_ld_moved - load moved in current iteration | |
6902 | * ld_moved - cumulative load moved across iterations | |
6903 | */ | |
163122b7 | 6904 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
6905 | |
6906 | /* | |
163122b7 KT |
6907 | * We've detached some tasks from busiest_rq. Every |
6908 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
6909 | * unlock busiest->lock, and we are able to be sure | |
6910 | * that nobody can manipulate the tasks in parallel. | |
6911 | * See task_rq_lock() family for the details. | |
1e3c88bd | 6912 | */ |
163122b7 KT |
6913 | |
6914 | raw_spin_unlock(&busiest->lock); | |
6915 | ||
6916 | if (cur_ld_moved) { | |
6917 | attach_tasks(&env); | |
6918 | ld_moved += cur_ld_moved; | |
6919 | } | |
6920 | ||
1e3c88bd | 6921 | local_irq_restore(flags); |
88b8dac0 | 6922 | |
f1cd0858 JK |
6923 | if (env.flags & LBF_NEED_BREAK) { |
6924 | env.flags &= ~LBF_NEED_BREAK; | |
6925 | goto more_balance; | |
6926 | } | |
6927 | ||
88b8dac0 SV |
6928 | /* |
6929 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
6930 | * us and move them to an alternate dst_cpu in our sched_group | |
6931 | * where they can run. The upper limit on how many times we | |
6932 | * iterate on same src_cpu is dependent on number of cpus in our | |
6933 | * sched_group. | |
6934 | * | |
6935 | * This changes load balance semantics a bit on who can move | |
6936 | * load to a given_cpu. In addition to the given_cpu itself | |
6937 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
6938 | * nohz-idle), we now have balance_cpu in a position to move | |
6939 | * load to given_cpu. In rare situations, this may cause | |
6940 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
6941 | * _independently_ and at _same_ time to move some load to | |
6942 | * given_cpu) causing exceess load to be moved to given_cpu. | |
6943 | * This however should not happen so much in practice and | |
6944 | * moreover subsequent load balance cycles should correct the | |
6945 | * excess load moved. | |
6946 | */ | |
6263322c | 6947 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 6948 | |
7aff2e3a VD |
6949 | /* Prevent to re-select dst_cpu via env's cpus */ |
6950 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
6951 | ||
78feefc5 | 6952 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 6953 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 6954 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
6955 | env.loop = 0; |
6956 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 6957 | |
88b8dac0 SV |
6958 | /* |
6959 | * Go back to "more_balance" rather than "redo" since we | |
6960 | * need to continue with same src_cpu. | |
6961 | */ | |
6962 | goto more_balance; | |
6963 | } | |
1e3c88bd | 6964 | |
6263322c PZ |
6965 | /* |
6966 | * We failed to reach balance because of affinity. | |
6967 | */ | |
6968 | if (sd_parent) { | |
63b2ca30 | 6969 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 6970 | |
afdeee05 | 6971 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 6972 | *group_imbalance = 1; |
6263322c PZ |
6973 | } |
6974 | ||
1e3c88bd | 6975 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 6976 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 6977 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
6978 | if (!cpumask_empty(cpus)) { |
6979 | env.loop = 0; | |
6980 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 6981 | goto redo; |
bbf18b19 | 6982 | } |
afdeee05 | 6983 | goto out_all_pinned; |
1e3c88bd PZ |
6984 | } |
6985 | } | |
6986 | ||
6987 | if (!ld_moved) { | |
6988 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
6989 | /* |
6990 | * Increment the failure counter only on periodic balance. | |
6991 | * We do not want newidle balance, which can be very | |
6992 | * frequent, pollute the failure counter causing | |
6993 | * excessive cache_hot migrations and active balances. | |
6994 | */ | |
6995 | if (idle != CPU_NEWLY_IDLE) | |
6996 | sd->nr_balance_failed++; | |
1e3c88bd | 6997 | |
bd939f45 | 6998 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
6999 | raw_spin_lock_irqsave(&busiest->lock, flags); |
7000 | ||
969c7921 TH |
7001 | /* don't kick the active_load_balance_cpu_stop, |
7002 | * if the curr task on busiest cpu can't be | |
7003 | * moved to this_cpu | |
1e3c88bd PZ |
7004 | */ |
7005 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 7006 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
7007 | raw_spin_unlock_irqrestore(&busiest->lock, |
7008 | flags); | |
8e45cb54 | 7009 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
7010 | goto out_one_pinned; |
7011 | } | |
7012 | ||
969c7921 TH |
7013 | /* |
7014 | * ->active_balance synchronizes accesses to | |
7015 | * ->active_balance_work. Once set, it's cleared | |
7016 | * only after active load balance is finished. | |
7017 | */ | |
1e3c88bd PZ |
7018 | if (!busiest->active_balance) { |
7019 | busiest->active_balance = 1; | |
7020 | busiest->push_cpu = this_cpu; | |
7021 | active_balance = 1; | |
7022 | } | |
7023 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 7024 | |
bd939f45 | 7025 | if (active_balance) { |
969c7921 TH |
7026 | stop_one_cpu_nowait(cpu_of(busiest), |
7027 | active_load_balance_cpu_stop, busiest, | |
7028 | &busiest->active_balance_work); | |
bd939f45 | 7029 | } |
1e3c88bd PZ |
7030 | |
7031 | /* | |
7032 | * We've kicked active balancing, reset the failure | |
7033 | * counter. | |
7034 | */ | |
7035 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
7036 | } | |
7037 | } else | |
7038 | sd->nr_balance_failed = 0; | |
7039 | ||
7040 | if (likely(!active_balance)) { | |
7041 | /* We were unbalanced, so reset the balancing interval */ | |
7042 | sd->balance_interval = sd->min_interval; | |
7043 | } else { | |
7044 | /* | |
7045 | * If we've begun active balancing, start to back off. This | |
7046 | * case may not be covered by the all_pinned logic if there | |
7047 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 7048 | * detach_tasks). |
1e3c88bd PZ |
7049 | */ |
7050 | if (sd->balance_interval < sd->max_interval) | |
7051 | sd->balance_interval *= 2; | |
7052 | } | |
7053 | ||
1e3c88bd PZ |
7054 | goto out; |
7055 | ||
7056 | out_balanced: | |
afdeee05 VG |
7057 | /* |
7058 | * We reach balance although we may have faced some affinity | |
7059 | * constraints. Clear the imbalance flag if it was set. | |
7060 | */ | |
7061 | if (sd_parent) { | |
7062 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | |
7063 | ||
7064 | if (*group_imbalance) | |
7065 | *group_imbalance = 0; | |
7066 | } | |
7067 | ||
7068 | out_all_pinned: | |
7069 | /* | |
7070 | * We reach balance because all tasks are pinned at this level so | |
7071 | * we can't migrate them. Let the imbalance flag set so parent level | |
7072 | * can try to migrate them. | |
7073 | */ | |
1e3c88bd PZ |
7074 | schedstat_inc(sd, lb_balanced[idle]); |
7075 | ||
7076 | sd->nr_balance_failed = 0; | |
7077 | ||
7078 | out_one_pinned: | |
7079 | /* tune up the balancing interval */ | |
8e45cb54 | 7080 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 7081 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
7082 | (sd->balance_interval < sd->max_interval)) |
7083 | sd->balance_interval *= 2; | |
7084 | ||
46e49b38 | 7085 | ld_moved = 0; |
1e3c88bd | 7086 | out: |
1e3c88bd PZ |
7087 | return ld_moved; |
7088 | } | |
7089 | ||
52a08ef1 JL |
7090 | static inline unsigned long |
7091 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
7092 | { | |
7093 | unsigned long interval = sd->balance_interval; | |
7094 | ||
7095 | if (cpu_busy) | |
7096 | interval *= sd->busy_factor; | |
7097 | ||
7098 | /* scale ms to jiffies */ | |
7099 | interval = msecs_to_jiffies(interval); | |
7100 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
7101 | ||
7102 | return interval; | |
7103 | } | |
7104 | ||
7105 | static inline void | |
7106 | update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) | |
7107 | { | |
7108 | unsigned long interval, next; | |
7109 | ||
7110 | interval = get_sd_balance_interval(sd, cpu_busy); | |
7111 | next = sd->last_balance + interval; | |
7112 | ||
7113 | if (time_after(*next_balance, next)) | |
7114 | *next_balance = next; | |
7115 | } | |
7116 | ||
1e3c88bd PZ |
7117 | /* |
7118 | * idle_balance is called by schedule() if this_cpu is about to become | |
7119 | * idle. Attempts to pull tasks from other CPUs. | |
7120 | */ | |
6e83125c | 7121 | static int idle_balance(struct rq *this_rq) |
1e3c88bd | 7122 | { |
52a08ef1 JL |
7123 | unsigned long next_balance = jiffies + HZ; |
7124 | int this_cpu = this_rq->cpu; | |
1e3c88bd PZ |
7125 | struct sched_domain *sd; |
7126 | int pulled_task = 0; | |
9bd721c5 | 7127 | u64 curr_cost = 0; |
1e3c88bd | 7128 | |
6e83125c | 7129 | idle_enter_fair(this_rq); |
0e5b5337 | 7130 | |
6e83125c PZ |
7131 | /* |
7132 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
7133 | * measure the duration of idle_balance() as idle time. | |
7134 | */ | |
7135 | this_rq->idle_stamp = rq_clock(this_rq); | |
7136 | ||
4486edd1 TC |
7137 | if (this_rq->avg_idle < sysctl_sched_migration_cost || |
7138 | !this_rq->rd->overload) { | |
52a08ef1 JL |
7139 | rcu_read_lock(); |
7140 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
7141 | if (sd) | |
7142 | update_next_balance(sd, 0, &next_balance); | |
7143 | rcu_read_unlock(); | |
7144 | ||
6e83125c | 7145 | goto out; |
52a08ef1 | 7146 | } |
1e3c88bd | 7147 | |
f492e12e PZ |
7148 | /* |
7149 | * Drop the rq->lock, but keep IRQ/preempt disabled. | |
7150 | */ | |
7151 | raw_spin_unlock(&this_rq->lock); | |
7152 | ||
48a16753 | 7153 | update_blocked_averages(this_cpu); |
dce840a0 | 7154 | rcu_read_lock(); |
1e3c88bd | 7155 | for_each_domain(this_cpu, sd) { |
23f0d209 | 7156 | int continue_balancing = 1; |
9bd721c5 | 7157 | u64 t0, domain_cost; |
1e3c88bd PZ |
7158 | |
7159 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
7160 | continue; | |
7161 | ||
52a08ef1 JL |
7162 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
7163 | update_next_balance(sd, 0, &next_balance); | |
9bd721c5 | 7164 | break; |
52a08ef1 | 7165 | } |
9bd721c5 | 7166 | |
f492e12e | 7167 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
7168 | t0 = sched_clock_cpu(this_cpu); |
7169 | ||
f492e12e | 7170 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
7171 | sd, CPU_NEWLY_IDLE, |
7172 | &continue_balancing); | |
9bd721c5 JL |
7173 | |
7174 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
7175 | if (domain_cost > sd->max_newidle_lb_cost) | |
7176 | sd->max_newidle_lb_cost = domain_cost; | |
7177 | ||
7178 | curr_cost += domain_cost; | |
f492e12e | 7179 | } |
1e3c88bd | 7180 | |
52a08ef1 | 7181 | update_next_balance(sd, 0, &next_balance); |
39a4d9ca JL |
7182 | |
7183 | /* | |
7184 | * Stop searching for tasks to pull if there are | |
7185 | * now runnable tasks on this rq. | |
7186 | */ | |
7187 | if (pulled_task || this_rq->nr_running > 0) | |
1e3c88bd | 7188 | break; |
1e3c88bd | 7189 | } |
dce840a0 | 7190 | rcu_read_unlock(); |
f492e12e PZ |
7191 | |
7192 | raw_spin_lock(&this_rq->lock); | |
7193 | ||
0e5b5337 JL |
7194 | if (curr_cost > this_rq->max_idle_balance_cost) |
7195 | this_rq->max_idle_balance_cost = curr_cost; | |
7196 | ||
e5fc6611 | 7197 | /* |
0e5b5337 JL |
7198 | * While browsing the domains, we released the rq lock, a task could |
7199 | * have been enqueued in the meantime. Since we're not going idle, | |
7200 | * pretend we pulled a task. | |
e5fc6611 | 7201 | */ |
0e5b5337 | 7202 | if (this_rq->cfs.h_nr_running && !pulled_task) |
6e83125c | 7203 | pulled_task = 1; |
e5fc6611 | 7204 | |
52a08ef1 JL |
7205 | out: |
7206 | /* Move the next balance forward */ | |
7207 | if (time_after(this_rq->next_balance, next_balance)) | |
1e3c88bd | 7208 | this_rq->next_balance = next_balance; |
9bd721c5 | 7209 | |
e4aa358b | 7210 | /* Is there a task of a high priority class? */ |
46383648 | 7211 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
e4aa358b KT |
7212 | pulled_task = -1; |
7213 | ||
7214 | if (pulled_task) { | |
7215 | idle_exit_fair(this_rq); | |
6e83125c | 7216 | this_rq->idle_stamp = 0; |
e4aa358b | 7217 | } |
6e83125c | 7218 | |
3c4017c1 | 7219 | return pulled_task; |
1e3c88bd PZ |
7220 | } |
7221 | ||
7222 | /* | |
969c7921 TH |
7223 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
7224 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
7225 | * least 1 task to be running on each physical CPU where possible, and | |
7226 | * avoids physical / logical imbalances. | |
1e3c88bd | 7227 | */ |
969c7921 | 7228 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 7229 | { |
969c7921 TH |
7230 | struct rq *busiest_rq = data; |
7231 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 7232 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 7233 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 7234 | struct sched_domain *sd; |
e5673f28 | 7235 | struct task_struct *p = NULL; |
969c7921 TH |
7236 | |
7237 | raw_spin_lock_irq(&busiest_rq->lock); | |
7238 | ||
7239 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
7240 | if (unlikely(busiest_cpu != smp_processor_id() || | |
7241 | !busiest_rq->active_balance)) | |
7242 | goto out_unlock; | |
1e3c88bd PZ |
7243 | |
7244 | /* Is there any task to move? */ | |
7245 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 7246 | goto out_unlock; |
1e3c88bd PZ |
7247 | |
7248 | /* | |
7249 | * This condition is "impossible", if it occurs | |
7250 | * we need to fix it. Originally reported by | |
7251 | * Bjorn Helgaas on a 128-cpu setup. | |
7252 | */ | |
7253 | BUG_ON(busiest_rq == target_rq); | |
7254 | ||
1e3c88bd | 7255 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 7256 | rcu_read_lock(); |
1e3c88bd PZ |
7257 | for_each_domain(target_cpu, sd) { |
7258 | if ((sd->flags & SD_LOAD_BALANCE) && | |
7259 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
7260 | break; | |
7261 | } | |
7262 | ||
7263 | if (likely(sd)) { | |
8e45cb54 PZ |
7264 | struct lb_env env = { |
7265 | .sd = sd, | |
ddcdf6e7 PZ |
7266 | .dst_cpu = target_cpu, |
7267 | .dst_rq = target_rq, | |
7268 | .src_cpu = busiest_rq->cpu, | |
7269 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
7270 | .idle = CPU_IDLE, |
7271 | }; | |
7272 | ||
1e3c88bd PZ |
7273 | schedstat_inc(sd, alb_count); |
7274 | ||
e5673f28 KT |
7275 | p = detach_one_task(&env); |
7276 | if (p) | |
1e3c88bd PZ |
7277 | schedstat_inc(sd, alb_pushed); |
7278 | else | |
7279 | schedstat_inc(sd, alb_failed); | |
7280 | } | |
dce840a0 | 7281 | rcu_read_unlock(); |
969c7921 TH |
7282 | out_unlock: |
7283 | busiest_rq->active_balance = 0; | |
e5673f28 KT |
7284 | raw_spin_unlock(&busiest_rq->lock); |
7285 | ||
7286 | if (p) | |
7287 | attach_one_task(target_rq, p); | |
7288 | ||
7289 | local_irq_enable(); | |
7290 | ||
969c7921 | 7291 | return 0; |
1e3c88bd PZ |
7292 | } |
7293 | ||
d987fc7f MG |
7294 | static inline int on_null_domain(struct rq *rq) |
7295 | { | |
7296 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
7297 | } | |
7298 | ||
3451d024 | 7299 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
7300 | /* |
7301 | * idle load balancing details | |
83cd4fe2 VP |
7302 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
7303 | * needed, they will kick the idle load balancer, which then does idle | |
7304 | * load balancing for all the idle CPUs. | |
7305 | */ | |
1e3c88bd | 7306 | static struct { |
83cd4fe2 | 7307 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 7308 | atomic_t nr_cpus; |
83cd4fe2 VP |
7309 | unsigned long next_balance; /* in jiffy units */ |
7310 | } nohz ____cacheline_aligned; | |
1e3c88bd | 7311 | |
3dd0337d | 7312 | static inline int find_new_ilb(void) |
1e3c88bd | 7313 | { |
0b005cf5 | 7314 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 7315 | |
786d6dc7 SS |
7316 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
7317 | return ilb; | |
7318 | ||
7319 | return nr_cpu_ids; | |
1e3c88bd | 7320 | } |
1e3c88bd | 7321 | |
83cd4fe2 VP |
7322 | /* |
7323 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
7324 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
7325 | * CPU (if there is one). | |
7326 | */ | |
0aeeeeba | 7327 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
7328 | { |
7329 | int ilb_cpu; | |
7330 | ||
7331 | nohz.next_balance++; | |
7332 | ||
3dd0337d | 7333 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 7334 | |
0b005cf5 SS |
7335 | if (ilb_cpu >= nr_cpu_ids) |
7336 | return; | |
83cd4fe2 | 7337 | |
cd490c5b | 7338 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
7339 | return; |
7340 | /* | |
7341 | * Use smp_send_reschedule() instead of resched_cpu(). | |
7342 | * This way we generate a sched IPI on the target cpu which | |
7343 | * is idle. And the softirq performing nohz idle load balance | |
7344 | * will be run before returning from the IPI. | |
7345 | */ | |
7346 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
7347 | return; |
7348 | } | |
7349 | ||
c1cc017c | 7350 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
7351 | { |
7352 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
d987fc7f MG |
7353 | /* |
7354 | * Completely isolated CPUs don't ever set, so we must test. | |
7355 | */ | |
7356 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | |
7357 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
7358 | atomic_dec(&nohz.nr_cpus); | |
7359 | } | |
71325960 SS |
7360 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
7361 | } | |
7362 | } | |
7363 | ||
69e1e811 SS |
7364 | static inline void set_cpu_sd_state_busy(void) |
7365 | { | |
7366 | struct sched_domain *sd; | |
37dc6b50 | 7367 | int cpu = smp_processor_id(); |
69e1e811 | 7368 | |
69e1e811 | 7369 | rcu_read_lock(); |
37dc6b50 | 7370 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
7371 | |
7372 | if (!sd || !sd->nohz_idle) | |
7373 | goto unlock; | |
7374 | sd->nohz_idle = 0; | |
7375 | ||
63b2ca30 | 7376 | atomic_inc(&sd->groups->sgc->nr_busy_cpus); |
25f55d9d | 7377 | unlock: |
69e1e811 SS |
7378 | rcu_read_unlock(); |
7379 | } | |
7380 | ||
7381 | void set_cpu_sd_state_idle(void) | |
7382 | { | |
7383 | struct sched_domain *sd; | |
37dc6b50 | 7384 | int cpu = smp_processor_id(); |
69e1e811 | 7385 | |
69e1e811 | 7386 | rcu_read_lock(); |
37dc6b50 | 7387 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
7388 | |
7389 | if (!sd || sd->nohz_idle) | |
7390 | goto unlock; | |
7391 | sd->nohz_idle = 1; | |
7392 | ||
63b2ca30 | 7393 | atomic_dec(&sd->groups->sgc->nr_busy_cpus); |
25f55d9d | 7394 | unlock: |
69e1e811 SS |
7395 | rcu_read_unlock(); |
7396 | } | |
7397 | ||
1e3c88bd | 7398 | /* |
c1cc017c | 7399 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 7400 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 7401 | */ |
c1cc017c | 7402 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 7403 | { |
71325960 SS |
7404 | /* |
7405 | * If this cpu is going down, then nothing needs to be done. | |
7406 | */ | |
7407 | if (!cpu_active(cpu)) | |
7408 | return; | |
7409 | ||
c1cc017c AS |
7410 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
7411 | return; | |
1e3c88bd | 7412 | |
d987fc7f MG |
7413 | /* |
7414 | * If we're a completely isolated CPU, we don't play. | |
7415 | */ | |
7416 | if (on_null_domain(cpu_rq(cpu))) | |
7417 | return; | |
7418 | ||
c1cc017c AS |
7419 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
7420 | atomic_inc(&nohz.nr_cpus); | |
7421 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 7422 | } |
71325960 | 7423 | |
0db0628d | 7424 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
7425 | unsigned long action, void *hcpu) |
7426 | { | |
7427 | switch (action & ~CPU_TASKS_FROZEN) { | |
7428 | case CPU_DYING: | |
c1cc017c | 7429 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
7430 | return NOTIFY_OK; |
7431 | default: | |
7432 | return NOTIFY_DONE; | |
7433 | } | |
7434 | } | |
1e3c88bd PZ |
7435 | #endif |
7436 | ||
7437 | static DEFINE_SPINLOCK(balancing); | |
7438 | ||
49c022e6 PZ |
7439 | /* |
7440 | * Scale the max load_balance interval with the number of CPUs in the system. | |
7441 | * This trades load-balance latency on larger machines for less cross talk. | |
7442 | */ | |
029632fb | 7443 | void update_max_interval(void) |
49c022e6 PZ |
7444 | { |
7445 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
7446 | } | |
7447 | ||
1e3c88bd PZ |
7448 | /* |
7449 | * It checks each scheduling domain to see if it is due to be balanced, | |
7450 | * and initiates a balancing operation if so. | |
7451 | * | |
b9b0853a | 7452 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 7453 | */ |
f7ed0a89 | 7454 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 7455 | { |
23f0d209 | 7456 | int continue_balancing = 1; |
f7ed0a89 | 7457 | int cpu = rq->cpu; |
1e3c88bd | 7458 | unsigned long interval; |
04f733b4 | 7459 | struct sched_domain *sd; |
1e3c88bd PZ |
7460 | /* Earliest time when we have to do rebalance again */ |
7461 | unsigned long next_balance = jiffies + 60*HZ; | |
7462 | int update_next_balance = 0; | |
f48627e6 JL |
7463 | int need_serialize, need_decay = 0; |
7464 | u64 max_cost = 0; | |
1e3c88bd | 7465 | |
48a16753 | 7466 | update_blocked_averages(cpu); |
2069dd75 | 7467 | |
dce840a0 | 7468 | rcu_read_lock(); |
1e3c88bd | 7469 | for_each_domain(cpu, sd) { |
f48627e6 JL |
7470 | /* |
7471 | * Decay the newidle max times here because this is a regular | |
7472 | * visit to all the domains. Decay ~1% per second. | |
7473 | */ | |
7474 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
7475 | sd->max_newidle_lb_cost = | |
7476 | (sd->max_newidle_lb_cost * 253) / 256; | |
7477 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
7478 | need_decay = 1; | |
7479 | } | |
7480 | max_cost += sd->max_newidle_lb_cost; | |
7481 | ||
1e3c88bd PZ |
7482 | if (!(sd->flags & SD_LOAD_BALANCE)) |
7483 | continue; | |
7484 | ||
f48627e6 JL |
7485 | /* |
7486 | * Stop the load balance at this level. There is another | |
7487 | * CPU in our sched group which is doing load balancing more | |
7488 | * actively. | |
7489 | */ | |
7490 | if (!continue_balancing) { | |
7491 | if (need_decay) | |
7492 | continue; | |
7493 | break; | |
7494 | } | |
7495 | ||
52a08ef1 | 7496 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
7497 | |
7498 | need_serialize = sd->flags & SD_SERIALIZE; | |
1e3c88bd PZ |
7499 | if (need_serialize) { |
7500 | if (!spin_trylock(&balancing)) | |
7501 | goto out; | |
7502 | } | |
7503 | ||
7504 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 7505 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 7506 | /* |
6263322c | 7507 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
7508 | * env->dst_cpu, so we can't know our idle |
7509 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 7510 | */ |
de5eb2dd | 7511 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
7512 | } |
7513 | sd->last_balance = jiffies; | |
52a08ef1 | 7514 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
7515 | } |
7516 | if (need_serialize) | |
7517 | spin_unlock(&balancing); | |
7518 | out: | |
7519 | if (time_after(next_balance, sd->last_balance + interval)) { | |
7520 | next_balance = sd->last_balance + interval; | |
7521 | update_next_balance = 1; | |
7522 | } | |
f48627e6 JL |
7523 | } |
7524 | if (need_decay) { | |
1e3c88bd | 7525 | /* |
f48627e6 JL |
7526 | * Ensure the rq-wide value also decays but keep it at a |
7527 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 7528 | */ |
f48627e6 JL |
7529 | rq->max_idle_balance_cost = |
7530 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 7531 | } |
dce840a0 | 7532 | rcu_read_unlock(); |
1e3c88bd PZ |
7533 | |
7534 | /* | |
7535 | * next_balance will be updated only when there is a need. | |
7536 | * When the cpu is attached to null domain for ex, it will not be | |
7537 | * updated. | |
7538 | */ | |
7539 | if (likely(update_next_balance)) | |
7540 | rq->next_balance = next_balance; | |
7541 | } | |
7542 | ||
3451d024 | 7543 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 7544 | /* |
3451d024 | 7545 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
7546 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
7547 | */ | |
208cb16b | 7548 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 7549 | { |
208cb16b | 7550 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
7551 | struct rq *rq; |
7552 | int balance_cpu; | |
7553 | ||
1c792db7 SS |
7554 | if (idle != CPU_IDLE || |
7555 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
7556 | goto end; | |
83cd4fe2 VP |
7557 | |
7558 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 7559 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
7560 | continue; |
7561 | ||
7562 | /* | |
7563 | * If this cpu gets work to do, stop the load balancing | |
7564 | * work being done for other cpus. Next load | |
7565 | * balancing owner will pick it up. | |
7566 | */ | |
1c792db7 | 7567 | if (need_resched()) |
83cd4fe2 | 7568 | break; |
83cd4fe2 | 7569 | |
5ed4f1d9 VG |
7570 | rq = cpu_rq(balance_cpu); |
7571 | ||
ed61bbc6 TC |
7572 | /* |
7573 | * If time for next balance is due, | |
7574 | * do the balance. | |
7575 | */ | |
7576 | if (time_after_eq(jiffies, rq->next_balance)) { | |
7577 | raw_spin_lock_irq(&rq->lock); | |
7578 | update_rq_clock(rq); | |
7579 | update_idle_cpu_load(rq); | |
7580 | raw_spin_unlock_irq(&rq->lock); | |
7581 | rebalance_domains(rq, CPU_IDLE); | |
7582 | } | |
83cd4fe2 | 7583 | |
83cd4fe2 VP |
7584 | if (time_after(this_rq->next_balance, rq->next_balance)) |
7585 | this_rq->next_balance = rq->next_balance; | |
7586 | } | |
7587 | nohz.next_balance = this_rq->next_balance; | |
1c792db7 SS |
7588 | end: |
7589 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
7590 | } |
7591 | ||
7592 | /* | |
0b005cf5 SS |
7593 | * Current heuristic for kicking the idle load balancer in the presence |
7594 | * of an idle cpu is the system. | |
7595 | * - This rq has more than one task. | |
7596 | * - At any scheduler domain level, this cpu's scheduler group has multiple | |
63b2ca30 | 7597 | * busy cpu's exceeding the group's capacity. |
0b005cf5 SS |
7598 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler |
7599 | * domain span are idle. | |
83cd4fe2 | 7600 | */ |
4a725627 | 7601 | static inline int nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
7602 | { |
7603 | unsigned long now = jiffies; | |
0b005cf5 | 7604 | struct sched_domain *sd; |
63b2ca30 | 7605 | struct sched_group_capacity *sgc; |
4a725627 | 7606 | int nr_busy, cpu = rq->cpu; |
83cd4fe2 | 7607 | |
4a725627 | 7608 | if (unlikely(rq->idle_balance)) |
83cd4fe2 VP |
7609 | return 0; |
7610 | ||
1c792db7 SS |
7611 | /* |
7612 | * We may be recently in ticked or tickless idle mode. At the first | |
7613 | * busy tick after returning from idle, we will update the busy stats. | |
7614 | */ | |
69e1e811 | 7615 | set_cpu_sd_state_busy(); |
c1cc017c | 7616 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
7617 | |
7618 | /* | |
7619 | * None are in tickless mode and hence no need for NOHZ idle load | |
7620 | * balancing. | |
7621 | */ | |
7622 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
7623 | return 0; | |
1c792db7 SS |
7624 | |
7625 | if (time_before(now, nohz.next_balance)) | |
83cd4fe2 VP |
7626 | return 0; |
7627 | ||
0b005cf5 SS |
7628 | if (rq->nr_running >= 2) |
7629 | goto need_kick; | |
83cd4fe2 | 7630 | |
067491b7 | 7631 | rcu_read_lock(); |
37dc6b50 | 7632 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
83cd4fe2 | 7633 | |
37dc6b50 | 7634 | if (sd) { |
63b2ca30 NP |
7635 | sgc = sd->groups->sgc; |
7636 | nr_busy = atomic_read(&sgc->nr_busy_cpus); | |
0b005cf5 | 7637 | |
37dc6b50 | 7638 | if (nr_busy > 1) |
067491b7 | 7639 | goto need_kick_unlock; |
83cd4fe2 | 7640 | } |
37dc6b50 PM |
7641 | |
7642 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); | |
7643 | ||
7644 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, | |
7645 | sched_domain_span(sd)) < cpu)) | |
7646 | goto need_kick_unlock; | |
7647 | ||
067491b7 | 7648 | rcu_read_unlock(); |
83cd4fe2 | 7649 | return 0; |
067491b7 PZ |
7650 | |
7651 | need_kick_unlock: | |
7652 | rcu_read_unlock(); | |
0b005cf5 SS |
7653 | need_kick: |
7654 | return 1; | |
83cd4fe2 VP |
7655 | } |
7656 | #else | |
208cb16b | 7657 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
7658 | #endif |
7659 | ||
7660 | /* | |
7661 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
7662 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
7663 | */ | |
1e3c88bd PZ |
7664 | static void run_rebalance_domains(struct softirq_action *h) |
7665 | { | |
208cb16b | 7666 | struct rq *this_rq = this_rq(); |
6eb57e0d | 7667 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
7668 | CPU_IDLE : CPU_NOT_IDLE; |
7669 | ||
f7ed0a89 | 7670 | rebalance_domains(this_rq, idle); |
1e3c88bd | 7671 | |
1e3c88bd | 7672 | /* |
83cd4fe2 | 7673 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd PZ |
7674 | * balancing on behalf of the other idle cpus whose ticks are |
7675 | * stopped. | |
7676 | */ | |
208cb16b | 7677 | nohz_idle_balance(this_rq, idle); |
1e3c88bd PZ |
7678 | } |
7679 | ||
1e3c88bd PZ |
7680 | /* |
7681 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 7682 | */ |
7caff66f | 7683 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 7684 | { |
1e3c88bd | 7685 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
7686 | if (unlikely(on_null_domain(rq))) |
7687 | return; | |
7688 | ||
7689 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 7690 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 7691 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 7692 | if (nohz_kick_needed(rq)) |
0aeeeeba | 7693 | nohz_balancer_kick(); |
83cd4fe2 | 7694 | #endif |
1e3c88bd PZ |
7695 | } |
7696 | ||
0bcdcf28 CE |
7697 | static void rq_online_fair(struct rq *rq) |
7698 | { | |
7699 | update_sysctl(); | |
0e59bdae KT |
7700 | |
7701 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
7702 | } |
7703 | ||
7704 | static void rq_offline_fair(struct rq *rq) | |
7705 | { | |
7706 | update_sysctl(); | |
a4c96ae3 PB |
7707 | |
7708 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
7709 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
7710 | } |
7711 | ||
55e12e5e | 7712 | #endif /* CONFIG_SMP */ |
e1d1484f | 7713 | |
bf0f6f24 IM |
7714 | /* |
7715 | * scheduler tick hitting a task of our scheduling class: | |
7716 | */ | |
8f4d37ec | 7717 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
7718 | { |
7719 | struct cfs_rq *cfs_rq; | |
7720 | struct sched_entity *se = &curr->se; | |
7721 | ||
7722 | for_each_sched_entity(se) { | |
7723 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 7724 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 7725 | } |
18bf2805 | 7726 | |
10e84b97 | 7727 | if (numabalancing_enabled) |
cbee9f88 | 7728 | task_tick_numa(rq, curr); |
3d59eebc | 7729 | |
18bf2805 | 7730 | update_rq_runnable_avg(rq, 1); |
bf0f6f24 IM |
7731 | } |
7732 | ||
7733 | /* | |
cd29fe6f PZ |
7734 | * called on fork with the child task as argument from the parent's context |
7735 | * - child not yet on the tasklist | |
7736 | * - preemption disabled | |
bf0f6f24 | 7737 | */ |
cd29fe6f | 7738 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 7739 | { |
4fc420c9 DN |
7740 | struct cfs_rq *cfs_rq; |
7741 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 7742 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
7743 | struct rq *rq = this_rq(); |
7744 | unsigned long flags; | |
7745 | ||
05fa785c | 7746 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 7747 | |
861d034e PZ |
7748 | update_rq_clock(rq); |
7749 | ||
4fc420c9 DN |
7750 | cfs_rq = task_cfs_rq(current); |
7751 | curr = cfs_rq->curr; | |
7752 | ||
6c9a27f5 DN |
7753 | /* |
7754 | * Not only the cpu but also the task_group of the parent might have | |
7755 | * been changed after parent->se.parent,cfs_rq were copied to | |
7756 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
7757 | * of child point to valid ones. | |
7758 | */ | |
7759 | rcu_read_lock(); | |
7760 | __set_task_cpu(p, this_cpu); | |
7761 | rcu_read_unlock(); | |
bf0f6f24 | 7762 | |
7109c442 | 7763 | update_curr(cfs_rq); |
cd29fe6f | 7764 | |
b5d9d734 MG |
7765 | if (curr) |
7766 | se->vruntime = curr->vruntime; | |
aeb73b04 | 7767 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 7768 | |
cd29fe6f | 7769 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 7770 | /* |
edcb60a3 IM |
7771 | * Upon rescheduling, sched_class::put_prev_task() will place |
7772 | * 'current' within the tree based on its new key value. | |
7773 | */ | |
4d78e7b6 | 7774 | swap(curr->vruntime, se->vruntime); |
8875125e | 7775 | resched_curr(rq); |
4d78e7b6 | 7776 | } |
bf0f6f24 | 7777 | |
88ec22d3 PZ |
7778 | se->vruntime -= cfs_rq->min_vruntime; |
7779 | ||
05fa785c | 7780 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
7781 | } |
7782 | ||
cb469845 SR |
7783 | /* |
7784 | * Priority of the task has changed. Check to see if we preempt | |
7785 | * the current task. | |
7786 | */ | |
da7a735e PZ |
7787 | static void |
7788 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 7789 | { |
da0c1e65 | 7790 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
7791 | return; |
7792 | ||
cb469845 SR |
7793 | /* |
7794 | * Reschedule if we are currently running on this runqueue and | |
7795 | * our priority decreased, or if we are not currently running on | |
7796 | * this runqueue and our priority is higher than the current's | |
7797 | */ | |
da7a735e | 7798 | if (rq->curr == p) { |
cb469845 | 7799 | if (p->prio > oldprio) |
8875125e | 7800 | resched_curr(rq); |
cb469845 | 7801 | } else |
15afe09b | 7802 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7803 | } |
7804 | ||
da7a735e PZ |
7805 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
7806 | { | |
7807 | struct sched_entity *se = &p->se; | |
7808 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7809 | ||
7810 | /* | |
791c9e02 | 7811 | * Ensure the task's vruntime is normalized, so that when it's |
da7a735e PZ |
7812 | * switched back to the fair class the enqueue_entity(.flags=0) will |
7813 | * do the right thing. | |
7814 | * | |
da0c1e65 KT |
7815 | * If it's queued, then the dequeue_entity(.flags=0) will already |
7816 | * have normalized the vruntime, if it's !queued, then only when | |
da7a735e PZ |
7817 | * the task is sleeping will it still have non-normalized vruntime. |
7818 | */ | |
da0c1e65 | 7819 | if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) { |
da7a735e PZ |
7820 | /* |
7821 | * Fix up our vruntime so that the current sleep doesn't | |
7822 | * cause 'unlimited' sleep bonus. | |
7823 | */ | |
7824 | place_entity(cfs_rq, se, 0); | |
7825 | se->vruntime -= cfs_rq->min_vruntime; | |
7826 | } | |
9ee474f5 | 7827 | |
141965c7 | 7828 | #ifdef CONFIG_SMP |
9ee474f5 PT |
7829 | /* |
7830 | * Remove our load from contribution when we leave sched_fair | |
7831 | * and ensure we don't carry in an old decay_count if we | |
7832 | * switch back. | |
7833 | */ | |
87e3c8ae KT |
7834 | if (se->avg.decay_count) { |
7835 | __synchronize_entity_decay(se); | |
7836 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | |
9ee474f5 PT |
7837 | } |
7838 | #endif | |
da7a735e PZ |
7839 | } |
7840 | ||
cb469845 SR |
7841 | /* |
7842 | * We switched to the sched_fair class. | |
7843 | */ | |
da7a735e | 7844 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
cb469845 | 7845 | { |
eb7a59b2 | 7846 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f36c019c | 7847 | struct sched_entity *se = &p->se; |
eb7a59b2 M |
7848 | /* |
7849 | * Since the real-depth could have been changed (only FAIR | |
7850 | * class maintain depth value), reset depth properly. | |
7851 | */ | |
7852 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
7853 | #endif | |
da0c1e65 | 7854 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
7855 | return; |
7856 | ||
cb469845 SR |
7857 | /* |
7858 | * We were most likely switched from sched_rt, so | |
7859 | * kick off the schedule if running, otherwise just see | |
7860 | * if we can still preempt the current task. | |
7861 | */ | |
da7a735e | 7862 | if (rq->curr == p) |
8875125e | 7863 | resched_curr(rq); |
cb469845 | 7864 | else |
15afe09b | 7865 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
7866 | } |
7867 | ||
83b699ed SV |
7868 | /* Account for a task changing its policy or group. |
7869 | * | |
7870 | * This routine is mostly called to set cfs_rq->curr field when a task | |
7871 | * migrates between groups/classes. | |
7872 | */ | |
7873 | static void set_curr_task_fair(struct rq *rq) | |
7874 | { | |
7875 | struct sched_entity *se = &rq->curr->se; | |
7876 | ||
ec12cb7f PT |
7877 | for_each_sched_entity(se) { |
7878 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7879 | ||
7880 | set_next_entity(cfs_rq, se); | |
7881 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
7882 | account_cfs_rq_runtime(cfs_rq, 0); | |
7883 | } | |
83b699ed SV |
7884 | } |
7885 | ||
029632fb PZ |
7886 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
7887 | { | |
7888 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
7889 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
7890 | #ifndef CONFIG_64BIT | |
7891 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
7892 | #endif | |
141965c7 | 7893 | #ifdef CONFIG_SMP |
9ee474f5 | 7894 | atomic64_set(&cfs_rq->decay_counter, 1); |
2509940f | 7895 | atomic_long_set(&cfs_rq->removed_load, 0); |
9ee474f5 | 7896 | #endif |
029632fb PZ |
7897 | } |
7898 | ||
810b3817 | 7899 | #ifdef CONFIG_FAIR_GROUP_SCHED |
da0c1e65 | 7900 | static void task_move_group_fair(struct task_struct *p, int queued) |
810b3817 | 7901 | { |
fed14d45 | 7902 | struct sched_entity *se = &p->se; |
aff3e498 | 7903 | struct cfs_rq *cfs_rq; |
fed14d45 | 7904 | |
b2b5ce02 PZ |
7905 | /* |
7906 | * If the task was not on the rq at the time of this cgroup movement | |
7907 | * it must have been asleep, sleeping tasks keep their ->vruntime | |
7908 | * absolute on their old rq until wakeup (needed for the fair sleeper | |
7909 | * bonus in place_entity()). | |
7910 | * | |
7911 | * If it was on the rq, we've just 'preempted' it, which does convert | |
7912 | * ->vruntime to a relative base. | |
7913 | * | |
7914 | * Make sure both cases convert their relative position when migrating | |
7915 | * to another cgroup's rq. This does somewhat interfere with the | |
7916 | * fair sleeper stuff for the first placement, but who cares. | |
7917 | */ | |
7ceff013 | 7918 | /* |
da0c1e65 | 7919 | * When !queued, vruntime of the task has usually NOT been normalized. |
7ceff013 DN |
7920 | * But there are some cases where it has already been normalized: |
7921 | * | |
7922 | * - Moving a forked child which is waiting for being woken up by | |
7923 | * wake_up_new_task(). | |
62af3783 DN |
7924 | * - Moving a task which has been woken up by try_to_wake_up() and |
7925 | * waiting for actually being woken up by sched_ttwu_pending(). | |
7ceff013 DN |
7926 | * |
7927 | * To prevent boost or penalty in the new cfs_rq caused by delta | |
7928 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | |
7929 | */ | |
da0c1e65 KT |
7930 | if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING)) |
7931 | queued = 1; | |
7ceff013 | 7932 | |
da0c1e65 | 7933 | if (!queued) |
fed14d45 | 7934 | se->vruntime -= cfs_rq_of(se)->min_vruntime; |
b2b5ce02 | 7935 | set_task_rq(p, task_cpu(p)); |
fed14d45 | 7936 | se->depth = se->parent ? se->parent->depth + 1 : 0; |
da0c1e65 | 7937 | if (!queued) { |
fed14d45 PZ |
7938 | cfs_rq = cfs_rq_of(se); |
7939 | se->vruntime += cfs_rq->min_vruntime; | |
aff3e498 PT |
7940 | #ifdef CONFIG_SMP |
7941 | /* | |
7942 | * migrate_task_rq_fair() will have removed our previous | |
7943 | * contribution, but we must synchronize for ongoing future | |
7944 | * decay. | |
7945 | */ | |
fed14d45 PZ |
7946 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); |
7947 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | |
aff3e498 PT |
7948 | #endif |
7949 | } | |
810b3817 | 7950 | } |
029632fb PZ |
7951 | |
7952 | void free_fair_sched_group(struct task_group *tg) | |
7953 | { | |
7954 | int i; | |
7955 | ||
7956 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7957 | ||
7958 | for_each_possible_cpu(i) { | |
7959 | if (tg->cfs_rq) | |
7960 | kfree(tg->cfs_rq[i]); | |
7961 | if (tg->se) | |
7962 | kfree(tg->se[i]); | |
7963 | } | |
7964 | ||
7965 | kfree(tg->cfs_rq); | |
7966 | kfree(tg->se); | |
7967 | } | |
7968 | ||
7969 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
7970 | { | |
7971 | struct cfs_rq *cfs_rq; | |
7972 | struct sched_entity *se; | |
7973 | int i; | |
7974 | ||
7975 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
7976 | if (!tg->cfs_rq) | |
7977 | goto err; | |
7978 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
7979 | if (!tg->se) | |
7980 | goto err; | |
7981 | ||
7982 | tg->shares = NICE_0_LOAD; | |
7983 | ||
7984 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
7985 | ||
7986 | for_each_possible_cpu(i) { | |
7987 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
7988 | GFP_KERNEL, cpu_to_node(i)); | |
7989 | if (!cfs_rq) | |
7990 | goto err; | |
7991 | ||
7992 | se = kzalloc_node(sizeof(struct sched_entity), | |
7993 | GFP_KERNEL, cpu_to_node(i)); | |
7994 | if (!se) | |
7995 | goto err_free_rq; | |
7996 | ||
7997 | init_cfs_rq(cfs_rq); | |
7998 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
7999 | } | |
8000 | ||
8001 | return 1; | |
8002 | ||
8003 | err_free_rq: | |
8004 | kfree(cfs_rq); | |
8005 | err: | |
8006 | return 0; | |
8007 | } | |
8008 | ||
8009 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8010 | { | |
8011 | struct rq *rq = cpu_rq(cpu); | |
8012 | unsigned long flags; | |
8013 | ||
8014 | /* | |
8015 | * Only empty task groups can be destroyed; so we can speculatively | |
8016 | * check on_list without danger of it being re-added. | |
8017 | */ | |
8018 | if (!tg->cfs_rq[cpu]->on_list) | |
8019 | return; | |
8020 | ||
8021 | raw_spin_lock_irqsave(&rq->lock, flags); | |
8022 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
8023 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8024 | } | |
8025 | ||
8026 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
8027 | struct sched_entity *se, int cpu, | |
8028 | struct sched_entity *parent) | |
8029 | { | |
8030 | struct rq *rq = cpu_rq(cpu); | |
8031 | ||
8032 | cfs_rq->tg = tg; | |
8033 | cfs_rq->rq = rq; | |
029632fb PZ |
8034 | init_cfs_rq_runtime(cfs_rq); |
8035 | ||
8036 | tg->cfs_rq[cpu] = cfs_rq; | |
8037 | tg->se[cpu] = se; | |
8038 | ||
8039 | /* se could be NULL for root_task_group */ | |
8040 | if (!se) | |
8041 | return; | |
8042 | ||
fed14d45 | 8043 | if (!parent) { |
029632fb | 8044 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
8045 | se->depth = 0; |
8046 | } else { | |
029632fb | 8047 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
8048 | se->depth = parent->depth + 1; |
8049 | } | |
029632fb PZ |
8050 | |
8051 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
8052 | /* guarantee group entities always have weight */ |
8053 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
8054 | se->parent = parent; |
8055 | } | |
8056 | ||
8057 | static DEFINE_MUTEX(shares_mutex); | |
8058 | ||
8059 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
8060 | { | |
8061 | int i; | |
8062 | unsigned long flags; | |
8063 | ||
8064 | /* | |
8065 | * We can't change the weight of the root cgroup. | |
8066 | */ | |
8067 | if (!tg->se[0]) | |
8068 | return -EINVAL; | |
8069 | ||
8070 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
8071 | ||
8072 | mutex_lock(&shares_mutex); | |
8073 | if (tg->shares == shares) | |
8074 | goto done; | |
8075 | ||
8076 | tg->shares = shares; | |
8077 | for_each_possible_cpu(i) { | |
8078 | struct rq *rq = cpu_rq(i); | |
8079 | struct sched_entity *se; | |
8080 | ||
8081 | se = tg->se[i]; | |
8082 | /* Propagate contribution to hierarchy */ | |
8083 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
8084 | |
8085 | /* Possible calls to update_curr() need rq clock */ | |
8086 | update_rq_clock(rq); | |
17bc14b7 | 8087 | for_each_sched_entity(se) |
029632fb PZ |
8088 | update_cfs_shares(group_cfs_rq(se)); |
8089 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8090 | } | |
8091 | ||
8092 | done: | |
8093 | mutex_unlock(&shares_mutex); | |
8094 | return 0; | |
8095 | } | |
8096 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
8097 | ||
8098 | void free_fair_sched_group(struct task_group *tg) { } | |
8099 | ||
8100 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8101 | { | |
8102 | return 1; | |
8103 | } | |
8104 | ||
8105 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
8106 | ||
8107 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
8108 | ||
810b3817 | 8109 | |
6d686f45 | 8110 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
8111 | { |
8112 | struct sched_entity *se = &task->se; | |
0d721cea PW |
8113 | unsigned int rr_interval = 0; |
8114 | ||
8115 | /* | |
8116 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
8117 | * idle runqueue: | |
8118 | */ | |
0d721cea | 8119 | if (rq->cfs.load.weight) |
a59f4e07 | 8120 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
8121 | |
8122 | return rr_interval; | |
8123 | } | |
8124 | ||
bf0f6f24 IM |
8125 | /* |
8126 | * All the scheduling class methods: | |
8127 | */ | |
029632fb | 8128 | const struct sched_class fair_sched_class = { |
5522d5d5 | 8129 | .next = &idle_sched_class, |
bf0f6f24 IM |
8130 | .enqueue_task = enqueue_task_fair, |
8131 | .dequeue_task = dequeue_task_fair, | |
8132 | .yield_task = yield_task_fair, | |
d95f4122 | 8133 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 8134 | |
2e09bf55 | 8135 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
8136 | |
8137 | .pick_next_task = pick_next_task_fair, | |
8138 | .put_prev_task = put_prev_task_fair, | |
8139 | ||
681f3e68 | 8140 | #ifdef CONFIG_SMP |
4ce72a2c | 8141 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 8142 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 8143 | |
0bcdcf28 CE |
8144 | .rq_online = rq_online_fair, |
8145 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
8146 | |
8147 | .task_waking = task_waking_fair, | |
681f3e68 | 8148 | #endif |
bf0f6f24 | 8149 | |
83b699ed | 8150 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 8151 | .task_tick = task_tick_fair, |
cd29fe6f | 8152 | .task_fork = task_fork_fair, |
cb469845 SR |
8153 | |
8154 | .prio_changed = prio_changed_fair, | |
da7a735e | 8155 | .switched_from = switched_from_fair, |
cb469845 | 8156 | .switched_to = switched_to_fair, |
810b3817 | 8157 | |
0d721cea PW |
8158 | .get_rr_interval = get_rr_interval_fair, |
8159 | ||
6e998916 SG |
8160 | .update_curr = update_curr_fair, |
8161 | ||
810b3817 | 8162 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 8163 | .task_move_group = task_move_group_fair, |
810b3817 | 8164 | #endif |
bf0f6f24 IM |
8165 | }; |
8166 | ||
8167 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 8168 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 8169 | { |
bf0f6f24 IM |
8170 | struct cfs_rq *cfs_rq; |
8171 | ||
5973e5b9 | 8172 | rcu_read_lock(); |
c3b64f1e | 8173 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 8174 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 8175 | rcu_read_unlock(); |
bf0f6f24 IM |
8176 | } |
8177 | #endif | |
029632fb PZ |
8178 | |
8179 | __init void init_sched_fair_class(void) | |
8180 | { | |
8181 | #ifdef CONFIG_SMP | |
8182 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
8183 | ||
3451d024 | 8184 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 8185 | nohz.next_balance = jiffies; |
029632fb | 8186 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 8187 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
8188 | #endif |
8189 | #endif /* SMP */ | |
8190 | ||
8191 | } |