]>
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 | |
90eec103 | 20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra |
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 | */ | |
58ac93e4 | 144 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 145 | { |
58ac93e4 | 146 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
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 | ||
3d4b47b4 PZ |
286 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
287 | { | |
288 | if (!cfs_rq->on_list) { | |
67e86250 PT |
289 | /* |
290 | * Ensure we either appear before our parent (if already | |
291 | * enqueued) or force our parent to appear after us when it is | |
292 | * enqueued. The fact that we always enqueue bottom-up | |
293 | * reduces this to two cases. | |
294 | */ | |
295 | if (cfs_rq->tg->parent && | |
296 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
297 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
298 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
299 | } else { | |
300 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 302 | } |
3d4b47b4 PZ |
303 | |
304 | cfs_rq->on_list = 1; | |
305 | } | |
306 | } | |
307 | ||
308 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
309 | { | |
310 | if (cfs_rq->on_list) { | |
311 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
312 | cfs_rq->on_list = 0; | |
313 | } | |
314 | } | |
315 | ||
b758149c PZ |
316 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
317 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
318 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
319 | ||
320 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 321 | static inline struct cfs_rq * |
b758149c PZ |
322 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
323 | { | |
324 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 325 | return se->cfs_rq; |
b758149c | 326 | |
fed14d45 | 327 | return NULL; |
b758149c PZ |
328 | } |
329 | ||
330 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
331 | { | |
332 | return se->parent; | |
333 | } | |
334 | ||
464b7527 PZ |
335 | static void |
336 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
337 | { | |
338 | int se_depth, pse_depth; | |
339 | ||
340 | /* | |
341 | * preemption test can be made between sibling entities who are in the | |
342 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
343 | * both tasks until we find their ancestors who are siblings of common | |
344 | * parent. | |
345 | */ | |
346 | ||
347 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
348 | se_depth = (*se)->depth; |
349 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
350 | |
351 | while (se_depth > pse_depth) { | |
352 | se_depth--; | |
353 | *se = parent_entity(*se); | |
354 | } | |
355 | ||
356 | while (pse_depth > se_depth) { | |
357 | pse_depth--; | |
358 | *pse = parent_entity(*pse); | |
359 | } | |
360 | ||
361 | while (!is_same_group(*se, *pse)) { | |
362 | *se = parent_entity(*se); | |
363 | *pse = parent_entity(*pse); | |
364 | } | |
365 | } | |
366 | ||
8f48894f PZ |
367 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
368 | ||
369 | static inline struct task_struct *task_of(struct sched_entity *se) | |
370 | { | |
371 | return container_of(se, struct task_struct, se); | |
372 | } | |
bf0f6f24 | 373 | |
62160e3f IM |
374 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
375 | { | |
376 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
377 | } |
378 | ||
379 | #define entity_is_task(se) 1 | |
380 | ||
b758149c PZ |
381 | #define for_each_sched_entity(se) \ |
382 | for (; se; se = NULL) | |
bf0f6f24 | 383 | |
b758149c | 384 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 385 | { |
b758149c | 386 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
387 | } |
388 | ||
b758149c PZ |
389 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
390 | { | |
391 | struct task_struct *p = task_of(se); | |
392 | struct rq *rq = task_rq(p); | |
393 | ||
394 | return &rq->cfs; | |
395 | } | |
396 | ||
397 | /* runqueue "owned" by this group */ | |
398 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
399 | { | |
400 | return NULL; | |
401 | } | |
402 | ||
3d4b47b4 PZ |
403 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
404 | { | |
405 | } | |
406 | ||
407 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
408 | { | |
409 | } | |
410 | ||
b758149c PZ |
411 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
412 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
413 | ||
b758149c PZ |
414 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
415 | { | |
416 | return NULL; | |
417 | } | |
418 | ||
464b7527 PZ |
419 | static inline void |
420 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
421 | { | |
422 | } | |
423 | ||
b758149c PZ |
424 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
425 | ||
6c16a6dc | 426 | static __always_inline |
9dbdb155 | 427 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
428 | |
429 | /************************************************************** | |
430 | * Scheduling class tree data structure manipulation methods: | |
431 | */ | |
432 | ||
1bf08230 | 433 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 434 | { |
1bf08230 | 435 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 436 | if (delta > 0) |
1bf08230 | 437 | max_vruntime = vruntime; |
02e0431a | 438 | |
1bf08230 | 439 | return max_vruntime; |
02e0431a PZ |
440 | } |
441 | ||
0702e3eb | 442 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
443 | { |
444 | s64 delta = (s64)(vruntime - min_vruntime); | |
445 | if (delta < 0) | |
446 | min_vruntime = vruntime; | |
447 | ||
448 | return min_vruntime; | |
449 | } | |
450 | ||
54fdc581 FC |
451 | static inline int entity_before(struct sched_entity *a, |
452 | struct sched_entity *b) | |
453 | { | |
454 | return (s64)(a->vruntime - b->vruntime) < 0; | |
455 | } | |
456 | ||
1af5f730 PZ |
457 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
458 | { | |
459 | u64 vruntime = cfs_rq->min_vruntime; | |
460 | ||
461 | if (cfs_rq->curr) | |
462 | vruntime = cfs_rq->curr->vruntime; | |
463 | ||
464 | if (cfs_rq->rb_leftmost) { | |
465 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
466 | struct sched_entity, | |
467 | run_node); | |
468 | ||
e17036da | 469 | if (!cfs_rq->curr) |
1af5f730 PZ |
470 | vruntime = se->vruntime; |
471 | else | |
472 | vruntime = min_vruntime(vruntime, se->vruntime); | |
473 | } | |
474 | ||
1bf08230 | 475 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 476 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
477 | #ifndef CONFIG_64BIT |
478 | smp_wmb(); | |
479 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
480 | #endif | |
1af5f730 PZ |
481 | } |
482 | ||
bf0f6f24 IM |
483 | /* |
484 | * Enqueue an entity into the rb-tree: | |
485 | */ | |
0702e3eb | 486 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
487 | { |
488 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
489 | struct rb_node *parent = NULL; | |
490 | struct sched_entity *entry; | |
bf0f6f24 IM |
491 | int leftmost = 1; |
492 | ||
493 | /* | |
494 | * Find the right place in the rbtree: | |
495 | */ | |
496 | while (*link) { | |
497 | parent = *link; | |
498 | entry = rb_entry(parent, struct sched_entity, run_node); | |
499 | /* | |
500 | * We dont care about collisions. Nodes with | |
501 | * the same key stay together. | |
502 | */ | |
2bd2d6f2 | 503 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
504 | link = &parent->rb_left; |
505 | } else { | |
506 | link = &parent->rb_right; | |
507 | leftmost = 0; | |
508 | } | |
509 | } | |
510 | ||
511 | /* | |
512 | * Maintain a cache of leftmost tree entries (it is frequently | |
513 | * used): | |
514 | */ | |
1af5f730 | 515 | if (leftmost) |
57cb499d | 516 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
517 | |
518 | rb_link_node(&se->run_node, parent, link); | |
519 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
520 | } |
521 | ||
0702e3eb | 522 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 523 | { |
3fe69747 PZ |
524 | if (cfs_rq->rb_leftmost == &se->run_node) { |
525 | struct rb_node *next_node; | |
3fe69747 PZ |
526 | |
527 | next_node = rb_next(&se->run_node); | |
528 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 529 | } |
e9acbff6 | 530 | |
bf0f6f24 | 531 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
532 | } |
533 | ||
029632fb | 534 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 535 | { |
f4b6755f PZ |
536 | struct rb_node *left = cfs_rq->rb_leftmost; |
537 | ||
538 | if (!left) | |
539 | return NULL; | |
540 | ||
541 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
542 | } |
543 | ||
ac53db59 RR |
544 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
545 | { | |
546 | struct rb_node *next = rb_next(&se->run_node); | |
547 | ||
548 | if (!next) | |
549 | return NULL; | |
550 | ||
551 | return rb_entry(next, struct sched_entity, run_node); | |
552 | } | |
553 | ||
554 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 555 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 556 | { |
7eee3e67 | 557 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 558 | |
70eee74b BS |
559 | if (!last) |
560 | return NULL; | |
7eee3e67 IM |
561 | |
562 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
563 | } |
564 | ||
bf0f6f24 IM |
565 | /************************************************************** |
566 | * Scheduling class statistics methods: | |
567 | */ | |
568 | ||
acb4a848 | 569 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 570 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
571 | loff_t *ppos) |
572 | { | |
8d65af78 | 573 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 574 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
575 | |
576 | if (ret || !write) | |
577 | return ret; | |
578 | ||
579 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
580 | sysctl_sched_min_granularity); | |
581 | ||
acb4a848 CE |
582 | #define WRT_SYSCTL(name) \ |
583 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
584 | WRT_SYSCTL(sched_min_granularity); | |
585 | WRT_SYSCTL(sched_latency); | |
586 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
587 | #undef WRT_SYSCTL |
588 | ||
b2be5e96 PZ |
589 | return 0; |
590 | } | |
591 | #endif | |
647e7cac | 592 | |
a7be37ac | 593 | /* |
f9c0b095 | 594 | * delta /= w |
a7be37ac | 595 | */ |
9dbdb155 | 596 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 597 | { |
f9c0b095 | 598 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 599 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
600 | |
601 | return delta; | |
602 | } | |
603 | ||
647e7cac IM |
604 | /* |
605 | * The idea is to set a period in which each task runs once. | |
606 | * | |
532b1858 | 607 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
608 | * this period because otherwise the slices get too small. |
609 | * | |
610 | * p = (nr <= nl) ? l : l*nr/nl | |
611 | */ | |
4d78e7b6 PZ |
612 | static u64 __sched_period(unsigned long nr_running) |
613 | { | |
8e2b0bf3 BF |
614 | if (unlikely(nr_running > sched_nr_latency)) |
615 | return nr_running * sysctl_sched_min_granularity; | |
616 | else | |
617 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
618 | } |
619 | ||
647e7cac IM |
620 | /* |
621 | * We calculate the wall-time slice from the period by taking a part | |
622 | * proportional to the weight. | |
623 | * | |
f9c0b095 | 624 | * s = p*P[w/rw] |
647e7cac | 625 | */ |
6d0f0ebd | 626 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 627 | { |
0a582440 | 628 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 629 | |
0a582440 | 630 | for_each_sched_entity(se) { |
6272d68c | 631 | struct load_weight *load; |
3104bf03 | 632 | struct load_weight lw; |
6272d68c LM |
633 | |
634 | cfs_rq = cfs_rq_of(se); | |
635 | load = &cfs_rq->load; | |
f9c0b095 | 636 | |
0a582440 | 637 | if (unlikely(!se->on_rq)) { |
3104bf03 | 638 | lw = cfs_rq->load; |
0a582440 MG |
639 | |
640 | update_load_add(&lw, se->load.weight); | |
641 | load = &lw; | |
642 | } | |
9dbdb155 | 643 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
644 | } |
645 | return slice; | |
bf0f6f24 IM |
646 | } |
647 | ||
647e7cac | 648 | /* |
660cc00f | 649 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 650 | * |
f9c0b095 | 651 | * vs = s/w |
647e7cac | 652 | */ |
f9c0b095 | 653 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 654 | { |
f9c0b095 | 655 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
656 | } |
657 | ||
a75cdaa9 | 658 | #ifdef CONFIG_SMP |
ba7e5a27 | 659 | static int select_idle_sibling(struct task_struct *p, int cpu); |
fb13c7ee MG |
660 | static unsigned long task_h_load(struct task_struct *p); |
661 | ||
9d89c257 YD |
662 | /* |
663 | * We choose a half-life close to 1 scheduling period. | |
84fb5a18 LY |
664 | * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are |
665 | * dependent on this value. | |
9d89c257 YD |
666 | */ |
667 | #define LOAD_AVG_PERIOD 32 | |
668 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
84fb5a18 | 669 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */ |
a75cdaa9 | 670 | |
540247fb YD |
671 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
672 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 673 | { |
540247fb | 674 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 675 | |
9d89c257 YD |
676 | sa->last_update_time = 0; |
677 | /* | |
678 | * sched_avg's period_contrib should be strictly less then 1024, so | |
679 | * we give it 1023 to make sure it is almost a period (1024us), and | |
680 | * will definitely be update (after enqueue). | |
681 | */ | |
682 | sa->period_contrib = 1023; | |
540247fb | 683 | sa->load_avg = scale_load_down(se->load.weight); |
9d89c257 YD |
684 | sa->load_sum = sa->load_avg * LOAD_AVG_MAX; |
685 | sa->util_avg = scale_load_down(SCHED_LOAD_SCALE); | |
006cdf02 | 686 | sa->util_sum = sa->util_avg * LOAD_AVG_MAX; |
9d89c257 | 687 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 688 | } |
7ea241af YD |
689 | |
690 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq); | |
691 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq); | |
a75cdaa9 | 692 | #else |
540247fb | 693 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
694 | { |
695 | } | |
696 | #endif | |
697 | ||
bf0f6f24 | 698 | /* |
9dbdb155 | 699 | * Update the current task's runtime statistics. |
bf0f6f24 | 700 | */ |
b7cc0896 | 701 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 702 | { |
429d43bc | 703 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 704 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 705 | u64 delta_exec; |
bf0f6f24 IM |
706 | |
707 | if (unlikely(!curr)) | |
708 | return; | |
709 | ||
9dbdb155 PZ |
710 | delta_exec = now - curr->exec_start; |
711 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 712 | return; |
bf0f6f24 | 713 | |
8ebc91d9 | 714 | curr->exec_start = now; |
d842de87 | 715 | |
9dbdb155 PZ |
716 | schedstat_set(curr->statistics.exec_max, |
717 | max(delta_exec, curr->statistics.exec_max)); | |
718 | ||
719 | curr->sum_exec_runtime += delta_exec; | |
720 | schedstat_add(cfs_rq, exec_clock, delta_exec); | |
721 | ||
722 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
723 | update_min_vruntime(cfs_rq); | |
724 | ||
d842de87 SV |
725 | if (entity_is_task(curr)) { |
726 | struct task_struct *curtask = task_of(curr); | |
727 | ||
f977bb49 | 728 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 729 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 730 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 731 | } |
ec12cb7f PT |
732 | |
733 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
734 | } |
735 | ||
6e998916 SG |
736 | static void update_curr_fair(struct rq *rq) |
737 | { | |
738 | update_curr(cfs_rq_of(&rq->curr->se)); | |
739 | } | |
740 | ||
3ea94de1 | 741 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 | 742 | static inline void |
5870db5b | 743 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 744 | { |
3ea94de1 JP |
745 | u64 wait_start = rq_clock(rq_of(cfs_rq)); |
746 | ||
747 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
748 | likely(wait_start > se->statistics.wait_start)) | |
749 | wait_start -= se->statistics.wait_start; | |
750 | ||
751 | se->statistics.wait_start = wait_start; | |
bf0f6f24 IM |
752 | } |
753 | ||
3ea94de1 JP |
754 | static void |
755 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
756 | { | |
757 | struct task_struct *p; | |
758 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start; | |
759 | ||
760 | if (entity_is_task(se)) { | |
761 | p = task_of(se); | |
762 | if (task_on_rq_migrating(p)) { | |
763 | /* | |
764 | * Preserve migrating task's wait time so wait_start | |
765 | * time stamp can be adjusted to accumulate wait time | |
766 | * prior to migration. | |
767 | */ | |
768 | se->statistics.wait_start = delta; | |
769 | return; | |
770 | } | |
771 | trace_sched_stat_wait(p, delta); | |
772 | } | |
773 | ||
774 | se->statistics.wait_max = max(se->statistics.wait_max, delta); | |
775 | se->statistics.wait_count++; | |
776 | se->statistics.wait_sum += delta; | |
777 | se->statistics.wait_start = 0; | |
778 | } | |
779 | #else | |
780 | static inline void | |
781 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
782 | { | |
783 | } | |
784 | ||
785 | static inline void | |
786 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
787 | { | |
788 | } | |
789 | #endif | |
790 | ||
bf0f6f24 IM |
791 | /* |
792 | * Task is being enqueued - update stats: | |
793 | */ | |
d2417e5a | 794 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 795 | { |
bf0f6f24 IM |
796 | /* |
797 | * Are we enqueueing a waiting task? (for current tasks | |
798 | * a dequeue/enqueue event is a NOP) | |
799 | */ | |
429d43bc | 800 | if (se != cfs_rq->curr) |
5870db5b | 801 | update_stats_wait_start(cfs_rq, se); |
bf0f6f24 IM |
802 | } |
803 | ||
bf0f6f24 | 804 | static inline void |
19b6a2e3 | 805 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 806 | { |
bf0f6f24 IM |
807 | /* |
808 | * Mark the end of the wait period if dequeueing a | |
809 | * waiting task: | |
810 | */ | |
429d43bc | 811 | if (se != cfs_rq->curr) |
9ef0a961 | 812 | update_stats_wait_end(cfs_rq, se); |
bf0f6f24 IM |
813 | } |
814 | ||
815 | /* | |
816 | * We are picking a new current task - update its stats: | |
817 | */ | |
818 | static inline void | |
79303e9e | 819 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
820 | { |
821 | /* | |
822 | * We are starting a new run period: | |
823 | */ | |
78becc27 | 824 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
825 | } |
826 | ||
bf0f6f24 IM |
827 | /************************************************** |
828 | * Scheduling class queueing methods: | |
829 | */ | |
830 | ||
cbee9f88 PZ |
831 | #ifdef CONFIG_NUMA_BALANCING |
832 | /* | |
598f0ec0 MG |
833 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
834 | * calculated based on the tasks virtual memory size and | |
835 | * numa_balancing_scan_size. | |
cbee9f88 | 836 | */ |
598f0ec0 MG |
837 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
838 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
839 | |
840 | /* Portion of address space to scan in MB */ | |
841 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 842 | |
4b96a29b PZ |
843 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
844 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
845 | ||
598f0ec0 MG |
846 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
847 | { | |
848 | unsigned long rss = 0; | |
849 | unsigned long nr_scan_pages; | |
850 | ||
851 | /* | |
852 | * Calculations based on RSS as non-present and empty pages are skipped | |
853 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
854 | * on resident pages | |
855 | */ | |
856 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
857 | rss = get_mm_rss(p->mm); | |
858 | if (!rss) | |
859 | rss = nr_scan_pages; | |
860 | ||
861 | rss = round_up(rss, nr_scan_pages); | |
862 | return rss / nr_scan_pages; | |
863 | } | |
864 | ||
865 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
866 | #define MAX_SCAN_WINDOW 2560 | |
867 | ||
868 | static unsigned int task_scan_min(struct task_struct *p) | |
869 | { | |
316c1608 | 870 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
871 | unsigned int scan, floor; |
872 | unsigned int windows = 1; | |
873 | ||
64192658 KT |
874 | if (scan_size < MAX_SCAN_WINDOW) |
875 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
876 | floor = 1000 / windows; |
877 | ||
878 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
879 | return max_t(unsigned int, floor, scan); | |
880 | } | |
881 | ||
882 | static unsigned int task_scan_max(struct task_struct *p) | |
883 | { | |
884 | unsigned int smin = task_scan_min(p); | |
885 | unsigned int smax; | |
886 | ||
887 | /* Watch for min being lower than max due to floor calculations */ | |
888 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
889 | return max(smin, smax); | |
890 | } | |
891 | ||
0ec8aa00 PZ |
892 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
893 | { | |
894 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
895 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
896 | } | |
897 | ||
898 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
899 | { | |
900 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
901 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
902 | } | |
903 | ||
8c8a743c PZ |
904 | struct numa_group { |
905 | atomic_t refcount; | |
906 | ||
907 | spinlock_t lock; /* nr_tasks, tasks */ | |
908 | int nr_tasks; | |
e29cf08b | 909 | pid_t gid; |
8c8a743c PZ |
910 | |
911 | struct rcu_head rcu; | |
20e07dea | 912 | nodemask_t active_nodes; |
989348b5 | 913 | unsigned long total_faults; |
7e2703e6 RR |
914 | /* |
915 | * Faults_cpu is used to decide whether memory should move | |
916 | * towards the CPU. As a consequence, these stats are weighted | |
917 | * more by CPU use than by memory faults. | |
918 | */ | |
50ec8a40 | 919 | unsigned long *faults_cpu; |
989348b5 | 920 | unsigned long faults[0]; |
8c8a743c PZ |
921 | }; |
922 | ||
be1e4e76 RR |
923 | /* Shared or private faults. */ |
924 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
925 | ||
926 | /* Memory and CPU locality */ | |
927 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
928 | ||
929 | /* Averaged statistics, and temporary buffers. */ | |
930 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
931 | ||
e29cf08b MG |
932 | pid_t task_numa_group_id(struct task_struct *p) |
933 | { | |
934 | return p->numa_group ? p->numa_group->gid : 0; | |
935 | } | |
936 | ||
44dba3d5 IM |
937 | /* |
938 | * The averaged statistics, shared & private, memory & cpu, | |
939 | * occupy the first half of the array. The second half of the | |
940 | * array is for current counters, which are averaged into the | |
941 | * first set by task_numa_placement. | |
942 | */ | |
943 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 944 | { |
44dba3d5 | 945 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
946 | } |
947 | ||
948 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
949 | { | |
44dba3d5 | 950 | if (!p->numa_faults) |
ac8e895b MG |
951 | return 0; |
952 | ||
44dba3d5 IM |
953 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
954 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
955 | } |
956 | ||
83e1d2cd MG |
957 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
958 | { | |
959 | if (!p->numa_group) | |
960 | return 0; | |
961 | ||
44dba3d5 IM |
962 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
963 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
964 | } |
965 | ||
20e07dea RR |
966 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
967 | { | |
44dba3d5 IM |
968 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
969 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
970 | } |
971 | ||
6c6b1193 RR |
972 | /* Handle placement on systems where not all nodes are directly connected. */ |
973 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
974 | int maxdist, bool task) | |
975 | { | |
976 | unsigned long score = 0; | |
977 | int node; | |
978 | ||
979 | /* | |
980 | * All nodes are directly connected, and the same distance | |
981 | * from each other. No need for fancy placement algorithms. | |
982 | */ | |
983 | if (sched_numa_topology_type == NUMA_DIRECT) | |
984 | return 0; | |
985 | ||
986 | /* | |
987 | * This code is called for each node, introducing N^2 complexity, | |
988 | * which should be ok given the number of nodes rarely exceeds 8. | |
989 | */ | |
990 | for_each_online_node(node) { | |
991 | unsigned long faults; | |
992 | int dist = node_distance(nid, node); | |
993 | ||
994 | /* | |
995 | * The furthest away nodes in the system are not interesting | |
996 | * for placement; nid was already counted. | |
997 | */ | |
998 | if (dist == sched_max_numa_distance || node == nid) | |
999 | continue; | |
1000 | ||
1001 | /* | |
1002 | * On systems with a backplane NUMA topology, compare groups | |
1003 | * of nodes, and move tasks towards the group with the most | |
1004 | * memory accesses. When comparing two nodes at distance | |
1005 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1006 | * of each group. Skip other nodes. | |
1007 | */ | |
1008 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
1009 | dist > maxdist) | |
1010 | continue; | |
1011 | ||
1012 | /* Add up the faults from nearby nodes. */ | |
1013 | if (task) | |
1014 | faults = task_faults(p, node); | |
1015 | else | |
1016 | faults = group_faults(p, node); | |
1017 | ||
1018 | /* | |
1019 | * On systems with a glueless mesh NUMA topology, there are | |
1020 | * no fixed "groups of nodes". Instead, nodes that are not | |
1021 | * directly connected bounce traffic through intermediate | |
1022 | * nodes; a numa_group can occupy any set of nodes. | |
1023 | * The further away a node is, the less the faults count. | |
1024 | * This seems to result in good task placement. | |
1025 | */ | |
1026 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1027 | faults *= (sched_max_numa_distance - dist); | |
1028 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1029 | } | |
1030 | ||
1031 | score += faults; | |
1032 | } | |
1033 | ||
1034 | return score; | |
1035 | } | |
1036 | ||
83e1d2cd MG |
1037 | /* |
1038 | * These return the fraction of accesses done by a particular task, or | |
1039 | * task group, on a particular numa node. The group weight is given a | |
1040 | * larger multiplier, in order to group tasks together that are almost | |
1041 | * evenly spread out between numa nodes. | |
1042 | */ | |
7bd95320 RR |
1043 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1044 | int dist) | |
83e1d2cd | 1045 | { |
7bd95320 | 1046 | unsigned long faults, total_faults; |
83e1d2cd | 1047 | |
44dba3d5 | 1048 | if (!p->numa_faults) |
83e1d2cd MG |
1049 | return 0; |
1050 | ||
1051 | total_faults = p->total_numa_faults; | |
1052 | ||
1053 | if (!total_faults) | |
1054 | return 0; | |
1055 | ||
7bd95320 | 1056 | faults = task_faults(p, nid); |
6c6b1193 RR |
1057 | faults += score_nearby_nodes(p, nid, dist, true); |
1058 | ||
7bd95320 | 1059 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1060 | } |
1061 | ||
7bd95320 RR |
1062 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1063 | int dist) | |
83e1d2cd | 1064 | { |
7bd95320 RR |
1065 | unsigned long faults, total_faults; |
1066 | ||
1067 | if (!p->numa_group) | |
1068 | return 0; | |
1069 | ||
1070 | total_faults = p->numa_group->total_faults; | |
1071 | ||
1072 | if (!total_faults) | |
83e1d2cd MG |
1073 | return 0; |
1074 | ||
7bd95320 | 1075 | faults = group_faults(p, nid); |
6c6b1193 RR |
1076 | faults += score_nearby_nodes(p, nid, dist, false); |
1077 | ||
7bd95320 | 1078 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1079 | } |
1080 | ||
10f39042 RR |
1081 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1082 | int src_nid, int dst_cpu) | |
1083 | { | |
1084 | struct numa_group *ng = p->numa_group; | |
1085 | int dst_nid = cpu_to_node(dst_cpu); | |
1086 | int last_cpupid, this_cpupid; | |
1087 | ||
1088 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
1089 | ||
1090 | /* | |
1091 | * Multi-stage node selection is used in conjunction with a periodic | |
1092 | * migration fault to build a temporal task<->page relation. By using | |
1093 | * a two-stage filter we remove short/unlikely relations. | |
1094 | * | |
1095 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1096 | * a task's usage of a particular page (n_p) per total usage of this | |
1097 | * page (n_t) (in a given time-span) to a probability. | |
1098 | * | |
1099 | * Our periodic faults will sample this probability and getting the | |
1100 | * same result twice in a row, given these samples are fully | |
1101 | * independent, is then given by P(n)^2, provided our sample period | |
1102 | * is sufficiently short compared to the usage pattern. | |
1103 | * | |
1104 | * This quadric squishes small probabilities, making it less likely we | |
1105 | * act on an unlikely task<->page relation. | |
1106 | */ | |
1107 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
1108 | if (!cpupid_pid_unset(last_cpupid) && | |
1109 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1110 | return false; | |
1111 | ||
1112 | /* Always allow migrate on private faults */ | |
1113 | if (cpupid_match_pid(p, last_cpupid)) | |
1114 | return true; | |
1115 | ||
1116 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1117 | if (!ng) | |
1118 | return true; | |
1119 | ||
1120 | /* | |
1121 | * Do not migrate if the destination is not a node that | |
1122 | * is actively used by this numa group. | |
1123 | */ | |
1124 | if (!node_isset(dst_nid, ng->active_nodes)) | |
1125 | return false; | |
1126 | ||
1127 | /* | |
1128 | * Source is a node that is not actively used by this | |
1129 | * numa group, while the destination is. Migrate. | |
1130 | */ | |
1131 | if (!node_isset(src_nid, ng->active_nodes)) | |
1132 | return true; | |
1133 | ||
1134 | /* | |
1135 | * Both source and destination are nodes in active | |
1136 | * use by this numa group. Maximize memory bandwidth | |
1137 | * by migrating from more heavily used groups, to less | |
1138 | * heavily used ones, spreading the load around. | |
1139 | * Use a 1/4 hysteresis to avoid spurious page movement. | |
1140 | */ | |
1141 | return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); | |
1142 | } | |
1143 | ||
e6628d5b | 1144 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
1145 | static unsigned long source_load(int cpu, int type); |
1146 | static unsigned long target_load(int cpu, int type); | |
ced549fa | 1147 | static unsigned long capacity_of(int cpu); |
58d081b5 MG |
1148 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); |
1149 | ||
fb13c7ee | 1150 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1151 | struct numa_stats { |
fb13c7ee | 1152 | unsigned long nr_running; |
58d081b5 | 1153 | unsigned long load; |
fb13c7ee MG |
1154 | |
1155 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1156 | unsigned long compute_capacity; |
fb13c7ee MG |
1157 | |
1158 | /* Approximate capacity in terms of runnable tasks on a node */ | |
5ef20ca1 | 1159 | unsigned long task_capacity; |
1b6a7495 | 1160 | int has_free_capacity; |
58d081b5 | 1161 | }; |
e6628d5b | 1162 | |
fb13c7ee MG |
1163 | /* |
1164 | * XXX borrowed from update_sg_lb_stats | |
1165 | */ | |
1166 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1167 | { | |
83d7f242 RR |
1168 | int smt, cpu, cpus = 0; |
1169 | unsigned long capacity; | |
fb13c7ee MG |
1170 | |
1171 | memset(ns, 0, sizeof(*ns)); | |
1172 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1173 | struct rq *rq = cpu_rq(cpu); | |
1174 | ||
1175 | ns->nr_running += rq->nr_running; | |
1176 | ns->load += weighted_cpuload(cpu); | |
ced549fa | 1177 | ns->compute_capacity += capacity_of(cpu); |
5eca82a9 PZ |
1178 | |
1179 | cpus++; | |
fb13c7ee MG |
1180 | } |
1181 | ||
5eca82a9 PZ |
1182 | /* |
1183 | * If we raced with hotplug and there are no CPUs left in our mask | |
1184 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1185 | * not find this node attractive. | |
1186 | * | |
1b6a7495 NP |
1187 | * We'll either bail at !has_free_capacity, or we'll detect a huge |
1188 | * imbalance and bail there. | |
5eca82a9 PZ |
1189 | */ |
1190 | if (!cpus) | |
1191 | return; | |
1192 | ||
83d7f242 RR |
1193 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ |
1194 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); | |
1195 | capacity = cpus / smt; /* cores */ | |
1196 | ||
1197 | ns->task_capacity = min_t(unsigned, capacity, | |
1198 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); | |
1b6a7495 | 1199 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); |
fb13c7ee MG |
1200 | } |
1201 | ||
58d081b5 MG |
1202 | struct task_numa_env { |
1203 | struct task_struct *p; | |
e6628d5b | 1204 | |
58d081b5 MG |
1205 | int src_cpu, src_nid; |
1206 | int dst_cpu, dst_nid; | |
e6628d5b | 1207 | |
58d081b5 | 1208 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1209 | |
40ea2b42 | 1210 | int imbalance_pct; |
7bd95320 | 1211 | int dist; |
fb13c7ee MG |
1212 | |
1213 | struct task_struct *best_task; | |
1214 | long best_imp; | |
58d081b5 MG |
1215 | int best_cpu; |
1216 | }; | |
1217 | ||
fb13c7ee MG |
1218 | static void task_numa_assign(struct task_numa_env *env, |
1219 | struct task_struct *p, long imp) | |
1220 | { | |
1221 | if (env->best_task) | |
1222 | put_task_struct(env->best_task); | |
fb13c7ee MG |
1223 | |
1224 | env->best_task = p; | |
1225 | env->best_imp = imp; | |
1226 | env->best_cpu = env->dst_cpu; | |
1227 | } | |
1228 | ||
28a21745 | 1229 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1230 | struct task_numa_env *env) |
1231 | { | |
e4991b24 RR |
1232 | long imb, old_imb; |
1233 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1234 | long src_capacity, dst_capacity; |
1235 | ||
1236 | /* | |
1237 | * The load is corrected for the CPU capacity available on each node. | |
1238 | * | |
1239 | * src_load dst_load | |
1240 | * ------------ vs --------- | |
1241 | * src_capacity dst_capacity | |
1242 | */ | |
1243 | src_capacity = env->src_stats.compute_capacity; | |
1244 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 RR |
1245 | |
1246 | /* We care about the slope of the imbalance, not the direction. */ | |
e4991b24 RR |
1247 | if (dst_load < src_load) |
1248 | swap(dst_load, src_load); | |
e63da036 RR |
1249 | |
1250 | /* Is the difference below the threshold? */ | |
e4991b24 RR |
1251 | imb = dst_load * src_capacity * 100 - |
1252 | src_load * dst_capacity * env->imbalance_pct; | |
e63da036 RR |
1253 | if (imb <= 0) |
1254 | return false; | |
1255 | ||
1256 | /* | |
1257 | * The imbalance is above the allowed threshold. | |
e4991b24 | 1258 | * Compare it with the old imbalance. |
e63da036 | 1259 | */ |
28a21745 | 1260 | orig_src_load = env->src_stats.load; |
e4991b24 | 1261 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1262 | |
e4991b24 RR |
1263 | if (orig_dst_load < orig_src_load) |
1264 | swap(orig_dst_load, orig_src_load); | |
e63da036 | 1265 | |
e4991b24 RR |
1266 | old_imb = orig_dst_load * src_capacity * 100 - |
1267 | orig_src_load * dst_capacity * env->imbalance_pct; | |
1268 | ||
1269 | /* Would this change make things worse? */ | |
1270 | return (imb > old_imb); | |
e63da036 RR |
1271 | } |
1272 | ||
fb13c7ee MG |
1273 | /* |
1274 | * This checks if the overall compute and NUMA accesses of the system would | |
1275 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1276 | * into account that it might be best if task running on the dst_cpu should | |
1277 | * be exchanged with the source task | |
1278 | */ | |
887c290e RR |
1279 | static void task_numa_compare(struct task_numa_env *env, |
1280 | long taskimp, long groupimp) | |
fb13c7ee MG |
1281 | { |
1282 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1283 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1284 | struct task_struct *cur; | |
28a21745 | 1285 | long src_load, dst_load; |
fb13c7ee | 1286 | long load; |
1c5d3eb3 | 1287 | long imp = env->p->numa_group ? groupimp : taskimp; |
0132c3e1 | 1288 | long moveimp = imp; |
7bd95320 | 1289 | int dist = env->dist; |
1dff76b9 | 1290 | bool assigned = false; |
fb13c7ee MG |
1291 | |
1292 | rcu_read_lock(); | |
1effd9f1 KT |
1293 | |
1294 | raw_spin_lock_irq(&dst_rq->lock); | |
1295 | cur = dst_rq->curr; | |
1296 | /* | |
1dff76b9 | 1297 | * No need to move the exiting task or idle task. |
1effd9f1 KT |
1298 | */ |
1299 | if ((cur->flags & PF_EXITING) || is_idle_task(cur)) | |
fb13c7ee | 1300 | cur = NULL; |
1dff76b9 GG |
1301 | else { |
1302 | /* | |
1303 | * The task_struct must be protected here to protect the | |
1304 | * p->numa_faults access in the task_weight since the | |
1305 | * numa_faults could already be freed in the following path: | |
1306 | * finish_task_switch() | |
1307 | * --> put_task_struct() | |
1308 | * --> __put_task_struct() | |
1309 | * --> task_numa_free() | |
1310 | */ | |
1311 | get_task_struct(cur); | |
1312 | } | |
1313 | ||
1effd9f1 | 1314 | raw_spin_unlock_irq(&dst_rq->lock); |
fb13c7ee | 1315 | |
7af68335 PZ |
1316 | /* |
1317 | * Because we have preemption enabled we can get migrated around and | |
1318 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1319 | */ | |
1320 | if (cur == env->p) | |
1321 | goto unlock; | |
1322 | ||
fb13c7ee MG |
1323 | /* |
1324 | * "imp" is the fault differential for the source task between the | |
1325 | * source and destination node. Calculate the total differential for | |
1326 | * the source task and potential destination task. The more negative | |
1327 | * the value is, the more rmeote accesses that would be expected to | |
1328 | * be incurred if the tasks were swapped. | |
1329 | */ | |
1330 | if (cur) { | |
1331 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1332 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1333 | goto unlock; | |
1334 | ||
887c290e RR |
1335 | /* |
1336 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1337 | * in any group then look only at task weights. |
887c290e | 1338 | */ |
ca28aa53 | 1339 | if (cur->numa_group == env->p->numa_group) { |
7bd95320 RR |
1340 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1341 | task_weight(cur, env->dst_nid, dist); | |
ca28aa53 RR |
1342 | /* |
1343 | * Add some hysteresis to prevent swapping the | |
1344 | * tasks within a group over tiny differences. | |
1345 | */ | |
1346 | if (cur->numa_group) | |
1347 | imp -= imp/16; | |
887c290e | 1348 | } else { |
ca28aa53 RR |
1349 | /* |
1350 | * Compare the group weights. If a task is all by | |
1351 | * itself (not part of a group), use the task weight | |
1352 | * instead. | |
1353 | */ | |
ca28aa53 | 1354 | if (cur->numa_group) |
7bd95320 RR |
1355 | imp += group_weight(cur, env->src_nid, dist) - |
1356 | group_weight(cur, env->dst_nid, dist); | |
ca28aa53 | 1357 | else |
7bd95320 RR |
1358 | imp += task_weight(cur, env->src_nid, dist) - |
1359 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1360 | } |
fb13c7ee MG |
1361 | } |
1362 | ||
0132c3e1 | 1363 | if (imp <= env->best_imp && moveimp <= env->best_imp) |
fb13c7ee MG |
1364 | goto unlock; |
1365 | ||
1366 | if (!cur) { | |
1367 | /* Is there capacity at our destination? */ | |
b932c03c | 1368 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && |
1b6a7495 | 1369 | !env->dst_stats.has_free_capacity) |
fb13c7ee MG |
1370 | goto unlock; |
1371 | ||
1372 | goto balance; | |
1373 | } | |
1374 | ||
1375 | /* Balance doesn't matter much if we're running a task per cpu */ | |
0132c3e1 RR |
1376 | if (imp > env->best_imp && src_rq->nr_running == 1 && |
1377 | dst_rq->nr_running == 1) | |
fb13c7ee MG |
1378 | goto assign; |
1379 | ||
1380 | /* | |
1381 | * In the overloaded case, try and keep the load balanced. | |
1382 | */ | |
1383 | balance: | |
e720fff6 PZ |
1384 | load = task_h_load(env->p); |
1385 | dst_load = env->dst_stats.load + load; | |
1386 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1387 | |
0132c3e1 RR |
1388 | if (moveimp > imp && moveimp > env->best_imp) { |
1389 | /* | |
1390 | * If the improvement from just moving env->p direction is | |
1391 | * better than swapping tasks around, check if a move is | |
1392 | * possible. Store a slightly smaller score than moveimp, | |
1393 | * so an actually idle CPU will win. | |
1394 | */ | |
1395 | if (!load_too_imbalanced(src_load, dst_load, env)) { | |
1396 | imp = moveimp - 1; | |
1dff76b9 | 1397 | put_task_struct(cur); |
0132c3e1 RR |
1398 | cur = NULL; |
1399 | goto assign; | |
1400 | } | |
1401 | } | |
1402 | ||
1403 | if (imp <= env->best_imp) | |
1404 | goto unlock; | |
1405 | ||
fb13c7ee | 1406 | if (cur) { |
e720fff6 PZ |
1407 | load = task_h_load(cur); |
1408 | dst_load -= load; | |
1409 | src_load += load; | |
fb13c7ee MG |
1410 | } |
1411 | ||
28a21745 | 1412 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1413 | goto unlock; |
1414 | ||
ba7e5a27 RR |
1415 | /* |
1416 | * One idle CPU per node is evaluated for a task numa move. | |
1417 | * Call select_idle_sibling to maybe find a better one. | |
1418 | */ | |
1419 | if (!cur) | |
1420 | env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); | |
1421 | ||
fb13c7ee | 1422 | assign: |
1dff76b9 | 1423 | assigned = true; |
fb13c7ee MG |
1424 | task_numa_assign(env, cur, imp); |
1425 | unlock: | |
1426 | rcu_read_unlock(); | |
1dff76b9 GG |
1427 | /* |
1428 | * The dst_rq->curr isn't assigned. The protection for task_struct is | |
1429 | * finished. | |
1430 | */ | |
1431 | if (cur && !assigned) | |
1432 | put_task_struct(cur); | |
fb13c7ee MG |
1433 | } |
1434 | ||
887c290e RR |
1435 | static void task_numa_find_cpu(struct task_numa_env *env, |
1436 | long taskimp, long groupimp) | |
2c8a50aa MG |
1437 | { |
1438 | int cpu; | |
1439 | ||
1440 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1441 | /* Skip this CPU if the source task cannot migrate */ | |
1442 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1443 | continue; | |
1444 | ||
1445 | env->dst_cpu = cpu; | |
887c290e | 1446 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1447 | } |
1448 | } | |
1449 | ||
6f9aad0b RR |
1450 | /* Only move tasks to a NUMA node less busy than the current node. */ |
1451 | static bool numa_has_capacity(struct task_numa_env *env) | |
1452 | { | |
1453 | struct numa_stats *src = &env->src_stats; | |
1454 | struct numa_stats *dst = &env->dst_stats; | |
1455 | ||
1456 | if (src->has_free_capacity && !dst->has_free_capacity) | |
1457 | return false; | |
1458 | ||
1459 | /* | |
1460 | * Only consider a task move if the source has a higher load | |
1461 | * than the destination, corrected for CPU capacity on each node. | |
1462 | * | |
1463 | * src->load dst->load | |
1464 | * --------------------- vs --------------------- | |
1465 | * src->compute_capacity dst->compute_capacity | |
1466 | */ | |
44dcb04f SD |
1467 | if (src->load * dst->compute_capacity * env->imbalance_pct > |
1468 | ||
1469 | dst->load * src->compute_capacity * 100) | |
6f9aad0b RR |
1470 | return true; |
1471 | ||
1472 | return false; | |
1473 | } | |
1474 | ||
58d081b5 MG |
1475 | static int task_numa_migrate(struct task_struct *p) |
1476 | { | |
58d081b5 MG |
1477 | struct task_numa_env env = { |
1478 | .p = p, | |
fb13c7ee | 1479 | |
58d081b5 | 1480 | .src_cpu = task_cpu(p), |
b32e86b4 | 1481 | .src_nid = task_node(p), |
fb13c7ee MG |
1482 | |
1483 | .imbalance_pct = 112, | |
1484 | ||
1485 | .best_task = NULL, | |
1486 | .best_imp = 0, | |
1487 | .best_cpu = -1 | |
58d081b5 MG |
1488 | }; |
1489 | struct sched_domain *sd; | |
887c290e | 1490 | unsigned long taskweight, groupweight; |
7bd95320 | 1491 | int nid, ret, dist; |
887c290e | 1492 | long taskimp, groupimp; |
e6628d5b | 1493 | |
58d081b5 | 1494 | /* |
fb13c7ee MG |
1495 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1496 | * imbalance and would be the first to start moving tasks about. | |
1497 | * | |
1498 | * And we want to avoid any moving of tasks about, as that would create | |
1499 | * random movement of tasks -- counter the numa conditions we're trying | |
1500 | * to satisfy here. | |
58d081b5 MG |
1501 | */ |
1502 | rcu_read_lock(); | |
fb13c7ee | 1503 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1504 | if (sd) |
1505 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1506 | rcu_read_unlock(); |
1507 | ||
46a73e8a RR |
1508 | /* |
1509 | * Cpusets can break the scheduler domain tree into smaller | |
1510 | * balance domains, some of which do not cross NUMA boundaries. | |
1511 | * Tasks that are "trapped" in such domains cannot be migrated | |
1512 | * elsewhere, so there is no point in (re)trying. | |
1513 | */ | |
1514 | if (unlikely(!sd)) { | |
de1b301a | 1515 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1516 | return -EINVAL; |
1517 | } | |
1518 | ||
2c8a50aa | 1519 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1520 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1521 | taskweight = task_weight(p, env.src_nid, dist); | |
1522 | groupweight = group_weight(p, env.src_nid, dist); | |
1523 | update_numa_stats(&env.src_stats, env.src_nid); | |
1524 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1525 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1526 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1527 | |
a43455a1 | 1528 | /* Try to find a spot on the preferred nid. */ |
6f9aad0b RR |
1529 | if (numa_has_capacity(&env)) |
1530 | task_numa_find_cpu(&env, taskimp, groupimp); | |
e1dda8a7 | 1531 | |
9de05d48 RR |
1532 | /* |
1533 | * Look at other nodes in these cases: | |
1534 | * - there is no space available on the preferred_nid | |
1535 | * - the task is part of a numa_group that is interleaved across | |
1536 | * multiple NUMA nodes; in order to better consolidate the group, | |
1537 | * we need to check other locations. | |
1538 | */ | |
1539 | if (env.best_cpu == -1 || (p->numa_group && | |
1540 | nodes_weight(p->numa_group->active_nodes) > 1)) { | |
2c8a50aa MG |
1541 | for_each_online_node(nid) { |
1542 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1543 | continue; | |
58d081b5 | 1544 | |
7bd95320 | 1545 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1546 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1547 | dist != env.dist) { | |
1548 | taskweight = task_weight(p, env.src_nid, dist); | |
1549 | groupweight = group_weight(p, env.src_nid, dist); | |
1550 | } | |
7bd95320 | 1551 | |
83e1d2cd | 1552 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1553 | taskimp = task_weight(p, nid, dist) - taskweight; |
1554 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1555 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1556 | continue; |
1557 | ||
7bd95320 | 1558 | env.dist = dist; |
2c8a50aa MG |
1559 | env.dst_nid = nid; |
1560 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
6f9aad0b RR |
1561 | if (numa_has_capacity(&env)) |
1562 | task_numa_find_cpu(&env, taskimp, groupimp); | |
58d081b5 MG |
1563 | } |
1564 | } | |
1565 | ||
68d1b02a RR |
1566 | /* |
1567 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1568 | * and is migrating into one of the workload's active nodes, remember | |
1569 | * this node as the task's preferred numa node, so the workload can | |
1570 | * settle down. | |
1571 | * A task that migrated to a second choice node will be better off | |
1572 | * trying for a better one later. Do not set the preferred node here. | |
1573 | */ | |
db015dae RR |
1574 | if (p->numa_group) { |
1575 | if (env.best_cpu == -1) | |
1576 | nid = env.src_nid; | |
1577 | else | |
1578 | nid = env.dst_nid; | |
1579 | ||
1580 | if (node_isset(nid, p->numa_group->active_nodes)) | |
1581 | sched_setnuma(p, env.dst_nid); | |
1582 | } | |
1583 | ||
1584 | /* No better CPU than the current one was found. */ | |
1585 | if (env.best_cpu == -1) | |
1586 | return -EAGAIN; | |
0ec8aa00 | 1587 | |
04bb2f94 RR |
1588 | /* |
1589 | * Reset the scan period if the task is being rescheduled on an | |
1590 | * alternative node to recheck if the tasks is now properly placed. | |
1591 | */ | |
1592 | p->numa_scan_period = task_scan_min(p); | |
1593 | ||
fb13c7ee | 1594 | if (env.best_task == NULL) { |
286549dc MG |
1595 | ret = migrate_task_to(p, env.best_cpu); |
1596 | if (ret != 0) | |
1597 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1598 | return ret; |
1599 | } | |
1600 | ||
1601 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1602 | if (ret != 0) |
1603 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1604 | put_task_struct(env.best_task); |
1605 | return ret; | |
e6628d5b MG |
1606 | } |
1607 | ||
6b9a7460 MG |
1608 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1609 | static void numa_migrate_preferred(struct task_struct *p) | |
1610 | { | |
5085e2a3 RR |
1611 | unsigned long interval = HZ; |
1612 | ||
2739d3ee | 1613 | /* This task has no NUMA fault statistics yet */ |
44dba3d5 | 1614 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
6b9a7460 MG |
1615 | return; |
1616 | ||
2739d3ee | 1617 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 RR |
1618 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1619 | p->numa_migrate_retry = jiffies + interval; | |
2739d3ee RR |
1620 | |
1621 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1622 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1623 | return; |
1624 | ||
1625 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1626 | task_numa_migrate(p); |
6b9a7460 MG |
1627 | } |
1628 | ||
20e07dea RR |
1629 | /* |
1630 | * Find the nodes on which the workload is actively running. We do this by | |
1631 | * tracking the nodes from which NUMA hinting faults are triggered. This can | |
1632 | * be different from the set of nodes where the workload's memory is currently | |
1633 | * located. | |
1634 | * | |
1635 | * The bitmask is used to make smarter decisions on when to do NUMA page | |
1636 | * migrations, To prevent flip-flopping, and excessive page migrations, nodes | |
1637 | * are added when they cause over 6/16 of the maximum number of faults, but | |
1638 | * only removed when they drop below 3/16. | |
1639 | */ | |
1640 | static void update_numa_active_node_mask(struct numa_group *numa_group) | |
1641 | { | |
1642 | unsigned long faults, max_faults = 0; | |
1643 | int nid; | |
1644 | ||
1645 | for_each_online_node(nid) { | |
1646 | faults = group_faults_cpu(numa_group, nid); | |
1647 | if (faults > max_faults) | |
1648 | max_faults = faults; | |
1649 | } | |
1650 | ||
1651 | for_each_online_node(nid) { | |
1652 | faults = group_faults_cpu(numa_group, nid); | |
1653 | if (!node_isset(nid, numa_group->active_nodes)) { | |
1654 | if (faults > max_faults * 6 / 16) | |
1655 | node_set(nid, numa_group->active_nodes); | |
1656 | } else if (faults < max_faults * 3 / 16) | |
1657 | node_clear(nid, numa_group->active_nodes); | |
1658 | } | |
1659 | } | |
1660 | ||
04bb2f94 RR |
1661 | /* |
1662 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1663 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1664 | * period will be for the next scan window. If local/(local+remote) ratio is |
1665 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1666 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1667 | */ |
1668 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1669 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1670 | |
1671 | /* | |
1672 | * Increase the scan period (slow down scanning) if the majority of | |
1673 | * our memory is already on our local node, or if the majority of | |
1674 | * the page accesses are shared with other processes. | |
1675 | * Otherwise, decrease the scan period. | |
1676 | */ | |
1677 | static void update_task_scan_period(struct task_struct *p, | |
1678 | unsigned long shared, unsigned long private) | |
1679 | { | |
1680 | unsigned int period_slot; | |
1681 | int ratio; | |
1682 | int diff; | |
1683 | ||
1684 | unsigned long remote = p->numa_faults_locality[0]; | |
1685 | unsigned long local = p->numa_faults_locality[1]; | |
1686 | ||
1687 | /* | |
1688 | * If there were no record hinting faults then either the task is | |
1689 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
1690 | * to automatic numa balancing. Related to that, if there were failed |
1691 | * migration then it implies we are migrating too quickly or the local | |
1692 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 1693 | */ |
074c2381 | 1694 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
1695 | p->numa_scan_period = min(p->numa_scan_period_max, |
1696 | p->numa_scan_period << 1); | |
1697 | ||
1698 | p->mm->numa_next_scan = jiffies + | |
1699 | msecs_to_jiffies(p->numa_scan_period); | |
1700 | ||
1701 | return; | |
1702 | } | |
1703 | ||
1704 | /* | |
1705 | * Prepare to scale scan period relative to the current period. | |
1706 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1707 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1708 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1709 | */ | |
1710 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1711 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1712 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1713 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1714 | if (!slot) | |
1715 | slot = 1; | |
1716 | diff = slot * period_slot; | |
1717 | } else { | |
1718 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1719 | ||
1720 | /* | |
1721 | * Scale scan rate increases based on sharing. There is an | |
1722 | * inverse relationship between the degree of sharing and | |
1723 | * the adjustment made to the scanning period. Broadly | |
1724 | * speaking the intent is that there is little point | |
1725 | * scanning faster if shared accesses dominate as it may | |
1726 | * simply bounce migrations uselessly | |
1727 | */ | |
2847c90e | 1728 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); |
04bb2f94 RR |
1729 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; |
1730 | } | |
1731 | ||
1732 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1733 | task_scan_min(p), task_scan_max(p)); | |
1734 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1735 | } | |
1736 | ||
7e2703e6 RR |
1737 | /* |
1738 | * Get the fraction of time the task has been running since the last | |
1739 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1740 | * decays those on a 32ms period, which is orders of magnitude off | |
1741 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1742 | * stats only if the task is so new there are no NUMA statistics yet. | |
1743 | */ | |
1744 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1745 | { | |
1746 | u64 runtime, delta, now; | |
1747 | /* Use the start of this time slice to avoid calculations. */ | |
1748 | now = p->se.exec_start; | |
1749 | runtime = p->se.sum_exec_runtime; | |
1750 | ||
1751 | if (p->last_task_numa_placement) { | |
1752 | delta = runtime - p->last_sum_exec_runtime; | |
1753 | *period = now - p->last_task_numa_placement; | |
1754 | } else { | |
9d89c257 YD |
1755 | delta = p->se.avg.load_sum / p->se.load.weight; |
1756 | *period = LOAD_AVG_MAX; | |
7e2703e6 RR |
1757 | } |
1758 | ||
1759 | p->last_sum_exec_runtime = runtime; | |
1760 | p->last_task_numa_placement = now; | |
1761 | ||
1762 | return delta; | |
1763 | } | |
1764 | ||
54009416 RR |
1765 | /* |
1766 | * Determine the preferred nid for a task in a numa_group. This needs to | |
1767 | * be done in a way that produces consistent results with group_weight, | |
1768 | * otherwise workloads might not converge. | |
1769 | */ | |
1770 | static int preferred_group_nid(struct task_struct *p, int nid) | |
1771 | { | |
1772 | nodemask_t nodes; | |
1773 | int dist; | |
1774 | ||
1775 | /* Direct connections between all NUMA nodes. */ | |
1776 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1777 | return nid; | |
1778 | ||
1779 | /* | |
1780 | * On a system with glueless mesh NUMA topology, group_weight | |
1781 | * scores nodes according to the number of NUMA hinting faults on | |
1782 | * both the node itself, and on nearby nodes. | |
1783 | */ | |
1784 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1785 | unsigned long score, max_score = 0; | |
1786 | int node, max_node = nid; | |
1787 | ||
1788 | dist = sched_max_numa_distance; | |
1789 | ||
1790 | for_each_online_node(node) { | |
1791 | score = group_weight(p, node, dist); | |
1792 | if (score > max_score) { | |
1793 | max_score = score; | |
1794 | max_node = node; | |
1795 | } | |
1796 | } | |
1797 | return max_node; | |
1798 | } | |
1799 | ||
1800 | /* | |
1801 | * Finding the preferred nid in a system with NUMA backplane | |
1802 | * interconnect topology is more involved. The goal is to locate | |
1803 | * tasks from numa_groups near each other in the system, and | |
1804 | * untangle workloads from different sides of the system. This requires | |
1805 | * searching down the hierarchy of node groups, recursively searching | |
1806 | * inside the highest scoring group of nodes. The nodemask tricks | |
1807 | * keep the complexity of the search down. | |
1808 | */ | |
1809 | nodes = node_online_map; | |
1810 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
1811 | unsigned long max_faults = 0; | |
81907478 | 1812 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
1813 | int a, b; |
1814 | ||
1815 | /* Are there nodes at this distance from each other? */ | |
1816 | if (!find_numa_distance(dist)) | |
1817 | continue; | |
1818 | ||
1819 | for_each_node_mask(a, nodes) { | |
1820 | unsigned long faults = 0; | |
1821 | nodemask_t this_group; | |
1822 | nodes_clear(this_group); | |
1823 | ||
1824 | /* Sum group's NUMA faults; includes a==b case. */ | |
1825 | for_each_node_mask(b, nodes) { | |
1826 | if (node_distance(a, b) < dist) { | |
1827 | faults += group_faults(p, b); | |
1828 | node_set(b, this_group); | |
1829 | node_clear(b, nodes); | |
1830 | } | |
1831 | } | |
1832 | ||
1833 | /* Remember the top group. */ | |
1834 | if (faults > max_faults) { | |
1835 | max_faults = faults; | |
1836 | max_group = this_group; | |
1837 | /* | |
1838 | * subtle: at the smallest distance there is | |
1839 | * just one node left in each "group", the | |
1840 | * winner is the preferred nid. | |
1841 | */ | |
1842 | nid = a; | |
1843 | } | |
1844 | } | |
1845 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
1846 | if (!max_faults) |
1847 | break; | |
54009416 RR |
1848 | nodes = max_group; |
1849 | } | |
1850 | return nid; | |
1851 | } | |
1852 | ||
cbee9f88 PZ |
1853 | static void task_numa_placement(struct task_struct *p) |
1854 | { | |
83e1d2cd MG |
1855 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1856 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 1857 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
1858 | unsigned long total_faults; |
1859 | u64 runtime, period; | |
7dbd13ed | 1860 | spinlock_t *group_lock = NULL; |
cbee9f88 | 1861 | |
7e5a2c17 JL |
1862 | /* |
1863 | * The p->mm->numa_scan_seq field gets updated without | |
1864 | * exclusive access. Use READ_ONCE() here to ensure | |
1865 | * that the field is read in a single access: | |
1866 | */ | |
316c1608 | 1867 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
1868 | if (p->numa_scan_seq == seq) |
1869 | return; | |
1870 | p->numa_scan_seq = seq; | |
598f0ec0 | 1871 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 1872 | |
7e2703e6 RR |
1873 | total_faults = p->numa_faults_locality[0] + |
1874 | p->numa_faults_locality[1]; | |
1875 | runtime = numa_get_avg_runtime(p, &period); | |
1876 | ||
7dbd13ed MG |
1877 | /* If the task is part of a group prevent parallel updates to group stats */ |
1878 | if (p->numa_group) { | |
1879 | group_lock = &p->numa_group->lock; | |
60e69eed | 1880 | spin_lock_irq(group_lock); |
7dbd13ed MG |
1881 | } |
1882 | ||
688b7585 MG |
1883 | /* Find the node with the highest number of faults */ |
1884 | for_each_online_node(nid) { | |
44dba3d5 IM |
1885 | /* Keep track of the offsets in numa_faults array */ |
1886 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 1887 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 1888 | int priv; |
745d6147 | 1889 | |
be1e4e76 | 1890 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 1891 | long diff, f_diff, f_weight; |
8c8a743c | 1892 | |
44dba3d5 IM |
1893 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
1894 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
1895 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
1896 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 1897 | |
ac8e895b | 1898 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
1899 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
1900 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
1901 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 1902 | |
7e2703e6 RR |
1903 | /* |
1904 | * Normalize the faults_from, so all tasks in a group | |
1905 | * count according to CPU use, instead of by the raw | |
1906 | * number of faults. Tasks with little runtime have | |
1907 | * little over-all impact on throughput, and thus their | |
1908 | * faults are less important. | |
1909 | */ | |
1910 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 1911 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 1912 | (total_faults + 1); |
44dba3d5 IM |
1913 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
1914 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 1915 | |
44dba3d5 IM |
1916 | p->numa_faults[mem_idx] += diff; |
1917 | p->numa_faults[cpu_idx] += f_diff; | |
1918 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 1919 | p->total_numa_faults += diff; |
8c8a743c | 1920 | if (p->numa_group) { |
44dba3d5 IM |
1921 | /* |
1922 | * safe because we can only change our own group | |
1923 | * | |
1924 | * mem_idx represents the offset for a given | |
1925 | * nid and priv in a specific region because it | |
1926 | * is at the beginning of the numa_faults array. | |
1927 | */ | |
1928 | p->numa_group->faults[mem_idx] += diff; | |
1929 | p->numa_group->faults_cpu[mem_idx] += f_diff; | |
989348b5 | 1930 | p->numa_group->total_faults += diff; |
44dba3d5 | 1931 | group_faults += p->numa_group->faults[mem_idx]; |
8c8a743c | 1932 | } |
ac8e895b MG |
1933 | } |
1934 | ||
688b7585 MG |
1935 | if (faults > max_faults) { |
1936 | max_faults = faults; | |
1937 | max_nid = nid; | |
1938 | } | |
83e1d2cd MG |
1939 | |
1940 | if (group_faults > max_group_faults) { | |
1941 | max_group_faults = group_faults; | |
1942 | max_group_nid = nid; | |
1943 | } | |
1944 | } | |
1945 | ||
04bb2f94 RR |
1946 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1947 | ||
7dbd13ed | 1948 | if (p->numa_group) { |
20e07dea | 1949 | update_numa_active_node_mask(p->numa_group); |
60e69eed | 1950 | spin_unlock_irq(group_lock); |
54009416 | 1951 | max_nid = preferred_group_nid(p, max_group_nid); |
688b7585 MG |
1952 | } |
1953 | ||
bb97fc31 RR |
1954 | if (max_faults) { |
1955 | /* Set the new preferred node */ | |
1956 | if (max_nid != p->numa_preferred_nid) | |
1957 | sched_setnuma(p, max_nid); | |
1958 | ||
1959 | if (task_node(p) != p->numa_preferred_nid) | |
1960 | numa_migrate_preferred(p); | |
3a7053b3 | 1961 | } |
cbee9f88 PZ |
1962 | } |
1963 | ||
8c8a743c PZ |
1964 | static inline int get_numa_group(struct numa_group *grp) |
1965 | { | |
1966 | return atomic_inc_not_zero(&grp->refcount); | |
1967 | } | |
1968 | ||
1969 | static inline void put_numa_group(struct numa_group *grp) | |
1970 | { | |
1971 | if (atomic_dec_and_test(&grp->refcount)) | |
1972 | kfree_rcu(grp, rcu); | |
1973 | } | |
1974 | ||
3e6a9418 MG |
1975 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1976 | int *priv) | |
8c8a743c PZ |
1977 | { |
1978 | struct numa_group *grp, *my_grp; | |
1979 | struct task_struct *tsk; | |
1980 | bool join = false; | |
1981 | int cpu = cpupid_to_cpu(cpupid); | |
1982 | int i; | |
1983 | ||
1984 | if (unlikely(!p->numa_group)) { | |
1985 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 1986 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
1987 | |
1988 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
1989 | if (!grp) | |
1990 | return; | |
1991 | ||
1992 | atomic_set(&grp->refcount, 1); | |
1993 | spin_lock_init(&grp->lock); | |
e29cf08b | 1994 | grp->gid = p->pid; |
50ec8a40 | 1995 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
1996 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
1997 | nr_node_ids; | |
8c8a743c | 1998 | |
20e07dea RR |
1999 | node_set(task_node(current), grp->active_nodes); |
2000 | ||
be1e4e76 | 2001 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2002 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2003 | |
989348b5 | 2004 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2005 | |
8c8a743c PZ |
2006 | grp->nr_tasks++; |
2007 | rcu_assign_pointer(p->numa_group, grp); | |
2008 | } | |
2009 | ||
2010 | rcu_read_lock(); | |
316c1608 | 2011 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2012 | |
2013 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2014 | goto no_join; |
8c8a743c PZ |
2015 | |
2016 | grp = rcu_dereference(tsk->numa_group); | |
2017 | if (!grp) | |
3354781a | 2018 | goto no_join; |
8c8a743c PZ |
2019 | |
2020 | my_grp = p->numa_group; | |
2021 | if (grp == my_grp) | |
3354781a | 2022 | goto no_join; |
8c8a743c PZ |
2023 | |
2024 | /* | |
2025 | * Only join the other group if its bigger; if we're the bigger group, | |
2026 | * the other task will join us. | |
2027 | */ | |
2028 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2029 | goto no_join; |
8c8a743c PZ |
2030 | |
2031 | /* | |
2032 | * Tie-break on the grp address. | |
2033 | */ | |
2034 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2035 | goto no_join; |
8c8a743c | 2036 | |
dabe1d99 RR |
2037 | /* Always join threads in the same process. */ |
2038 | if (tsk->mm == current->mm) | |
2039 | join = true; | |
2040 | ||
2041 | /* Simple filter to avoid false positives due to PID collisions */ | |
2042 | if (flags & TNF_SHARED) | |
2043 | join = true; | |
8c8a743c | 2044 | |
3e6a9418 MG |
2045 | /* Update priv based on whether false sharing was detected */ |
2046 | *priv = !join; | |
2047 | ||
dabe1d99 | 2048 | if (join && !get_numa_group(grp)) |
3354781a | 2049 | goto no_join; |
8c8a743c | 2050 | |
8c8a743c PZ |
2051 | rcu_read_unlock(); |
2052 | ||
2053 | if (!join) | |
2054 | return; | |
2055 | ||
60e69eed MG |
2056 | BUG_ON(irqs_disabled()); |
2057 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2058 | |
be1e4e76 | 2059 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2060 | my_grp->faults[i] -= p->numa_faults[i]; |
2061 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2062 | } |
989348b5 MG |
2063 | my_grp->total_faults -= p->total_numa_faults; |
2064 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2065 | |
8c8a743c PZ |
2066 | my_grp->nr_tasks--; |
2067 | grp->nr_tasks++; | |
2068 | ||
2069 | spin_unlock(&my_grp->lock); | |
60e69eed | 2070 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2071 | |
2072 | rcu_assign_pointer(p->numa_group, grp); | |
2073 | ||
2074 | put_numa_group(my_grp); | |
3354781a PZ |
2075 | return; |
2076 | ||
2077 | no_join: | |
2078 | rcu_read_unlock(); | |
2079 | return; | |
8c8a743c PZ |
2080 | } |
2081 | ||
2082 | void task_numa_free(struct task_struct *p) | |
2083 | { | |
2084 | struct numa_group *grp = p->numa_group; | |
44dba3d5 | 2085 | void *numa_faults = p->numa_faults; |
e9dd685c SR |
2086 | unsigned long flags; |
2087 | int i; | |
8c8a743c PZ |
2088 | |
2089 | if (grp) { | |
e9dd685c | 2090 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2091 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2092 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2093 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2094 | |
8c8a743c | 2095 | grp->nr_tasks--; |
e9dd685c | 2096 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2097 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2098 | put_numa_group(grp); |
2099 | } | |
2100 | ||
44dba3d5 | 2101 | p->numa_faults = NULL; |
82727018 | 2102 | kfree(numa_faults); |
8c8a743c PZ |
2103 | } |
2104 | ||
cbee9f88 PZ |
2105 | /* |
2106 | * Got a PROT_NONE fault for a page on @node. | |
2107 | */ | |
58b46da3 | 2108 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2109 | { |
2110 | struct task_struct *p = current; | |
6688cc05 | 2111 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2112 | int cpu_node = task_node(current); |
792568ec | 2113 | int local = !!(flags & TNF_FAULT_LOCAL); |
ac8e895b | 2114 | int priv; |
cbee9f88 | 2115 | |
2a595721 | 2116 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2117 | return; |
2118 | ||
9ff1d9ff MG |
2119 | /* for example, ksmd faulting in a user's mm */ |
2120 | if (!p->mm) | |
2121 | return; | |
2122 | ||
f809ca9a | 2123 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2124 | if (unlikely(!p->numa_faults)) { |
2125 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2126 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2127 | |
44dba3d5 IM |
2128 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2129 | if (!p->numa_faults) | |
f809ca9a | 2130 | return; |
745d6147 | 2131 | |
83e1d2cd | 2132 | p->total_numa_faults = 0; |
04bb2f94 | 2133 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2134 | } |
cbee9f88 | 2135 | |
8c8a743c PZ |
2136 | /* |
2137 | * First accesses are treated as private, otherwise consider accesses | |
2138 | * to be private if the accessing pid has not changed | |
2139 | */ | |
2140 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2141 | priv = 1; | |
2142 | } else { | |
2143 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2144 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2145 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2146 | } |
2147 | ||
792568ec RR |
2148 | /* |
2149 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2150 | * occurs wholly within the set of nodes that the workload is | |
2151 | * actively using should be counted as local. This allows the | |
2152 | * scan rate to slow down when a workload has settled down. | |
2153 | */ | |
2154 | if (!priv && !local && p->numa_group && | |
2155 | node_isset(cpu_node, p->numa_group->active_nodes) && | |
2156 | node_isset(mem_node, p->numa_group->active_nodes)) | |
2157 | local = 1; | |
2158 | ||
cbee9f88 | 2159 | task_numa_placement(p); |
f809ca9a | 2160 | |
2739d3ee RR |
2161 | /* |
2162 | * Retry task to preferred node migration periodically, in case it | |
2163 | * case it previously failed, or the scheduler moved us. | |
2164 | */ | |
2165 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
2166 | numa_migrate_preferred(p); |
2167 | ||
b32e86b4 IM |
2168 | if (migrated) |
2169 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2170 | if (flags & TNF_MIGRATE_FAIL) |
2171 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2172 | |
44dba3d5 IM |
2173 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2174 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2175 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2176 | } |
2177 | ||
6e5fb223 PZ |
2178 | static void reset_ptenuma_scan(struct task_struct *p) |
2179 | { | |
7e5a2c17 JL |
2180 | /* |
2181 | * We only did a read acquisition of the mmap sem, so | |
2182 | * p->mm->numa_scan_seq is written to without exclusive access | |
2183 | * and the update is not guaranteed to be atomic. That's not | |
2184 | * much of an issue though, since this is just used for | |
2185 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2186 | * expensive, to avoid any form of compiler optimizations: | |
2187 | */ | |
316c1608 | 2188 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2189 | p->mm->numa_scan_offset = 0; |
2190 | } | |
2191 | ||
cbee9f88 PZ |
2192 | /* |
2193 | * The expensive part of numa migration is done from task_work context. | |
2194 | * Triggered from task_tick_numa(). | |
2195 | */ | |
2196 | void task_numa_work(struct callback_head *work) | |
2197 | { | |
2198 | unsigned long migrate, next_scan, now = jiffies; | |
2199 | struct task_struct *p = current; | |
2200 | struct mm_struct *mm = p->mm; | |
51170840 | 2201 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2202 | struct vm_area_struct *vma; |
9f40604c | 2203 | unsigned long start, end; |
598f0ec0 | 2204 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2205 | long pages, virtpages; |
cbee9f88 PZ |
2206 | |
2207 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
2208 | ||
2209 | work->next = work; /* protect against double add */ | |
2210 | /* | |
2211 | * Who cares about NUMA placement when they're dying. | |
2212 | * | |
2213 | * NOTE: make sure not to dereference p->mm before this check, | |
2214 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2215 | * without p->mm even though we still had it when we enqueued this | |
2216 | * work. | |
2217 | */ | |
2218 | if (p->flags & PF_EXITING) | |
2219 | return; | |
2220 | ||
930aa174 | 2221 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2222 | mm->numa_next_scan = now + |
2223 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2224 | } |
2225 | ||
cbee9f88 PZ |
2226 | /* |
2227 | * Enforce maximal scan/migration frequency.. | |
2228 | */ | |
2229 | migrate = mm->numa_next_scan; | |
2230 | if (time_before(now, migrate)) | |
2231 | return; | |
2232 | ||
598f0ec0 MG |
2233 | if (p->numa_scan_period == 0) { |
2234 | p->numa_scan_period_max = task_scan_max(p); | |
2235 | p->numa_scan_period = task_scan_min(p); | |
2236 | } | |
cbee9f88 | 2237 | |
fb003b80 | 2238 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2239 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2240 | return; | |
2241 | ||
19a78d11 PZ |
2242 | /* |
2243 | * Delay this task enough that another task of this mm will likely win | |
2244 | * the next time around. | |
2245 | */ | |
2246 | p->node_stamp += 2 * TICK_NSEC; | |
2247 | ||
9f40604c MG |
2248 | start = mm->numa_scan_offset; |
2249 | pages = sysctl_numa_balancing_scan_size; | |
2250 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2251 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2252 | if (!pages) |
2253 | return; | |
cbee9f88 | 2254 | |
4620f8c1 | 2255 | |
6e5fb223 | 2256 | down_read(&mm->mmap_sem); |
9f40604c | 2257 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2258 | if (!vma) { |
2259 | reset_ptenuma_scan(p); | |
9f40604c | 2260 | start = 0; |
6e5fb223 PZ |
2261 | vma = mm->mmap; |
2262 | } | |
9f40604c | 2263 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2264 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2265 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2266 | continue; |
6b79c57b | 2267 | } |
6e5fb223 | 2268 | |
4591ce4f MG |
2269 | /* |
2270 | * Shared library pages mapped by multiple processes are not | |
2271 | * migrated as it is expected they are cache replicated. Avoid | |
2272 | * hinting faults in read-only file-backed mappings or the vdso | |
2273 | * as migrating the pages will be of marginal benefit. | |
2274 | */ | |
2275 | if (!vma->vm_mm || | |
2276 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2277 | continue; | |
2278 | ||
3c67f474 MG |
2279 | /* |
2280 | * Skip inaccessible VMAs to avoid any confusion between | |
2281 | * PROT_NONE and NUMA hinting ptes | |
2282 | */ | |
2283 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2284 | continue; | |
4591ce4f | 2285 | |
9f40604c MG |
2286 | do { |
2287 | start = max(start, vma->vm_start); | |
2288 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2289 | end = min(end, vma->vm_end); | |
4620f8c1 | 2290 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2291 | |
2292 | /* | |
4620f8c1 RR |
2293 | * Try to scan sysctl_numa_balancing_size worth of |
2294 | * hpages that have at least one present PTE that | |
2295 | * is not already pte-numa. If the VMA contains | |
2296 | * areas that are unused or already full of prot_numa | |
2297 | * PTEs, scan up to virtpages, to skip through those | |
2298 | * areas faster. | |
598f0ec0 MG |
2299 | */ |
2300 | if (nr_pte_updates) | |
2301 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2302 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2303 | |
9f40604c | 2304 | start = end; |
4620f8c1 | 2305 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2306 | goto out; |
3cf1962c RR |
2307 | |
2308 | cond_resched(); | |
9f40604c | 2309 | } while (end != vma->vm_end); |
cbee9f88 | 2310 | } |
6e5fb223 | 2311 | |
9f40604c | 2312 | out: |
6e5fb223 | 2313 | /* |
c69307d5 PZ |
2314 | * It is possible to reach the end of the VMA list but the last few |
2315 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2316 | * would find the !migratable VMA on the next scan but not reset the | |
2317 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2318 | */ |
2319 | if (vma) | |
9f40604c | 2320 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2321 | else |
2322 | reset_ptenuma_scan(p); | |
2323 | up_read(&mm->mmap_sem); | |
51170840 RR |
2324 | |
2325 | /* | |
2326 | * Make sure tasks use at least 32x as much time to run other code | |
2327 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2328 | * Usually update_task_scan_period slows down scanning enough; on an | |
2329 | * overloaded system we need to limit overhead on a per task basis. | |
2330 | */ | |
2331 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2332 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2333 | p->node_stamp += 32 * diff; | |
2334 | } | |
cbee9f88 PZ |
2335 | } |
2336 | ||
2337 | /* | |
2338 | * Drive the periodic memory faults.. | |
2339 | */ | |
2340 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2341 | { | |
2342 | struct callback_head *work = &curr->numa_work; | |
2343 | u64 period, now; | |
2344 | ||
2345 | /* | |
2346 | * We don't care about NUMA placement if we don't have memory. | |
2347 | */ | |
2348 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2349 | return; | |
2350 | ||
2351 | /* | |
2352 | * Using runtime rather than walltime has the dual advantage that | |
2353 | * we (mostly) drive the selection from busy threads and that the | |
2354 | * task needs to have done some actual work before we bother with | |
2355 | * NUMA placement. | |
2356 | */ | |
2357 | now = curr->se.sum_exec_runtime; | |
2358 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2359 | ||
25b3e5a3 | 2360 | if (now > curr->node_stamp + period) { |
4b96a29b | 2361 | if (!curr->node_stamp) |
598f0ec0 | 2362 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 2363 | curr->node_stamp += period; |
cbee9f88 PZ |
2364 | |
2365 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2366 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2367 | task_work_add(curr, work, true); | |
2368 | } | |
2369 | } | |
2370 | } | |
2371 | #else | |
2372 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2373 | { | |
2374 | } | |
0ec8aa00 PZ |
2375 | |
2376 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2377 | { | |
2378 | } | |
2379 | ||
2380 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2381 | { | |
2382 | } | |
cbee9f88 PZ |
2383 | #endif /* CONFIG_NUMA_BALANCING */ |
2384 | ||
30cfdcfc DA |
2385 | static void |
2386 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2387 | { | |
2388 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2389 | if (!parent_entity(se)) |
029632fb | 2390 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2391 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2392 | if (entity_is_task(se)) { |
2393 | struct rq *rq = rq_of(cfs_rq); | |
2394 | ||
2395 | account_numa_enqueue(rq, task_of(se)); | |
2396 | list_add(&se->group_node, &rq->cfs_tasks); | |
2397 | } | |
367456c7 | 2398 | #endif |
30cfdcfc | 2399 | cfs_rq->nr_running++; |
30cfdcfc DA |
2400 | } |
2401 | ||
2402 | static void | |
2403 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2404 | { | |
2405 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2406 | if (!parent_entity(se)) |
029632fb | 2407 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
0ec8aa00 PZ |
2408 | if (entity_is_task(se)) { |
2409 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2410 | list_del_init(&se->group_node); |
0ec8aa00 | 2411 | } |
30cfdcfc | 2412 | cfs_rq->nr_running--; |
30cfdcfc DA |
2413 | } |
2414 | ||
3ff6dcac YZ |
2415 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2416 | # ifdef CONFIG_SMP | |
cf5f0acf PZ |
2417 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
2418 | { | |
2419 | long tg_weight; | |
2420 | ||
2421 | /* | |
9d89c257 YD |
2422 | * Use this CPU's real-time load instead of the last load contribution |
2423 | * as the updating of the contribution is delayed, and we will use the | |
2424 | * the real-time load to calc the share. See update_tg_load_avg(). | |
cf5f0acf | 2425 | */ |
bf5b986e | 2426 | tg_weight = atomic_long_read(&tg->load_avg); |
9d89c257 | 2427 | tg_weight -= cfs_rq->tg_load_avg_contrib; |
fde7d22e | 2428 | tg_weight += cfs_rq->load.weight; |
cf5f0acf PZ |
2429 | |
2430 | return tg_weight; | |
2431 | } | |
2432 | ||
6d5ab293 | 2433 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac | 2434 | { |
cf5f0acf | 2435 | long tg_weight, load, shares; |
3ff6dcac | 2436 | |
cf5f0acf | 2437 | tg_weight = calc_tg_weight(tg, cfs_rq); |
fde7d22e | 2438 | load = cfs_rq->load.weight; |
3ff6dcac | 2439 | |
3ff6dcac | 2440 | shares = (tg->shares * load); |
cf5f0acf PZ |
2441 | if (tg_weight) |
2442 | shares /= tg_weight; | |
3ff6dcac YZ |
2443 | |
2444 | if (shares < MIN_SHARES) | |
2445 | shares = MIN_SHARES; | |
2446 | if (shares > tg->shares) | |
2447 | shares = tg->shares; | |
2448 | ||
2449 | return shares; | |
2450 | } | |
3ff6dcac | 2451 | # else /* CONFIG_SMP */ |
6d5ab293 | 2452 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
2453 | { |
2454 | return tg->shares; | |
2455 | } | |
3ff6dcac | 2456 | # endif /* CONFIG_SMP */ |
2069dd75 PZ |
2457 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2458 | unsigned long weight) | |
2459 | { | |
19e5eebb PT |
2460 | if (se->on_rq) { |
2461 | /* commit outstanding execution time */ | |
2462 | if (cfs_rq->curr == se) | |
2463 | update_curr(cfs_rq); | |
2069dd75 | 2464 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 2465 | } |
2069dd75 PZ |
2466 | |
2467 | update_load_set(&se->load, weight); | |
2468 | ||
2469 | if (se->on_rq) | |
2470 | account_entity_enqueue(cfs_rq, se); | |
2471 | } | |
2472 | ||
82958366 PT |
2473 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2474 | ||
6d5ab293 | 2475 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2476 | { |
2477 | struct task_group *tg; | |
2478 | struct sched_entity *se; | |
3ff6dcac | 2479 | long shares; |
2069dd75 | 2480 | |
2069dd75 PZ |
2481 | tg = cfs_rq->tg; |
2482 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 2483 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 2484 | return; |
3ff6dcac YZ |
2485 | #ifndef CONFIG_SMP |
2486 | if (likely(se->load.weight == tg->shares)) | |
2487 | return; | |
2488 | #endif | |
6d5ab293 | 2489 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
2490 | |
2491 | reweight_entity(cfs_rq_of(se), se, shares); | |
2492 | } | |
2493 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 2494 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2495 | { |
2496 | } | |
2497 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
2498 | ||
141965c7 | 2499 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
2500 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ |
2501 | static const u32 runnable_avg_yN_inv[] = { | |
2502 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
2503 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
2504 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
2505 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
2506 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
2507 | 0x85aac367, 0x82cd8698, | |
2508 | }; | |
2509 | ||
2510 | /* | |
2511 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
2512 | * over-estimates when re-combining. | |
2513 | */ | |
2514 | static const u32 runnable_avg_yN_sum[] = { | |
2515 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
2516 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
2517 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
2518 | }; | |
2519 | ||
9d85f21c PT |
2520 | /* |
2521 | * Approximate: | |
2522 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
2523 | */ | |
2524 | static __always_inline u64 decay_load(u64 val, u64 n) | |
2525 | { | |
5b51f2f8 PT |
2526 | unsigned int local_n; |
2527 | ||
2528 | if (!n) | |
2529 | return val; | |
2530 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
2531 | return 0; | |
2532 | ||
2533 | /* after bounds checking we can collapse to 32-bit */ | |
2534 | local_n = n; | |
2535 | ||
2536 | /* | |
2537 | * As y^PERIOD = 1/2, we can combine | |
9c58c79a ZZ |
2538 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) |
2539 | * With a look-up table which covers y^n (n<PERIOD) | |
5b51f2f8 PT |
2540 | * |
2541 | * To achieve constant time decay_load. | |
2542 | */ | |
2543 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2544 | val >>= local_n / LOAD_AVG_PERIOD; | |
2545 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2546 | } |
2547 | ||
9d89c257 YD |
2548 | val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32); |
2549 | return val; | |
5b51f2f8 PT |
2550 | } |
2551 | ||
2552 | /* | |
2553 | * For updates fully spanning n periods, the contribution to runnable | |
2554 | * average will be: \Sum 1024*y^n | |
2555 | * | |
2556 | * We can compute this reasonably efficiently by combining: | |
2557 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2558 | */ | |
2559 | static u32 __compute_runnable_contrib(u64 n) | |
2560 | { | |
2561 | u32 contrib = 0; | |
2562 | ||
2563 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2564 | return runnable_avg_yN_sum[n]; | |
2565 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2566 | return LOAD_AVG_MAX; | |
2567 | ||
2568 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | |
2569 | do { | |
2570 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | |
2571 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | |
2572 | ||
2573 | n -= LOAD_AVG_PERIOD; | |
2574 | } while (n > LOAD_AVG_PERIOD); | |
2575 | ||
2576 | contrib = decay_load(contrib, n); | |
2577 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2578 | } |
2579 | ||
006cdf02 PZ |
2580 | #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10 |
2581 | #error "load tracking assumes 2^10 as unit" | |
2582 | #endif | |
2583 | ||
54a21385 | 2584 | #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) |
e0f5f3af | 2585 | |
9d85f21c PT |
2586 | /* |
2587 | * We can represent the historical contribution to runnable average as the | |
2588 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2589 | * history into segments of approximately 1ms (1024us); label the segment that | |
2590 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2591 | * | |
2592 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2593 | * p0 p1 p2 | |
2594 | * (now) (~1ms ago) (~2ms ago) | |
2595 | * | |
2596 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2597 | * | |
2598 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2599 | * following representation of historical load: | |
2600 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2601 | * | |
2602 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2603 | * y^32 = 0.5 | |
2604 | * | |
2605 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2606 | * approximately half as much as the contribution to load within the last ms | |
2607 | * (u_0). | |
2608 | * | |
2609 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2610 | * sum again by y is sufficient to update: | |
2611 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2612 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2613 | */ | |
9d89c257 YD |
2614 | static __always_inline int |
2615 | __update_load_avg(u64 now, int cpu, struct sched_avg *sa, | |
13962234 | 2616 | unsigned long weight, int running, struct cfs_rq *cfs_rq) |
9d85f21c | 2617 | { |
e0f5f3af | 2618 | u64 delta, scaled_delta, periods; |
9d89c257 | 2619 | u32 contrib; |
6115c793 | 2620 | unsigned int delta_w, scaled_delta_w, decayed = 0; |
6f2b0452 | 2621 | unsigned long scale_freq, scale_cpu; |
9d85f21c | 2622 | |
9d89c257 | 2623 | delta = now - sa->last_update_time; |
9d85f21c PT |
2624 | /* |
2625 | * This should only happen when time goes backwards, which it | |
2626 | * unfortunately does during sched clock init when we swap over to TSC. | |
2627 | */ | |
2628 | if ((s64)delta < 0) { | |
9d89c257 | 2629 | sa->last_update_time = now; |
9d85f21c PT |
2630 | return 0; |
2631 | } | |
2632 | ||
2633 | /* | |
2634 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2635 | * approximation of 1us and fast to compute. | |
2636 | */ | |
2637 | delta >>= 10; | |
2638 | if (!delta) | |
2639 | return 0; | |
9d89c257 | 2640 | sa->last_update_time = now; |
9d85f21c | 2641 | |
6f2b0452 DE |
2642 | scale_freq = arch_scale_freq_capacity(NULL, cpu); |
2643 | scale_cpu = arch_scale_cpu_capacity(NULL, cpu); | |
2644 | ||
9d85f21c | 2645 | /* delta_w is the amount already accumulated against our next period */ |
9d89c257 | 2646 | delta_w = sa->period_contrib; |
9d85f21c | 2647 | if (delta + delta_w >= 1024) { |
9d85f21c PT |
2648 | decayed = 1; |
2649 | ||
9d89c257 YD |
2650 | /* how much left for next period will start over, we don't know yet */ |
2651 | sa->period_contrib = 0; | |
2652 | ||
9d85f21c PT |
2653 | /* |
2654 | * Now that we know we're crossing a period boundary, figure | |
2655 | * out how much from delta we need to complete the current | |
2656 | * period and accrue it. | |
2657 | */ | |
2658 | delta_w = 1024 - delta_w; | |
54a21385 | 2659 | scaled_delta_w = cap_scale(delta_w, scale_freq); |
13962234 | 2660 | if (weight) { |
e0f5f3af DE |
2661 | sa->load_sum += weight * scaled_delta_w; |
2662 | if (cfs_rq) { | |
2663 | cfs_rq->runnable_load_sum += | |
2664 | weight * scaled_delta_w; | |
2665 | } | |
13962234 | 2666 | } |
36ee28e4 | 2667 | if (running) |
006cdf02 | 2668 | sa->util_sum += scaled_delta_w * scale_cpu; |
5b51f2f8 PT |
2669 | |
2670 | delta -= delta_w; | |
2671 | ||
2672 | /* Figure out how many additional periods this update spans */ | |
2673 | periods = delta / 1024; | |
2674 | delta %= 1024; | |
2675 | ||
9d89c257 | 2676 | sa->load_sum = decay_load(sa->load_sum, periods + 1); |
13962234 YD |
2677 | if (cfs_rq) { |
2678 | cfs_rq->runnable_load_sum = | |
2679 | decay_load(cfs_rq->runnable_load_sum, periods + 1); | |
2680 | } | |
9d89c257 | 2681 | sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1); |
5b51f2f8 PT |
2682 | |
2683 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
9d89c257 | 2684 | contrib = __compute_runnable_contrib(periods); |
54a21385 | 2685 | contrib = cap_scale(contrib, scale_freq); |
13962234 | 2686 | if (weight) { |
9d89c257 | 2687 | sa->load_sum += weight * contrib; |
13962234 YD |
2688 | if (cfs_rq) |
2689 | cfs_rq->runnable_load_sum += weight * contrib; | |
2690 | } | |
36ee28e4 | 2691 | if (running) |
006cdf02 | 2692 | sa->util_sum += contrib * scale_cpu; |
9d85f21c PT |
2693 | } |
2694 | ||
2695 | /* Remainder of delta accrued against u_0` */ | |
54a21385 | 2696 | scaled_delta = cap_scale(delta, scale_freq); |
13962234 | 2697 | if (weight) { |
e0f5f3af | 2698 | sa->load_sum += weight * scaled_delta; |
13962234 | 2699 | if (cfs_rq) |
e0f5f3af | 2700 | cfs_rq->runnable_load_sum += weight * scaled_delta; |
13962234 | 2701 | } |
36ee28e4 | 2702 | if (running) |
006cdf02 | 2703 | sa->util_sum += scaled_delta * scale_cpu; |
9ee474f5 | 2704 | |
9d89c257 | 2705 | sa->period_contrib += delta; |
9ee474f5 | 2706 | |
9d89c257 YD |
2707 | if (decayed) { |
2708 | sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX); | |
13962234 YD |
2709 | if (cfs_rq) { |
2710 | cfs_rq->runnable_load_avg = | |
2711 | div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX); | |
2712 | } | |
006cdf02 | 2713 | sa->util_avg = sa->util_sum / LOAD_AVG_MAX; |
9d89c257 | 2714 | } |
aff3e498 | 2715 | |
9d89c257 | 2716 | return decayed; |
9ee474f5 PT |
2717 | } |
2718 | ||
c566e8e9 | 2719 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bb17f655 | 2720 | /* |
9d89c257 YD |
2721 | * Updating tg's load_avg is necessary before update_cfs_share (which is done) |
2722 | * and effective_load (which is not done because it is too costly). | |
bb17f655 | 2723 | */ |
9d89c257 | 2724 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 2725 | { |
9d89c257 | 2726 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 2727 | |
aa0b7ae0 WL |
2728 | /* |
2729 | * No need to update load_avg for root_task_group as it is not used. | |
2730 | */ | |
2731 | if (cfs_rq->tg == &root_task_group) | |
2732 | return; | |
2733 | ||
9d89c257 YD |
2734 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
2735 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
2736 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 2737 | } |
8165e145 | 2738 | } |
f5f9739d | 2739 | |
ad936d86 BP |
2740 | /* |
2741 | * Called within set_task_rq() right before setting a task's cpu. The | |
2742 | * caller only guarantees p->pi_lock is held; no other assumptions, | |
2743 | * including the state of rq->lock, should be made. | |
2744 | */ | |
2745 | void set_task_rq_fair(struct sched_entity *se, | |
2746 | struct cfs_rq *prev, struct cfs_rq *next) | |
2747 | { | |
2748 | if (!sched_feat(ATTACH_AGE_LOAD)) | |
2749 | return; | |
2750 | ||
2751 | /* | |
2752 | * We are supposed to update the task to "current" time, then its up to | |
2753 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
2754 | * getting what current time is, so simply throw away the out-of-date | |
2755 | * time. This will result in the wakee task is less decayed, but giving | |
2756 | * the wakee more load sounds not bad. | |
2757 | */ | |
2758 | if (se->avg.last_update_time && prev) { | |
2759 | u64 p_last_update_time; | |
2760 | u64 n_last_update_time; | |
2761 | ||
2762 | #ifndef CONFIG_64BIT | |
2763 | u64 p_last_update_time_copy; | |
2764 | u64 n_last_update_time_copy; | |
2765 | ||
2766 | do { | |
2767 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
2768 | n_last_update_time_copy = next->load_last_update_time_copy; | |
2769 | ||
2770 | smp_rmb(); | |
2771 | ||
2772 | p_last_update_time = prev->avg.last_update_time; | |
2773 | n_last_update_time = next->avg.last_update_time; | |
2774 | ||
2775 | } while (p_last_update_time != p_last_update_time_copy || | |
2776 | n_last_update_time != n_last_update_time_copy); | |
2777 | #else | |
2778 | p_last_update_time = prev->avg.last_update_time; | |
2779 | n_last_update_time = next->avg.last_update_time; | |
2780 | #endif | |
2781 | __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)), | |
2782 | &se->avg, 0, 0, NULL); | |
2783 | se->avg.last_update_time = n_last_update_time; | |
2784 | } | |
2785 | } | |
6e83125c | 2786 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
9d89c257 | 2787 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
6e83125c | 2788 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 2789 | |
9d89c257 | 2790 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
8165e145 | 2791 | |
9d89c257 YD |
2792 | /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */ |
2793 | static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) | |
2dac754e | 2794 | { |
9d89c257 | 2795 | struct sched_avg *sa = &cfs_rq->avg; |
3e386d56 | 2796 | int decayed, removed = 0; |
2dac754e | 2797 | |
9d89c257 | 2798 | if (atomic_long_read(&cfs_rq->removed_load_avg)) { |
9e0e83a1 | 2799 | s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); |
9d89c257 YD |
2800 | sa->load_avg = max_t(long, sa->load_avg - r, 0); |
2801 | sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0); | |
3e386d56 | 2802 | removed = 1; |
8165e145 | 2803 | } |
2dac754e | 2804 | |
9d89c257 YD |
2805 | if (atomic_long_read(&cfs_rq->removed_util_avg)) { |
2806 | long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0); | |
2807 | sa->util_avg = max_t(long, sa->util_avg - r, 0); | |
006cdf02 | 2808 | sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0); |
9d89c257 | 2809 | } |
36ee28e4 | 2810 | |
9d89c257 | 2811 | decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa, |
13962234 | 2812 | scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq); |
36ee28e4 | 2813 | |
9d89c257 YD |
2814 | #ifndef CONFIG_64BIT |
2815 | smp_wmb(); | |
2816 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
2817 | #endif | |
36ee28e4 | 2818 | |
3e386d56 | 2819 | return decayed || removed; |
9ee474f5 PT |
2820 | } |
2821 | ||
9d89c257 YD |
2822 | /* Update task and its cfs_rq load average */ |
2823 | static inline void update_load_avg(struct sched_entity *se, int update_tg) | |
9d85f21c | 2824 | { |
2dac754e | 2825 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
9d89c257 | 2826 | u64 now = cfs_rq_clock_task(cfs_rq); |
a05e8c51 | 2827 | int cpu = cpu_of(rq_of(cfs_rq)); |
2dac754e | 2828 | |
f1b17280 | 2829 | /* |
9d89c257 YD |
2830 | * Track task load average for carrying it to new CPU after migrated, and |
2831 | * track group sched_entity load average for task_h_load calc in migration | |
f1b17280 | 2832 | */ |
9d89c257 | 2833 | __update_load_avg(now, cpu, &se->avg, |
a05e8c51 BP |
2834 | se->on_rq * scale_load_down(se->load.weight), |
2835 | cfs_rq->curr == se, NULL); | |
f1b17280 | 2836 | |
9d89c257 YD |
2837 | if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg) |
2838 | update_tg_load_avg(cfs_rq, 0); | |
9ee474f5 PT |
2839 | } |
2840 | ||
a05e8c51 BP |
2841 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2842 | { | |
a9280514 PZ |
2843 | if (!sched_feat(ATTACH_AGE_LOAD)) |
2844 | goto skip_aging; | |
2845 | ||
6efdb105 BP |
2846 | /* |
2847 | * If we got migrated (either between CPUs or between cgroups) we'll | |
2848 | * have aged the average right before clearing @last_update_time. | |
2849 | */ | |
2850 | if (se->avg.last_update_time) { | |
2851 | __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)), | |
2852 | &se->avg, 0, 0, NULL); | |
2853 | ||
2854 | /* | |
2855 | * XXX: we could have just aged the entire load away if we've been | |
2856 | * absent from the fair class for too long. | |
2857 | */ | |
2858 | } | |
2859 | ||
a9280514 | 2860 | skip_aging: |
a05e8c51 BP |
2861 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
2862 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
2863 | cfs_rq->avg.load_sum += se->avg.load_sum; | |
2864 | cfs_rq->avg.util_avg += se->avg.util_avg; | |
2865 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
2866 | } | |
2867 | ||
2868 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2869 | { | |
2870 | __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)), | |
2871 | &se->avg, se->on_rq * scale_load_down(se->load.weight), | |
2872 | cfs_rq->curr == se, NULL); | |
2873 | ||
2874 | cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0); | |
2875 | cfs_rq->avg.load_sum = max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0); | |
2876 | cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0); | |
2877 | cfs_rq->avg.util_sum = max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0); | |
2878 | } | |
2879 | ||
9d89c257 YD |
2880 | /* Add the load generated by se into cfs_rq's load average */ |
2881 | static inline void | |
2882 | enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
9ee474f5 | 2883 | { |
9d89c257 YD |
2884 | struct sched_avg *sa = &se->avg; |
2885 | u64 now = cfs_rq_clock_task(cfs_rq); | |
a05e8c51 | 2886 | int migrated, decayed; |
9ee474f5 | 2887 | |
a05e8c51 BP |
2888 | migrated = !sa->last_update_time; |
2889 | if (!migrated) { | |
9d89c257 | 2890 | __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa, |
13962234 YD |
2891 | se->on_rq * scale_load_down(se->load.weight), |
2892 | cfs_rq->curr == se, NULL); | |
aff3e498 | 2893 | } |
c566e8e9 | 2894 | |
9d89c257 | 2895 | decayed = update_cfs_rq_load_avg(now, cfs_rq); |
18bf2805 | 2896 | |
13962234 YD |
2897 | cfs_rq->runnable_load_avg += sa->load_avg; |
2898 | cfs_rq->runnable_load_sum += sa->load_sum; | |
2899 | ||
a05e8c51 BP |
2900 | if (migrated) |
2901 | attach_entity_load_avg(cfs_rq, se); | |
9ee474f5 | 2902 | |
9d89c257 YD |
2903 | if (decayed || migrated) |
2904 | update_tg_load_avg(cfs_rq, 0); | |
2dac754e PT |
2905 | } |
2906 | ||
13962234 YD |
2907 | /* Remove the runnable load generated by se from cfs_rq's runnable load average */ |
2908 | static inline void | |
2909 | dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2910 | { | |
2911 | update_load_avg(se, 1); | |
2912 | ||
2913 | cfs_rq->runnable_load_avg = | |
2914 | max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0); | |
2915 | cfs_rq->runnable_load_sum = | |
a05e8c51 | 2916 | max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0); |
13962234 YD |
2917 | } |
2918 | ||
9d89c257 | 2919 | #ifndef CONFIG_64BIT |
0905f04e YD |
2920 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
2921 | { | |
9d89c257 | 2922 | u64 last_update_time_copy; |
0905f04e | 2923 | u64 last_update_time; |
9ee474f5 | 2924 | |
9d89c257 YD |
2925 | do { |
2926 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
2927 | smp_rmb(); | |
2928 | last_update_time = cfs_rq->avg.last_update_time; | |
2929 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
2930 | |
2931 | return last_update_time; | |
2932 | } | |
9d89c257 | 2933 | #else |
0905f04e YD |
2934 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
2935 | { | |
2936 | return cfs_rq->avg.last_update_time; | |
2937 | } | |
9d89c257 YD |
2938 | #endif |
2939 | ||
0905f04e YD |
2940 | /* |
2941 | * Task first catches up with cfs_rq, and then subtract | |
2942 | * itself from the cfs_rq (task must be off the queue now). | |
2943 | */ | |
2944 | void remove_entity_load_avg(struct sched_entity *se) | |
2945 | { | |
2946 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2947 | u64 last_update_time; | |
2948 | ||
2949 | /* | |
2950 | * Newly created task or never used group entity should not be removed | |
2951 | * from its (source) cfs_rq | |
2952 | */ | |
2953 | if (se->avg.last_update_time == 0) | |
2954 | return; | |
2955 | ||
2956 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
2957 | ||
13962234 | 2958 | __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL); |
9d89c257 YD |
2959 | atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg); |
2960 | atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg); | |
2dac754e | 2961 | } |
642dbc39 | 2962 | |
7ea241af YD |
2963 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) |
2964 | { | |
2965 | return cfs_rq->runnable_load_avg; | |
2966 | } | |
2967 | ||
2968 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) | |
2969 | { | |
2970 | return cfs_rq->avg.load_avg; | |
2971 | } | |
2972 | ||
6e83125c PZ |
2973 | static int idle_balance(struct rq *this_rq); |
2974 | ||
38033c37 PZ |
2975 | #else /* CONFIG_SMP */ |
2976 | ||
9d89c257 YD |
2977 | static inline void update_load_avg(struct sched_entity *se, int update_tg) {} |
2978 | static inline void | |
2979 | enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
13962234 YD |
2980 | static inline void |
2981 | dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
9d89c257 | 2982 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 2983 | |
a05e8c51 BP |
2984 | static inline void |
2985 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
2986 | static inline void | |
2987 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
2988 | ||
6e83125c PZ |
2989 | static inline int idle_balance(struct rq *rq) |
2990 | { | |
2991 | return 0; | |
2992 | } | |
2993 | ||
38033c37 | 2994 | #endif /* CONFIG_SMP */ |
9d85f21c | 2995 | |
2396af69 | 2996 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 2997 | { |
bf0f6f24 | 2998 | #ifdef CONFIG_SCHEDSTATS |
e414314c PZ |
2999 | struct task_struct *tsk = NULL; |
3000 | ||
3001 | if (entity_is_task(se)) | |
3002 | tsk = task_of(se); | |
3003 | ||
41acab88 | 3004 | if (se->statistics.sleep_start) { |
78becc27 | 3005 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
bf0f6f24 IM |
3006 | |
3007 | if ((s64)delta < 0) | |
3008 | delta = 0; | |
3009 | ||
41acab88 LDM |
3010 | if (unlikely(delta > se->statistics.sleep_max)) |
3011 | se->statistics.sleep_max = delta; | |
bf0f6f24 | 3012 | |
8c79a045 | 3013 | se->statistics.sleep_start = 0; |
41acab88 | 3014 | se->statistics.sum_sleep_runtime += delta; |
9745512c | 3015 | |
768d0c27 | 3016 | if (tsk) { |
e414314c | 3017 | account_scheduler_latency(tsk, delta >> 10, 1); |
768d0c27 PZ |
3018 | trace_sched_stat_sleep(tsk, delta); |
3019 | } | |
bf0f6f24 | 3020 | } |
41acab88 | 3021 | if (se->statistics.block_start) { |
78becc27 | 3022 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
bf0f6f24 IM |
3023 | |
3024 | if ((s64)delta < 0) | |
3025 | delta = 0; | |
3026 | ||
41acab88 LDM |
3027 | if (unlikely(delta > se->statistics.block_max)) |
3028 | se->statistics.block_max = delta; | |
bf0f6f24 | 3029 | |
8c79a045 | 3030 | se->statistics.block_start = 0; |
41acab88 | 3031 | se->statistics.sum_sleep_runtime += delta; |
30084fbd | 3032 | |
e414314c | 3033 | if (tsk) { |
8f0dfc34 | 3034 | if (tsk->in_iowait) { |
41acab88 LDM |
3035 | se->statistics.iowait_sum += delta; |
3036 | se->statistics.iowait_count++; | |
768d0c27 | 3037 | trace_sched_stat_iowait(tsk, delta); |
8f0dfc34 AV |
3038 | } |
3039 | ||
b781a602 AV |
3040 | trace_sched_stat_blocked(tsk, delta); |
3041 | ||
e414314c PZ |
3042 | /* |
3043 | * Blocking time is in units of nanosecs, so shift by | |
3044 | * 20 to get a milliseconds-range estimation of the | |
3045 | * amount of time that the task spent sleeping: | |
3046 | */ | |
3047 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
3048 | profile_hits(SLEEP_PROFILING, | |
3049 | (void *)get_wchan(tsk), | |
3050 | delta >> 20); | |
3051 | } | |
3052 | account_scheduler_latency(tsk, delta >> 10, 0); | |
30084fbd | 3053 | } |
bf0f6f24 IM |
3054 | } |
3055 | #endif | |
3056 | } | |
3057 | ||
ddc97297 PZ |
3058 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3059 | { | |
3060 | #ifdef CONFIG_SCHED_DEBUG | |
3061 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
3062 | ||
3063 | if (d < 0) | |
3064 | d = -d; | |
3065 | ||
3066 | if (d > 3*sysctl_sched_latency) | |
3067 | schedstat_inc(cfs_rq, nr_spread_over); | |
3068 | #endif | |
3069 | } | |
3070 | ||
aeb73b04 PZ |
3071 | static void |
3072 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
3073 | { | |
1af5f730 | 3074 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 3075 | |
2cb8600e PZ |
3076 | /* |
3077 | * The 'current' period is already promised to the current tasks, | |
3078 | * however the extra weight of the new task will slow them down a | |
3079 | * little, place the new task so that it fits in the slot that | |
3080 | * stays open at the end. | |
3081 | */ | |
94dfb5e7 | 3082 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 3083 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 3084 | |
a2e7a7eb | 3085 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 3086 | if (!initial) { |
a2e7a7eb | 3087 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 3088 | |
a2e7a7eb MG |
3089 | /* |
3090 | * Halve their sleep time's effect, to allow | |
3091 | * for a gentler effect of sleepers: | |
3092 | */ | |
3093 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
3094 | thresh >>= 1; | |
51e0304c | 3095 | |
a2e7a7eb | 3096 | vruntime -= thresh; |
aeb73b04 PZ |
3097 | } |
3098 | ||
b5d9d734 | 3099 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3100 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3101 | } |
3102 | ||
d3d9dc33 PT |
3103 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3104 | ||
bf0f6f24 | 3105 | static void |
88ec22d3 | 3106 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3107 | { |
88ec22d3 PZ |
3108 | /* |
3109 | * Update the normalized vruntime before updating min_vruntime | |
0fc576d5 | 3110 | * through calling update_curr(). |
88ec22d3 | 3111 | */ |
371fd7e7 | 3112 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
88ec22d3 PZ |
3113 | se->vruntime += cfs_rq->min_vruntime; |
3114 | ||
bf0f6f24 | 3115 | /* |
a2a2d680 | 3116 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3117 | */ |
b7cc0896 | 3118 | update_curr(cfs_rq); |
9d89c257 | 3119 | enqueue_entity_load_avg(cfs_rq, se); |
17bc14b7 LT |
3120 | account_entity_enqueue(cfs_rq, se); |
3121 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 3122 | |
88ec22d3 | 3123 | if (flags & ENQUEUE_WAKEUP) { |
aeb73b04 | 3124 | place_entity(cfs_rq, se, 0); |
2396af69 | 3125 | enqueue_sleeper(cfs_rq, se); |
e9acbff6 | 3126 | } |
bf0f6f24 | 3127 | |
d2417e5a | 3128 | update_stats_enqueue(cfs_rq, se); |
ddc97297 | 3129 | check_spread(cfs_rq, se); |
83b699ed SV |
3130 | if (se != cfs_rq->curr) |
3131 | __enqueue_entity(cfs_rq, se); | |
2069dd75 | 3132 | se->on_rq = 1; |
3d4b47b4 | 3133 | |
d3d9dc33 | 3134 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 3135 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
3136 | check_enqueue_throttle(cfs_rq); |
3137 | } | |
bf0f6f24 IM |
3138 | } |
3139 | ||
2c13c919 | 3140 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 3141 | { |
2c13c919 RR |
3142 | for_each_sched_entity(se) { |
3143 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3144 | if (cfs_rq->last != se) |
2c13c919 | 3145 | break; |
f1044799 PZ |
3146 | |
3147 | cfs_rq->last = NULL; | |
2c13c919 RR |
3148 | } |
3149 | } | |
2002c695 | 3150 | |
2c13c919 RR |
3151 | static void __clear_buddies_next(struct sched_entity *se) |
3152 | { | |
3153 | for_each_sched_entity(se) { | |
3154 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3155 | if (cfs_rq->next != se) |
2c13c919 | 3156 | break; |
f1044799 PZ |
3157 | |
3158 | cfs_rq->next = NULL; | |
2c13c919 | 3159 | } |
2002c695 PZ |
3160 | } |
3161 | ||
ac53db59 RR |
3162 | static void __clear_buddies_skip(struct sched_entity *se) |
3163 | { | |
3164 | for_each_sched_entity(se) { | |
3165 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3166 | if (cfs_rq->skip != se) |
ac53db59 | 3167 | break; |
f1044799 PZ |
3168 | |
3169 | cfs_rq->skip = NULL; | |
ac53db59 RR |
3170 | } |
3171 | } | |
3172 | ||
a571bbea PZ |
3173 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3174 | { | |
2c13c919 RR |
3175 | if (cfs_rq->last == se) |
3176 | __clear_buddies_last(se); | |
3177 | ||
3178 | if (cfs_rq->next == se) | |
3179 | __clear_buddies_next(se); | |
ac53db59 RR |
3180 | |
3181 | if (cfs_rq->skip == se) | |
3182 | __clear_buddies_skip(se); | |
a571bbea PZ |
3183 | } |
3184 | ||
6c16a6dc | 3185 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 3186 | |
bf0f6f24 | 3187 | static void |
371fd7e7 | 3188 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3189 | { |
a2a2d680 DA |
3190 | /* |
3191 | * Update run-time statistics of the 'current'. | |
3192 | */ | |
3193 | update_curr(cfs_rq); | |
13962234 | 3194 | dequeue_entity_load_avg(cfs_rq, se); |
a2a2d680 | 3195 | |
19b6a2e3 | 3196 | update_stats_dequeue(cfs_rq, se); |
371fd7e7 | 3197 | if (flags & DEQUEUE_SLEEP) { |
67e9fb2a | 3198 | #ifdef CONFIG_SCHEDSTATS |
bf0f6f24 IM |
3199 | if (entity_is_task(se)) { |
3200 | struct task_struct *tsk = task_of(se); | |
3201 | ||
3202 | if (tsk->state & TASK_INTERRUPTIBLE) | |
78becc27 | 3203 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 3204 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
78becc27 | 3205 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
bf0f6f24 | 3206 | } |
db36cc7d | 3207 | #endif |
67e9fb2a PZ |
3208 | } |
3209 | ||
2002c695 | 3210 | clear_buddies(cfs_rq, se); |
4793241b | 3211 | |
83b699ed | 3212 | if (se != cfs_rq->curr) |
30cfdcfc | 3213 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 3214 | se->on_rq = 0; |
30cfdcfc | 3215 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
3216 | |
3217 | /* | |
3218 | * Normalize the entity after updating the min_vruntime because the | |
3219 | * update can refer to the ->curr item and we need to reflect this | |
3220 | * movement in our normalized position. | |
3221 | */ | |
371fd7e7 | 3222 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 3223 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 3224 | |
d8b4986d PT |
3225 | /* return excess runtime on last dequeue */ |
3226 | return_cfs_rq_runtime(cfs_rq); | |
3227 | ||
1e876231 | 3228 | update_min_vruntime(cfs_rq); |
17bc14b7 | 3229 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
3230 | } |
3231 | ||
3232 | /* | |
3233 | * Preempt the current task with a newly woken task if needed: | |
3234 | */ | |
7c92e54f | 3235 | static void |
2e09bf55 | 3236 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 3237 | { |
11697830 | 3238 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
3239 | struct sched_entity *se; |
3240 | s64 delta; | |
11697830 | 3241 | |
6d0f0ebd | 3242 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 3243 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 3244 | if (delta_exec > ideal_runtime) { |
8875125e | 3245 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
3246 | /* |
3247 | * The current task ran long enough, ensure it doesn't get | |
3248 | * re-elected due to buddy favours. | |
3249 | */ | |
3250 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
3251 | return; |
3252 | } | |
3253 | ||
3254 | /* | |
3255 | * Ensure that a task that missed wakeup preemption by a | |
3256 | * narrow margin doesn't have to wait for a full slice. | |
3257 | * This also mitigates buddy induced latencies under load. | |
3258 | */ | |
f685ceac MG |
3259 | if (delta_exec < sysctl_sched_min_granularity) |
3260 | return; | |
3261 | ||
f4cfb33e WX |
3262 | se = __pick_first_entity(cfs_rq); |
3263 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 3264 | |
f4cfb33e WX |
3265 | if (delta < 0) |
3266 | return; | |
d7d82944 | 3267 | |
f4cfb33e | 3268 | if (delta > ideal_runtime) |
8875125e | 3269 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
3270 | } |
3271 | ||
83b699ed | 3272 | static void |
8494f412 | 3273 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 3274 | { |
83b699ed SV |
3275 | /* 'current' is not kept within the tree. */ |
3276 | if (se->on_rq) { | |
3277 | /* | |
3278 | * Any task has to be enqueued before it get to execute on | |
3279 | * a CPU. So account for the time it spent waiting on the | |
3280 | * runqueue. | |
3281 | */ | |
3282 | update_stats_wait_end(cfs_rq, se); | |
3283 | __dequeue_entity(cfs_rq, se); | |
9d89c257 | 3284 | update_load_avg(se, 1); |
83b699ed SV |
3285 | } |
3286 | ||
79303e9e | 3287 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 3288 | cfs_rq->curr = se; |
eba1ed4b IM |
3289 | #ifdef CONFIG_SCHEDSTATS |
3290 | /* | |
3291 | * Track our maximum slice length, if the CPU's load is at | |
3292 | * least twice that of our own weight (i.e. dont track it | |
3293 | * when there are only lesser-weight tasks around): | |
3294 | */ | |
495eca49 | 3295 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
41acab88 | 3296 | se->statistics.slice_max = max(se->statistics.slice_max, |
eba1ed4b IM |
3297 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
3298 | } | |
3299 | #endif | |
4a55b450 | 3300 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
3301 | } |
3302 | ||
3f3a4904 PZ |
3303 | static int |
3304 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
3305 | ||
ac53db59 RR |
3306 | /* |
3307 | * Pick the next process, keeping these things in mind, in this order: | |
3308 | * 1) keep things fair between processes/task groups | |
3309 | * 2) pick the "next" process, since someone really wants that to run | |
3310 | * 3) pick the "last" process, for cache locality | |
3311 | * 4) do not run the "skip" process, if something else is available | |
3312 | */ | |
678d5718 PZ |
3313 | static struct sched_entity * |
3314 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 3315 | { |
678d5718 PZ |
3316 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
3317 | struct sched_entity *se; | |
3318 | ||
3319 | /* | |
3320 | * If curr is set we have to see if its left of the leftmost entity | |
3321 | * still in the tree, provided there was anything in the tree at all. | |
3322 | */ | |
3323 | if (!left || (curr && entity_before(curr, left))) | |
3324 | left = curr; | |
3325 | ||
3326 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 3327 | |
ac53db59 RR |
3328 | /* |
3329 | * Avoid running the skip buddy, if running something else can | |
3330 | * be done without getting too unfair. | |
3331 | */ | |
3332 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
3333 | struct sched_entity *second; |
3334 | ||
3335 | if (se == curr) { | |
3336 | second = __pick_first_entity(cfs_rq); | |
3337 | } else { | |
3338 | second = __pick_next_entity(se); | |
3339 | if (!second || (curr && entity_before(curr, second))) | |
3340 | second = curr; | |
3341 | } | |
3342 | ||
ac53db59 RR |
3343 | if (second && wakeup_preempt_entity(second, left) < 1) |
3344 | se = second; | |
3345 | } | |
aa2ac252 | 3346 | |
f685ceac MG |
3347 | /* |
3348 | * Prefer last buddy, try to return the CPU to a preempted task. | |
3349 | */ | |
3350 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
3351 | se = cfs_rq->last; | |
3352 | ||
ac53db59 RR |
3353 | /* |
3354 | * Someone really wants this to run. If it's not unfair, run it. | |
3355 | */ | |
3356 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
3357 | se = cfs_rq->next; | |
3358 | ||
f685ceac | 3359 | clear_buddies(cfs_rq, se); |
4793241b PZ |
3360 | |
3361 | return se; | |
aa2ac252 PZ |
3362 | } |
3363 | ||
678d5718 | 3364 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 3365 | |
ab6cde26 | 3366 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
3367 | { |
3368 | /* | |
3369 | * If still on the runqueue then deactivate_task() | |
3370 | * was not called and update_curr() has to be done: | |
3371 | */ | |
3372 | if (prev->on_rq) | |
b7cc0896 | 3373 | update_curr(cfs_rq); |
bf0f6f24 | 3374 | |
d3d9dc33 PT |
3375 | /* throttle cfs_rqs exceeding runtime */ |
3376 | check_cfs_rq_runtime(cfs_rq); | |
3377 | ||
ddc97297 | 3378 | check_spread(cfs_rq, prev); |
30cfdcfc | 3379 | if (prev->on_rq) { |
5870db5b | 3380 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
3381 | /* Put 'current' back into the tree. */ |
3382 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 3383 | /* in !on_rq case, update occurred at dequeue */ |
9d89c257 | 3384 | update_load_avg(prev, 0); |
30cfdcfc | 3385 | } |
429d43bc | 3386 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
3387 | } |
3388 | ||
8f4d37ec PZ |
3389 | static void |
3390 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 3391 | { |
bf0f6f24 | 3392 | /* |
30cfdcfc | 3393 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3394 | */ |
30cfdcfc | 3395 | update_curr(cfs_rq); |
bf0f6f24 | 3396 | |
9d85f21c PT |
3397 | /* |
3398 | * Ensure that runnable average is periodically updated. | |
3399 | */ | |
9d89c257 | 3400 | update_load_avg(curr, 1); |
bf0bd948 | 3401 | update_cfs_shares(cfs_rq); |
9d85f21c | 3402 | |
8f4d37ec PZ |
3403 | #ifdef CONFIG_SCHED_HRTICK |
3404 | /* | |
3405 | * queued ticks are scheduled to match the slice, so don't bother | |
3406 | * validating it and just reschedule. | |
3407 | */ | |
983ed7a6 | 3408 | if (queued) { |
8875125e | 3409 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
3410 | return; |
3411 | } | |
8f4d37ec PZ |
3412 | /* |
3413 | * don't let the period tick interfere with the hrtick preemption | |
3414 | */ | |
3415 | if (!sched_feat(DOUBLE_TICK) && | |
3416 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
3417 | return; | |
3418 | #endif | |
3419 | ||
2c2efaed | 3420 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 3421 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
3422 | } |
3423 | ||
ab84d31e PT |
3424 | |
3425 | /************************************************** | |
3426 | * CFS bandwidth control machinery | |
3427 | */ | |
3428 | ||
3429 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
3430 | |
3431 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 3432 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
3433 | |
3434 | static inline bool cfs_bandwidth_used(void) | |
3435 | { | |
c5905afb | 3436 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
3437 | } |
3438 | ||
1ee14e6c | 3439 | void cfs_bandwidth_usage_inc(void) |
029632fb | 3440 | { |
1ee14e6c BS |
3441 | static_key_slow_inc(&__cfs_bandwidth_used); |
3442 | } | |
3443 | ||
3444 | void cfs_bandwidth_usage_dec(void) | |
3445 | { | |
3446 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
3447 | } |
3448 | #else /* HAVE_JUMP_LABEL */ | |
3449 | static bool cfs_bandwidth_used(void) | |
3450 | { | |
3451 | return true; | |
3452 | } | |
3453 | ||
1ee14e6c BS |
3454 | void cfs_bandwidth_usage_inc(void) {} |
3455 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
3456 | #endif /* HAVE_JUMP_LABEL */ |
3457 | ||
ab84d31e PT |
3458 | /* |
3459 | * default period for cfs group bandwidth. | |
3460 | * default: 0.1s, units: nanoseconds | |
3461 | */ | |
3462 | static inline u64 default_cfs_period(void) | |
3463 | { | |
3464 | return 100000000ULL; | |
3465 | } | |
ec12cb7f PT |
3466 | |
3467 | static inline u64 sched_cfs_bandwidth_slice(void) | |
3468 | { | |
3469 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
3470 | } | |
3471 | ||
a9cf55b2 PT |
3472 | /* |
3473 | * Replenish runtime according to assigned quota and update expiration time. | |
3474 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
3475 | * additional synchronization around rq->lock. | |
3476 | * | |
3477 | * requires cfs_b->lock | |
3478 | */ | |
029632fb | 3479 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
3480 | { |
3481 | u64 now; | |
3482 | ||
3483 | if (cfs_b->quota == RUNTIME_INF) | |
3484 | return; | |
3485 | ||
3486 | now = sched_clock_cpu(smp_processor_id()); | |
3487 | cfs_b->runtime = cfs_b->quota; | |
3488 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
3489 | } | |
3490 | ||
029632fb PZ |
3491 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3492 | { | |
3493 | return &tg->cfs_bandwidth; | |
3494 | } | |
3495 | ||
f1b17280 PT |
3496 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
3497 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
3498 | { | |
3499 | if (unlikely(cfs_rq->throttle_count)) | |
3500 | return cfs_rq->throttled_clock_task; | |
3501 | ||
78becc27 | 3502 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
3503 | } |
3504 | ||
85dac906 PT |
3505 | /* returns 0 on failure to allocate runtime */ |
3506 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
3507 | { |
3508 | struct task_group *tg = cfs_rq->tg; | |
3509 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 3510 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
3511 | |
3512 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
3513 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
3514 | ||
3515 | raw_spin_lock(&cfs_b->lock); | |
3516 | if (cfs_b->quota == RUNTIME_INF) | |
3517 | amount = min_amount; | |
58088ad0 | 3518 | else { |
77a4d1a1 | 3519 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
3520 | |
3521 | if (cfs_b->runtime > 0) { | |
3522 | amount = min(cfs_b->runtime, min_amount); | |
3523 | cfs_b->runtime -= amount; | |
3524 | cfs_b->idle = 0; | |
3525 | } | |
ec12cb7f | 3526 | } |
a9cf55b2 | 3527 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
3528 | raw_spin_unlock(&cfs_b->lock); |
3529 | ||
3530 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
3531 | /* |
3532 | * we may have advanced our local expiration to account for allowed | |
3533 | * spread between our sched_clock and the one on which runtime was | |
3534 | * issued. | |
3535 | */ | |
3536 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
3537 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
3538 | |
3539 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
3540 | } |
3541 | ||
a9cf55b2 PT |
3542 | /* |
3543 | * Note: This depends on the synchronization provided by sched_clock and the | |
3544 | * fact that rq->clock snapshots this value. | |
3545 | */ | |
3546 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 3547 | { |
a9cf55b2 | 3548 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
3549 | |
3550 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 3551 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
3552 | return; |
3553 | ||
a9cf55b2 PT |
3554 | if (cfs_rq->runtime_remaining < 0) |
3555 | return; | |
3556 | ||
3557 | /* | |
3558 | * If the local deadline has passed we have to consider the | |
3559 | * possibility that our sched_clock is 'fast' and the global deadline | |
3560 | * has not truly expired. | |
3561 | * | |
3562 | * Fortunately we can check determine whether this the case by checking | |
51f2176d BS |
3563 | * whether the global deadline has advanced. It is valid to compare |
3564 | * cfs_b->runtime_expires without any locks since we only care about | |
3565 | * exact equality, so a partial write will still work. | |
a9cf55b2 PT |
3566 | */ |
3567 | ||
51f2176d | 3568 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { |
a9cf55b2 PT |
3569 | /* extend local deadline, drift is bounded above by 2 ticks */ |
3570 | cfs_rq->runtime_expires += TICK_NSEC; | |
3571 | } else { | |
3572 | /* global deadline is ahead, expiration has passed */ | |
3573 | cfs_rq->runtime_remaining = 0; | |
3574 | } | |
3575 | } | |
3576 | ||
9dbdb155 | 3577 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
3578 | { |
3579 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3580 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3581 | expire_cfs_rq_runtime(cfs_rq); |
3582 | ||
3583 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3584 | return; |
3585 | ||
85dac906 PT |
3586 | /* |
3587 | * if we're unable to extend our runtime we resched so that the active | |
3588 | * hierarchy can be throttled | |
3589 | */ | |
3590 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 3591 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
3592 | } |
3593 | ||
6c16a6dc | 3594 | static __always_inline |
9dbdb155 | 3595 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 3596 | { |
56f570e5 | 3597 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3598 | return; |
3599 | ||
3600 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3601 | } | |
3602 | ||
85dac906 PT |
3603 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3604 | { | |
56f570e5 | 3605 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3606 | } |
3607 | ||
64660c86 PT |
3608 | /* check whether cfs_rq, or any parent, is throttled */ |
3609 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3610 | { | |
56f570e5 | 3611 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3612 | } |
3613 | ||
3614 | /* | |
3615 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3616 | * dest_cpu are members of a throttled hierarchy when performing group | |
3617 | * load-balance operations. | |
3618 | */ | |
3619 | static inline int throttled_lb_pair(struct task_group *tg, | |
3620 | int src_cpu, int dest_cpu) | |
3621 | { | |
3622 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3623 | ||
3624 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3625 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3626 | ||
3627 | return throttled_hierarchy(src_cfs_rq) || | |
3628 | throttled_hierarchy(dest_cfs_rq); | |
3629 | } | |
3630 | ||
3631 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3632 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3633 | { | |
3634 | struct rq *rq = data; | |
3635 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3636 | ||
3637 | cfs_rq->throttle_count--; | |
3638 | #ifdef CONFIG_SMP | |
3639 | if (!cfs_rq->throttle_count) { | |
f1b17280 | 3640 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3641 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3642 | cfs_rq->throttled_clock_task; |
64660c86 PT |
3643 | } |
3644 | #endif | |
3645 | ||
3646 | return 0; | |
3647 | } | |
3648 | ||
3649 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3650 | { | |
3651 | struct rq *rq = data; | |
3652 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3653 | ||
82958366 PT |
3654 | /* group is entering throttled state, stop time */ |
3655 | if (!cfs_rq->throttle_count) | |
78becc27 | 3656 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3657 | cfs_rq->throttle_count++; |
3658 | ||
3659 | return 0; | |
3660 | } | |
3661 | ||
d3d9dc33 | 3662 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3663 | { |
3664 | struct rq *rq = rq_of(cfs_rq); | |
3665 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3666 | struct sched_entity *se; | |
3667 | long task_delta, dequeue = 1; | |
77a4d1a1 | 3668 | bool empty; |
85dac906 PT |
3669 | |
3670 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3671 | ||
f1b17280 | 3672 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3673 | rcu_read_lock(); |
3674 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3675 | rcu_read_unlock(); | |
85dac906 PT |
3676 | |
3677 | task_delta = cfs_rq->h_nr_running; | |
3678 | for_each_sched_entity(se) { | |
3679 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3680 | /* throttled entity or throttle-on-deactivate */ | |
3681 | if (!se->on_rq) | |
3682 | break; | |
3683 | ||
3684 | if (dequeue) | |
3685 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3686 | qcfs_rq->h_nr_running -= task_delta; | |
3687 | ||
3688 | if (qcfs_rq->load.weight) | |
3689 | dequeue = 0; | |
3690 | } | |
3691 | ||
3692 | if (!se) | |
72465447 | 3693 | sub_nr_running(rq, task_delta); |
85dac906 PT |
3694 | |
3695 | cfs_rq->throttled = 1; | |
78becc27 | 3696 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 3697 | raw_spin_lock(&cfs_b->lock); |
d49db342 | 3698 | empty = list_empty(&cfs_b->throttled_cfs_rq); |
77a4d1a1 | 3699 | |
c06f04c7 BS |
3700 | /* |
3701 | * Add to the _head_ of the list, so that an already-started | |
3702 | * distribute_cfs_runtime will not see us | |
3703 | */ | |
3704 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
77a4d1a1 PZ |
3705 | |
3706 | /* | |
3707 | * If we're the first throttled task, make sure the bandwidth | |
3708 | * timer is running. | |
3709 | */ | |
3710 | if (empty) | |
3711 | start_cfs_bandwidth(cfs_b); | |
3712 | ||
85dac906 PT |
3713 | raw_spin_unlock(&cfs_b->lock); |
3714 | } | |
3715 | ||
029632fb | 3716 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3717 | { |
3718 | struct rq *rq = rq_of(cfs_rq); | |
3719 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3720 | struct sched_entity *se; | |
3721 | int enqueue = 1; | |
3722 | long task_delta; | |
3723 | ||
22b958d8 | 3724 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3725 | |
3726 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3727 | |
3728 | update_rq_clock(rq); | |
3729 | ||
671fd9da | 3730 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3731 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3732 | list_del_rcu(&cfs_rq->throttled_list); |
3733 | raw_spin_unlock(&cfs_b->lock); | |
3734 | ||
64660c86 PT |
3735 | /* update hierarchical throttle state */ |
3736 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3737 | ||
671fd9da PT |
3738 | if (!cfs_rq->load.weight) |
3739 | return; | |
3740 | ||
3741 | task_delta = cfs_rq->h_nr_running; | |
3742 | for_each_sched_entity(se) { | |
3743 | if (se->on_rq) | |
3744 | enqueue = 0; | |
3745 | ||
3746 | cfs_rq = cfs_rq_of(se); | |
3747 | if (enqueue) | |
3748 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3749 | cfs_rq->h_nr_running += task_delta; | |
3750 | ||
3751 | if (cfs_rq_throttled(cfs_rq)) | |
3752 | break; | |
3753 | } | |
3754 | ||
3755 | if (!se) | |
72465447 | 3756 | add_nr_running(rq, task_delta); |
671fd9da PT |
3757 | |
3758 | /* determine whether we need to wake up potentially idle cpu */ | |
3759 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
8875125e | 3760 | resched_curr(rq); |
671fd9da PT |
3761 | } |
3762 | ||
3763 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
3764 | u64 remaining, u64 expires) | |
3765 | { | |
3766 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
3767 | u64 runtime; |
3768 | u64 starting_runtime = remaining; | |
671fd9da PT |
3769 | |
3770 | rcu_read_lock(); | |
3771 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
3772 | throttled_list) { | |
3773 | struct rq *rq = rq_of(cfs_rq); | |
3774 | ||
3775 | raw_spin_lock(&rq->lock); | |
3776 | if (!cfs_rq_throttled(cfs_rq)) | |
3777 | goto next; | |
3778 | ||
3779 | runtime = -cfs_rq->runtime_remaining + 1; | |
3780 | if (runtime > remaining) | |
3781 | runtime = remaining; | |
3782 | remaining -= runtime; | |
3783 | ||
3784 | cfs_rq->runtime_remaining += runtime; | |
3785 | cfs_rq->runtime_expires = expires; | |
3786 | ||
3787 | /* we check whether we're throttled above */ | |
3788 | if (cfs_rq->runtime_remaining > 0) | |
3789 | unthrottle_cfs_rq(cfs_rq); | |
3790 | ||
3791 | next: | |
3792 | raw_spin_unlock(&rq->lock); | |
3793 | ||
3794 | if (!remaining) | |
3795 | break; | |
3796 | } | |
3797 | rcu_read_unlock(); | |
3798 | ||
c06f04c7 | 3799 | return starting_runtime - remaining; |
671fd9da PT |
3800 | } |
3801 | ||
58088ad0 PT |
3802 | /* |
3803 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
3804 | * cfs_rqs as appropriate. If there has been no activity within the last | |
3805 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
3806 | * used to track this state. | |
3807 | */ | |
3808 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
3809 | { | |
671fd9da | 3810 | u64 runtime, runtime_expires; |
51f2176d | 3811 | int throttled; |
58088ad0 | 3812 | |
58088ad0 PT |
3813 | /* no need to continue the timer with no bandwidth constraint */ |
3814 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 3815 | goto out_deactivate; |
58088ad0 | 3816 | |
671fd9da | 3817 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 3818 | cfs_b->nr_periods += overrun; |
671fd9da | 3819 | |
51f2176d BS |
3820 | /* |
3821 | * idle depends on !throttled (for the case of a large deficit), and if | |
3822 | * we're going inactive then everything else can be deferred | |
3823 | */ | |
3824 | if (cfs_b->idle && !throttled) | |
3825 | goto out_deactivate; | |
a9cf55b2 PT |
3826 | |
3827 | __refill_cfs_bandwidth_runtime(cfs_b); | |
3828 | ||
671fd9da PT |
3829 | if (!throttled) { |
3830 | /* mark as potentially idle for the upcoming period */ | |
3831 | cfs_b->idle = 1; | |
51f2176d | 3832 | return 0; |
671fd9da PT |
3833 | } |
3834 | ||
e8da1b18 NR |
3835 | /* account preceding periods in which throttling occurred */ |
3836 | cfs_b->nr_throttled += overrun; | |
3837 | ||
671fd9da | 3838 | runtime_expires = cfs_b->runtime_expires; |
671fd9da PT |
3839 | |
3840 | /* | |
c06f04c7 BS |
3841 | * This check is repeated as we are holding onto the new bandwidth while |
3842 | * we unthrottle. This can potentially race with an unthrottled group | |
3843 | * trying to acquire new bandwidth from the global pool. This can result | |
3844 | * in us over-using our runtime if it is all used during this loop, but | |
3845 | * only by limited amounts in that extreme case. | |
671fd9da | 3846 | */ |
c06f04c7 BS |
3847 | while (throttled && cfs_b->runtime > 0) { |
3848 | runtime = cfs_b->runtime; | |
671fd9da PT |
3849 | raw_spin_unlock(&cfs_b->lock); |
3850 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
3851 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
3852 | runtime_expires); | |
3853 | raw_spin_lock(&cfs_b->lock); | |
3854 | ||
3855 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
c06f04c7 BS |
3856 | |
3857 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | |
671fd9da | 3858 | } |
58088ad0 | 3859 | |
671fd9da PT |
3860 | /* |
3861 | * While we are ensured activity in the period following an | |
3862 | * unthrottle, this also covers the case in which the new bandwidth is | |
3863 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
3864 | * timer to remain active while there are any throttled entities.) | |
3865 | */ | |
3866 | cfs_b->idle = 0; | |
58088ad0 | 3867 | |
51f2176d BS |
3868 | return 0; |
3869 | ||
3870 | out_deactivate: | |
51f2176d | 3871 | return 1; |
58088ad0 | 3872 | } |
d3d9dc33 | 3873 | |
d8b4986d PT |
3874 | /* a cfs_rq won't donate quota below this amount */ |
3875 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
3876 | /* minimum remaining period time to redistribute slack quota */ | |
3877 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
3878 | /* how long we wait to gather additional slack before distributing */ | |
3879 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
3880 | ||
db06e78c BS |
3881 | /* |
3882 | * Are we near the end of the current quota period? | |
3883 | * | |
3884 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 3885 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
3886 | * migrate_hrtimers, base is never cleared, so we are fine. |
3887 | */ | |
d8b4986d PT |
3888 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
3889 | { | |
3890 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
3891 | u64 remaining; | |
3892 | ||
3893 | /* if the call-back is running a quota refresh is already occurring */ | |
3894 | if (hrtimer_callback_running(refresh_timer)) | |
3895 | return 1; | |
3896 | ||
3897 | /* is a quota refresh about to occur? */ | |
3898 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
3899 | if (remaining < min_expire) | |
3900 | return 1; | |
3901 | ||
3902 | return 0; | |
3903 | } | |
3904 | ||
3905 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
3906 | { | |
3907 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
3908 | ||
3909 | /* if there's a quota refresh soon don't bother with slack */ | |
3910 | if (runtime_refresh_within(cfs_b, min_left)) | |
3911 | return; | |
3912 | ||
4cfafd30 PZ |
3913 | hrtimer_start(&cfs_b->slack_timer, |
3914 | ns_to_ktime(cfs_bandwidth_slack_period), | |
3915 | HRTIMER_MODE_REL); | |
d8b4986d PT |
3916 | } |
3917 | ||
3918 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
3919 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3920 | { | |
3921 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3922 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
3923 | ||
3924 | if (slack_runtime <= 0) | |
3925 | return; | |
3926 | ||
3927 | raw_spin_lock(&cfs_b->lock); | |
3928 | if (cfs_b->quota != RUNTIME_INF && | |
3929 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
3930 | cfs_b->runtime += slack_runtime; | |
3931 | ||
3932 | /* we are under rq->lock, defer unthrottling using a timer */ | |
3933 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
3934 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
3935 | start_cfs_slack_bandwidth(cfs_b); | |
3936 | } | |
3937 | raw_spin_unlock(&cfs_b->lock); | |
3938 | ||
3939 | /* even if it's not valid for return we don't want to try again */ | |
3940 | cfs_rq->runtime_remaining -= slack_runtime; | |
3941 | } | |
3942 | ||
3943 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
3944 | { | |
56f570e5 PT |
3945 | if (!cfs_bandwidth_used()) |
3946 | return; | |
3947 | ||
fccfdc6f | 3948 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
3949 | return; |
3950 | ||
3951 | __return_cfs_rq_runtime(cfs_rq); | |
3952 | } | |
3953 | ||
3954 | /* | |
3955 | * This is done with a timer (instead of inline with bandwidth return) since | |
3956 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
3957 | */ | |
3958 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
3959 | { | |
3960 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
3961 | u64 expires; | |
3962 | ||
3963 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
3964 | raw_spin_lock(&cfs_b->lock); |
3965 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
3966 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 3967 | return; |
db06e78c | 3968 | } |
d8b4986d | 3969 | |
c06f04c7 | 3970 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 3971 | runtime = cfs_b->runtime; |
c06f04c7 | 3972 | |
d8b4986d PT |
3973 | expires = cfs_b->runtime_expires; |
3974 | raw_spin_unlock(&cfs_b->lock); | |
3975 | ||
3976 | if (!runtime) | |
3977 | return; | |
3978 | ||
3979 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
3980 | ||
3981 | raw_spin_lock(&cfs_b->lock); | |
3982 | if (expires == cfs_b->runtime_expires) | |
c06f04c7 | 3983 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
d8b4986d PT |
3984 | raw_spin_unlock(&cfs_b->lock); |
3985 | } | |
3986 | ||
d3d9dc33 PT |
3987 | /* |
3988 | * When a group wakes up we want to make sure that its quota is not already | |
3989 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
3990 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
3991 | */ | |
3992 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
3993 | { | |
56f570e5 PT |
3994 | if (!cfs_bandwidth_used()) |
3995 | return; | |
3996 | ||
d3d9dc33 PT |
3997 | /* an active group must be handled by the update_curr()->put() path */ |
3998 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
3999 | return; | |
4000 | ||
4001 | /* ensure the group is not already throttled */ | |
4002 | if (cfs_rq_throttled(cfs_rq)) | |
4003 | return; | |
4004 | ||
4005 | /* update runtime allocation */ | |
4006 | account_cfs_rq_runtime(cfs_rq, 0); | |
4007 | if (cfs_rq->runtime_remaining <= 0) | |
4008 | throttle_cfs_rq(cfs_rq); | |
4009 | } | |
4010 | ||
4011 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | |
678d5718 | 4012 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 4013 | { |
56f570e5 | 4014 | if (!cfs_bandwidth_used()) |
678d5718 | 4015 | return false; |
56f570e5 | 4016 | |
d3d9dc33 | 4017 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 4018 | return false; |
d3d9dc33 PT |
4019 | |
4020 | /* | |
4021 | * it's possible for a throttled entity to be forced into a running | |
4022 | * state (e.g. set_curr_task), in this case we're finished. | |
4023 | */ | |
4024 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 4025 | return true; |
d3d9dc33 PT |
4026 | |
4027 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 4028 | return true; |
d3d9dc33 | 4029 | } |
029632fb | 4030 | |
029632fb PZ |
4031 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
4032 | { | |
4033 | struct cfs_bandwidth *cfs_b = | |
4034 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 4035 | |
029632fb PZ |
4036 | do_sched_cfs_slack_timer(cfs_b); |
4037 | ||
4038 | return HRTIMER_NORESTART; | |
4039 | } | |
4040 | ||
4041 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
4042 | { | |
4043 | struct cfs_bandwidth *cfs_b = | |
4044 | container_of(timer, struct cfs_bandwidth, period_timer); | |
029632fb PZ |
4045 | int overrun; |
4046 | int idle = 0; | |
4047 | ||
51f2176d | 4048 | raw_spin_lock(&cfs_b->lock); |
029632fb | 4049 | for (;;) { |
77a4d1a1 | 4050 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
4051 | if (!overrun) |
4052 | break; | |
4053 | ||
4054 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
4055 | } | |
4cfafd30 PZ |
4056 | if (idle) |
4057 | cfs_b->period_active = 0; | |
51f2176d | 4058 | raw_spin_unlock(&cfs_b->lock); |
029632fb PZ |
4059 | |
4060 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
4061 | } | |
4062 | ||
4063 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4064 | { | |
4065 | raw_spin_lock_init(&cfs_b->lock); | |
4066 | cfs_b->runtime = 0; | |
4067 | cfs_b->quota = RUNTIME_INF; | |
4068 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
4069 | ||
4070 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 4071 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
4072 | cfs_b->period_timer.function = sched_cfs_period_timer; |
4073 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
4074 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
4075 | } | |
4076 | ||
4077 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4078 | { | |
4079 | cfs_rq->runtime_enabled = 0; | |
4080 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
4081 | } | |
4082 | ||
77a4d1a1 | 4083 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 4084 | { |
4cfafd30 | 4085 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 4086 | |
4cfafd30 PZ |
4087 | if (!cfs_b->period_active) { |
4088 | cfs_b->period_active = 1; | |
4089 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); | |
4090 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); | |
4091 | } | |
029632fb PZ |
4092 | } |
4093 | ||
4094 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4095 | { | |
7f1a169b TH |
4096 | /* init_cfs_bandwidth() was not called */ |
4097 | if (!cfs_b->throttled_cfs_rq.next) | |
4098 | return; | |
4099 | ||
029632fb PZ |
4100 | hrtimer_cancel(&cfs_b->period_timer); |
4101 | hrtimer_cancel(&cfs_b->slack_timer); | |
4102 | } | |
4103 | ||
0e59bdae KT |
4104 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
4105 | { | |
4106 | struct cfs_rq *cfs_rq; | |
4107 | ||
4108 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
4109 | struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; | |
4110 | ||
4111 | raw_spin_lock(&cfs_b->lock); | |
4112 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
4113 | raw_spin_unlock(&cfs_b->lock); | |
4114 | } | |
4115 | } | |
4116 | ||
38dc3348 | 4117 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
4118 | { |
4119 | struct cfs_rq *cfs_rq; | |
4120 | ||
4121 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
029632fb PZ |
4122 | if (!cfs_rq->runtime_enabled) |
4123 | continue; | |
4124 | ||
4125 | /* | |
4126 | * clock_task is not advancing so we just need to make sure | |
4127 | * there's some valid quota amount | |
4128 | */ | |
51f2176d | 4129 | cfs_rq->runtime_remaining = 1; |
0e59bdae KT |
4130 | /* |
4131 | * Offline rq is schedulable till cpu is completely disabled | |
4132 | * in take_cpu_down(), so we prevent new cfs throttling here. | |
4133 | */ | |
4134 | cfs_rq->runtime_enabled = 0; | |
4135 | ||
029632fb PZ |
4136 | if (cfs_rq_throttled(cfs_rq)) |
4137 | unthrottle_cfs_rq(cfs_rq); | |
4138 | } | |
4139 | } | |
4140 | ||
4141 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
4142 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
4143 | { | |
78becc27 | 4144 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
4145 | } |
4146 | ||
9dbdb155 | 4147 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 4148 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 4149 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
6c16a6dc | 4150 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
4151 | |
4152 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
4153 | { | |
4154 | return 0; | |
4155 | } | |
64660c86 PT |
4156 | |
4157 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4158 | { | |
4159 | return 0; | |
4160 | } | |
4161 | ||
4162 | static inline int throttled_lb_pair(struct task_group *tg, | |
4163 | int src_cpu, int dest_cpu) | |
4164 | { | |
4165 | return 0; | |
4166 | } | |
029632fb PZ |
4167 | |
4168 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
4169 | ||
4170 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
4171 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
4172 | #endif |
4173 | ||
029632fb PZ |
4174 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4175 | { | |
4176 | return NULL; | |
4177 | } | |
4178 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 4179 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 4180 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
4181 | |
4182 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
4183 | ||
bf0f6f24 IM |
4184 | /************************************************** |
4185 | * CFS operations on tasks: | |
4186 | */ | |
4187 | ||
8f4d37ec PZ |
4188 | #ifdef CONFIG_SCHED_HRTICK |
4189 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4190 | { | |
8f4d37ec PZ |
4191 | struct sched_entity *se = &p->se; |
4192 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4193 | ||
4194 | WARN_ON(task_rq(p) != rq); | |
4195 | ||
b39e66ea | 4196 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
4197 | u64 slice = sched_slice(cfs_rq, se); |
4198 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
4199 | s64 delta = slice - ran; | |
4200 | ||
4201 | if (delta < 0) { | |
4202 | if (rq->curr == p) | |
8875125e | 4203 | resched_curr(rq); |
8f4d37ec PZ |
4204 | return; |
4205 | } | |
31656519 | 4206 | hrtick_start(rq, delta); |
8f4d37ec PZ |
4207 | } |
4208 | } | |
a4c2f00f PZ |
4209 | |
4210 | /* | |
4211 | * called from enqueue/dequeue and updates the hrtick when the | |
4212 | * current task is from our class and nr_running is low enough | |
4213 | * to matter. | |
4214 | */ | |
4215 | static void hrtick_update(struct rq *rq) | |
4216 | { | |
4217 | struct task_struct *curr = rq->curr; | |
4218 | ||
b39e66ea | 4219 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
4220 | return; |
4221 | ||
4222 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
4223 | hrtick_start_fair(rq, curr); | |
4224 | } | |
55e12e5e | 4225 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
4226 | static inline void |
4227 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4228 | { | |
4229 | } | |
a4c2f00f PZ |
4230 | |
4231 | static inline void hrtick_update(struct rq *rq) | |
4232 | { | |
4233 | } | |
8f4d37ec PZ |
4234 | #endif |
4235 | ||
bf0f6f24 IM |
4236 | /* |
4237 | * The enqueue_task method is called before nr_running is | |
4238 | * increased. Here we update the fair scheduling stats and | |
4239 | * then put the task into the rbtree: | |
4240 | */ | |
ea87bb78 | 4241 | static void |
371fd7e7 | 4242 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4243 | { |
4244 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4245 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
4246 | |
4247 | for_each_sched_entity(se) { | |
62fb1851 | 4248 | if (se->on_rq) |
bf0f6f24 IM |
4249 | break; |
4250 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 4251 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
4252 | |
4253 | /* | |
4254 | * end evaluation on encountering a throttled cfs_rq | |
4255 | * | |
4256 | * note: in the case of encountering a throttled cfs_rq we will | |
4257 | * post the final h_nr_running increment below. | |
4258 | */ | |
4259 | if (cfs_rq_throttled(cfs_rq)) | |
4260 | break; | |
953bfcd1 | 4261 | cfs_rq->h_nr_running++; |
85dac906 | 4262 | |
88ec22d3 | 4263 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 4264 | } |
8f4d37ec | 4265 | |
2069dd75 | 4266 | for_each_sched_entity(se) { |
0f317143 | 4267 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4268 | cfs_rq->h_nr_running++; |
2069dd75 | 4269 | |
85dac906 PT |
4270 | if (cfs_rq_throttled(cfs_rq)) |
4271 | break; | |
4272 | ||
9d89c257 | 4273 | update_load_avg(se, 1); |
17bc14b7 | 4274 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
4275 | } |
4276 | ||
cd126afe | 4277 | if (!se) |
72465447 | 4278 | add_nr_running(rq, 1); |
cd126afe | 4279 | |
a4c2f00f | 4280 | hrtick_update(rq); |
bf0f6f24 IM |
4281 | } |
4282 | ||
2f36825b VP |
4283 | static void set_next_buddy(struct sched_entity *se); |
4284 | ||
bf0f6f24 IM |
4285 | /* |
4286 | * The dequeue_task method is called before nr_running is | |
4287 | * decreased. We remove the task from the rbtree and | |
4288 | * update the fair scheduling stats: | |
4289 | */ | |
371fd7e7 | 4290 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4291 | { |
4292 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4293 | struct sched_entity *se = &p->se; |
2f36825b | 4294 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
4295 | |
4296 | for_each_sched_entity(se) { | |
4297 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 4298 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
4299 | |
4300 | /* | |
4301 | * end evaluation on encountering a throttled cfs_rq | |
4302 | * | |
4303 | * note: in the case of encountering a throttled cfs_rq we will | |
4304 | * post the final h_nr_running decrement below. | |
4305 | */ | |
4306 | if (cfs_rq_throttled(cfs_rq)) | |
4307 | break; | |
953bfcd1 | 4308 | cfs_rq->h_nr_running--; |
2069dd75 | 4309 | |
bf0f6f24 | 4310 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b VP |
4311 | if (cfs_rq->load.weight) { |
4312 | /* | |
4313 | * Bias pick_next to pick a task from this cfs_rq, as | |
4314 | * p is sleeping when it is within its sched_slice. | |
4315 | */ | |
4316 | if (task_sleep && parent_entity(se)) | |
4317 | set_next_buddy(parent_entity(se)); | |
9598c82d PT |
4318 | |
4319 | /* avoid re-evaluating load for this entity */ | |
4320 | se = parent_entity(se); | |
bf0f6f24 | 4321 | break; |
2f36825b | 4322 | } |
371fd7e7 | 4323 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 4324 | } |
8f4d37ec | 4325 | |
2069dd75 | 4326 | for_each_sched_entity(se) { |
0f317143 | 4327 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4328 | cfs_rq->h_nr_running--; |
2069dd75 | 4329 | |
85dac906 PT |
4330 | if (cfs_rq_throttled(cfs_rq)) |
4331 | break; | |
4332 | ||
9d89c257 | 4333 | update_load_avg(se, 1); |
17bc14b7 | 4334 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
4335 | } |
4336 | ||
cd126afe | 4337 | if (!se) |
72465447 | 4338 | sub_nr_running(rq, 1); |
cd126afe | 4339 | |
a4c2f00f | 4340 | hrtick_update(rq); |
bf0f6f24 IM |
4341 | } |
4342 | ||
e7693a36 | 4343 | #ifdef CONFIG_SMP |
3289bdb4 PZ |
4344 | |
4345 | /* | |
4346 | * per rq 'load' arrray crap; XXX kill this. | |
4347 | */ | |
4348 | ||
4349 | /* | |
d937cdc5 | 4350 | * The exact cpuload calculated at every tick would be: |
3289bdb4 | 4351 | * |
d937cdc5 PZ |
4352 | * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load |
4353 | * | |
4354 | * If a cpu misses updates for n ticks (as it was idle) and update gets | |
4355 | * called on the n+1-th tick when cpu may be busy, then we have: | |
4356 | * | |
4357 | * load_n = (1 - 1/2^i)^n * load_0 | |
4358 | * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load | |
3289bdb4 PZ |
4359 | * |
4360 | * decay_load_missed() below does efficient calculation of | |
3289bdb4 | 4361 | * |
d937cdc5 PZ |
4362 | * load' = (1 - 1/2^i)^n * load |
4363 | * | |
4364 | * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors. | |
4365 | * This allows us to precompute the above in said factors, thereby allowing the | |
4366 | * reduction of an arbitrary n in O(log_2 n) steps. (See also | |
4367 | * fixed_power_int()) | |
3289bdb4 | 4368 | * |
d937cdc5 | 4369 | * The calculation is approximated on a 128 point scale. |
3289bdb4 PZ |
4370 | */ |
4371 | #define DEGRADE_SHIFT 7 | |
d937cdc5 PZ |
4372 | |
4373 | static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; | |
4374 | static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { | |
4375 | { 0, 0, 0, 0, 0, 0, 0, 0 }, | |
4376 | { 64, 32, 8, 0, 0, 0, 0, 0 }, | |
4377 | { 96, 72, 40, 12, 1, 0, 0, 0 }, | |
4378 | { 112, 98, 75, 43, 15, 1, 0, 0 }, | |
4379 | { 120, 112, 98, 76, 45, 16, 2, 0 } | |
4380 | }; | |
3289bdb4 PZ |
4381 | |
4382 | /* | |
4383 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog | |
4384 | * would be when CPU is idle and so we just decay the old load without | |
4385 | * adding any new load. | |
4386 | */ | |
4387 | static unsigned long | |
4388 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) | |
4389 | { | |
4390 | int j = 0; | |
4391 | ||
4392 | if (!missed_updates) | |
4393 | return load; | |
4394 | ||
4395 | if (missed_updates >= degrade_zero_ticks[idx]) | |
4396 | return 0; | |
4397 | ||
4398 | if (idx == 1) | |
4399 | return load >> missed_updates; | |
4400 | ||
4401 | while (missed_updates) { | |
4402 | if (missed_updates % 2) | |
4403 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; | |
4404 | ||
4405 | missed_updates >>= 1; | |
4406 | j++; | |
4407 | } | |
4408 | return load; | |
4409 | } | |
4410 | ||
59543275 BP |
4411 | /** |
4412 | * __update_cpu_load - update the rq->cpu_load[] statistics | |
4413 | * @this_rq: The rq to update statistics for | |
4414 | * @this_load: The current load | |
4415 | * @pending_updates: The number of missed updates | |
4416 | * @active: !0 for NOHZ_FULL | |
4417 | * | |
3289bdb4 | 4418 | * Update rq->cpu_load[] statistics. This function is usually called every |
59543275 BP |
4419 | * scheduler tick (TICK_NSEC). |
4420 | * | |
4421 | * This function computes a decaying average: | |
4422 | * | |
4423 | * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load | |
4424 | * | |
4425 | * Because of NOHZ it might not get called on every tick which gives need for | |
4426 | * the @pending_updates argument. | |
4427 | * | |
4428 | * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1 | |
4429 | * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load | |
4430 | * = A * (A * load[i]_n-2 + B) + B | |
4431 | * = A * (A * (A * load[i]_n-3 + B) + B) + B | |
4432 | * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B | |
4433 | * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B | |
4434 | * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B | |
4435 | * = (1 - 1/2^i)^n * (load[i]_0 - load) + load | |
4436 | * | |
4437 | * In the above we've assumed load_n := load, which is true for NOHZ_FULL as | |
4438 | * any change in load would have resulted in the tick being turned back on. | |
4439 | * | |
4440 | * For regular NOHZ, this reduces to: | |
4441 | * | |
4442 | * load[i]_n = (1 - 1/2^i)^n * load[i]_0 | |
4443 | * | |
4444 | * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra | |
4445 | * term. See the @active paramter. | |
3289bdb4 PZ |
4446 | */ |
4447 | static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, | |
59543275 | 4448 | unsigned long pending_updates, int active) |
3289bdb4 | 4449 | { |
59543275 | 4450 | unsigned long tickless_load = active ? this_rq->cpu_load[0] : 0; |
3289bdb4 PZ |
4451 | int i, scale; |
4452 | ||
4453 | this_rq->nr_load_updates++; | |
4454 | ||
4455 | /* Update our load: */ | |
4456 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ | |
4457 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
4458 | unsigned long old_load, new_load; | |
4459 | ||
4460 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
4461 | ||
59543275 | 4462 | old_load = this_rq->cpu_load[i] - tickless_load; |
3289bdb4 | 4463 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
59543275 | 4464 | old_load += tickless_load; |
3289bdb4 PZ |
4465 | new_load = this_load; |
4466 | /* | |
4467 | * Round up the averaging division if load is increasing. This | |
4468 | * prevents us from getting stuck on 9 if the load is 10, for | |
4469 | * example. | |
4470 | */ | |
4471 | if (new_load > old_load) | |
4472 | new_load += scale - 1; | |
4473 | ||
4474 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; | |
4475 | } | |
4476 | ||
4477 | sched_avg_update(this_rq); | |
4478 | } | |
4479 | ||
7ea241af YD |
4480 | /* Used instead of source_load when we know the type == 0 */ |
4481 | static unsigned long weighted_cpuload(const int cpu) | |
4482 | { | |
4483 | return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs); | |
4484 | } | |
4485 | ||
3289bdb4 PZ |
4486 | #ifdef CONFIG_NO_HZ_COMMON |
4487 | /* | |
4488 | * There is no sane way to deal with nohz on smp when using jiffies because the | |
4489 | * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading | |
4490 | * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. | |
4491 | * | |
4492 | * Therefore we cannot use the delta approach from the regular tick since that | |
4493 | * would seriously skew the load calculation. However we'll make do for those | |
4494 | * updates happening while idle (nohz_idle_balance) or coming out of idle | |
4495 | * (tick_nohz_idle_exit). | |
4496 | * | |
4497 | * This means we might still be one tick off for nohz periods. | |
4498 | */ | |
4499 | ||
4500 | /* | |
4501 | * Called from nohz_idle_balance() to update the load ratings before doing the | |
4502 | * idle balance. | |
4503 | */ | |
4504 | static void update_idle_cpu_load(struct rq *this_rq) | |
4505 | { | |
316c1608 | 4506 | unsigned long curr_jiffies = READ_ONCE(jiffies); |
7ea241af | 4507 | unsigned long load = weighted_cpuload(cpu_of(this_rq)); |
3289bdb4 PZ |
4508 | unsigned long pending_updates; |
4509 | ||
4510 | /* | |
4511 | * bail if there's load or we're actually up-to-date. | |
4512 | */ | |
4513 | if (load || curr_jiffies == this_rq->last_load_update_tick) | |
4514 | return; | |
4515 | ||
4516 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
4517 | this_rq->last_load_update_tick = curr_jiffies; | |
4518 | ||
59543275 | 4519 | __update_cpu_load(this_rq, load, pending_updates, 0); |
3289bdb4 PZ |
4520 | } |
4521 | ||
4522 | /* | |
4523 | * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. | |
4524 | */ | |
525705d1 | 4525 | void update_cpu_load_nohz(int active) |
3289bdb4 PZ |
4526 | { |
4527 | struct rq *this_rq = this_rq(); | |
316c1608 | 4528 | unsigned long curr_jiffies = READ_ONCE(jiffies); |
525705d1 | 4529 | unsigned long load = active ? weighted_cpuload(cpu_of(this_rq)) : 0; |
3289bdb4 PZ |
4530 | unsigned long pending_updates; |
4531 | ||
4532 | if (curr_jiffies == this_rq->last_load_update_tick) | |
4533 | return; | |
4534 | ||
4535 | raw_spin_lock(&this_rq->lock); | |
4536 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
4537 | if (pending_updates) { | |
4538 | this_rq->last_load_update_tick = curr_jiffies; | |
4539 | /* | |
525705d1 BP |
4540 | * In the regular NOHZ case, we were idle, this means load 0. |
4541 | * In the NOHZ_FULL case, we were non-idle, we should consider | |
4542 | * its weighted load. | |
3289bdb4 | 4543 | */ |
525705d1 | 4544 | __update_cpu_load(this_rq, load, pending_updates, active); |
3289bdb4 PZ |
4545 | } |
4546 | raw_spin_unlock(&this_rq->lock); | |
4547 | } | |
4548 | #endif /* CONFIG_NO_HZ */ | |
4549 | ||
4550 | /* | |
4551 | * Called from scheduler_tick() | |
4552 | */ | |
4553 | void update_cpu_load_active(struct rq *this_rq) | |
4554 | { | |
7ea241af | 4555 | unsigned long load = weighted_cpuload(cpu_of(this_rq)); |
3289bdb4 PZ |
4556 | /* |
4557 | * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). | |
4558 | */ | |
4559 | this_rq->last_load_update_tick = jiffies; | |
59543275 | 4560 | __update_cpu_load(this_rq, load, 1, 1); |
3289bdb4 PZ |
4561 | } |
4562 | ||
029632fb PZ |
4563 | /* |
4564 | * Return a low guess at the load of a migration-source cpu weighted | |
4565 | * according to the scheduling class and "nice" value. | |
4566 | * | |
4567 | * We want to under-estimate the load of migration sources, to | |
4568 | * balance conservatively. | |
4569 | */ | |
4570 | static unsigned long source_load(int cpu, int type) | |
4571 | { | |
4572 | struct rq *rq = cpu_rq(cpu); | |
4573 | unsigned long total = weighted_cpuload(cpu); | |
4574 | ||
4575 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4576 | return total; | |
4577 | ||
4578 | return min(rq->cpu_load[type-1], total); | |
4579 | } | |
4580 | ||
4581 | /* | |
4582 | * Return a high guess at the load of a migration-target cpu weighted | |
4583 | * according to the scheduling class and "nice" value. | |
4584 | */ | |
4585 | static unsigned long target_load(int cpu, int type) | |
4586 | { | |
4587 | struct rq *rq = cpu_rq(cpu); | |
4588 | unsigned long total = weighted_cpuload(cpu); | |
4589 | ||
4590 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4591 | return total; | |
4592 | ||
4593 | return max(rq->cpu_load[type-1], total); | |
4594 | } | |
4595 | ||
ced549fa | 4596 | static unsigned long capacity_of(int cpu) |
029632fb | 4597 | { |
ced549fa | 4598 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
4599 | } |
4600 | ||
ca6d75e6 VG |
4601 | static unsigned long capacity_orig_of(int cpu) |
4602 | { | |
4603 | return cpu_rq(cpu)->cpu_capacity_orig; | |
4604 | } | |
4605 | ||
029632fb PZ |
4606 | static unsigned long cpu_avg_load_per_task(int cpu) |
4607 | { | |
4608 | struct rq *rq = cpu_rq(cpu); | |
316c1608 | 4609 | unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); |
7ea241af | 4610 | unsigned long load_avg = weighted_cpuload(cpu); |
029632fb PZ |
4611 | |
4612 | if (nr_running) | |
b92486cb | 4613 | return load_avg / nr_running; |
029632fb PZ |
4614 | |
4615 | return 0; | |
4616 | } | |
4617 | ||
62470419 MW |
4618 | static void record_wakee(struct task_struct *p) |
4619 | { | |
4620 | /* | |
4621 | * Rough decay (wiping) for cost saving, don't worry | |
4622 | * about the boundary, really active task won't care | |
4623 | * about the loss. | |
4624 | */ | |
2538d960 | 4625 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { |
096aa338 | 4626 | current->wakee_flips >>= 1; |
62470419 MW |
4627 | current->wakee_flip_decay_ts = jiffies; |
4628 | } | |
4629 | ||
4630 | if (current->last_wakee != p) { | |
4631 | current->last_wakee = p; | |
4632 | current->wakee_flips++; | |
4633 | } | |
4634 | } | |
098fb9db | 4635 | |
74f8e4b2 | 4636 | static void task_waking_fair(struct task_struct *p) |
88ec22d3 PZ |
4637 | { |
4638 | struct sched_entity *se = &p->se; | |
4639 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3fe1698b PZ |
4640 | u64 min_vruntime; |
4641 | ||
4642 | #ifndef CONFIG_64BIT | |
4643 | u64 min_vruntime_copy; | |
88ec22d3 | 4644 | |
3fe1698b PZ |
4645 | do { |
4646 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
4647 | smp_rmb(); | |
4648 | min_vruntime = cfs_rq->min_vruntime; | |
4649 | } while (min_vruntime != min_vruntime_copy); | |
4650 | #else | |
4651 | min_vruntime = cfs_rq->min_vruntime; | |
4652 | #endif | |
88ec22d3 | 4653 | |
3fe1698b | 4654 | se->vruntime -= min_vruntime; |
62470419 | 4655 | record_wakee(p); |
88ec22d3 PZ |
4656 | } |
4657 | ||
bb3469ac | 4658 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
4659 | /* |
4660 | * effective_load() calculates the load change as seen from the root_task_group | |
4661 | * | |
4662 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
4663 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
4664 | * can calculate the shift in shares. | |
cf5f0acf PZ |
4665 | * |
4666 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
4667 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
4668 | * total group weight. | |
4669 | * | |
4670 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
4671 | * distribution (s_i) using: | |
4672 | * | |
4673 | * s_i = rw_i / \Sum rw_j (1) | |
4674 | * | |
4675 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
4676 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
4677 | * shares distribution (s_i): | |
4678 | * | |
4679 | * rw_i = { 2, 4, 1, 0 } | |
4680 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
4681 | * | |
4682 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
4683 | * task used to run on and the CPU the waker is running on), we need to | |
4684 | * compute the effect of waking a task on either CPU and, in case of a sync | |
4685 | * wakeup, compute the effect of the current task going to sleep. | |
4686 | * | |
4687 | * So for a change of @wl to the local @cpu with an overall group weight change | |
4688 | * of @wl we can compute the new shares distribution (s'_i) using: | |
4689 | * | |
4690 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
4691 | * | |
4692 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
4693 | * differences in waking a task to CPU 0. The additional task changes the | |
4694 | * weight and shares distributions like: | |
4695 | * | |
4696 | * rw'_i = { 3, 4, 1, 0 } | |
4697 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
4698 | * | |
4699 | * We can then compute the difference in effective weight by using: | |
4700 | * | |
4701 | * dw_i = S * (s'_i - s_i) (3) | |
4702 | * | |
4703 | * Where 'S' is the group weight as seen by its parent. | |
4704 | * | |
4705 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
4706 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
4707 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 4708 | */ |
2069dd75 | 4709 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 4710 | { |
4be9daaa | 4711 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 4712 | |
9722c2da | 4713 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
4714 | return wl; |
4715 | ||
4be9daaa | 4716 | for_each_sched_entity(se) { |
cf5f0acf | 4717 | long w, W; |
4be9daaa | 4718 | |
977dda7c | 4719 | tg = se->my_q->tg; |
bb3469ac | 4720 | |
cf5f0acf PZ |
4721 | /* |
4722 | * W = @wg + \Sum rw_j | |
4723 | */ | |
4724 | W = wg + calc_tg_weight(tg, se->my_q); | |
4be9daaa | 4725 | |
cf5f0acf PZ |
4726 | /* |
4727 | * w = rw_i + @wl | |
4728 | */ | |
7ea241af | 4729 | w = cfs_rq_load_avg(se->my_q) + wl; |
940959e9 | 4730 | |
cf5f0acf PZ |
4731 | /* |
4732 | * wl = S * s'_i; see (2) | |
4733 | */ | |
4734 | if (W > 0 && w < W) | |
32a8df4e | 4735 | wl = (w * (long)tg->shares) / W; |
977dda7c PT |
4736 | else |
4737 | wl = tg->shares; | |
940959e9 | 4738 | |
cf5f0acf PZ |
4739 | /* |
4740 | * Per the above, wl is the new se->load.weight value; since | |
4741 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
4742 | * calc_cfs_shares(). | |
4743 | */ | |
977dda7c PT |
4744 | if (wl < MIN_SHARES) |
4745 | wl = MIN_SHARES; | |
cf5f0acf PZ |
4746 | |
4747 | /* | |
4748 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
4749 | */ | |
9d89c257 | 4750 | wl -= se->avg.load_avg; |
cf5f0acf PZ |
4751 | |
4752 | /* | |
4753 | * Recursively apply this logic to all parent groups to compute | |
4754 | * the final effective load change on the root group. Since | |
4755 | * only the @tg group gets extra weight, all parent groups can | |
4756 | * only redistribute existing shares. @wl is the shift in shares | |
4757 | * resulting from this level per the above. | |
4758 | */ | |
4be9daaa | 4759 | wg = 0; |
4be9daaa | 4760 | } |
bb3469ac | 4761 | |
4be9daaa | 4762 | return wl; |
bb3469ac PZ |
4763 | } |
4764 | #else | |
4be9daaa | 4765 | |
58d081b5 | 4766 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 4767 | { |
83378269 | 4768 | return wl; |
bb3469ac | 4769 | } |
4be9daaa | 4770 | |
bb3469ac PZ |
4771 | #endif |
4772 | ||
63b0e9ed MG |
4773 | /* |
4774 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
4775 | * A waker of many should wake a different task than the one last awakened | |
4776 | * at a frequency roughly N times higher than one of its wakees. In order | |
4777 | * to determine whether we should let the load spread vs consolodating to | |
4778 | * shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
4779 | * partner, and a factor of lls_size higher frequency in the other. With | |
4780 | * both conditions met, we can be relatively sure that the relationship is | |
4781 | * non-monogamous, with partner count exceeding socket size. Waker/wakee | |
4782 | * being client/server, worker/dispatcher, interrupt source or whatever is | |
4783 | * irrelevant, spread criteria is apparent partner count exceeds socket size. | |
4784 | */ | |
62470419 MW |
4785 | static int wake_wide(struct task_struct *p) |
4786 | { | |
63b0e9ed MG |
4787 | unsigned int master = current->wakee_flips; |
4788 | unsigned int slave = p->wakee_flips; | |
7d9ffa89 | 4789 | int factor = this_cpu_read(sd_llc_size); |
62470419 | 4790 | |
63b0e9ed MG |
4791 | if (master < slave) |
4792 | swap(master, slave); | |
4793 | if (slave < factor || master < slave * factor) | |
4794 | return 0; | |
4795 | return 1; | |
62470419 MW |
4796 | } |
4797 | ||
c88d5910 | 4798 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
098fb9db | 4799 | { |
e37b6a7b | 4800 | s64 this_load, load; |
bd61c98f | 4801 | s64 this_eff_load, prev_eff_load; |
c88d5910 | 4802 | int idx, this_cpu, prev_cpu; |
c88d5910 | 4803 | struct task_group *tg; |
83378269 | 4804 | unsigned long weight; |
b3137bc8 | 4805 | int balanced; |
098fb9db | 4806 | |
c88d5910 PZ |
4807 | idx = sd->wake_idx; |
4808 | this_cpu = smp_processor_id(); | |
4809 | prev_cpu = task_cpu(p); | |
4810 | load = source_load(prev_cpu, idx); | |
4811 | this_load = target_load(this_cpu, idx); | |
098fb9db | 4812 | |
b3137bc8 MG |
4813 | /* |
4814 | * If sync wakeup then subtract the (maximum possible) | |
4815 | * effect of the currently running task from the load | |
4816 | * of the current CPU: | |
4817 | */ | |
83378269 PZ |
4818 | if (sync) { |
4819 | tg = task_group(current); | |
9d89c257 | 4820 | weight = current->se.avg.load_avg; |
83378269 | 4821 | |
c88d5910 | 4822 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
4823 | load += effective_load(tg, prev_cpu, 0, -weight); |
4824 | } | |
b3137bc8 | 4825 | |
83378269 | 4826 | tg = task_group(p); |
9d89c257 | 4827 | weight = p->se.avg.load_avg; |
b3137bc8 | 4828 | |
71a29aa7 PZ |
4829 | /* |
4830 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
4831 | * due to the sync cause above having dropped this_load to 0, we'll |
4832 | * always have an imbalance, but there's really nothing you can do | |
4833 | * about that, so that's good too. | |
71a29aa7 PZ |
4834 | * |
4835 | * Otherwise check if either cpus are near enough in load to allow this | |
4836 | * task to be woken on this_cpu. | |
4837 | */ | |
bd61c98f VG |
4838 | this_eff_load = 100; |
4839 | this_eff_load *= capacity_of(prev_cpu); | |
e51fd5e2 | 4840 | |
bd61c98f VG |
4841 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; |
4842 | prev_eff_load *= capacity_of(this_cpu); | |
e51fd5e2 | 4843 | |
bd61c98f | 4844 | if (this_load > 0) { |
e51fd5e2 PZ |
4845 | this_eff_load *= this_load + |
4846 | effective_load(tg, this_cpu, weight, weight); | |
4847 | ||
e51fd5e2 | 4848 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); |
bd61c98f | 4849 | } |
e51fd5e2 | 4850 | |
bd61c98f | 4851 | balanced = this_eff_load <= prev_eff_load; |
098fb9db | 4852 | |
41acab88 | 4853 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
098fb9db | 4854 | |
05bfb65f VG |
4855 | if (!balanced) |
4856 | return 0; | |
098fb9db | 4857 | |
05bfb65f VG |
4858 | schedstat_inc(sd, ttwu_move_affine); |
4859 | schedstat_inc(p, se.statistics.nr_wakeups_affine); | |
4860 | ||
4861 | return 1; | |
098fb9db IM |
4862 | } |
4863 | ||
aaee1203 PZ |
4864 | /* |
4865 | * find_idlest_group finds and returns the least busy CPU group within the | |
4866 | * domain. | |
4867 | */ | |
4868 | static struct sched_group * | |
78e7ed53 | 4869 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 4870 | int this_cpu, int sd_flag) |
e7693a36 | 4871 | { |
b3bd3de6 | 4872 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 4873 | unsigned long min_load = ULONG_MAX, this_load = 0; |
c44f2a02 | 4874 | int load_idx = sd->forkexec_idx; |
aaee1203 | 4875 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 4876 | |
c44f2a02 VG |
4877 | if (sd_flag & SD_BALANCE_WAKE) |
4878 | load_idx = sd->wake_idx; | |
4879 | ||
aaee1203 PZ |
4880 | do { |
4881 | unsigned long load, avg_load; | |
4882 | int local_group; | |
4883 | int i; | |
e7693a36 | 4884 | |
aaee1203 PZ |
4885 | /* Skip over this group if it has no CPUs allowed */ |
4886 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 4887 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
4888 | continue; |
4889 | ||
4890 | local_group = cpumask_test_cpu(this_cpu, | |
4891 | sched_group_cpus(group)); | |
4892 | ||
4893 | /* Tally up the load of all CPUs in the group */ | |
4894 | avg_load = 0; | |
4895 | ||
4896 | for_each_cpu(i, sched_group_cpus(group)) { | |
4897 | /* Bias balancing toward cpus of our domain */ | |
4898 | if (local_group) | |
4899 | load = source_load(i, load_idx); | |
4900 | else | |
4901 | load = target_load(i, load_idx); | |
4902 | ||
4903 | avg_load += load; | |
4904 | } | |
4905 | ||
63b2ca30 | 4906 | /* Adjust by relative CPU capacity of the group */ |
ca8ce3d0 | 4907 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; |
aaee1203 PZ |
4908 | |
4909 | if (local_group) { | |
4910 | this_load = avg_load; | |
aaee1203 PZ |
4911 | } else if (avg_load < min_load) { |
4912 | min_load = avg_load; | |
4913 | idlest = group; | |
4914 | } | |
4915 | } while (group = group->next, group != sd->groups); | |
4916 | ||
4917 | if (!idlest || 100*this_load < imbalance*min_load) | |
4918 | return NULL; | |
4919 | return idlest; | |
4920 | } | |
4921 | ||
4922 | /* | |
4923 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
4924 | */ | |
4925 | static int | |
4926 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
4927 | { | |
4928 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
4929 | unsigned int min_exit_latency = UINT_MAX; |
4930 | u64 latest_idle_timestamp = 0; | |
4931 | int least_loaded_cpu = this_cpu; | |
4932 | int shallowest_idle_cpu = -1; | |
aaee1203 PZ |
4933 | int i; |
4934 | ||
4935 | /* Traverse only the allowed CPUs */ | |
fa17b507 | 4936 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
83a0a96a NP |
4937 | if (idle_cpu(i)) { |
4938 | struct rq *rq = cpu_rq(i); | |
4939 | struct cpuidle_state *idle = idle_get_state(rq); | |
4940 | if (idle && idle->exit_latency < min_exit_latency) { | |
4941 | /* | |
4942 | * We give priority to a CPU whose idle state | |
4943 | * has the smallest exit latency irrespective | |
4944 | * of any idle timestamp. | |
4945 | */ | |
4946 | min_exit_latency = idle->exit_latency; | |
4947 | latest_idle_timestamp = rq->idle_stamp; | |
4948 | shallowest_idle_cpu = i; | |
4949 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
4950 | rq->idle_stamp > latest_idle_timestamp) { | |
4951 | /* | |
4952 | * If equal or no active idle state, then | |
4953 | * the most recently idled CPU might have | |
4954 | * a warmer cache. | |
4955 | */ | |
4956 | latest_idle_timestamp = rq->idle_stamp; | |
4957 | shallowest_idle_cpu = i; | |
4958 | } | |
9f96742a | 4959 | } else if (shallowest_idle_cpu == -1) { |
83a0a96a NP |
4960 | load = weighted_cpuload(i); |
4961 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
4962 | min_load = load; | |
4963 | least_loaded_cpu = i; | |
4964 | } | |
e7693a36 GH |
4965 | } |
4966 | } | |
4967 | ||
83a0a96a | 4968 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 4969 | } |
e7693a36 | 4970 | |
a50bde51 PZ |
4971 | /* |
4972 | * Try and locate an idle CPU in the sched_domain. | |
4973 | */ | |
99bd5e2f | 4974 | static int select_idle_sibling(struct task_struct *p, int target) |
a50bde51 | 4975 | { |
99bd5e2f | 4976 | struct sched_domain *sd; |
37407ea7 | 4977 | struct sched_group *sg; |
e0a79f52 | 4978 | int i = task_cpu(p); |
a50bde51 | 4979 | |
e0a79f52 MG |
4980 | if (idle_cpu(target)) |
4981 | return target; | |
99bd5e2f SS |
4982 | |
4983 | /* | |
e0a79f52 | 4984 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 4985 | */ |
e0a79f52 MG |
4986 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4987 | return i; | |
a50bde51 PZ |
4988 | |
4989 | /* | |
37407ea7 | 4990 | * Otherwise, iterate the domains and find an elegible idle cpu. |
a50bde51 | 4991 | */ |
518cd623 | 4992 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 4993 | for_each_lower_domain(sd) { |
37407ea7 LT |
4994 | sg = sd->groups; |
4995 | do { | |
4996 | if (!cpumask_intersects(sched_group_cpus(sg), | |
4997 | tsk_cpus_allowed(p))) | |
4998 | goto next; | |
4999 | ||
5000 | for_each_cpu(i, sched_group_cpus(sg)) { | |
e0a79f52 | 5001 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
5002 | goto next; |
5003 | } | |
970e1789 | 5004 | |
37407ea7 LT |
5005 | target = cpumask_first_and(sched_group_cpus(sg), |
5006 | tsk_cpus_allowed(p)); | |
5007 | goto done; | |
5008 | next: | |
5009 | sg = sg->next; | |
5010 | } while (sg != sd->groups); | |
5011 | } | |
5012 | done: | |
a50bde51 PZ |
5013 | return target; |
5014 | } | |
231678b7 | 5015 | |
8bb5b00c | 5016 | /* |
9e91d61d | 5017 | * cpu_util returns the amount of capacity of a CPU that is used by CFS |
8bb5b00c | 5018 | * tasks. The unit of the return value must be the one of capacity so we can |
9e91d61d DE |
5019 | * compare the utilization with the capacity of the CPU that is available for |
5020 | * CFS task (ie cpu_capacity). | |
231678b7 DE |
5021 | * |
5022 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
5023 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
5024 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
5025 | * capacity_orig is the cpu_capacity available at the highest frequency | |
5026 | * (arch_scale_freq_capacity()). | |
5027 | * The utilization of a CPU converges towards a sum equal to or less than the | |
5028 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
5029 | * the running time on this CPU scaled by capacity_curr. | |
5030 | * | |
5031 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even | |
5032 | * higher than capacity_orig because of unfortunate rounding in | |
5033 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
5034 | * the average stabilizes with the new running time. We need to check that the | |
5035 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
5036 | * necessary. Without utilization capping, a group could be seen as overloaded | |
5037 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
5038 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
5039 | * capacity_orig) as it useful for predicting the capacity required after task | |
5040 | * migrations (scheduler-driven DVFS). | |
8bb5b00c | 5041 | */ |
9e91d61d | 5042 | static int cpu_util(int cpu) |
8bb5b00c | 5043 | { |
9e91d61d | 5044 | unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg; |
8bb5b00c VG |
5045 | unsigned long capacity = capacity_orig_of(cpu); |
5046 | ||
231678b7 | 5047 | return (util >= capacity) ? capacity : util; |
8bb5b00c | 5048 | } |
a50bde51 | 5049 | |
aaee1203 | 5050 | /* |
de91b9cb MR |
5051 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
5052 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
5053 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 5054 | * |
de91b9cb MR |
5055 | * Balances load by selecting the idlest cpu in the idlest group, or under |
5056 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 5057 | * |
de91b9cb | 5058 | * Returns the target cpu number. |
aaee1203 PZ |
5059 | * |
5060 | * preempt must be disabled. | |
5061 | */ | |
0017d735 | 5062 | static int |
ac66f547 | 5063 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 5064 | { |
29cd8bae | 5065 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 5066 | int cpu = smp_processor_id(); |
63b0e9ed | 5067 | int new_cpu = prev_cpu; |
99bd5e2f | 5068 | int want_affine = 0; |
5158f4e4 | 5069 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 5070 | |
a8edd075 | 5071 | if (sd_flag & SD_BALANCE_WAKE) |
63b0e9ed | 5072 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); |
aaee1203 | 5073 | |
dce840a0 | 5074 | rcu_read_lock(); |
aaee1203 | 5075 | for_each_domain(cpu, tmp) { |
e4f42888 | 5076 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
63b0e9ed | 5077 | break; |
e4f42888 | 5078 | |
fe3bcfe1 | 5079 | /* |
99bd5e2f SS |
5080 | * If both cpu and prev_cpu are part of this domain, |
5081 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 5082 | */ |
99bd5e2f SS |
5083 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
5084 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
5085 | affine_sd = tmp; | |
29cd8bae | 5086 | break; |
f03542a7 | 5087 | } |
29cd8bae | 5088 | |
f03542a7 | 5089 | if (tmp->flags & sd_flag) |
29cd8bae | 5090 | sd = tmp; |
63b0e9ed MG |
5091 | else if (!want_affine) |
5092 | break; | |
29cd8bae PZ |
5093 | } |
5094 | ||
63b0e9ed MG |
5095 | if (affine_sd) { |
5096 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
5097 | if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) | |
5098 | new_cpu = cpu; | |
8b911acd | 5099 | } |
e7693a36 | 5100 | |
63b0e9ed MG |
5101 | if (!sd) { |
5102 | if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */ | |
5103 | new_cpu = select_idle_sibling(p, new_cpu); | |
5104 | ||
5105 | } else while (sd) { | |
aaee1203 | 5106 | struct sched_group *group; |
c88d5910 | 5107 | int weight; |
098fb9db | 5108 | |
0763a660 | 5109 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
5110 | sd = sd->child; |
5111 | continue; | |
5112 | } | |
098fb9db | 5113 | |
c44f2a02 | 5114 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
5115 | if (!group) { |
5116 | sd = sd->child; | |
5117 | continue; | |
5118 | } | |
4ae7d5ce | 5119 | |
d7c33c49 | 5120 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
5121 | if (new_cpu == -1 || new_cpu == cpu) { |
5122 | /* Now try balancing at a lower domain level of cpu */ | |
5123 | sd = sd->child; | |
5124 | continue; | |
e7693a36 | 5125 | } |
aaee1203 PZ |
5126 | |
5127 | /* Now try balancing at a lower domain level of new_cpu */ | |
5128 | cpu = new_cpu; | |
669c55e9 | 5129 | weight = sd->span_weight; |
aaee1203 PZ |
5130 | sd = NULL; |
5131 | for_each_domain(cpu, tmp) { | |
669c55e9 | 5132 | if (weight <= tmp->span_weight) |
aaee1203 | 5133 | break; |
0763a660 | 5134 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
5135 | sd = tmp; |
5136 | } | |
5137 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 5138 | } |
dce840a0 | 5139 | rcu_read_unlock(); |
e7693a36 | 5140 | |
c88d5910 | 5141 | return new_cpu; |
e7693a36 | 5142 | } |
0a74bef8 PT |
5143 | |
5144 | /* | |
5145 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
5146 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
525628c7 | 5147 | * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 5148 | */ |
5a4fd036 | 5149 | static void migrate_task_rq_fair(struct task_struct *p) |
0a74bef8 | 5150 | { |
aff3e498 | 5151 | /* |
9d89c257 YD |
5152 | * We are supposed to update the task to "current" time, then its up to date |
5153 | * and ready to go to new CPU/cfs_rq. But we have difficulty in getting | |
5154 | * what current time is, so simply throw away the out-of-date time. This | |
5155 | * will result in the wakee task is less decayed, but giving the wakee more | |
5156 | * load sounds not bad. | |
aff3e498 | 5157 | */ |
9d89c257 YD |
5158 | remove_entity_load_avg(&p->se); |
5159 | ||
5160 | /* Tell new CPU we are migrated */ | |
5161 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
5162 | |
5163 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 5164 | p->se.exec_start = 0; |
0a74bef8 | 5165 | } |
12695578 YD |
5166 | |
5167 | static void task_dead_fair(struct task_struct *p) | |
5168 | { | |
5169 | remove_entity_load_avg(&p->se); | |
5170 | } | |
e7693a36 GH |
5171 | #endif /* CONFIG_SMP */ |
5172 | ||
e52fb7c0 PZ |
5173 | static unsigned long |
5174 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
5175 | { |
5176 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
5177 | ||
5178 | /* | |
e52fb7c0 PZ |
5179 | * Since its curr running now, convert the gran from real-time |
5180 | * to virtual-time in his units. | |
13814d42 MG |
5181 | * |
5182 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
5183 | * they get preempted easier. That is, if 'se' < 'curr' then | |
5184 | * the resulting gran will be larger, therefore penalizing the | |
5185 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
5186 | * be smaller, again penalizing the lighter task. | |
5187 | * | |
5188 | * This is especially important for buddies when the leftmost | |
5189 | * task is higher priority than the buddy. | |
0bbd3336 | 5190 | */ |
f4ad9bd2 | 5191 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
5192 | } |
5193 | ||
464b7527 PZ |
5194 | /* |
5195 | * Should 'se' preempt 'curr'. | |
5196 | * | |
5197 | * |s1 | |
5198 | * |s2 | |
5199 | * |s3 | |
5200 | * g | |
5201 | * |<--->|c | |
5202 | * | |
5203 | * w(c, s1) = -1 | |
5204 | * w(c, s2) = 0 | |
5205 | * w(c, s3) = 1 | |
5206 | * | |
5207 | */ | |
5208 | static int | |
5209 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
5210 | { | |
5211 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
5212 | ||
5213 | if (vdiff <= 0) | |
5214 | return -1; | |
5215 | ||
e52fb7c0 | 5216 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
5217 | if (vdiff > gran) |
5218 | return 1; | |
5219 | ||
5220 | return 0; | |
5221 | } | |
5222 | ||
02479099 PZ |
5223 | static void set_last_buddy(struct sched_entity *se) |
5224 | { | |
69c80f3e VP |
5225 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
5226 | return; | |
5227 | ||
5228 | for_each_sched_entity(se) | |
5229 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
5230 | } |
5231 | ||
5232 | static void set_next_buddy(struct sched_entity *se) | |
5233 | { | |
69c80f3e VP |
5234 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
5235 | return; | |
5236 | ||
5237 | for_each_sched_entity(se) | |
5238 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
5239 | } |
5240 | ||
ac53db59 RR |
5241 | static void set_skip_buddy(struct sched_entity *se) |
5242 | { | |
69c80f3e VP |
5243 | for_each_sched_entity(se) |
5244 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
5245 | } |
5246 | ||
bf0f6f24 IM |
5247 | /* |
5248 | * Preempt the current task with a newly woken task if needed: | |
5249 | */ | |
5a9b86f6 | 5250 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
5251 | { |
5252 | struct task_struct *curr = rq->curr; | |
8651a86c | 5253 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 5254 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 5255 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 5256 | int next_buddy_marked = 0; |
bf0f6f24 | 5257 | |
4ae7d5ce IM |
5258 | if (unlikely(se == pse)) |
5259 | return; | |
5260 | ||
5238cdd3 | 5261 | /* |
163122b7 | 5262 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
5263 | * unconditionally check_prempt_curr() after an enqueue (which may have |
5264 | * lead to a throttle). This both saves work and prevents false | |
5265 | * next-buddy nomination below. | |
5266 | */ | |
5267 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
5268 | return; | |
5269 | ||
2f36825b | 5270 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 5271 | set_next_buddy(pse); |
2f36825b VP |
5272 | next_buddy_marked = 1; |
5273 | } | |
57fdc26d | 5274 | |
aec0a514 BR |
5275 | /* |
5276 | * We can come here with TIF_NEED_RESCHED already set from new task | |
5277 | * wake up path. | |
5238cdd3 PT |
5278 | * |
5279 | * Note: this also catches the edge-case of curr being in a throttled | |
5280 | * group (e.g. via set_curr_task), since update_curr() (in the | |
5281 | * enqueue of curr) will have resulted in resched being set. This | |
5282 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
5283 | * below. | |
aec0a514 BR |
5284 | */ |
5285 | if (test_tsk_need_resched(curr)) | |
5286 | return; | |
5287 | ||
a2f5c9ab DH |
5288 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
5289 | if (unlikely(curr->policy == SCHED_IDLE) && | |
5290 | likely(p->policy != SCHED_IDLE)) | |
5291 | goto preempt; | |
5292 | ||
91c234b4 | 5293 | /* |
a2f5c9ab DH |
5294 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
5295 | * is driven by the tick): | |
91c234b4 | 5296 | */ |
8ed92e51 | 5297 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 5298 | return; |
bf0f6f24 | 5299 | |
464b7527 | 5300 | find_matching_se(&se, &pse); |
9bbd7374 | 5301 | update_curr(cfs_rq_of(se)); |
002f128b | 5302 | BUG_ON(!pse); |
2f36825b VP |
5303 | if (wakeup_preempt_entity(se, pse) == 1) { |
5304 | /* | |
5305 | * Bias pick_next to pick the sched entity that is | |
5306 | * triggering this preemption. | |
5307 | */ | |
5308 | if (!next_buddy_marked) | |
5309 | set_next_buddy(pse); | |
3a7e73a2 | 5310 | goto preempt; |
2f36825b | 5311 | } |
464b7527 | 5312 | |
3a7e73a2 | 5313 | return; |
a65ac745 | 5314 | |
3a7e73a2 | 5315 | preempt: |
8875125e | 5316 | resched_curr(rq); |
3a7e73a2 PZ |
5317 | /* |
5318 | * Only set the backward buddy when the current task is still | |
5319 | * on the rq. This can happen when a wakeup gets interleaved | |
5320 | * with schedule on the ->pre_schedule() or idle_balance() | |
5321 | * point, either of which can * drop the rq lock. | |
5322 | * | |
5323 | * Also, during early boot the idle thread is in the fair class, | |
5324 | * for obvious reasons its a bad idea to schedule back to it. | |
5325 | */ | |
5326 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
5327 | return; | |
5328 | ||
5329 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
5330 | set_last_buddy(se); | |
bf0f6f24 IM |
5331 | } |
5332 | ||
606dba2e PZ |
5333 | static struct task_struct * |
5334 | pick_next_task_fair(struct rq *rq, struct task_struct *prev) | |
bf0f6f24 IM |
5335 | { |
5336 | struct cfs_rq *cfs_rq = &rq->cfs; | |
5337 | struct sched_entity *se; | |
678d5718 | 5338 | struct task_struct *p; |
37e117c0 | 5339 | int new_tasks; |
678d5718 | 5340 | |
6e83125c | 5341 | again: |
678d5718 PZ |
5342 | #ifdef CONFIG_FAIR_GROUP_SCHED |
5343 | if (!cfs_rq->nr_running) | |
38033c37 | 5344 | goto idle; |
678d5718 | 5345 | |
3f1d2a31 | 5346 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
5347 | goto simple; |
5348 | ||
5349 | /* | |
5350 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
5351 | * likely that a next task is from the same cgroup as the current. | |
5352 | * | |
5353 | * Therefore attempt to avoid putting and setting the entire cgroup | |
5354 | * hierarchy, only change the part that actually changes. | |
5355 | */ | |
5356 | ||
5357 | do { | |
5358 | struct sched_entity *curr = cfs_rq->curr; | |
5359 | ||
5360 | /* | |
5361 | * Since we got here without doing put_prev_entity() we also | |
5362 | * have to consider cfs_rq->curr. If it is still a runnable | |
5363 | * entity, update_curr() will update its vruntime, otherwise | |
5364 | * forget we've ever seen it. | |
5365 | */ | |
54d27365 BS |
5366 | if (curr) { |
5367 | if (curr->on_rq) | |
5368 | update_curr(cfs_rq); | |
5369 | else | |
5370 | curr = NULL; | |
678d5718 | 5371 | |
54d27365 BS |
5372 | /* |
5373 | * This call to check_cfs_rq_runtime() will do the | |
5374 | * throttle and dequeue its entity in the parent(s). | |
5375 | * Therefore the 'simple' nr_running test will indeed | |
5376 | * be correct. | |
5377 | */ | |
5378 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
5379 | goto simple; | |
5380 | } | |
678d5718 PZ |
5381 | |
5382 | se = pick_next_entity(cfs_rq, curr); | |
5383 | cfs_rq = group_cfs_rq(se); | |
5384 | } while (cfs_rq); | |
5385 | ||
5386 | p = task_of(se); | |
5387 | ||
5388 | /* | |
5389 | * Since we haven't yet done put_prev_entity and if the selected task | |
5390 | * is a different task than we started out with, try and touch the | |
5391 | * least amount of cfs_rqs. | |
5392 | */ | |
5393 | if (prev != p) { | |
5394 | struct sched_entity *pse = &prev->se; | |
5395 | ||
5396 | while (!(cfs_rq = is_same_group(se, pse))) { | |
5397 | int se_depth = se->depth; | |
5398 | int pse_depth = pse->depth; | |
5399 | ||
5400 | if (se_depth <= pse_depth) { | |
5401 | put_prev_entity(cfs_rq_of(pse), pse); | |
5402 | pse = parent_entity(pse); | |
5403 | } | |
5404 | if (se_depth >= pse_depth) { | |
5405 | set_next_entity(cfs_rq_of(se), se); | |
5406 | se = parent_entity(se); | |
5407 | } | |
5408 | } | |
5409 | ||
5410 | put_prev_entity(cfs_rq, pse); | |
5411 | set_next_entity(cfs_rq, se); | |
5412 | } | |
5413 | ||
5414 | if (hrtick_enabled(rq)) | |
5415 | hrtick_start_fair(rq, p); | |
5416 | ||
5417 | return p; | |
5418 | simple: | |
5419 | cfs_rq = &rq->cfs; | |
5420 | #endif | |
bf0f6f24 | 5421 | |
36ace27e | 5422 | if (!cfs_rq->nr_running) |
38033c37 | 5423 | goto idle; |
bf0f6f24 | 5424 | |
3f1d2a31 | 5425 | put_prev_task(rq, prev); |
606dba2e | 5426 | |
bf0f6f24 | 5427 | do { |
678d5718 | 5428 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 5429 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
5430 | cfs_rq = group_cfs_rq(se); |
5431 | } while (cfs_rq); | |
5432 | ||
8f4d37ec | 5433 | p = task_of(se); |
678d5718 | 5434 | |
b39e66ea MG |
5435 | if (hrtick_enabled(rq)) |
5436 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
5437 | |
5438 | return p; | |
38033c37 PZ |
5439 | |
5440 | idle: | |
cbce1a68 PZ |
5441 | /* |
5442 | * This is OK, because current is on_cpu, which avoids it being picked | |
5443 | * for load-balance and preemption/IRQs are still disabled avoiding | |
5444 | * further scheduler activity on it and we're being very careful to | |
5445 | * re-start the picking loop. | |
5446 | */ | |
5447 | lockdep_unpin_lock(&rq->lock); | |
e4aa358b | 5448 | new_tasks = idle_balance(rq); |
cbce1a68 | 5449 | lockdep_pin_lock(&rq->lock); |
37e117c0 PZ |
5450 | /* |
5451 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
5452 | * possible for any higher priority task to appear. In that case we | |
5453 | * must re-start the pick_next_entity() loop. | |
5454 | */ | |
e4aa358b | 5455 | if (new_tasks < 0) |
37e117c0 PZ |
5456 | return RETRY_TASK; |
5457 | ||
e4aa358b | 5458 | if (new_tasks > 0) |
38033c37 | 5459 | goto again; |
38033c37 PZ |
5460 | |
5461 | return NULL; | |
bf0f6f24 IM |
5462 | } |
5463 | ||
5464 | /* | |
5465 | * Account for a descheduled task: | |
5466 | */ | |
31ee529c | 5467 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
5468 | { |
5469 | struct sched_entity *se = &prev->se; | |
5470 | struct cfs_rq *cfs_rq; | |
5471 | ||
5472 | for_each_sched_entity(se) { | |
5473 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 5474 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
5475 | } |
5476 | } | |
5477 | ||
ac53db59 RR |
5478 | /* |
5479 | * sched_yield() is very simple | |
5480 | * | |
5481 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
5482 | */ | |
5483 | static void yield_task_fair(struct rq *rq) | |
5484 | { | |
5485 | struct task_struct *curr = rq->curr; | |
5486 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
5487 | struct sched_entity *se = &curr->se; | |
5488 | ||
5489 | /* | |
5490 | * Are we the only task in the tree? | |
5491 | */ | |
5492 | if (unlikely(rq->nr_running == 1)) | |
5493 | return; | |
5494 | ||
5495 | clear_buddies(cfs_rq, se); | |
5496 | ||
5497 | if (curr->policy != SCHED_BATCH) { | |
5498 | update_rq_clock(rq); | |
5499 | /* | |
5500 | * Update run-time statistics of the 'current'. | |
5501 | */ | |
5502 | update_curr(cfs_rq); | |
916671c0 MG |
5503 | /* |
5504 | * Tell update_rq_clock() that we've just updated, | |
5505 | * so we don't do microscopic update in schedule() | |
5506 | * and double the fastpath cost. | |
5507 | */ | |
9edfbfed | 5508 | rq_clock_skip_update(rq, true); |
ac53db59 RR |
5509 | } |
5510 | ||
5511 | set_skip_buddy(se); | |
5512 | } | |
5513 | ||
d95f4122 MG |
5514 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
5515 | { | |
5516 | struct sched_entity *se = &p->se; | |
5517 | ||
5238cdd3 PT |
5518 | /* throttled hierarchies are not runnable */ |
5519 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
5520 | return false; |
5521 | ||
5522 | /* Tell the scheduler that we'd really like pse to run next. */ | |
5523 | set_next_buddy(se); | |
5524 | ||
d95f4122 MG |
5525 | yield_task_fair(rq); |
5526 | ||
5527 | return true; | |
5528 | } | |
5529 | ||
681f3e68 | 5530 | #ifdef CONFIG_SMP |
bf0f6f24 | 5531 | /************************************************** |
e9c84cb8 PZ |
5532 | * Fair scheduling class load-balancing methods. |
5533 | * | |
5534 | * BASICS | |
5535 | * | |
5536 | * The purpose of load-balancing is to achieve the same basic fairness the | |
5537 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
5538 | * time to each task. This is expressed in the following equation: | |
5539 | * | |
5540 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
5541 | * | |
5542 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
5543 | * W_i,0 is defined as: | |
5544 | * | |
5545 | * W_i,0 = \Sum_j w_i,j (2) | |
5546 | * | |
5547 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
5548 | * is derived from the nice value as per prio_to_weight[]. | |
5549 | * | |
5550 | * The weight average is an exponential decay average of the instantaneous | |
5551 | * weight: | |
5552 | * | |
5553 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
5554 | * | |
ced549fa | 5555 | * C_i is the compute capacity of cpu i, typically it is the |
e9c84cb8 PZ |
5556 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
5557 | * can also include other factors [XXX]. | |
5558 | * | |
5559 | * To achieve this balance we define a measure of imbalance which follows | |
5560 | * directly from (1): | |
5561 | * | |
ced549fa | 5562 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
5563 | * |
5564 | * We them move tasks around to minimize the imbalance. In the continuous | |
5565 | * function space it is obvious this converges, in the discrete case we get | |
5566 | * a few fun cases generally called infeasible weight scenarios. | |
5567 | * | |
5568 | * [XXX expand on: | |
5569 | * - infeasible weights; | |
5570 | * - local vs global optima in the discrete case. ] | |
5571 | * | |
5572 | * | |
5573 | * SCHED DOMAINS | |
5574 | * | |
5575 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
5576 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
5577 | * topology where each level pairs two lower groups (or better). This results | |
5578 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
5579 | * tree to only the first of the previous level and we decrease the frequency | |
5580 | * of load-balance at each level inv. proportional to the number of cpus in | |
5581 | * the groups. | |
5582 | * | |
5583 | * This yields: | |
5584 | * | |
5585 | * log_2 n 1 n | |
5586 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
5587 | * i = 0 2^i 2^i | |
5588 | * `- size of each group | |
5589 | * | | `- number of cpus doing load-balance | |
5590 | * | `- freq | |
5591 | * `- sum over all levels | |
5592 | * | |
5593 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
5594 | * this makes (5) the runtime complexity of the balancer. | |
5595 | * | |
5596 | * An important property here is that each CPU is still (indirectly) connected | |
5597 | * to every other cpu in at most O(log n) steps: | |
5598 | * | |
5599 | * The adjacency matrix of the resulting graph is given by: | |
5600 | * | |
5601 | * log_2 n | |
5602 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | |
5603 | * k = 0 | |
5604 | * | |
5605 | * And you'll find that: | |
5606 | * | |
5607 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
5608 | * | |
5609 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
5610 | * The task movement gives a factor of O(m), giving a convergence complexity | |
5611 | * of: | |
5612 | * | |
5613 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
5614 | * | |
5615 | * | |
5616 | * WORK CONSERVING | |
5617 | * | |
5618 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
5619 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
5620 | * tree itself instead of relying on other CPUs to bring it work. | |
5621 | * | |
5622 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
5623 | * time. | |
5624 | * | |
5625 | * [XXX more?] | |
5626 | * | |
5627 | * | |
5628 | * CGROUPS | |
5629 | * | |
5630 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
5631 | * | |
5632 | * s_k,i | |
5633 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
5634 | * S_k | |
5635 | * | |
5636 | * Where | |
5637 | * | |
5638 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
5639 | * | |
5640 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
5641 | * | |
5642 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
5643 | * property. | |
5644 | * | |
5645 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
5646 | * rewrite all of this once again.] | |
5647 | */ | |
bf0f6f24 | 5648 | |
ed387b78 HS |
5649 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
5650 | ||
0ec8aa00 PZ |
5651 | enum fbq_type { regular, remote, all }; |
5652 | ||
ddcdf6e7 | 5653 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 5654 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
5655 | #define LBF_DST_PINNED 0x04 |
5656 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
5657 | |
5658 | struct lb_env { | |
5659 | struct sched_domain *sd; | |
5660 | ||
ddcdf6e7 | 5661 | struct rq *src_rq; |
85c1e7da | 5662 | int src_cpu; |
ddcdf6e7 PZ |
5663 | |
5664 | int dst_cpu; | |
5665 | struct rq *dst_rq; | |
5666 | ||
88b8dac0 SV |
5667 | struct cpumask *dst_grpmask; |
5668 | int new_dst_cpu; | |
ddcdf6e7 | 5669 | enum cpu_idle_type idle; |
bd939f45 | 5670 | long imbalance; |
b9403130 MW |
5671 | /* The set of CPUs under consideration for load-balancing */ |
5672 | struct cpumask *cpus; | |
5673 | ||
ddcdf6e7 | 5674 | unsigned int flags; |
367456c7 PZ |
5675 | |
5676 | unsigned int loop; | |
5677 | unsigned int loop_break; | |
5678 | unsigned int loop_max; | |
0ec8aa00 PZ |
5679 | |
5680 | enum fbq_type fbq_type; | |
163122b7 | 5681 | struct list_head tasks; |
ddcdf6e7 PZ |
5682 | }; |
5683 | ||
029632fb PZ |
5684 | /* |
5685 | * Is this task likely cache-hot: | |
5686 | */ | |
5d5e2b1b | 5687 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
5688 | { |
5689 | s64 delta; | |
5690 | ||
e5673f28 KT |
5691 | lockdep_assert_held(&env->src_rq->lock); |
5692 | ||
029632fb PZ |
5693 | if (p->sched_class != &fair_sched_class) |
5694 | return 0; | |
5695 | ||
5696 | if (unlikely(p->policy == SCHED_IDLE)) | |
5697 | return 0; | |
5698 | ||
5699 | /* | |
5700 | * Buddy candidates are cache hot: | |
5701 | */ | |
5d5e2b1b | 5702 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
5703 | (&p->se == cfs_rq_of(&p->se)->next || |
5704 | &p->se == cfs_rq_of(&p->se)->last)) | |
5705 | return 1; | |
5706 | ||
5707 | if (sysctl_sched_migration_cost == -1) | |
5708 | return 1; | |
5709 | if (sysctl_sched_migration_cost == 0) | |
5710 | return 0; | |
5711 | ||
5d5e2b1b | 5712 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
5713 | |
5714 | return delta < (s64)sysctl_sched_migration_cost; | |
5715 | } | |
5716 | ||
3a7053b3 | 5717 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 5718 | /* |
2a1ed24c SD |
5719 | * Returns 1, if task migration degrades locality |
5720 | * Returns 0, if task migration improves locality i.e migration preferred. | |
5721 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 5722 | */ |
2a1ed24c | 5723 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 5724 | { |
b1ad065e | 5725 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
c1ceac62 | 5726 | unsigned long src_faults, dst_faults; |
3a7053b3 MG |
5727 | int src_nid, dst_nid; |
5728 | ||
2a595721 | 5729 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
5730 | return -1; |
5731 | ||
c3b9bc5b | 5732 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 5733 | return -1; |
7a0f3083 MG |
5734 | |
5735 | src_nid = cpu_to_node(env->src_cpu); | |
5736 | dst_nid = cpu_to_node(env->dst_cpu); | |
5737 | ||
83e1d2cd | 5738 | if (src_nid == dst_nid) |
2a1ed24c | 5739 | return -1; |
7a0f3083 | 5740 | |
2a1ed24c SD |
5741 | /* Migrating away from the preferred node is always bad. */ |
5742 | if (src_nid == p->numa_preferred_nid) { | |
5743 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
5744 | return 1; | |
5745 | else | |
5746 | return -1; | |
5747 | } | |
b1ad065e | 5748 | |
c1ceac62 RR |
5749 | /* Encourage migration to the preferred node. */ |
5750 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 5751 | return 0; |
b1ad065e | 5752 | |
c1ceac62 RR |
5753 | if (numa_group) { |
5754 | src_faults = group_faults(p, src_nid); | |
5755 | dst_faults = group_faults(p, dst_nid); | |
5756 | } else { | |
5757 | src_faults = task_faults(p, src_nid); | |
5758 | dst_faults = task_faults(p, dst_nid); | |
b1ad065e RR |
5759 | } |
5760 | ||
c1ceac62 | 5761 | return dst_faults < src_faults; |
7a0f3083 MG |
5762 | } |
5763 | ||
3a7053b3 | 5764 | #else |
2a1ed24c | 5765 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
5766 | struct lb_env *env) |
5767 | { | |
2a1ed24c | 5768 | return -1; |
7a0f3083 | 5769 | } |
3a7053b3 MG |
5770 | #endif |
5771 | ||
1e3c88bd PZ |
5772 | /* |
5773 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
5774 | */ | |
5775 | static | |
8e45cb54 | 5776 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 5777 | { |
2a1ed24c | 5778 | int tsk_cache_hot; |
e5673f28 KT |
5779 | |
5780 | lockdep_assert_held(&env->src_rq->lock); | |
5781 | ||
1e3c88bd PZ |
5782 | /* |
5783 | * We do not migrate tasks that are: | |
d3198084 | 5784 | * 1) throttled_lb_pair, or |
1e3c88bd | 5785 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
5786 | * 3) running (obviously), or |
5787 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 5788 | */ |
d3198084 JK |
5789 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
5790 | return 0; | |
5791 | ||
ddcdf6e7 | 5792 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 5793 | int cpu; |
88b8dac0 | 5794 | |
41acab88 | 5795 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 5796 | |
6263322c PZ |
5797 | env->flags |= LBF_SOME_PINNED; |
5798 | ||
88b8dac0 SV |
5799 | /* |
5800 | * Remember if this task can be migrated to any other cpu in | |
5801 | * our sched_group. We may want to revisit it if we couldn't | |
5802 | * meet load balance goals by pulling other tasks on src_cpu. | |
5803 | * | |
5804 | * Also avoid computing new_dst_cpu if we have already computed | |
5805 | * one in current iteration. | |
5806 | */ | |
6263322c | 5807 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
5808 | return 0; |
5809 | ||
e02e60c1 JK |
5810 | /* Prevent to re-select dst_cpu via env's cpus */ |
5811 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
5812 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 5813 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
5814 | env->new_dst_cpu = cpu; |
5815 | break; | |
5816 | } | |
88b8dac0 | 5817 | } |
e02e60c1 | 5818 | |
1e3c88bd PZ |
5819 | return 0; |
5820 | } | |
88b8dac0 SV |
5821 | |
5822 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 5823 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 5824 | |
ddcdf6e7 | 5825 | if (task_running(env->src_rq, p)) { |
41acab88 | 5826 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
5827 | return 0; |
5828 | } | |
5829 | ||
5830 | /* | |
5831 | * Aggressive migration if: | |
3a7053b3 MG |
5832 | * 1) destination numa is preferred |
5833 | * 2) task is cache cold, or | |
5834 | * 3) too many balance attempts have failed. | |
1e3c88bd | 5835 | */ |
2a1ed24c SD |
5836 | tsk_cache_hot = migrate_degrades_locality(p, env); |
5837 | if (tsk_cache_hot == -1) | |
5838 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 5839 | |
2a1ed24c | 5840 | if (tsk_cache_hot <= 0 || |
7a96c231 | 5841 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 5842 | if (tsk_cache_hot == 1) { |
3a7053b3 MG |
5843 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
5844 | schedstat_inc(p, se.statistics.nr_forced_migrations); | |
5845 | } | |
1e3c88bd PZ |
5846 | return 1; |
5847 | } | |
5848 | ||
4e2dcb73 ZH |
5849 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
5850 | return 0; | |
1e3c88bd PZ |
5851 | } |
5852 | ||
897c395f | 5853 | /* |
163122b7 KT |
5854 | * detach_task() -- detach the task for the migration specified in env |
5855 | */ | |
5856 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
5857 | { | |
5858 | lockdep_assert_held(&env->src_rq->lock); | |
5859 | ||
163122b7 | 5860 | p->on_rq = TASK_ON_RQ_MIGRATING; |
3ea94de1 | 5861 | deactivate_task(env->src_rq, p, 0); |
163122b7 KT |
5862 | set_task_cpu(p, env->dst_cpu); |
5863 | } | |
5864 | ||
897c395f | 5865 | /* |
e5673f28 | 5866 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 5867 | * part of active balancing operations within "domain". |
897c395f | 5868 | * |
e5673f28 | 5869 | * Returns a task if successful and NULL otherwise. |
897c395f | 5870 | */ |
e5673f28 | 5871 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f PZ |
5872 | { |
5873 | struct task_struct *p, *n; | |
897c395f | 5874 | |
e5673f28 KT |
5875 | lockdep_assert_held(&env->src_rq->lock); |
5876 | ||
367456c7 | 5877 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
5878 | if (!can_migrate_task(p, env)) |
5879 | continue; | |
897c395f | 5880 | |
163122b7 | 5881 | detach_task(p, env); |
e5673f28 | 5882 | |
367456c7 | 5883 | /* |
e5673f28 | 5884 | * Right now, this is only the second place where |
163122b7 | 5885 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 5886 | * so we can safely collect stats here rather than |
163122b7 | 5887 | * inside detach_tasks(). |
367456c7 PZ |
5888 | */ |
5889 | schedstat_inc(env->sd, lb_gained[env->idle]); | |
e5673f28 | 5890 | return p; |
897c395f | 5891 | } |
e5673f28 | 5892 | return NULL; |
897c395f PZ |
5893 | } |
5894 | ||
eb95308e PZ |
5895 | static const unsigned int sched_nr_migrate_break = 32; |
5896 | ||
5d6523eb | 5897 | /* |
163122b7 KT |
5898 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
5899 | * busiest_rq, as part of a balancing operation within domain "sd". | |
5d6523eb | 5900 | * |
163122b7 | 5901 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 5902 | */ |
163122b7 | 5903 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 5904 | { |
5d6523eb PZ |
5905 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
5906 | struct task_struct *p; | |
367456c7 | 5907 | unsigned long load; |
163122b7 KT |
5908 | int detached = 0; |
5909 | ||
5910 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 5911 | |
bd939f45 | 5912 | if (env->imbalance <= 0) |
5d6523eb | 5913 | return 0; |
1e3c88bd | 5914 | |
5d6523eb | 5915 | while (!list_empty(tasks)) { |
985d3a4c YD |
5916 | /* |
5917 | * We don't want to steal all, otherwise we may be treated likewise, | |
5918 | * which could at worst lead to a livelock crash. | |
5919 | */ | |
5920 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
5921 | break; | |
5922 | ||
5d6523eb | 5923 | p = list_first_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 5924 | |
367456c7 PZ |
5925 | env->loop++; |
5926 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 5927 | if (env->loop > env->loop_max) |
367456c7 | 5928 | break; |
5d6523eb PZ |
5929 | |
5930 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 5931 | if (env->loop > env->loop_break) { |
eb95308e | 5932 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 5933 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 5934 | break; |
a195f004 | 5935 | } |
1e3c88bd | 5936 | |
d3198084 | 5937 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
5938 | goto next; |
5939 | ||
5940 | load = task_h_load(p); | |
5d6523eb | 5941 | |
eb95308e | 5942 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
5943 | goto next; |
5944 | ||
bd939f45 | 5945 | if ((load / 2) > env->imbalance) |
367456c7 | 5946 | goto next; |
1e3c88bd | 5947 | |
163122b7 KT |
5948 | detach_task(p, env); |
5949 | list_add(&p->se.group_node, &env->tasks); | |
5950 | ||
5951 | detached++; | |
bd939f45 | 5952 | env->imbalance -= load; |
1e3c88bd PZ |
5953 | |
5954 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
5955 | /* |
5956 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 5957 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
5958 | * the critical section. |
5959 | */ | |
5d6523eb | 5960 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 5961 | break; |
1e3c88bd PZ |
5962 | #endif |
5963 | ||
ee00e66f PZ |
5964 | /* |
5965 | * We only want to steal up to the prescribed amount of | |
5966 | * weighted load. | |
5967 | */ | |
bd939f45 | 5968 | if (env->imbalance <= 0) |
ee00e66f | 5969 | break; |
367456c7 PZ |
5970 | |
5971 | continue; | |
5972 | next: | |
5d6523eb | 5973 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 5974 | } |
5d6523eb | 5975 | |
1e3c88bd | 5976 | /* |
163122b7 KT |
5977 | * Right now, this is one of only two places we collect this stat |
5978 | * so we can safely collect detach_one_task() stats here rather | |
5979 | * than inside detach_one_task(). | |
1e3c88bd | 5980 | */ |
163122b7 | 5981 | schedstat_add(env->sd, lb_gained[env->idle], detached); |
1e3c88bd | 5982 | |
163122b7 KT |
5983 | return detached; |
5984 | } | |
5985 | ||
5986 | /* | |
5987 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
5988 | */ | |
5989 | static void attach_task(struct rq *rq, struct task_struct *p) | |
5990 | { | |
5991 | lockdep_assert_held(&rq->lock); | |
5992 | ||
5993 | BUG_ON(task_rq(p) != rq); | |
163122b7 | 5994 | activate_task(rq, p, 0); |
3ea94de1 | 5995 | p->on_rq = TASK_ON_RQ_QUEUED; |
163122b7 KT |
5996 | check_preempt_curr(rq, p, 0); |
5997 | } | |
5998 | ||
5999 | /* | |
6000 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
6001 | * its new rq. | |
6002 | */ | |
6003 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
6004 | { | |
6005 | raw_spin_lock(&rq->lock); | |
6006 | attach_task(rq, p); | |
6007 | raw_spin_unlock(&rq->lock); | |
6008 | } | |
6009 | ||
6010 | /* | |
6011 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
6012 | * new rq. | |
6013 | */ | |
6014 | static void attach_tasks(struct lb_env *env) | |
6015 | { | |
6016 | struct list_head *tasks = &env->tasks; | |
6017 | struct task_struct *p; | |
6018 | ||
6019 | raw_spin_lock(&env->dst_rq->lock); | |
6020 | ||
6021 | while (!list_empty(tasks)) { | |
6022 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
6023 | list_del_init(&p->se.group_node); | |
1e3c88bd | 6024 | |
163122b7 KT |
6025 | attach_task(env->dst_rq, p); |
6026 | } | |
6027 | ||
6028 | raw_spin_unlock(&env->dst_rq->lock); | |
1e3c88bd PZ |
6029 | } |
6030 | ||
230059de | 6031 | #ifdef CONFIG_FAIR_GROUP_SCHED |
48a16753 | 6032 | static void update_blocked_averages(int cpu) |
9e3081ca | 6033 | { |
9e3081ca | 6034 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
6035 | struct cfs_rq *cfs_rq; |
6036 | unsigned long flags; | |
9e3081ca | 6037 | |
48a16753 PT |
6038 | raw_spin_lock_irqsave(&rq->lock, flags); |
6039 | update_rq_clock(rq); | |
9d89c257 | 6040 | |
9763b67f PZ |
6041 | /* |
6042 | * Iterates the task_group tree in a bottom up fashion, see | |
6043 | * list_add_leaf_cfs_rq() for details. | |
6044 | */ | |
64660c86 | 6045 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
9d89c257 YD |
6046 | /* throttled entities do not contribute to load */ |
6047 | if (throttled_hierarchy(cfs_rq)) | |
6048 | continue; | |
48a16753 | 6049 | |
9d89c257 YD |
6050 | if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq)) |
6051 | update_tg_load_avg(cfs_rq, 0); | |
6052 | } | |
48a16753 | 6053 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
9e3081ca PZ |
6054 | } |
6055 | ||
9763b67f | 6056 | /* |
68520796 | 6057 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
6058 | * This needs to be done in a top-down fashion because the load of a child |
6059 | * group is a fraction of its parents load. | |
6060 | */ | |
68520796 | 6061 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 6062 | { |
68520796 VD |
6063 | struct rq *rq = rq_of(cfs_rq); |
6064 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 6065 | unsigned long now = jiffies; |
68520796 | 6066 | unsigned long load; |
a35b6466 | 6067 | |
68520796 | 6068 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
6069 | return; |
6070 | ||
68520796 VD |
6071 | cfs_rq->h_load_next = NULL; |
6072 | for_each_sched_entity(se) { | |
6073 | cfs_rq = cfs_rq_of(se); | |
6074 | cfs_rq->h_load_next = se; | |
6075 | if (cfs_rq->last_h_load_update == now) | |
6076 | break; | |
6077 | } | |
a35b6466 | 6078 | |
68520796 | 6079 | if (!se) { |
7ea241af | 6080 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
6081 | cfs_rq->last_h_load_update = now; |
6082 | } | |
6083 | ||
6084 | while ((se = cfs_rq->h_load_next) != NULL) { | |
6085 | load = cfs_rq->h_load; | |
7ea241af YD |
6086 | load = div64_ul(load * se->avg.load_avg, |
6087 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
6088 | cfs_rq = group_cfs_rq(se); |
6089 | cfs_rq->h_load = load; | |
6090 | cfs_rq->last_h_load_update = now; | |
6091 | } | |
9763b67f PZ |
6092 | } |
6093 | ||
367456c7 | 6094 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 6095 | { |
367456c7 | 6096 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 6097 | |
68520796 | 6098 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 6099 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 6100 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
6101 | } |
6102 | #else | |
48a16753 | 6103 | static inline void update_blocked_averages(int cpu) |
9e3081ca | 6104 | { |
6c1d47c0 VG |
6105 | struct rq *rq = cpu_rq(cpu); |
6106 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6107 | unsigned long flags; | |
6108 | ||
6109 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6110 | update_rq_clock(rq); | |
6111 | update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq); | |
6112 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9e3081ca PZ |
6113 | } |
6114 | ||
367456c7 | 6115 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 6116 | { |
9d89c257 | 6117 | return p->se.avg.load_avg; |
1e3c88bd | 6118 | } |
230059de | 6119 | #endif |
1e3c88bd | 6120 | |
1e3c88bd | 6121 | /********** Helpers for find_busiest_group ************************/ |
caeb178c RR |
6122 | |
6123 | enum group_type { | |
6124 | group_other = 0, | |
6125 | group_imbalanced, | |
6126 | group_overloaded, | |
6127 | }; | |
6128 | ||
1e3c88bd PZ |
6129 | /* |
6130 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
6131 | */ | |
6132 | struct sg_lb_stats { | |
6133 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
6134 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 6135 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 6136 | unsigned long load_per_task; |
63b2ca30 | 6137 | unsigned long group_capacity; |
9e91d61d | 6138 | unsigned long group_util; /* Total utilization of the group */ |
147c5fc2 | 6139 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
147c5fc2 PZ |
6140 | unsigned int idle_cpus; |
6141 | unsigned int group_weight; | |
caeb178c | 6142 | enum group_type group_type; |
ea67821b | 6143 | int group_no_capacity; |
0ec8aa00 PZ |
6144 | #ifdef CONFIG_NUMA_BALANCING |
6145 | unsigned int nr_numa_running; | |
6146 | unsigned int nr_preferred_running; | |
6147 | #endif | |
1e3c88bd PZ |
6148 | }; |
6149 | ||
56cf515b JK |
6150 | /* |
6151 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
6152 | * during load balancing. | |
6153 | */ | |
6154 | struct sd_lb_stats { | |
6155 | struct sched_group *busiest; /* Busiest group in this sd */ | |
6156 | struct sched_group *local; /* Local group in this sd */ | |
6157 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 6158 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
6159 | unsigned long avg_load; /* Average load across all groups in sd */ |
6160 | ||
56cf515b | 6161 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 6162 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
6163 | }; |
6164 | ||
147c5fc2 PZ |
6165 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
6166 | { | |
6167 | /* | |
6168 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
6169 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
6170 | * We must however clear busiest_stat::avg_load because | |
6171 | * update_sd_pick_busiest() reads this before assignment. | |
6172 | */ | |
6173 | *sds = (struct sd_lb_stats){ | |
6174 | .busiest = NULL, | |
6175 | .local = NULL, | |
6176 | .total_load = 0UL, | |
63b2ca30 | 6177 | .total_capacity = 0UL, |
147c5fc2 PZ |
6178 | .busiest_stat = { |
6179 | .avg_load = 0UL, | |
caeb178c RR |
6180 | .sum_nr_running = 0, |
6181 | .group_type = group_other, | |
147c5fc2 PZ |
6182 | }, |
6183 | }; | |
6184 | } | |
6185 | ||
1e3c88bd PZ |
6186 | /** |
6187 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
6188 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 6189 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
6190 | * |
6191 | * Return: The load index. | |
1e3c88bd PZ |
6192 | */ |
6193 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
6194 | enum cpu_idle_type idle) | |
6195 | { | |
6196 | int load_idx; | |
6197 | ||
6198 | switch (idle) { | |
6199 | case CPU_NOT_IDLE: | |
6200 | load_idx = sd->busy_idx; | |
6201 | break; | |
6202 | ||
6203 | case CPU_NEWLY_IDLE: | |
6204 | load_idx = sd->newidle_idx; | |
6205 | break; | |
6206 | default: | |
6207 | load_idx = sd->idle_idx; | |
6208 | break; | |
6209 | } | |
6210 | ||
6211 | return load_idx; | |
6212 | } | |
6213 | ||
ced549fa | 6214 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
6215 | { |
6216 | struct rq *rq = cpu_rq(cpu); | |
b5b4860d | 6217 | u64 total, used, age_stamp, avg; |
cadefd3d | 6218 | s64 delta; |
1e3c88bd | 6219 | |
b654f7de PZ |
6220 | /* |
6221 | * Since we're reading these variables without serialization make sure | |
6222 | * we read them once before doing sanity checks on them. | |
6223 | */ | |
316c1608 JL |
6224 | age_stamp = READ_ONCE(rq->age_stamp); |
6225 | avg = READ_ONCE(rq->rt_avg); | |
cebde6d6 | 6226 | delta = __rq_clock_broken(rq) - age_stamp; |
b654f7de | 6227 | |
cadefd3d PZ |
6228 | if (unlikely(delta < 0)) |
6229 | delta = 0; | |
6230 | ||
6231 | total = sched_avg_period() + delta; | |
aa483808 | 6232 | |
b5b4860d | 6233 | used = div_u64(avg, total); |
1e3c88bd | 6234 | |
b5b4860d VG |
6235 | if (likely(used < SCHED_CAPACITY_SCALE)) |
6236 | return SCHED_CAPACITY_SCALE - used; | |
1e3c88bd | 6237 | |
b5b4860d | 6238 | return 1; |
1e3c88bd PZ |
6239 | } |
6240 | ||
ced549fa | 6241 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 6242 | { |
8cd5601c | 6243 | unsigned long capacity = arch_scale_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
6244 | struct sched_group *sdg = sd->groups; |
6245 | ||
ca6d75e6 | 6246 | cpu_rq(cpu)->cpu_capacity_orig = capacity; |
9d5efe05 | 6247 | |
ced549fa | 6248 | capacity *= scale_rt_capacity(cpu); |
ca8ce3d0 | 6249 | capacity >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd | 6250 | |
ced549fa NP |
6251 | if (!capacity) |
6252 | capacity = 1; | |
1e3c88bd | 6253 | |
ced549fa NP |
6254 | cpu_rq(cpu)->cpu_capacity = capacity; |
6255 | sdg->sgc->capacity = capacity; | |
1e3c88bd PZ |
6256 | } |
6257 | ||
63b2ca30 | 6258 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
6259 | { |
6260 | struct sched_domain *child = sd->child; | |
6261 | struct sched_group *group, *sdg = sd->groups; | |
dc7ff76e | 6262 | unsigned long capacity; |
4ec4412e VG |
6263 | unsigned long interval; |
6264 | ||
6265 | interval = msecs_to_jiffies(sd->balance_interval); | |
6266 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 6267 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
6268 | |
6269 | if (!child) { | |
ced549fa | 6270 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
6271 | return; |
6272 | } | |
6273 | ||
dc7ff76e | 6274 | capacity = 0; |
1e3c88bd | 6275 | |
74a5ce20 PZ |
6276 | if (child->flags & SD_OVERLAP) { |
6277 | /* | |
6278 | * SD_OVERLAP domains cannot assume that child groups | |
6279 | * span the current group. | |
6280 | */ | |
6281 | ||
863bffc8 | 6282 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
63b2ca30 | 6283 | struct sched_group_capacity *sgc; |
9abf24d4 | 6284 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 6285 | |
9abf24d4 | 6286 | /* |
63b2ca30 | 6287 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
6288 | * gets here before we've attached the domains to the |
6289 | * runqueues. | |
6290 | * | |
ced549fa NP |
6291 | * Use capacity_of(), which is set irrespective of domains |
6292 | * in update_cpu_capacity(). | |
9abf24d4 | 6293 | * |
dc7ff76e | 6294 | * This avoids capacity from being 0 and |
9abf24d4 | 6295 | * causing divide-by-zero issues on boot. |
9abf24d4 SD |
6296 | */ |
6297 | if (unlikely(!rq->sd)) { | |
ced549fa | 6298 | capacity += capacity_of(cpu); |
9abf24d4 SD |
6299 | continue; |
6300 | } | |
863bffc8 | 6301 | |
63b2ca30 | 6302 | sgc = rq->sd->groups->sgc; |
63b2ca30 | 6303 | capacity += sgc->capacity; |
863bffc8 | 6304 | } |
74a5ce20 PZ |
6305 | } else { |
6306 | /* | |
6307 | * !SD_OVERLAP domains can assume that child groups | |
6308 | * span the current group. | |
6309 | */ | |
6310 | ||
6311 | group = child->groups; | |
6312 | do { | |
63b2ca30 | 6313 | capacity += group->sgc->capacity; |
74a5ce20 PZ |
6314 | group = group->next; |
6315 | } while (group != child->groups); | |
6316 | } | |
1e3c88bd | 6317 | |
63b2ca30 | 6318 | sdg->sgc->capacity = capacity; |
1e3c88bd PZ |
6319 | } |
6320 | ||
9d5efe05 | 6321 | /* |
ea67821b VG |
6322 | * Check whether the capacity of the rq has been noticeably reduced by side |
6323 | * activity. The imbalance_pct is used for the threshold. | |
6324 | * Return true is the capacity is reduced | |
9d5efe05 SV |
6325 | */ |
6326 | static inline int | |
ea67821b | 6327 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 6328 | { |
ea67821b VG |
6329 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
6330 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
6331 | } |
6332 | ||
30ce5dab PZ |
6333 | /* |
6334 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
6335 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
6336 | * | |
6337 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
6338 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
6339 | * Something like: | |
6340 | * | |
6341 | * { 0 1 2 3 } { 4 5 6 7 } | |
6342 | * * * * * | |
6343 | * | |
6344 | * If we were to balance group-wise we'd place two tasks in the first group and | |
6345 | * two tasks in the second group. Clearly this is undesired as it will overload | |
6346 | * cpu 3 and leave one of the cpus in the second group unused. | |
6347 | * | |
6348 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
6349 | * by noticing the lower domain failed to reach balance and had difficulty |
6350 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
6351 | * |
6352 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 6353 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 6354 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
6355 | * to create an effective group imbalance. |
6356 | * | |
6357 | * This is a somewhat tricky proposition since the next run might not find the | |
6358 | * group imbalance and decide the groups need to be balanced again. A most | |
6359 | * subtle and fragile situation. | |
6360 | */ | |
6361 | ||
6263322c | 6362 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 6363 | { |
63b2ca30 | 6364 | return group->sgc->imbalance; |
30ce5dab PZ |
6365 | } |
6366 | ||
b37d9316 | 6367 | /* |
ea67821b VG |
6368 | * group_has_capacity returns true if the group has spare capacity that could |
6369 | * be used by some tasks. | |
6370 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
6371 | * smaller than the number of CPUs or if the utilization is lower than the |
6372 | * available capacity for CFS tasks. | |
ea67821b VG |
6373 | * For the latter, we use a threshold to stabilize the state, to take into |
6374 | * account the variance of the tasks' load and to return true if the available | |
6375 | * capacity in meaningful for the load balancer. | |
6376 | * As an example, an available capacity of 1% can appear but it doesn't make | |
6377 | * any benefit for the load balance. | |
b37d9316 | 6378 | */ |
ea67821b VG |
6379 | static inline bool |
6380 | group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) | |
b37d9316 | 6381 | { |
ea67821b VG |
6382 | if (sgs->sum_nr_running < sgs->group_weight) |
6383 | return true; | |
c61037e9 | 6384 | |
ea67821b | 6385 | if ((sgs->group_capacity * 100) > |
9e91d61d | 6386 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 6387 | return true; |
b37d9316 | 6388 | |
ea67821b VG |
6389 | return false; |
6390 | } | |
6391 | ||
6392 | /* | |
6393 | * group_is_overloaded returns true if the group has more tasks than it can | |
6394 | * handle. | |
6395 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
6396 | * with the exact right number of tasks, has no more spare capacity but is not | |
6397 | * overloaded so both group_has_capacity and group_is_overloaded return | |
6398 | * false. | |
6399 | */ | |
6400 | static inline bool | |
6401 | group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) | |
6402 | { | |
6403 | if (sgs->sum_nr_running <= sgs->group_weight) | |
6404 | return false; | |
b37d9316 | 6405 | |
ea67821b | 6406 | if ((sgs->group_capacity * 100) < |
9e91d61d | 6407 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 6408 | return true; |
b37d9316 | 6409 | |
ea67821b | 6410 | return false; |
b37d9316 PZ |
6411 | } |
6412 | ||
79a89f92 LY |
6413 | static inline enum |
6414 | group_type group_classify(struct sched_group *group, | |
6415 | struct sg_lb_stats *sgs) | |
caeb178c | 6416 | { |
ea67821b | 6417 | if (sgs->group_no_capacity) |
caeb178c RR |
6418 | return group_overloaded; |
6419 | ||
6420 | if (sg_imbalanced(group)) | |
6421 | return group_imbalanced; | |
6422 | ||
6423 | return group_other; | |
6424 | } | |
6425 | ||
1e3c88bd PZ |
6426 | /** |
6427 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 6428 | * @env: The load balancing environment. |
1e3c88bd | 6429 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 6430 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 6431 | * @local_group: Does group contain this_cpu. |
1e3c88bd | 6432 | * @sgs: variable to hold the statistics for this group. |
cd3bd4e6 | 6433 | * @overload: Indicate more than one runnable task for any CPU. |
1e3c88bd | 6434 | */ |
bd939f45 PZ |
6435 | static inline void update_sg_lb_stats(struct lb_env *env, |
6436 | struct sched_group *group, int load_idx, | |
4486edd1 TC |
6437 | int local_group, struct sg_lb_stats *sgs, |
6438 | bool *overload) | |
1e3c88bd | 6439 | { |
30ce5dab | 6440 | unsigned long load; |
a426f99c | 6441 | int i, nr_running; |
1e3c88bd | 6442 | |
b72ff13c PZ |
6443 | memset(sgs, 0, sizeof(*sgs)); |
6444 | ||
b9403130 | 6445 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
6446 | struct rq *rq = cpu_rq(i); |
6447 | ||
1e3c88bd | 6448 | /* Bias balancing toward cpus of our domain */ |
6263322c | 6449 | if (local_group) |
04f733b4 | 6450 | load = target_load(i, load_idx); |
6263322c | 6451 | else |
1e3c88bd | 6452 | load = source_load(i, load_idx); |
1e3c88bd PZ |
6453 | |
6454 | sgs->group_load += load; | |
9e91d61d | 6455 | sgs->group_util += cpu_util(i); |
65fdac08 | 6456 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 6457 | |
a426f99c WL |
6458 | nr_running = rq->nr_running; |
6459 | if (nr_running > 1) | |
4486edd1 TC |
6460 | *overload = true; |
6461 | ||
0ec8aa00 PZ |
6462 | #ifdef CONFIG_NUMA_BALANCING |
6463 | sgs->nr_numa_running += rq->nr_numa_running; | |
6464 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
6465 | #endif | |
1e3c88bd | 6466 | sgs->sum_weighted_load += weighted_cpuload(i); |
a426f99c WL |
6467 | /* |
6468 | * No need to call idle_cpu() if nr_running is not 0 | |
6469 | */ | |
6470 | if (!nr_running && idle_cpu(i)) | |
aae6d3dd | 6471 | sgs->idle_cpus++; |
1e3c88bd PZ |
6472 | } |
6473 | ||
63b2ca30 NP |
6474 | /* Adjust by relative CPU capacity of the group */ |
6475 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 6476 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 6477 | |
dd5feea1 | 6478 | if (sgs->sum_nr_running) |
38d0f770 | 6479 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 6480 | |
aae6d3dd | 6481 | sgs->group_weight = group->group_weight; |
b37d9316 | 6482 | |
ea67821b | 6483 | sgs->group_no_capacity = group_is_overloaded(env, sgs); |
79a89f92 | 6484 | sgs->group_type = group_classify(group, sgs); |
1e3c88bd PZ |
6485 | } |
6486 | ||
532cb4c4 MN |
6487 | /** |
6488 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 6489 | * @env: The load balancing environment. |
532cb4c4 MN |
6490 | * @sds: sched_domain statistics |
6491 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 6492 | * @sgs: sched_group statistics |
532cb4c4 MN |
6493 | * |
6494 | * Determine if @sg is a busier group than the previously selected | |
6495 | * busiest group. | |
e69f6186 YB |
6496 | * |
6497 | * Return: %true if @sg is a busier group than the previously selected | |
6498 | * busiest group. %false otherwise. | |
532cb4c4 | 6499 | */ |
bd939f45 | 6500 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
6501 | struct sd_lb_stats *sds, |
6502 | struct sched_group *sg, | |
bd939f45 | 6503 | struct sg_lb_stats *sgs) |
532cb4c4 | 6504 | { |
caeb178c | 6505 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 6506 | |
caeb178c | 6507 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
6508 | return true; |
6509 | ||
caeb178c RR |
6510 | if (sgs->group_type < busiest->group_type) |
6511 | return false; | |
6512 | ||
6513 | if (sgs->avg_load <= busiest->avg_load) | |
6514 | return false; | |
6515 | ||
6516 | /* This is the busiest node in its class. */ | |
6517 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
6518 | return true; |
6519 | ||
6520 | /* | |
6521 | * ASYM_PACKING needs to move all the work to the lowest | |
6522 | * numbered CPUs in the group, therefore mark all groups | |
6523 | * higher than ourself as busy. | |
6524 | */ | |
caeb178c | 6525 | if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { |
532cb4c4 MN |
6526 | if (!sds->busiest) |
6527 | return true; | |
6528 | ||
6529 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | |
6530 | return true; | |
6531 | } | |
6532 | ||
6533 | return false; | |
6534 | } | |
6535 | ||
0ec8aa00 PZ |
6536 | #ifdef CONFIG_NUMA_BALANCING |
6537 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
6538 | { | |
6539 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
6540 | return regular; | |
6541 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
6542 | return remote; | |
6543 | return all; | |
6544 | } | |
6545 | ||
6546 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
6547 | { | |
6548 | if (rq->nr_running > rq->nr_numa_running) | |
6549 | return regular; | |
6550 | if (rq->nr_running > rq->nr_preferred_running) | |
6551 | return remote; | |
6552 | return all; | |
6553 | } | |
6554 | #else | |
6555 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
6556 | { | |
6557 | return all; | |
6558 | } | |
6559 | ||
6560 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
6561 | { | |
6562 | return regular; | |
6563 | } | |
6564 | #endif /* CONFIG_NUMA_BALANCING */ | |
6565 | ||
1e3c88bd | 6566 | /** |
461819ac | 6567 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 6568 | * @env: The load balancing environment. |
1e3c88bd PZ |
6569 | * @sds: variable to hold the statistics for this sched_domain. |
6570 | */ | |
0ec8aa00 | 6571 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6572 | { |
bd939f45 PZ |
6573 | struct sched_domain *child = env->sd->child; |
6574 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 6575 | struct sg_lb_stats tmp_sgs; |
1e3c88bd | 6576 | int load_idx, prefer_sibling = 0; |
4486edd1 | 6577 | bool overload = false; |
1e3c88bd PZ |
6578 | |
6579 | if (child && child->flags & SD_PREFER_SIBLING) | |
6580 | prefer_sibling = 1; | |
6581 | ||
bd939f45 | 6582 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
6583 | |
6584 | do { | |
56cf515b | 6585 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
6586 | int local_group; |
6587 | ||
bd939f45 | 6588 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
6589 | if (local_group) { |
6590 | sds->local = sg; | |
6591 | sgs = &sds->local_stat; | |
b72ff13c PZ |
6592 | |
6593 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
6594 | time_after_eq(jiffies, sg->sgc->next_update)) |
6595 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 6596 | } |
1e3c88bd | 6597 | |
4486edd1 TC |
6598 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, |
6599 | &overload); | |
1e3c88bd | 6600 | |
b72ff13c PZ |
6601 | if (local_group) |
6602 | goto next_group; | |
6603 | ||
1e3c88bd PZ |
6604 | /* |
6605 | * In case the child domain prefers tasks go to siblings | |
ea67821b | 6606 | * first, lower the sg capacity so that we'll try |
75dd321d NR |
6607 | * and move all the excess tasks away. We lower the capacity |
6608 | * of a group only if the local group has the capacity to fit | |
ea67821b VG |
6609 | * these excess tasks. The extra check prevents the case where |
6610 | * you always pull from the heaviest group when it is already | |
6611 | * under-utilized (possible with a large weight task outweighs | |
6612 | * the tasks on the system). | |
1e3c88bd | 6613 | */ |
b72ff13c | 6614 | if (prefer_sibling && sds->local && |
ea67821b VG |
6615 | group_has_capacity(env, &sds->local_stat) && |
6616 | (sgs->sum_nr_running > 1)) { | |
6617 | sgs->group_no_capacity = 1; | |
79a89f92 | 6618 | sgs->group_type = group_classify(sg, sgs); |
cb0b9f24 | 6619 | } |
1e3c88bd | 6620 | |
b72ff13c | 6621 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 6622 | sds->busiest = sg; |
56cf515b | 6623 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
6624 | } |
6625 | ||
b72ff13c PZ |
6626 | next_group: |
6627 | /* Now, start updating sd_lb_stats */ | |
6628 | sds->total_load += sgs->group_load; | |
63b2ca30 | 6629 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 6630 | |
532cb4c4 | 6631 | sg = sg->next; |
bd939f45 | 6632 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
6633 | |
6634 | if (env->sd->flags & SD_NUMA) | |
6635 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
6636 | |
6637 | if (!env->sd->parent) { | |
6638 | /* update overload indicator if we are at root domain */ | |
6639 | if (env->dst_rq->rd->overload != overload) | |
6640 | env->dst_rq->rd->overload = overload; | |
6641 | } | |
6642 | ||
532cb4c4 MN |
6643 | } |
6644 | ||
532cb4c4 MN |
6645 | /** |
6646 | * check_asym_packing - Check to see if the group is packed into the | |
6647 | * sched doman. | |
6648 | * | |
6649 | * This is primarily intended to used at the sibling level. Some | |
6650 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
6651 | * case of POWER7, it can move to lower SMT modes only when higher | |
6652 | * threads are idle. When in lower SMT modes, the threads will | |
6653 | * perform better since they share less core resources. Hence when we | |
6654 | * have idle threads, we want them to be the higher ones. | |
6655 | * | |
6656 | * This packing function is run on idle threads. It checks to see if | |
6657 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
6658 | * CPU number than the packing function is being run on. Here we are | |
6659 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
6660 | * number. | |
6661 | * | |
e69f6186 | 6662 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
6663 | * this CPU. The amount of the imbalance is returned in *imbalance. |
6664 | * | |
cd96891d | 6665 | * @env: The load balancing environment. |
532cb4c4 | 6666 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 6667 | */ |
bd939f45 | 6668 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
6669 | { |
6670 | int busiest_cpu; | |
6671 | ||
bd939f45 | 6672 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
6673 | return 0; |
6674 | ||
6675 | if (!sds->busiest) | |
6676 | return 0; | |
6677 | ||
6678 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 6679 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
6680 | return 0; |
6681 | ||
bd939f45 | 6682 | env->imbalance = DIV_ROUND_CLOSEST( |
63b2ca30 | 6683 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, |
ca8ce3d0 | 6684 | SCHED_CAPACITY_SCALE); |
bd939f45 | 6685 | |
532cb4c4 | 6686 | return 1; |
1e3c88bd PZ |
6687 | } |
6688 | ||
6689 | /** | |
6690 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
6691 | * amongst the groups of a sched_domain, during | |
6692 | * load balancing. | |
cd96891d | 6693 | * @env: The load balancing environment. |
1e3c88bd | 6694 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6695 | */ |
bd939f45 PZ |
6696 | static inline |
6697 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 6698 | { |
63b2ca30 | 6699 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 6700 | unsigned int imbn = 2; |
dd5feea1 | 6701 | unsigned long scaled_busy_load_per_task; |
56cf515b | 6702 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 6703 | |
56cf515b JK |
6704 | local = &sds->local_stat; |
6705 | busiest = &sds->busiest_stat; | |
1e3c88bd | 6706 | |
56cf515b JK |
6707 | if (!local->sum_nr_running) |
6708 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
6709 | else if (busiest->load_per_task > local->load_per_task) | |
6710 | imbn = 1; | |
dd5feea1 | 6711 | |
56cf515b | 6712 | scaled_busy_load_per_task = |
ca8ce3d0 | 6713 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 6714 | busiest->group_capacity; |
56cf515b | 6715 | |
3029ede3 VD |
6716 | if (busiest->avg_load + scaled_busy_load_per_task >= |
6717 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 6718 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6719 | return; |
6720 | } | |
6721 | ||
6722 | /* | |
6723 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 6724 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
6725 | * moving them. |
6726 | */ | |
6727 | ||
63b2ca30 | 6728 | capa_now += busiest->group_capacity * |
56cf515b | 6729 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 6730 | capa_now += local->group_capacity * |
56cf515b | 6731 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 6732 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
6733 | |
6734 | /* Amount of load we'd subtract */ | |
a2cd4260 | 6735 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 6736 | capa_move += busiest->group_capacity * |
56cf515b | 6737 | min(busiest->load_per_task, |
a2cd4260 | 6738 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 6739 | } |
1e3c88bd PZ |
6740 | |
6741 | /* Amount of load we'd add */ | |
63b2ca30 | 6742 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 6743 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
6744 | tmp = (busiest->avg_load * busiest->group_capacity) / |
6745 | local->group_capacity; | |
56cf515b | 6746 | } else { |
ca8ce3d0 | 6747 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 6748 | local->group_capacity; |
56cf515b | 6749 | } |
63b2ca30 | 6750 | capa_move += local->group_capacity * |
3ae11c90 | 6751 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 6752 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
6753 | |
6754 | /* Move if we gain throughput */ | |
63b2ca30 | 6755 | if (capa_move > capa_now) |
56cf515b | 6756 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
6757 | } |
6758 | ||
6759 | /** | |
6760 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
6761 | * groups of a given sched_domain during load balance. | |
bd939f45 | 6762 | * @env: load balance environment |
1e3c88bd | 6763 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 6764 | */ |
bd939f45 | 6765 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6766 | { |
dd5feea1 | 6767 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
6768 | struct sg_lb_stats *local, *busiest; |
6769 | ||
6770 | local = &sds->local_stat; | |
56cf515b | 6771 | busiest = &sds->busiest_stat; |
dd5feea1 | 6772 | |
caeb178c | 6773 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
6774 | /* |
6775 | * In the group_imb case we cannot rely on group-wide averages | |
6776 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
6777 | */ | |
56cf515b JK |
6778 | busiest->load_per_task = |
6779 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
6780 | } |
6781 | ||
1e3c88bd PZ |
6782 | /* |
6783 | * In the presence of smp nice balancing, certain scenarios can have | |
6784 | * max load less than avg load(as we skip the groups at or below | |
ced549fa | 6785 | * its cpu_capacity, while calculating max_load..) |
1e3c88bd | 6786 | */ |
b1885550 VD |
6787 | if (busiest->avg_load <= sds->avg_load || |
6788 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
6789 | env->imbalance = 0; |
6790 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
6791 | } |
6792 | ||
9a5d9ba6 PZ |
6793 | /* |
6794 | * If there aren't any idle cpus, avoid creating some. | |
6795 | */ | |
6796 | if (busiest->group_type == group_overloaded && | |
6797 | local->group_type == group_overloaded) { | |
ea67821b VG |
6798 | load_above_capacity = busiest->sum_nr_running * |
6799 | SCHED_LOAD_SCALE; | |
6800 | if (load_above_capacity > busiest->group_capacity) | |
6801 | load_above_capacity -= busiest->group_capacity; | |
6802 | else | |
6803 | load_above_capacity = ~0UL; | |
dd5feea1 SS |
6804 | } |
6805 | ||
6806 | /* | |
6807 | * We're trying to get all the cpus to the average_load, so we don't | |
6808 | * want to push ourselves above the average load, nor do we wish to | |
6809 | * reduce the max loaded cpu below the average load. At the same time, | |
6810 | * we also don't want to reduce the group load below the group capacity | |
6811 | * (so that we can implement power-savings policies etc). Thus we look | |
6812 | * for the minimum possible imbalance. | |
dd5feea1 | 6813 | */ |
30ce5dab | 6814 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
6815 | |
6816 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 6817 | env->imbalance = min( |
63b2ca30 NP |
6818 | max_pull * busiest->group_capacity, |
6819 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 6820 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
6821 | |
6822 | /* | |
6823 | * if *imbalance is less than the average load per runnable task | |
25985edc | 6824 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
6825 | * a think about bumping its value to force at least one task to be |
6826 | * moved | |
6827 | */ | |
56cf515b | 6828 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 6829 | return fix_small_imbalance(env, sds); |
1e3c88bd | 6830 | } |
fab47622 | 6831 | |
1e3c88bd PZ |
6832 | /******* find_busiest_group() helpers end here *********************/ |
6833 | ||
6834 | /** | |
6835 | * find_busiest_group - Returns the busiest group within the sched_domain | |
6836 | * if there is an imbalance. If there isn't an imbalance, and | |
6837 | * the user has opted for power-savings, it returns a group whose | |
6838 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | |
6839 | * such a group exists. | |
6840 | * | |
6841 | * Also calculates the amount of weighted load which should be moved | |
6842 | * to restore balance. | |
6843 | * | |
cd96891d | 6844 | * @env: The load balancing environment. |
1e3c88bd | 6845 | * |
e69f6186 | 6846 | * Return: - The busiest group if imbalance exists. |
1e3c88bd PZ |
6847 | * - If no imbalance and user has opted for power-savings balance, |
6848 | * return the least loaded group whose CPUs can be | |
6849 | * put to idle by rebalancing its tasks onto our group. | |
6850 | */ | |
56cf515b | 6851 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 6852 | { |
56cf515b | 6853 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
6854 | struct sd_lb_stats sds; |
6855 | ||
147c5fc2 | 6856 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
6857 | |
6858 | /* | |
6859 | * Compute the various statistics relavent for load balancing at | |
6860 | * this level. | |
6861 | */ | |
23f0d209 | 6862 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
6863 | local = &sds.local_stat; |
6864 | busiest = &sds.busiest_stat; | |
1e3c88bd | 6865 | |
ea67821b | 6866 | /* ASYM feature bypasses nice load balance check */ |
bd939f45 PZ |
6867 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
6868 | check_asym_packing(env, &sds)) | |
532cb4c4 MN |
6869 | return sds.busiest; |
6870 | ||
cc57aa8f | 6871 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 6872 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
6873 | goto out_balanced; |
6874 | ||
ca8ce3d0 NP |
6875 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
6876 | / sds.total_capacity; | |
b0432d8f | 6877 | |
866ab43e PZ |
6878 | /* |
6879 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 6880 | * work because they assume all things are equal, which typically |
866ab43e PZ |
6881 | * isn't true due to cpus_allowed constraints and the like. |
6882 | */ | |
caeb178c | 6883 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
6884 | goto force_balance; |
6885 | ||
cc57aa8f | 6886 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
ea67821b VG |
6887 | if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) && |
6888 | busiest->group_no_capacity) | |
fab47622 NR |
6889 | goto force_balance; |
6890 | ||
cc57aa8f | 6891 | /* |
9c58c79a | 6892 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
6893 | * don't try and pull any tasks. |
6894 | */ | |
56cf515b | 6895 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
6896 | goto out_balanced; |
6897 | ||
cc57aa8f PZ |
6898 | /* |
6899 | * Don't pull any tasks if this group is already above the domain | |
6900 | * average load. | |
6901 | */ | |
56cf515b | 6902 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
6903 | goto out_balanced; |
6904 | ||
bd939f45 | 6905 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 6906 | /* |
43f4d666 VG |
6907 | * This cpu is idle. If the busiest group is not overloaded |
6908 | * and there is no imbalance between this and busiest group | |
6909 | * wrt idle cpus, it is balanced. The imbalance becomes | |
6910 | * significant if the diff is greater than 1 otherwise we | |
6911 | * might end up to just move the imbalance on another group | |
aae6d3dd | 6912 | */ |
43f4d666 VG |
6913 | if ((busiest->group_type != group_overloaded) && |
6914 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 6915 | goto out_balanced; |
c186fafe PZ |
6916 | } else { |
6917 | /* | |
6918 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
6919 | * imbalance_pct to be conservative. | |
6920 | */ | |
56cf515b JK |
6921 | if (100 * busiest->avg_load <= |
6922 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 6923 | goto out_balanced; |
aae6d3dd | 6924 | } |
1e3c88bd | 6925 | |
fab47622 | 6926 | force_balance: |
1e3c88bd | 6927 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 6928 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
6929 | return sds.busiest; |
6930 | ||
6931 | out_balanced: | |
bd939f45 | 6932 | env->imbalance = 0; |
1e3c88bd PZ |
6933 | return NULL; |
6934 | } | |
6935 | ||
6936 | /* | |
6937 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
6938 | */ | |
bd939f45 | 6939 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 6940 | struct sched_group *group) |
1e3c88bd PZ |
6941 | { |
6942 | struct rq *busiest = NULL, *rq; | |
ced549fa | 6943 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
6944 | int i; |
6945 | ||
6906a408 | 6946 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
ea67821b | 6947 | unsigned long capacity, wl; |
0ec8aa00 PZ |
6948 | enum fbq_type rt; |
6949 | ||
6950 | rq = cpu_rq(i); | |
6951 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 6952 | |
0ec8aa00 PZ |
6953 | /* |
6954 | * We classify groups/runqueues into three groups: | |
6955 | * - regular: there are !numa tasks | |
6956 | * - remote: there are numa tasks that run on the 'wrong' node | |
6957 | * - all: there is no distinction | |
6958 | * | |
6959 | * In order to avoid migrating ideally placed numa tasks, | |
6960 | * ignore those when there's better options. | |
6961 | * | |
6962 | * If we ignore the actual busiest queue to migrate another | |
6963 | * task, the next balance pass can still reduce the busiest | |
6964 | * queue by moving tasks around inside the node. | |
6965 | * | |
6966 | * If we cannot move enough load due to this classification | |
6967 | * the next pass will adjust the group classification and | |
6968 | * allow migration of more tasks. | |
6969 | * | |
6970 | * Both cases only affect the total convergence complexity. | |
6971 | */ | |
6972 | if (rt > env->fbq_type) | |
6973 | continue; | |
6974 | ||
ced549fa | 6975 | capacity = capacity_of(i); |
9d5efe05 | 6976 | |
6e40f5bb | 6977 | wl = weighted_cpuload(i); |
1e3c88bd | 6978 | |
6e40f5bb TG |
6979 | /* |
6980 | * When comparing with imbalance, use weighted_cpuload() | |
ced549fa | 6981 | * which is not scaled with the cpu capacity. |
6e40f5bb | 6982 | */ |
ea67821b VG |
6983 | |
6984 | if (rq->nr_running == 1 && wl > env->imbalance && | |
6985 | !check_cpu_capacity(rq, env->sd)) | |
1e3c88bd PZ |
6986 | continue; |
6987 | ||
6e40f5bb TG |
6988 | /* |
6989 | * For the load comparisons with the other cpu's, consider | |
ced549fa NP |
6990 | * the weighted_cpuload() scaled with the cpu capacity, so |
6991 | * that the load can be moved away from the cpu that is | |
6992 | * potentially running at a lower capacity. | |
95a79b80 | 6993 | * |
ced549fa | 6994 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
95a79b80 | 6995 | * multiplication to rid ourselves of the division works out |
ced549fa NP |
6996 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
6997 | * our previous maximum. | |
6e40f5bb | 6998 | */ |
ced549fa | 6999 | if (wl * busiest_capacity > busiest_load * capacity) { |
95a79b80 | 7000 | busiest_load = wl; |
ced549fa | 7001 | busiest_capacity = capacity; |
1e3c88bd PZ |
7002 | busiest = rq; |
7003 | } | |
7004 | } | |
7005 | ||
7006 | return busiest; | |
7007 | } | |
7008 | ||
7009 | /* | |
7010 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
7011 | * so long as it is large enough. | |
7012 | */ | |
7013 | #define MAX_PINNED_INTERVAL 512 | |
7014 | ||
7015 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 7016 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 7017 | |
bd939f45 | 7018 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 7019 | { |
bd939f45 PZ |
7020 | struct sched_domain *sd = env->sd; |
7021 | ||
7022 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
7023 | |
7024 | /* | |
7025 | * ASYM_PACKING needs to force migrate tasks from busy but | |
7026 | * higher numbered CPUs in order to pack all tasks in the | |
7027 | * lowest numbered CPUs. | |
7028 | */ | |
bd939f45 | 7029 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 7030 | return 1; |
1af3ed3d PZ |
7031 | } |
7032 | ||
1aaf90a4 VG |
7033 | /* |
7034 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
7035 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
7036 | * because of other sched_class or IRQs if more capacity stays | |
7037 | * available on dst_cpu. | |
7038 | */ | |
7039 | if ((env->idle != CPU_NOT_IDLE) && | |
7040 | (env->src_rq->cfs.h_nr_running == 1)) { | |
7041 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
7042 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
7043 | return 1; | |
7044 | } | |
7045 | ||
1af3ed3d PZ |
7046 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
7047 | } | |
7048 | ||
969c7921 TH |
7049 | static int active_load_balance_cpu_stop(void *data); |
7050 | ||
23f0d209 JK |
7051 | static int should_we_balance(struct lb_env *env) |
7052 | { | |
7053 | struct sched_group *sg = env->sd->groups; | |
7054 | struct cpumask *sg_cpus, *sg_mask; | |
7055 | int cpu, balance_cpu = -1; | |
7056 | ||
7057 | /* | |
7058 | * In the newly idle case, we will allow all the cpu's | |
7059 | * to do the newly idle load balance. | |
7060 | */ | |
7061 | if (env->idle == CPU_NEWLY_IDLE) | |
7062 | return 1; | |
7063 | ||
7064 | sg_cpus = sched_group_cpus(sg); | |
7065 | sg_mask = sched_group_mask(sg); | |
7066 | /* Try to find first idle cpu */ | |
7067 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
7068 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
7069 | continue; | |
7070 | ||
7071 | balance_cpu = cpu; | |
7072 | break; | |
7073 | } | |
7074 | ||
7075 | if (balance_cpu == -1) | |
7076 | balance_cpu = group_balance_cpu(sg); | |
7077 | ||
7078 | /* | |
7079 | * First idle cpu or the first cpu(busiest) in this sched group | |
7080 | * is eligible for doing load balancing at this and above domains. | |
7081 | */ | |
b0cff9d8 | 7082 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
7083 | } |
7084 | ||
1e3c88bd PZ |
7085 | /* |
7086 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
7087 | * tasks if there is an imbalance. | |
7088 | */ | |
7089 | static int load_balance(int this_cpu, struct rq *this_rq, | |
7090 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 7091 | int *continue_balancing) |
1e3c88bd | 7092 | { |
88b8dac0 | 7093 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 7094 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 7095 | struct sched_group *group; |
1e3c88bd PZ |
7096 | struct rq *busiest; |
7097 | unsigned long flags; | |
4ba29684 | 7098 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 7099 | |
8e45cb54 PZ |
7100 | struct lb_env env = { |
7101 | .sd = sd, | |
ddcdf6e7 PZ |
7102 | .dst_cpu = this_cpu, |
7103 | .dst_rq = this_rq, | |
88b8dac0 | 7104 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 7105 | .idle = idle, |
eb95308e | 7106 | .loop_break = sched_nr_migrate_break, |
b9403130 | 7107 | .cpus = cpus, |
0ec8aa00 | 7108 | .fbq_type = all, |
163122b7 | 7109 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
7110 | }; |
7111 | ||
cfc03118 JK |
7112 | /* |
7113 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
7114 | * other cpus in our group | |
7115 | */ | |
e02e60c1 | 7116 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 7117 | env.dst_grpmask = NULL; |
cfc03118 | 7118 | |
1e3c88bd PZ |
7119 | cpumask_copy(cpus, cpu_active_mask); |
7120 | ||
1e3c88bd PZ |
7121 | schedstat_inc(sd, lb_count[idle]); |
7122 | ||
7123 | redo: | |
23f0d209 JK |
7124 | if (!should_we_balance(&env)) { |
7125 | *continue_balancing = 0; | |
1e3c88bd | 7126 | goto out_balanced; |
23f0d209 | 7127 | } |
1e3c88bd | 7128 | |
23f0d209 | 7129 | group = find_busiest_group(&env); |
1e3c88bd PZ |
7130 | if (!group) { |
7131 | schedstat_inc(sd, lb_nobusyg[idle]); | |
7132 | goto out_balanced; | |
7133 | } | |
7134 | ||
b9403130 | 7135 | busiest = find_busiest_queue(&env, group); |
1e3c88bd PZ |
7136 | if (!busiest) { |
7137 | schedstat_inc(sd, lb_nobusyq[idle]); | |
7138 | goto out_balanced; | |
7139 | } | |
7140 | ||
78feefc5 | 7141 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 7142 | |
bd939f45 | 7143 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
1e3c88bd | 7144 | |
1aaf90a4 VG |
7145 | env.src_cpu = busiest->cpu; |
7146 | env.src_rq = busiest; | |
7147 | ||
1e3c88bd PZ |
7148 | ld_moved = 0; |
7149 | if (busiest->nr_running > 1) { | |
7150 | /* | |
7151 | * Attempt to move tasks. If find_busiest_group has found | |
7152 | * an imbalance but busiest->nr_running <= 1, the group is | |
7153 | * still unbalanced. ld_moved simply stays zero, so it is | |
7154 | * correctly treated as an imbalance. | |
7155 | */ | |
8e45cb54 | 7156 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 7157 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 7158 | |
5d6523eb | 7159 | more_balance: |
163122b7 | 7160 | raw_spin_lock_irqsave(&busiest->lock, flags); |
88b8dac0 SV |
7161 | |
7162 | /* | |
7163 | * cur_ld_moved - load moved in current iteration | |
7164 | * ld_moved - cumulative load moved across iterations | |
7165 | */ | |
163122b7 | 7166 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
7167 | |
7168 | /* | |
163122b7 KT |
7169 | * We've detached some tasks from busiest_rq. Every |
7170 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
7171 | * unlock busiest->lock, and we are able to be sure | |
7172 | * that nobody can manipulate the tasks in parallel. | |
7173 | * See task_rq_lock() family for the details. | |
1e3c88bd | 7174 | */ |
163122b7 KT |
7175 | |
7176 | raw_spin_unlock(&busiest->lock); | |
7177 | ||
7178 | if (cur_ld_moved) { | |
7179 | attach_tasks(&env); | |
7180 | ld_moved += cur_ld_moved; | |
7181 | } | |
7182 | ||
1e3c88bd | 7183 | local_irq_restore(flags); |
88b8dac0 | 7184 | |
f1cd0858 JK |
7185 | if (env.flags & LBF_NEED_BREAK) { |
7186 | env.flags &= ~LBF_NEED_BREAK; | |
7187 | goto more_balance; | |
7188 | } | |
7189 | ||
88b8dac0 SV |
7190 | /* |
7191 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
7192 | * us and move them to an alternate dst_cpu in our sched_group | |
7193 | * where they can run. The upper limit on how many times we | |
7194 | * iterate on same src_cpu is dependent on number of cpus in our | |
7195 | * sched_group. | |
7196 | * | |
7197 | * This changes load balance semantics a bit on who can move | |
7198 | * load to a given_cpu. In addition to the given_cpu itself | |
7199 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
7200 | * nohz-idle), we now have balance_cpu in a position to move | |
7201 | * load to given_cpu. In rare situations, this may cause | |
7202 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
7203 | * _independently_ and at _same_ time to move some load to | |
7204 | * given_cpu) causing exceess load to be moved to given_cpu. | |
7205 | * This however should not happen so much in practice and | |
7206 | * moreover subsequent load balance cycles should correct the | |
7207 | * excess load moved. | |
7208 | */ | |
6263322c | 7209 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 7210 | |
7aff2e3a VD |
7211 | /* Prevent to re-select dst_cpu via env's cpus */ |
7212 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
7213 | ||
78feefc5 | 7214 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 7215 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 7216 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
7217 | env.loop = 0; |
7218 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 7219 | |
88b8dac0 SV |
7220 | /* |
7221 | * Go back to "more_balance" rather than "redo" since we | |
7222 | * need to continue with same src_cpu. | |
7223 | */ | |
7224 | goto more_balance; | |
7225 | } | |
1e3c88bd | 7226 | |
6263322c PZ |
7227 | /* |
7228 | * We failed to reach balance because of affinity. | |
7229 | */ | |
7230 | if (sd_parent) { | |
63b2ca30 | 7231 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 7232 | |
afdeee05 | 7233 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 7234 | *group_imbalance = 1; |
6263322c PZ |
7235 | } |
7236 | ||
1e3c88bd | 7237 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 7238 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 7239 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
7240 | if (!cpumask_empty(cpus)) { |
7241 | env.loop = 0; | |
7242 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 7243 | goto redo; |
bbf18b19 | 7244 | } |
afdeee05 | 7245 | goto out_all_pinned; |
1e3c88bd PZ |
7246 | } |
7247 | } | |
7248 | ||
7249 | if (!ld_moved) { | |
7250 | schedstat_inc(sd, lb_failed[idle]); | |
58b26c4c VP |
7251 | /* |
7252 | * Increment the failure counter only on periodic balance. | |
7253 | * We do not want newidle balance, which can be very | |
7254 | * frequent, pollute the failure counter causing | |
7255 | * excessive cache_hot migrations and active balances. | |
7256 | */ | |
7257 | if (idle != CPU_NEWLY_IDLE) | |
7258 | sd->nr_balance_failed++; | |
1e3c88bd | 7259 | |
bd939f45 | 7260 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
7261 | raw_spin_lock_irqsave(&busiest->lock, flags); |
7262 | ||
969c7921 TH |
7263 | /* don't kick the active_load_balance_cpu_stop, |
7264 | * if the curr task on busiest cpu can't be | |
7265 | * moved to this_cpu | |
1e3c88bd PZ |
7266 | */ |
7267 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 7268 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
7269 | raw_spin_unlock_irqrestore(&busiest->lock, |
7270 | flags); | |
8e45cb54 | 7271 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
7272 | goto out_one_pinned; |
7273 | } | |
7274 | ||
969c7921 TH |
7275 | /* |
7276 | * ->active_balance synchronizes accesses to | |
7277 | * ->active_balance_work. Once set, it's cleared | |
7278 | * only after active load balance is finished. | |
7279 | */ | |
1e3c88bd PZ |
7280 | if (!busiest->active_balance) { |
7281 | busiest->active_balance = 1; | |
7282 | busiest->push_cpu = this_cpu; | |
7283 | active_balance = 1; | |
7284 | } | |
7285 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 7286 | |
bd939f45 | 7287 | if (active_balance) { |
969c7921 TH |
7288 | stop_one_cpu_nowait(cpu_of(busiest), |
7289 | active_load_balance_cpu_stop, busiest, | |
7290 | &busiest->active_balance_work); | |
bd939f45 | 7291 | } |
1e3c88bd PZ |
7292 | |
7293 | /* | |
7294 | * We've kicked active balancing, reset the failure | |
7295 | * counter. | |
7296 | */ | |
7297 | sd->nr_balance_failed = sd->cache_nice_tries+1; | |
7298 | } | |
7299 | } else | |
7300 | sd->nr_balance_failed = 0; | |
7301 | ||
7302 | if (likely(!active_balance)) { | |
7303 | /* We were unbalanced, so reset the balancing interval */ | |
7304 | sd->balance_interval = sd->min_interval; | |
7305 | } else { | |
7306 | /* | |
7307 | * If we've begun active balancing, start to back off. This | |
7308 | * case may not be covered by the all_pinned logic if there | |
7309 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 7310 | * detach_tasks). |
1e3c88bd PZ |
7311 | */ |
7312 | if (sd->balance_interval < sd->max_interval) | |
7313 | sd->balance_interval *= 2; | |
7314 | } | |
7315 | ||
1e3c88bd PZ |
7316 | goto out; |
7317 | ||
7318 | out_balanced: | |
afdeee05 VG |
7319 | /* |
7320 | * We reach balance although we may have faced some affinity | |
7321 | * constraints. Clear the imbalance flag if it was set. | |
7322 | */ | |
7323 | if (sd_parent) { | |
7324 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | |
7325 | ||
7326 | if (*group_imbalance) | |
7327 | *group_imbalance = 0; | |
7328 | } | |
7329 | ||
7330 | out_all_pinned: | |
7331 | /* | |
7332 | * We reach balance because all tasks are pinned at this level so | |
7333 | * we can't migrate them. Let the imbalance flag set so parent level | |
7334 | * can try to migrate them. | |
7335 | */ | |
1e3c88bd PZ |
7336 | schedstat_inc(sd, lb_balanced[idle]); |
7337 | ||
7338 | sd->nr_balance_failed = 0; | |
7339 | ||
7340 | out_one_pinned: | |
7341 | /* tune up the balancing interval */ | |
8e45cb54 | 7342 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 7343 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
7344 | (sd->balance_interval < sd->max_interval)) |
7345 | sd->balance_interval *= 2; | |
7346 | ||
46e49b38 | 7347 | ld_moved = 0; |
1e3c88bd | 7348 | out: |
1e3c88bd PZ |
7349 | return ld_moved; |
7350 | } | |
7351 | ||
52a08ef1 JL |
7352 | static inline unsigned long |
7353 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
7354 | { | |
7355 | unsigned long interval = sd->balance_interval; | |
7356 | ||
7357 | if (cpu_busy) | |
7358 | interval *= sd->busy_factor; | |
7359 | ||
7360 | /* scale ms to jiffies */ | |
7361 | interval = msecs_to_jiffies(interval); | |
7362 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
7363 | ||
7364 | return interval; | |
7365 | } | |
7366 | ||
7367 | static inline void | |
7368 | update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) | |
7369 | { | |
7370 | unsigned long interval, next; | |
7371 | ||
7372 | interval = get_sd_balance_interval(sd, cpu_busy); | |
7373 | next = sd->last_balance + interval; | |
7374 | ||
7375 | if (time_after(*next_balance, next)) | |
7376 | *next_balance = next; | |
7377 | } | |
7378 | ||
1e3c88bd PZ |
7379 | /* |
7380 | * idle_balance is called by schedule() if this_cpu is about to become | |
7381 | * idle. Attempts to pull tasks from other CPUs. | |
7382 | */ | |
6e83125c | 7383 | static int idle_balance(struct rq *this_rq) |
1e3c88bd | 7384 | { |
52a08ef1 JL |
7385 | unsigned long next_balance = jiffies + HZ; |
7386 | int this_cpu = this_rq->cpu; | |
1e3c88bd PZ |
7387 | struct sched_domain *sd; |
7388 | int pulled_task = 0; | |
9bd721c5 | 7389 | u64 curr_cost = 0; |
1e3c88bd | 7390 | |
6e83125c PZ |
7391 | /* |
7392 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
7393 | * measure the duration of idle_balance() as idle time. | |
7394 | */ | |
7395 | this_rq->idle_stamp = rq_clock(this_rq); | |
7396 | ||
4486edd1 TC |
7397 | if (this_rq->avg_idle < sysctl_sched_migration_cost || |
7398 | !this_rq->rd->overload) { | |
52a08ef1 JL |
7399 | rcu_read_lock(); |
7400 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
7401 | if (sd) | |
7402 | update_next_balance(sd, 0, &next_balance); | |
7403 | rcu_read_unlock(); | |
7404 | ||
6e83125c | 7405 | goto out; |
52a08ef1 | 7406 | } |
1e3c88bd | 7407 | |
f492e12e PZ |
7408 | raw_spin_unlock(&this_rq->lock); |
7409 | ||
48a16753 | 7410 | update_blocked_averages(this_cpu); |
dce840a0 | 7411 | rcu_read_lock(); |
1e3c88bd | 7412 | for_each_domain(this_cpu, sd) { |
23f0d209 | 7413 | int continue_balancing = 1; |
9bd721c5 | 7414 | u64 t0, domain_cost; |
1e3c88bd PZ |
7415 | |
7416 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
7417 | continue; | |
7418 | ||
52a08ef1 JL |
7419 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
7420 | update_next_balance(sd, 0, &next_balance); | |
9bd721c5 | 7421 | break; |
52a08ef1 | 7422 | } |
9bd721c5 | 7423 | |
f492e12e | 7424 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
7425 | t0 = sched_clock_cpu(this_cpu); |
7426 | ||
f492e12e | 7427 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
7428 | sd, CPU_NEWLY_IDLE, |
7429 | &continue_balancing); | |
9bd721c5 JL |
7430 | |
7431 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
7432 | if (domain_cost > sd->max_newidle_lb_cost) | |
7433 | sd->max_newidle_lb_cost = domain_cost; | |
7434 | ||
7435 | curr_cost += domain_cost; | |
f492e12e | 7436 | } |
1e3c88bd | 7437 | |
52a08ef1 | 7438 | update_next_balance(sd, 0, &next_balance); |
39a4d9ca JL |
7439 | |
7440 | /* | |
7441 | * Stop searching for tasks to pull if there are | |
7442 | * now runnable tasks on this rq. | |
7443 | */ | |
7444 | if (pulled_task || this_rq->nr_running > 0) | |
1e3c88bd | 7445 | break; |
1e3c88bd | 7446 | } |
dce840a0 | 7447 | rcu_read_unlock(); |
f492e12e PZ |
7448 | |
7449 | raw_spin_lock(&this_rq->lock); | |
7450 | ||
0e5b5337 JL |
7451 | if (curr_cost > this_rq->max_idle_balance_cost) |
7452 | this_rq->max_idle_balance_cost = curr_cost; | |
7453 | ||
e5fc6611 | 7454 | /* |
0e5b5337 JL |
7455 | * While browsing the domains, we released the rq lock, a task could |
7456 | * have been enqueued in the meantime. Since we're not going idle, | |
7457 | * pretend we pulled a task. | |
e5fc6611 | 7458 | */ |
0e5b5337 | 7459 | if (this_rq->cfs.h_nr_running && !pulled_task) |
6e83125c | 7460 | pulled_task = 1; |
e5fc6611 | 7461 | |
52a08ef1 JL |
7462 | out: |
7463 | /* Move the next balance forward */ | |
7464 | if (time_after(this_rq->next_balance, next_balance)) | |
1e3c88bd | 7465 | this_rq->next_balance = next_balance; |
9bd721c5 | 7466 | |
e4aa358b | 7467 | /* Is there a task of a high priority class? */ |
46383648 | 7468 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
e4aa358b KT |
7469 | pulled_task = -1; |
7470 | ||
38c6ade2 | 7471 | if (pulled_task) |
6e83125c PZ |
7472 | this_rq->idle_stamp = 0; |
7473 | ||
3c4017c1 | 7474 | return pulled_task; |
1e3c88bd PZ |
7475 | } |
7476 | ||
7477 | /* | |
969c7921 TH |
7478 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
7479 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
7480 | * least 1 task to be running on each physical CPU where possible, and | |
7481 | * avoids physical / logical imbalances. | |
1e3c88bd | 7482 | */ |
969c7921 | 7483 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 7484 | { |
969c7921 TH |
7485 | struct rq *busiest_rq = data; |
7486 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 7487 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 7488 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 7489 | struct sched_domain *sd; |
e5673f28 | 7490 | struct task_struct *p = NULL; |
969c7921 TH |
7491 | |
7492 | raw_spin_lock_irq(&busiest_rq->lock); | |
7493 | ||
7494 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
7495 | if (unlikely(busiest_cpu != smp_processor_id() || | |
7496 | !busiest_rq->active_balance)) | |
7497 | goto out_unlock; | |
1e3c88bd PZ |
7498 | |
7499 | /* Is there any task to move? */ | |
7500 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 7501 | goto out_unlock; |
1e3c88bd PZ |
7502 | |
7503 | /* | |
7504 | * This condition is "impossible", if it occurs | |
7505 | * we need to fix it. Originally reported by | |
7506 | * Bjorn Helgaas on a 128-cpu setup. | |
7507 | */ | |
7508 | BUG_ON(busiest_rq == target_rq); | |
7509 | ||
1e3c88bd | 7510 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 7511 | rcu_read_lock(); |
1e3c88bd PZ |
7512 | for_each_domain(target_cpu, sd) { |
7513 | if ((sd->flags & SD_LOAD_BALANCE) && | |
7514 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
7515 | break; | |
7516 | } | |
7517 | ||
7518 | if (likely(sd)) { | |
8e45cb54 PZ |
7519 | struct lb_env env = { |
7520 | .sd = sd, | |
ddcdf6e7 PZ |
7521 | .dst_cpu = target_cpu, |
7522 | .dst_rq = target_rq, | |
7523 | .src_cpu = busiest_rq->cpu, | |
7524 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
7525 | .idle = CPU_IDLE, |
7526 | }; | |
7527 | ||
1e3c88bd PZ |
7528 | schedstat_inc(sd, alb_count); |
7529 | ||
e5673f28 KT |
7530 | p = detach_one_task(&env); |
7531 | if (p) | |
1e3c88bd PZ |
7532 | schedstat_inc(sd, alb_pushed); |
7533 | else | |
7534 | schedstat_inc(sd, alb_failed); | |
7535 | } | |
dce840a0 | 7536 | rcu_read_unlock(); |
969c7921 TH |
7537 | out_unlock: |
7538 | busiest_rq->active_balance = 0; | |
e5673f28 KT |
7539 | raw_spin_unlock(&busiest_rq->lock); |
7540 | ||
7541 | if (p) | |
7542 | attach_one_task(target_rq, p); | |
7543 | ||
7544 | local_irq_enable(); | |
7545 | ||
969c7921 | 7546 | return 0; |
1e3c88bd PZ |
7547 | } |
7548 | ||
d987fc7f MG |
7549 | static inline int on_null_domain(struct rq *rq) |
7550 | { | |
7551 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
7552 | } | |
7553 | ||
3451d024 | 7554 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
7555 | /* |
7556 | * idle load balancing details | |
83cd4fe2 VP |
7557 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
7558 | * needed, they will kick the idle load balancer, which then does idle | |
7559 | * load balancing for all the idle CPUs. | |
7560 | */ | |
1e3c88bd | 7561 | static struct { |
83cd4fe2 | 7562 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 7563 | atomic_t nr_cpus; |
83cd4fe2 VP |
7564 | unsigned long next_balance; /* in jiffy units */ |
7565 | } nohz ____cacheline_aligned; | |
1e3c88bd | 7566 | |
3dd0337d | 7567 | static inline int find_new_ilb(void) |
1e3c88bd | 7568 | { |
0b005cf5 | 7569 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 7570 | |
786d6dc7 SS |
7571 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
7572 | return ilb; | |
7573 | ||
7574 | return nr_cpu_ids; | |
1e3c88bd | 7575 | } |
1e3c88bd | 7576 | |
83cd4fe2 VP |
7577 | /* |
7578 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
7579 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
7580 | * CPU (if there is one). | |
7581 | */ | |
0aeeeeba | 7582 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
7583 | { |
7584 | int ilb_cpu; | |
7585 | ||
7586 | nohz.next_balance++; | |
7587 | ||
3dd0337d | 7588 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 7589 | |
0b005cf5 SS |
7590 | if (ilb_cpu >= nr_cpu_ids) |
7591 | return; | |
83cd4fe2 | 7592 | |
cd490c5b | 7593 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
7594 | return; |
7595 | /* | |
7596 | * Use smp_send_reschedule() instead of resched_cpu(). | |
7597 | * This way we generate a sched IPI on the target cpu which | |
7598 | * is idle. And the softirq performing nohz idle load balance | |
7599 | * will be run before returning from the IPI. | |
7600 | */ | |
7601 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
7602 | return; |
7603 | } | |
7604 | ||
c1cc017c | 7605 | static inline void nohz_balance_exit_idle(int cpu) |
71325960 SS |
7606 | { |
7607 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
d987fc7f MG |
7608 | /* |
7609 | * Completely isolated CPUs don't ever set, so we must test. | |
7610 | */ | |
7611 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | |
7612 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
7613 | atomic_dec(&nohz.nr_cpus); | |
7614 | } | |
71325960 SS |
7615 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
7616 | } | |
7617 | } | |
7618 | ||
69e1e811 SS |
7619 | static inline void set_cpu_sd_state_busy(void) |
7620 | { | |
7621 | struct sched_domain *sd; | |
37dc6b50 | 7622 | int cpu = smp_processor_id(); |
69e1e811 | 7623 | |
69e1e811 | 7624 | rcu_read_lock(); |
37dc6b50 | 7625 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
7626 | |
7627 | if (!sd || !sd->nohz_idle) | |
7628 | goto unlock; | |
7629 | sd->nohz_idle = 0; | |
7630 | ||
63b2ca30 | 7631 | atomic_inc(&sd->groups->sgc->nr_busy_cpus); |
25f55d9d | 7632 | unlock: |
69e1e811 SS |
7633 | rcu_read_unlock(); |
7634 | } | |
7635 | ||
7636 | void set_cpu_sd_state_idle(void) | |
7637 | { | |
7638 | struct sched_domain *sd; | |
37dc6b50 | 7639 | int cpu = smp_processor_id(); |
69e1e811 | 7640 | |
69e1e811 | 7641 | rcu_read_lock(); |
37dc6b50 | 7642 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
7643 | |
7644 | if (!sd || sd->nohz_idle) | |
7645 | goto unlock; | |
7646 | sd->nohz_idle = 1; | |
7647 | ||
63b2ca30 | 7648 | atomic_dec(&sd->groups->sgc->nr_busy_cpus); |
25f55d9d | 7649 | unlock: |
69e1e811 SS |
7650 | rcu_read_unlock(); |
7651 | } | |
7652 | ||
1e3c88bd | 7653 | /* |
c1cc017c | 7654 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 7655 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 7656 | */ |
c1cc017c | 7657 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 7658 | { |
71325960 SS |
7659 | /* |
7660 | * If this cpu is going down, then nothing needs to be done. | |
7661 | */ | |
7662 | if (!cpu_active(cpu)) | |
7663 | return; | |
7664 | ||
c1cc017c AS |
7665 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
7666 | return; | |
1e3c88bd | 7667 | |
d987fc7f MG |
7668 | /* |
7669 | * If we're a completely isolated CPU, we don't play. | |
7670 | */ | |
7671 | if (on_null_domain(cpu_rq(cpu))) | |
7672 | return; | |
7673 | ||
c1cc017c AS |
7674 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
7675 | atomic_inc(&nohz.nr_cpus); | |
7676 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd | 7677 | } |
71325960 | 7678 | |
0db0628d | 7679 | static int sched_ilb_notifier(struct notifier_block *nfb, |
71325960 SS |
7680 | unsigned long action, void *hcpu) |
7681 | { | |
7682 | switch (action & ~CPU_TASKS_FROZEN) { | |
7683 | case CPU_DYING: | |
c1cc017c | 7684 | nohz_balance_exit_idle(smp_processor_id()); |
71325960 SS |
7685 | return NOTIFY_OK; |
7686 | default: | |
7687 | return NOTIFY_DONE; | |
7688 | } | |
7689 | } | |
1e3c88bd PZ |
7690 | #endif |
7691 | ||
7692 | static DEFINE_SPINLOCK(balancing); | |
7693 | ||
49c022e6 PZ |
7694 | /* |
7695 | * Scale the max load_balance interval with the number of CPUs in the system. | |
7696 | * This trades load-balance latency on larger machines for less cross talk. | |
7697 | */ | |
029632fb | 7698 | void update_max_interval(void) |
49c022e6 PZ |
7699 | { |
7700 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
7701 | } | |
7702 | ||
1e3c88bd PZ |
7703 | /* |
7704 | * It checks each scheduling domain to see if it is due to be balanced, | |
7705 | * and initiates a balancing operation if so. | |
7706 | * | |
b9b0853a | 7707 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 7708 | */ |
f7ed0a89 | 7709 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 7710 | { |
23f0d209 | 7711 | int continue_balancing = 1; |
f7ed0a89 | 7712 | int cpu = rq->cpu; |
1e3c88bd | 7713 | unsigned long interval; |
04f733b4 | 7714 | struct sched_domain *sd; |
1e3c88bd PZ |
7715 | /* Earliest time when we have to do rebalance again */ |
7716 | unsigned long next_balance = jiffies + 60*HZ; | |
7717 | int update_next_balance = 0; | |
f48627e6 JL |
7718 | int need_serialize, need_decay = 0; |
7719 | u64 max_cost = 0; | |
1e3c88bd | 7720 | |
48a16753 | 7721 | update_blocked_averages(cpu); |
2069dd75 | 7722 | |
dce840a0 | 7723 | rcu_read_lock(); |
1e3c88bd | 7724 | for_each_domain(cpu, sd) { |
f48627e6 JL |
7725 | /* |
7726 | * Decay the newidle max times here because this is a regular | |
7727 | * visit to all the domains. Decay ~1% per second. | |
7728 | */ | |
7729 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
7730 | sd->max_newidle_lb_cost = | |
7731 | (sd->max_newidle_lb_cost * 253) / 256; | |
7732 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
7733 | need_decay = 1; | |
7734 | } | |
7735 | max_cost += sd->max_newidle_lb_cost; | |
7736 | ||
1e3c88bd PZ |
7737 | if (!(sd->flags & SD_LOAD_BALANCE)) |
7738 | continue; | |
7739 | ||
f48627e6 JL |
7740 | /* |
7741 | * Stop the load balance at this level. There is another | |
7742 | * CPU in our sched group which is doing load balancing more | |
7743 | * actively. | |
7744 | */ | |
7745 | if (!continue_balancing) { | |
7746 | if (need_decay) | |
7747 | continue; | |
7748 | break; | |
7749 | } | |
7750 | ||
52a08ef1 | 7751 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
7752 | |
7753 | need_serialize = sd->flags & SD_SERIALIZE; | |
1e3c88bd PZ |
7754 | if (need_serialize) { |
7755 | if (!spin_trylock(&balancing)) | |
7756 | goto out; | |
7757 | } | |
7758 | ||
7759 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 7760 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 7761 | /* |
6263322c | 7762 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
7763 | * env->dst_cpu, so we can't know our idle |
7764 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 7765 | */ |
de5eb2dd | 7766 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
7767 | } |
7768 | sd->last_balance = jiffies; | |
52a08ef1 | 7769 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
7770 | } |
7771 | if (need_serialize) | |
7772 | spin_unlock(&balancing); | |
7773 | out: | |
7774 | if (time_after(next_balance, sd->last_balance + interval)) { | |
7775 | next_balance = sd->last_balance + interval; | |
7776 | update_next_balance = 1; | |
7777 | } | |
f48627e6 JL |
7778 | } |
7779 | if (need_decay) { | |
1e3c88bd | 7780 | /* |
f48627e6 JL |
7781 | * Ensure the rq-wide value also decays but keep it at a |
7782 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 7783 | */ |
f48627e6 JL |
7784 | rq->max_idle_balance_cost = |
7785 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 7786 | } |
dce840a0 | 7787 | rcu_read_unlock(); |
1e3c88bd PZ |
7788 | |
7789 | /* | |
7790 | * next_balance will be updated only when there is a need. | |
7791 | * When the cpu is attached to null domain for ex, it will not be | |
7792 | * updated. | |
7793 | */ | |
c5afb6a8 | 7794 | if (likely(update_next_balance)) { |
1e3c88bd | 7795 | rq->next_balance = next_balance; |
c5afb6a8 VG |
7796 | |
7797 | #ifdef CONFIG_NO_HZ_COMMON | |
7798 | /* | |
7799 | * If this CPU has been elected to perform the nohz idle | |
7800 | * balance. Other idle CPUs have already rebalanced with | |
7801 | * nohz_idle_balance() and nohz.next_balance has been | |
7802 | * updated accordingly. This CPU is now running the idle load | |
7803 | * balance for itself and we need to update the | |
7804 | * nohz.next_balance accordingly. | |
7805 | */ | |
7806 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
7807 | nohz.next_balance = rq->next_balance; | |
7808 | #endif | |
7809 | } | |
1e3c88bd PZ |
7810 | } |
7811 | ||
3451d024 | 7812 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 7813 | /* |
3451d024 | 7814 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
7815 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
7816 | */ | |
208cb16b | 7817 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 7818 | { |
208cb16b | 7819 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
7820 | struct rq *rq; |
7821 | int balance_cpu; | |
c5afb6a8 VG |
7822 | /* Earliest time when we have to do rebalance again */ |
7823 | unsigned long next_balance = jiffies + 60*HZ; | |
7824 | int update_next_balance = 0; | |
83cd4fe2 | 7825 | |
1c792db7 SS |
7826 | if (idle != CPU_IDLE || |
7827 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
7828 | goto end; | |
83cd4fe2 VP |
7829 | |
7830 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 7831 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
7832 | continue; |
7833 | ||
7834 | /* | |
7835 | * If this cpu gets work to do, stop the load balancing | |
7836 | * work being done for other cpus. Next load | |
7837 | * balancing owner will pick it up. | |
7838 | */ | |
1c792db7 | 7839 | if (need_resched()) |
83cd4fe2 | 7840 | break; |
83cd4fe2 | 7841 | |
5ed4f1d9 VG |
7842 | rq = cpu_rq(balance_cpu); |
7843 | ||
ed61bbc6 TC |
7844 | /* |
7845 | * If time for next balance is due, | |
7846 | * do the balance. | |
7847 | */ | |
7848 | if (time_after_eq(jiffies, rq->next_balance)) { | |
7849 | raw_spin_lock_irq(&rq->lock); | |
7850 | update_rq_clock(rq); | |
7851 | update_idle_cpu_load(rq); | |
7852 | raw_spin_unlock_irq(&rq->lock); | |
7853 | rebalance_domains(rq, CPU_IDLE); | |
7854 | } | |
83cd4fe2 | 7855 | |
c5afb6a8 VG |
7856 | if (time_after(next_balance, rq->next_balance)) { |
7857 | next_balance = rq->next_balance; | |
7858 | update_next_balance = 1; | |
7859 | } | |
83cd4fe2 | 7860 | } |
c5afb6a8 VG |
7861 | |
7862 | /* | |
7863 | * next_balance will be updated only when there is a need. | |
7864 | * When the CPU is attached to null domain for ex, it will not be | |
7865 | * updated. | |
7866 | */ | |
7867 | if (likely(update_next_balance)) | |
7868 | nohz.next_balance = next_balance; | |
1c792db7 SS |
7869 | end: |
7870 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
7871 | } |
7872 | ||
7873 | /* | |
0b005cf5 | 7874 | * Current heuristic for kicking the idle load balancer in the presence |
1aaf90a4 | 7875 | * of an idle cpu in the system. |
0b005cf5 | 7876 | * - This rq has more than one task. |
1aaf90a4 VG |
7877 | * - This rq has at least one CFS task and the capacity of the CPU is |
7878 | * significantly reduced because of RT tasks or IRQs. | |
7879 | * - At parent of LLC scheduler domain level, this cpu's scheduler group has | |
7880 | * multiple busy cpu. | |
0b005cf5 SS |
7881 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler |
7882 | * domain span are idle. | |
83cd4fe2 | 7883 | */ |
1aaf90a4 | 7884 | static inline bool nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
7885 | { |
7886 | unsigned long now = jiffies; | |
0b005cf5 | 7887 | struct sched_domain *sd; |
63b2ca30 | 7888 | struct sched_group_capacity *sgc; |
4a725627 | 7889 | int nr_busy, cpu = rq->cpu; |
1aaf90a4 | 7890 | bool kick = false; |
83cd4fe2 | 7891 | |
4a725627 | 7892 | if (unlikely(rq->idle_balance)) |
1aaf90a4 | 7893 | return false; |
83cd4fe2 | 7894 | |
1c792db7 SS |
7895 | /* |
7896 | * We may be recently in ticked or tickless idle mode. At the first | |
7897 | * busy tick after returning from idle, we will update the busy stats. | |
7898 | */ | |
69e1e811 | 7899 | set_cpu_sd_state_busy(); |
c1cc017c | 7900 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
7901 | |
7902 | /* | |
7903 | * None are in tickless mode and hence no need for NOHZ idle load | |
7904 | * balancing. | |
7905 | */ | |
7906 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
1aaf90a4 | 7907 | return false; |
1c792db7 SS |
7908 | |
7909 | if (time_before(now, nohz.next_balance)) | |
1aaf90a4 | 7910 | return false; |
83cd4fe2 | 7911 | |
0b005cf5 | 7912 | if (rq->nr_running >= 2) |
1aaf90a4 | 7913 | return true; |
83cd4fe2 | 7914 | |
067491b7 | 7915 | rcu_read_lock(); |
37dc6b50 | 7916 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
37dc6b50 | 7917 | if (sd) { |
63b2ca30 NP |
7918 | sgc = sd->groups->sgc; |
7919 | nr_busy = atomic_read(&sgc->nr_busy_cpus); | |
0b005cf5 | 7920 | |
1aaf90a4 VG |
7921 | if (nr_busy > 1) { |
7922 | kick = true; | |
7923 | goto unlock; | |
7924 | } | |
7925 | ||
83cd4fe2 | 7926 | } |
37dc6b50 | 7927 | |
1aaf90a4 VG |
7928 | sd = rcu_dereference(rq->sd); |
7929 | if (sd) { | |
7930 | if ((rq->cfs.h_nr_running >= 1) && | |
7931 | check_cpu_capacity(rq, sd)) { | |
7932 | kick = true; | |
7933 | goto unlock; | |
7934 | } | |
7935 | } | |
37dc6b50 | 7936 | |
1aaf90a4 | 7937 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); |
37dc6b50 | 7938 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, |
1aaf90a4 VG |
7939 | sched_domain_span(sd)) < cpu)) { |
7940 | kick = true; | |
7941 | goto unlock; | |
7942 | } | |
067491b7 | 7943 | |
1aaf90a4 | 7944 | unlock: |
067491b7 | 7945 | rcu_read_unlock(); |
1aaf90a4 | 7946 | return kick; |
83cd4fe2 VP |
7947 | } |
7948 | #else | |
208cb16b | 7949 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
7950 | #endif |
7951 | ||
7952 | /* | |
7953 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
7954 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
7955 | */ | |
1e3c88bd PZ |
7956 | static void run_rebalance_domains(struct softirq_action *h) |
7957 | { | |
208cb16b | 7958 | struct rq *this_rq = this_rq(); |
6eb57e0d | 7959 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
7960 | CPU_IDLE : CPU_NOT_IDLE; |
7961 | ||
1e3c88bd | 7962 | /* |
83cd4fe2 | 7963 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd | 7964 | * balancing on behalf of the other idle cpus whose ticks are |
d4573c3e PM |
7965 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
7966 | * give the idle cpus a chance to load balance. Else we may | |
7967 | * load balance only within the local sched_domain hierarchy | |
7968 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 7969 | */ |
208cb16b | 7970 | nohz_idle_balance(this_rq, idle); |
d4573c3e | 7971 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
7972 | } |
7973 | ||
1e3c88bd PZ |
7974 | /* |
7975 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 7976 | */ |
7caff66f | 7977 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 7978 | { |
1e3c88bd | 7979 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
7980 | if (unlikely(on_null_domain(rq))) |
7981 | return; | |
7982 | ||
7983 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 7984 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 7985 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 7986 | if (nohz_kick_needed(rq)) |
0aeeeeba | 7987 | nohz_balancer_kick(); |
83cd4fe2 | 7988 | #endif |
1e3c88bd PZ |
7989 | } |
7990 | ||
0bcdcf28 CE |
7991 | static void rq_online_fair(struct rq *rq) |
7992 | { | |
7993 | update_sysctl(); | |
0e59bdae KT |
7994 | |
7995 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
7996 | } |
7997 | ||
7998 | static void rq_offline_fair(struct rq *rq) | |
7999 | { | |
8000 | update_sysctl(); | |
a4c96ae3 PB |
8001 | |
8002 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
8003 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
8004 | } |
8005 | ||
55e12e5e | 8006 | #endif /* CONFIG_SMP */ |
e1d1484f | 8007 | |
bf0f6f24 IM |
8008 | /* |
8009 | * scheduler tick hitting a task of our scheduling class: | |
8010 | */ | |
8f4d37ec | 8011 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
8012 | { |
8013 | struct cfs_rq *cfs_rq; | |
8014 | struct sched_entity *se = &curr->se; | |
8015 | ||
8016 | for_each_sched_entity(se) { | |
8017 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 8018 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 8019 | } |
18bf2805 | 8020 | |
b52da86e | 8021 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 8022 | task_tick_numa(rq, curr); |
bf0f6f24 IM |
8023 | } |
8024 | ||
8025 | /* | |
cd29fe6f PZ |
8026 | * called on fork with the child task as argument from the parent's context |
8027 | * - child not yet on the tasklist | |
8028 | * - preemption disabled | |
bf0f6f24 | 8029 | */ |
cd29fe6f | 8030 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 8031 | { |
4fc420c9 DN |
8032 | struct cfs_rq *cfs_rq; |
8033 | struct sched_entity *se = &p->se, *curr; | |
00bf7bfc | 8034 | int this_cpu = smp_processor_id(); |
cd29fe6f PZ |
8035 | struct rq *rq = this_rq(); |
8036 | unsigned long flags; | |
8037 | ||
05fa785c | 8038 | raw_spin_lock_irqsave(&rq->lock, flags); |
bf0f6f24 | 8039 | |
861d034e PZ |
8040 | update_rq_clock(rq); |
8041 | ||
4fc420c9 DN |
8042 | cfs_rq = task_cfs_rq(current); |
8043 | curr = cfs_rq->curr; | |
8044 | ||
6c9a27f5 DN |
8045 | /* |
8046 | * Not only the cpu but also the task_group of the parent might have | |
8047 | * been changed after parent->se.parent,cfs_rq were copied to | |
8048 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | |
8049 | * of child point to valid ones. | |
8050 | */ | |
8051 | rcu_read_lock(); | |
8052 | __set_task_cpu(p, this_cpu); | |
8053 | rcu_read_unlock(); | |
bf0f6f24 | 8054 | |
7109c442 | 8055 | update_curr(cfs_rq); |
cd29fe6f | 8056 | |
b5d9d734 MG |
8057 | if (curr) |
8058 | se->vruntime = curr->vruntime; | |
aeb73b04 | 8059 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 8060 | |
cd29fe6f | 8061 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 8062 | /* |
edcb60a3 IM |
8063 | * Upon rescheduling, sched_class::put_prev_task() will place |
8064 | * 'current' within the tree based on its new key value. | |
8065 | */ | |
4d78e7b6 | 8066 | swap(curr->vruntime, se->vruntime); |
8875125e | 8067 | resched_curr(rq); |
4d78e7b6 | 8068 | } |
bf0f6f24 | 8069 | |
88ec22d3 PZ |
8070 | se->vruntime -= cfs_rq->min_vruntime; |
8071 | ||
05fa785c | 8072 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
bf0f6f24 IM |
8073 | } |
8074 | ||
cb469845 SR |
8075 | /* |
8076 | * Priority of the task has changed. Check to see if we preempt | |
8077 | * the current task. | |
8078 | */ | |
da7a735e PZ |
8079 | static void |
8080 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 8081 | { |
da0c1e65 | 8082 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
8083 | return; |
8084 | ||
cb469845 SR |
8085 | /* |
8086 | * Reschedule if we are currently running on this runqueue and | |
8087 | * our priority decreased, or if we are not currently running on | |
8088 | * this runqueue and our priority is higher than the current's | |
8089 | */ | |
da7a735e | 8090 | if (rq->curr == p) { |
cb469845 | 8091 | if (p->prio > oldprio) |
8875125e | 8092 | resched_curr(rq); |
cb469845 | 8093 | } else |
15afe09b | 8094 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
8095 | } |
8096 | ||
daa59407 | 8097 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
8098 | { |
8099 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
8100 | |
8101 | /* | |
daa59407 BP |
8102 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
8103 | * the dequeue_entity(.flags=0) will already have normalized the | |
8104 | * vruntime. | |
8105 | */ | |
8106 | if (p->on_rq) | |
8107 | return true; | |
8108 | ||
8109 | /* | |
8110 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
8111 | * But there are some cases where it has already been normalized: | |
da7a735e | 8112 | * |
daa59407 BP |
8113 | * - A forked child which is waiting for being woken up by |
8114 | * wake_up_new_task(). | |
8115 | * - A task which has been woken up by try_to_wake_up() and | |
8116 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 8117 | */ |
daa59407 BP |
8118 | if (!se->sum_exec_runtime || p->state == TASK_WAKING) |
8119 | return true; | |
8120 | ||
8121 | return false; | |
8122 | } | |
8123 | ||
8124 | static void detach_task_cfs_rq(struct task_struct *p) | |
8125 | { | |
8126 | struct sched_entity *se = &p->se; | |
8127 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
8128 | ||
8129 | if (!vruntime_normalized(p)) { | |
da7a735e PZ |
8130 | /* |
8131 | * Fix up our vruntime so that the current sleep doesn't | |
8132 | * cause 'unlimited' sleep bonus. | |
8133 | */ | |
8134 | place_entity(cfs_rq, se, 0); | |
8135 | se->vruntime -= cfs_rq->min_vruntime; | |
8136 | } | |
9ee474f5 | 8137 | |
9d89c257 | 8138 | /* Catch up with the cfs_rq and remove our load when we leave */ |
a05e8c51 | 8139 | detach_entity_load_avg(cfs_rq, se); |
da7a735e PZ |
8140 | } |
8141 | ||
daa59407 | 8142 | static void attach_task_cfs_rq(struct task_struct *p) |
cb469845 | 8143 | { |
f36c019c | 8144 | struct sched_entity *se = &p->se; |
daa59407 | 8145 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
8146 | |
8147 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
8148 | /* |
8149 | * Since the real-depth could have been changed (only FAIR | |
8150 | * class maintain depth value), reset depth properly. | |
8151 | */ | |
8152 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
8153 | #endif | |
7855a35a | 8154 | |
6efdb105 | 8155 | /* Synchronize task with its cfs_rq */ |
daa59407 BP |
8156 | attach_entity_load_avg(cfs_rq, se); |
8157 | ||
8158 | if (!vruntime_normalized(p)) | |
8159 | se->vruntime += cfs_rq->min_vruntime; | |
8160 | } | |
6efdb105 | 8161 | |
daa59407 BP |
8162 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
8163 | { | |
8164 | detach_task_cfs_rq(p); | |
8165 | } | |
8166 | ||
8167 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
8168 | { | |
8169 | attach_task_cfs_rq(p); | |
7855a35a | 8170 | |
daa59407 | 8171 | if (task_on_rq_queued(p)) { |
7855a35a | 8172 | /* |
daa59407 BP |
8173 | * We were most likely switched from sched_rt, so |
8174 | * kick off the schedule if running, otherwise just see | |
8175 | * if we can still preempt the current task. | |
7855a35a | 8176 | */ |
daa59407 BP |
8177 | if (rq->curr == p) |
8178 | resched_curr(rq); | |
8179 | else | |
8180 | check_preempt_curr(rq, p, 0); | |
7855a35a | 8181 | } |
cb469845 SR |
8182 | } |
8183 | ||
83b699ed SV |
8184 | /* Account for a task changing its policy or group. |
8185 | * | |
8186 | * This routine is mostly called to set cfs_rq->curr field when a task | |
8187 | * migrates between groups/classes. | |
8188 | */ | |
8189 | static void set_curr_task_fair(struct rq *rq) | |
8190 | { | |
8191 | struct sched_entity *se = &rq->curr->se; | |
8192 | ||
ec12cb7f PT |
8193 | for_each_sched_entity(se) { |
8194 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
8195 | ||
8196 | set_next_entity(cfs_rq, se); | |
8197 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
8198 | account_cfs_rq_runtime(cfs_rq, 0); | |
8199 | } | |
83b699ed SV |
8200 | } |
8201 | ||
029632fb PZ |
8202 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
8203 | { | |
8204 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
8205 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
8206 | #ifndef CONFIG_64BIT | |
8207 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
8208 | #endif | |
141965c7 | 8209 | #ifdef CONFIG_SMP |
9d89c257 YD |
8210 | atomic_long_set(&cfs_rq->removed_load_avg, 0); |
8211 | atomic_long_set(&cfs_rq->removed_util_avg, 0); | |
9ee474f5 | 8212 | #endif |
029632fb PZ |
8213 | } |
8214 | ||
810b3817 | 8215 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bc54da21 | 8216 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 8217 | { |
daa59407 | 8218 | detach_task_cfs_rq(p); |
b2b5ce02 | 8219 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
8220 | |
8221 | #ifdef CONFIG_SMP | |
8222 | /* Tell se's cfs_rq has been changed -- migrated */ | |
8223 | p->se.avg.last_update_time = 0; | |
8224 | #endif | |
daa59407 | 8225 | attach_task_cfs_rq(p); |
810b3817 | 8226 | } |
029632fb PZ |
8227 | |
8228 | void free_fair_sched_group(struct task_group *tg) | |
8229 | { | |
8230 | int i; | |
8231 | ||
8232 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
8233 | ||
8234 | for_each_possible_cpu(i) { | |
8235 | if (tg->cfs_rq) | |
8236 | kfree(tg->cfs_rq[i]); | |
12695578 YD |
8237 | if (tg->se) { |
8238 | if (tg->se[i]) | |
8239 | remove_entity_load_avg(tg->se[i]); | |
029632fb | 8240 | kfree(tg->se[i]); |
12695578 | 8241 | } |
029632fb PZ |
8242 | } |
8243 | ||
8244 | kfree(tg->cfs_rq); | |
8245 | kfree(tg->se); | |
8246 | } | |
8247 | ||
8248 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8249 | { | |
8250 | struct cfs_rq *cfs_rq; | |
8251 | struct sched_entity *se; | |
8252 | int i; | |
8253 | ||
8254 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
8255 | if (!tg->cfs_rq) | |
8256 | goto err; | |
8257 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
8258 | if (!tg->se) | |
8259 | goto err; | |
8260 | ||
8261 | tg->shares = NICE_0_LOAD; | |
8262 | ||
8263 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
8264 | ||
8265 | for_each_possible_cpu(i) { | |
8266 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
8267 | GFP_KERNEL, cpu_to_node(i)); | |
8268 | if (!cfs_rq) | |
8269 | goto err; | |
8270 | ||
8271 | se = kzalloc_node(sizeof(struct sched_entity), | |
8272 | GFP_KERNEL, cpu_to_node(i)); | |
8273 | if (!se) | |
8274 | goto err_free_rq; | |
8275 | ||
8276 | init_cfs_rq(cfs_rq); | |
8277 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 8278 | init_entity_runnable_average(se); |
029632fb PZ |
8279 | } |
8280 | ||
8281 | return 1; | |
8282 | ||
8283 | err_free_rq: | |
8284 | kfree(cfs_rq); | |
8285 | err: | |
8286 | return 0; | |
8287 | } | |
8288 | ||
8289 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8290 | { | |
8291 | struct rq *rq = cpu_rq(cpu); | |
8292 | unsigned long flags; | |
8293 | ||
8294 | /* | |
8295 | * Only empty task groups can be destroyed; so we can speculatively | |
8296 | * check on_list without danger of it being re-added. | |
8297 | */ | |
8298 | if (!tg->cfs_rq[cpu]->on_list) | |
8299 | return; | |
8300 | ||
8301 | raw_spin_lock_irqsave(&rq->lock, flags); | |
8302 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
8303 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8304 | } | |
8305 | ||
8306 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
8307 | struct sched_entity *se, int cpu, | |
8308 | struct sched_entity *parent) | |
8309 | { | |
8310 | struct rq *rq = cpu_rq(cpu); | |
8311 | ||
8312 | cfs_rq->tg = tg; | |
8313 | cfs_rq->rq = rq; | |
029632fb PZ |
8314 | init_cfs_rq_runtime(cfs_rq); |
8315 | ||
8316 | tg->cfs_rq[cpu] = cfs_rq; | |
8317 | tg->se[cpu] = se; | |
8318 | ||
8319 | /* se could be NULL for root_task_group */ | |
8320 | if (!se) | |
8321 | return; | |
8322 | ||
fed14d45 | 8323 | if (!parent) { |
029632fb | 8324 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
8325 | se->depth = 0; |
8326 | } else { | |
029632fb | 8327 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
8328 | se->depth = parent->depth + 1; |
8329 | } | |
029632fb PZ |
8330 | |
8331 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
8332 | /* guarantee group entities always have weight */ |
8333 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
8334 | se->parent = parent; |
8335 | } | |
8336 | ||
8337 | static DEFINE_MUTEX(shares_mutex); | |
8338 | ||
8339 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
8340 | { | |
8341 | int i; | |
8342 | unsigned long flags; | |
8343 | ||
8344 | /* | |
8345 | * We can't change the weight of the root cgroup. | |
8346 | */ | |
8347 | if (!tg->se[0]) | |
8348 | return -EINVAL; | |
8349 | ||
8350 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
8351 | ||
8352 | mutex_lock(&shares_mutex); | |
8353 | if (tg->shares == shares) | |
8354 | goto done; | |
8355 | ||
8356 | tg->shares = shares; | |
8357 | for_each_possible_cpu(i) { | |
8358 | struct rq *rq = cpu_rq(i); | |
8359 | struct sched_entity *se; | |
8360 | ||
8361 | se = tg->se[i]; | |
8362 | /* Propagate contribution to hierarchy */ | |
8363 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
8364 | |
8365 | /* Possible calls to update_curr() need rq clock */ | |
8366 | update_rq_clock(rq); | |
17bc14b7 | 8367 | for_each_sched_entity(se) |
029632fb PZ |
8368 | update_cfs_shares(group_cfs_rq(se)); |
8369 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8370 | } | |
8371 | ||
8372 | done: | |
8373 | mutex_unlock(&shares_mutex); | |
8374 | return 0; | |
8375 | } | |
8376 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
8377 | ||
8378 | void free_fair_sched_group(struct task_group *tg) { } | |
8379 | ||
8380 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8381 | { | |
8382 | return 1; | |
8383 | } | |
8384 | ||
8385 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | |
8386 | ||
8387 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
8388 | ||
810b3817 | 8389 | |
6d686f45 | 8390 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
8391 | { |
8392 | struct sched_entity *se = &task->se; | |
0d721cea PW |
8393 | unsigned int rr_interval = 0; |
8394 | ||
8395 | /* | |
8396 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
8397 | * idle runqueue: | |
8398 | */ | |
0d721cea | 8399 | if (rq->cfs.load.weight) |
a59f4e07 | 8400 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
8401 | |
8402 | return rr_interval; | |
8403 | } | |
8404 | ||
bf0f6f24 IM |
8405 | /* |
8406 | * All the scheduling class methods: | |
8407 | */ | |
029632fb | 8408 | const struct sched_class fair_sched_class = { |
5522d5d5 | 8409 | .next = &idle_sched_class, |
bf0f6f24 IM |
8410 | .enqueue_task = enqueue_task_fair, |
8411 | .dequeue_task = dequeue_task_fair, | |
8412 | .yield_task = yield_task_fair, | |
d95f4122 | 8413 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 8414 | |
2e09bf55 | 8415 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
8416 | |
8417 | .pick_next_task = pick_next_task_fair, | |
8418 | .put_prev_task = put_prev_task_fair, | |
8419 | ||
681f3e68 | 8420 | #ifdef CONFIG_SMP |
4ce72a2c | 8421 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 8422 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 8423 | |
0bcdcf28 CE |
8424 | .rq_online = rq_online_fair, |
8425 | .rq_offline = rq_offline_fair, | |
88ec22d3 PZ |
8426 | |
8427 | .task_waking = task_waking_fair, | |
12695578 | 8428 | .task_dead = task_dead_fair, |
c5b28038 | 8429 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 8430 | #endif |
bf0f6f24 | 8431 | |
83b699ed | 8432 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 8433 | .task_tick = task_tick_fair, |
cd29fe6f | 8434 | .task_fork = task_fork_fair, |
cb469845 SR |
8435 | |
8436 | .prio_changed = prio_changed_fair, | |
da7a735e | 8437 | .switched_from = switched_from_fair, |
cb469845 | 8438 | .switched_to = switched_to_fair, |
810b3817 | 8439 | |
0d721cea PW |
8440 | .get_rr_interval = get_rr_interval_fair, |
8441 | ||
6e998916 SG |
8442 | .update_curr = update_curr_fair, |
8443 | ||
810b3817 | 8444 | #ifdef CONFIG_FAIR_GROUP_SCHED |
b2b5ce02 | 8445 | .task_move_group = task_move_group_fair, |
810b3817 | 8446 | #endif |
bf0f6f24 IM |
8447 | }; |
8448 | ||
8449 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 8450 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 8451 | { |
bf0f6f24 IM |
8452 | struct cfs_rq *cfs_rq; |
8453 | ||
5973e5b9 | 8454 | rcu_read_lock(); |
c3b64f1e | 8455 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 8456 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 8457 | rcu_read_unlock(); |
bf0f6f24 | 8458 | } |
397f2378 SD |
8459 | |
8460 | #ifdef CONFIG_NUMA_BALANCING | |
8461 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
8462 | { | |
8463 | int node; | |
8464 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
8465 | ||
8466 | for_each_online_node(node) { | |
8467 | if (p->numa_faults) { | |
8468 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
8469 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
8470 | } | |
8471 | if (p->numa_group) { | |
8472 | gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
8473 | gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
8474 | } | |
8475 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
8476 | } | |
8477 | } | |
8478 | #endif /* CONFIG_NUMA_BALANCING */ | |
8479 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
8480 | |
8481 | __init void init_sched_fair_class(void) | |
8482 | { | |
8483 | #ifdef CONFIG_SMP | |
8484 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
8485 | ||
3451d024 | 8486 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 8487 | nohz.next_balance = jiffies; |
029632fb | 8488 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
71325960 | 8489 | cpu_notifier(sched_ilb_notifier, 0); |
029632fb PZ |
8490 | #endif |
8491 | #endif /* SMP */ | |
8492 | ||
8493 | } |