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 | ||
1983a922 | 23 | #include <linux/sched.h> |
cb251765 | 24 | #include <linux/latencytop.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 | ||
3273163c MR |
117 | /* |
118 | * The margin used when comparing utilization with CPU capacity: | |
119 | * util * 1024 < capacity * margin | |
120 | */ | |
121 | unsigned int capacity_margin = 1280; /* ~20% */ | |
122 | ||
8527632d PG |
123 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
124 | { | |
125 | lw->weight += inc; | |
126 | lw->inv_weight = 0; | |
127 | } | |
128 | ||
129 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
130 | { | |
131 | lw->weight -= dec; | |
132 | lw->inv_weight = 0; | |
133 | } | |
134 | ||
135 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
136 | { | |
137 | lw->weight = w; | |
138 | lw->inv_weight = 0; | |
139 | } | |
140 | ||
029632fb PZ |
141 | /* |
142 | * Increase the granularity value when there are more CPUs, | |
143 | * because with more CPUs the 'effective latency' as visible | |
144 | * to users decreases. But the relationship is not linear, | |
145 | * so pick a second-best guess by going with the log2 of the | |
146 | * number of CPUs. | |
147 | * | |
148 | * This idea comes from the SD scheduler of Con Kolivas: | |
149 | */ | |
58ac93e4 | 150 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 151 | { |
58ac93e4 | 152 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
153 | unsigned int factor; |
154 | ||
155 | switch (sysctl_sched_tunable_scaling) { | |
156 | case SCHED_TUNABLESCALING_NONE: | |
157 | factor = 1; | |
158 | break; | |
159 | case SCHED_TUNABLESCALING_LINEAR: | |
160 | factor = cpus; | |
161 | break; | |
162 | case SCHED_TUNABLESCALING_LOG: | |
163 | default: | |
164 | factor = 1 + ilog2(cpus); | |
165 | break; | |
166 | } | |
167 | ||
168 | return factor; | |
169 | } | |
170 | ||
171 | static void update_sysctl(void) | |
172 | { | |
173 | unsigned int factor = get_update_sysctl_factor(); | |
174 | ||
175 | #define SET_SYSCTL(name) \ | |
176 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
177 | SET_SYSCTL(sched_min_granularity); | |
178 | SET_SYSCTL(sched_latency); | |
179 | SET_SYSCTL(sched_wakeup_granularity); | |
180 | #undef SET_SYSCTL | |
181 | } | |
182 | ||
183 | void sched_init_granularity(void) | |
184 | { | |
185 | update_sysctl(); | |
186 | } | |
187 | ||
9dbdb155 | 188 | #define WMULT_CONST (~0U) |
029632fb PZ |
189 | #define WMULT_SHIFT 32 |
190 | ||
9dbdb155 PZ |
191 | static void __update_inv_weight(struct load_weight *lw) |
192 | { | |
193 | unsigned long w; | |
194 | ||
195 | if (likely(lw->inv_weight)) | |
196 | return; | |
197 | ||
198 | w = scale_load_down(lw->weight); | |
199 | ||
200 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
201 | lw->inv_weight = 1; | |
202 | else if (unlikely(!w)) | |
203 | lw->inv_weight = WMULT_CONST; | |
204 | else | |
205 | lw->inv_weight = WMULT_CONST / w; | |
206 | } | |
029632fb PZ |
207 | |
208 | /* | |
9dbdb155 PZ |
209 | * delta_exec * weight / lw.weight |
210 | * OR | |
211 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
212 | * | |
1c3de5e1 | 213 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
214 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
215 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
216 | * | |
217 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
218 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 219 | */ |
9dbdb155 | 220 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 221 | { |
9dbdb155 PZ |
222 | u64 fact = scale_load_down(weight); |
223 | int shift = WMULT_SHIFT; | |
029632fb | 224 | |
9dbdb155 | 225 | __update_inv_weight(lw); |
029632fb | 226 | |
9dbdb155 PZ |
227 | if (unlikely(fact >> 32)) { |
228 | while (fact >> 32) { | |
229 | fact >>= 1; | |
230 | shift--; | |
231 | } | |
029632fb PZ |
232 | } |
233 | ||
9dbdb155 PZ |
234 | /* hint to use a 32x32->64 mul */ |
235 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 236 | |
9dbdb155 PZ |
237 | while (fact >> 32) { |
238 | fact >>= 1; | |
239 | shift--; | |
240 | } | |
029632fb | 241 | |
9dbdb155 | 242 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
243 | } |
244 | ||
245 | ||
246 | const struct sched_class fair_sched_class; | |
a4c2f00f | 247 | |
bf0f6f24 IM |
248 | /************************************************************** |
249 | * CFS operations on generic schedulable entities: | |
250 | */ | |
251 | ||
62160e3f | 252 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 253 | |
62160e3f | 254 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
255 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
256 | { | |
62160e3f | 257 | return cfs_rq->rq; |
bf0f6f24 IM |
258 | } |
259 | ||
62160e3f IM |
260 | /* An entity is a task if it doesn't "own" a runqueue */ |
261 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 262 | |
8f48894f PZ |
263 | static inline struct task_struct *task_of(struct sched_entity *se) |
264 | { | |
265 | #ifdef CONFIG_SCHED_DEBUG | |
266 | WARN_ON_ONCE(!entity_is_task(se)); | |
267 | #endif | |
268 | return container_of(se, struct task_struct, se); | |
269 | } | |
270 | ||
b758149c PZ |
271 | /* Walk up scheduling entities hierarchy */ |
272 | #define for_each_sched_entity(se) \ | |
273 | for (; se; se = se->parent) | |
274 | ||
275 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
276 | { | |
277 | return p->se.cfs_rq; | |
278 | } | |
279 | ||
280 | /* runqueue on which this entity is (to be) queued */ | |
281 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
282 | { | |
283 | return se->cfs_rq; | |
284 | } | |
285 | ||
286 | /* runqueue "owned" by this group */ | |
287 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
288 | { | |
289 | return grp->my_q; | |
290 | } | |
291 | ||
3d4b47b4 PZ |
292 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
293 | { | |
294 | if (!cfs_rq->on_list) { | |
67e86250 PT |
295 | /* |
296 | * Ensure we either appear before our parent (if already | |
297 | * enqueued) or force our parent to appear after us when it is | |
298 | * enqueued. The fact that we always enqueue bottom-up | |
299 | * reduces this to two cases. | |
300 | */ | |
301 | if (cfs_rq->tg->parent && | |
302 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | |
303 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
304 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | |
305 | } else { | |
306 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
3d4b47b4 | 307 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
67e86250 | 308 | } |
3d4b47b4 PZ |
309 | |
310 | cfs_rq->on_list = 1; | |
311 | } | |
312 | } | |
313 | ||
314 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
315 | { | |
316 | if (cfs_rq->on_list) { | |
317 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
318 | cfs_rq->on_list = 0; | |
319 | } | |
320 | } | |
321 | ||
b758149c PZ |
322 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
323 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
324 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
325 | ||
326 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 327 | static inline struct cfs_rq * |
b758149c PZ |
328 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
329 | { | |
330 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 331 | return se->cfs_rq; |
b758149c | 332 | |
fed14d45 | 333 | return NULL; |
b758149c PZ |
334 | } |
335 | ||
336 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
337 | { | |
338 | return se->parent; | |
339 | } | |
340 | ||
464b7527 PZ |
341 | static void |
342 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
343 | { | |
344 | int se_depth, pse_depth; | |
345 | ||
346 | /* | |
347 | * preemption test can be made between sibling entities who are in the | |
348 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
349 | * both tasks until we find their ancestors who are siblings of common | |
350 | * parent. | |
351 | */ | |
352 | ||
353 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
354 | se_depth = (*se)->depth; |
355 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
356 | |
357 | while (se_depth > pse_depth) { | |
358 | se_depth--; | |
359 | *se = parent_entity(*se); | |
360 | } | |
361 | ||
362 | while (pse_depth > se_depth) { | |
363 | pse_depth--; | |
364 | *pse = parent_entity(*pse); | |
365 | } | |
366 | ||
367 | while (!is_same_group(*se, *pse)) { | |
368 | *se = parent_entity(*se); | |
369 | *pse = parent_entity(*pse); | |
370 | } | |
371 | } | |
372 | ||
8f48894f PZ |
373 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
374 | ||
375 | static inline struct task_struct *task_of(struct sched_entity *se) | |
376 | { | |
377 | return container_of(se, struct task_struct, se); | |
378 | } | |
bf0f6f24 | 379 | |
62160e3f IM |
380 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
381 | { | |
382 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
383 | } |
384 | ||
385 | #define entity_is_task(se) 1 | |
386 | ||
b758149c PZ |
387 | #define for_each_sched_entity(se) \ |
388 | for (; se; se = NULL) | |
bf0f6f24 | 389 | |
b758149c | 390 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 391 | { |
b758149c | 392 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
393 | } |
394 | ||
b758149c PZ |
395 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
396 | { | |
397 | struct task_struct *p = task_of(se); | |
398 | struct rq *rq = task_rq(p); | |
399 | ||
400 | return &rq->cfs; | |
401 | } | |
402 | ||
403 | /* runqueue "owned" by this group */ | |
404 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
405 | { | |
406 | return NULL; | |
407 | } | |
408 | ||
3d4b47b4 PZ |
409 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
410 | { | |
411 | } | |
412 | ||
413 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
414 | { | |
415 | } | |
416 | ||
b758149c PZ |
417 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
418 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
419 | ||
b758149c PZ |
420 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
421 | { | |
422 | return NULL; | |
423 | } | |
424 | ||
464b7527 PZ |
425 | static inline void |
426 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
427 | { | |
428 | } | |
429 | ||
b758149c PZ |
430 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
431 | ||
6c16a6dc | 432 | static __always_inline |
9dbdb155 | 433 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
434 | |
435 | /************************************************************** | |
436 | * Scheduling class tree data structure manipulation methods: | |
437 | */ | |
438 | ||
1bf08230 | 439 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 440 | { |
1bf08230 | 441 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 442 | if (delta > 0) |
1bf08230 | 443 | max_vruntime = vruntime; |
02e0431a | 444 | |
1bf08230 | 445 | return max_vruntime; |
02e0431a PZ |
446 | } |
447 | ||
0702e3eb | 448 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
449 | { |
450 | s64 delta = (s64)(vruntime - min_vruntime); | |
451 | if (delta < 0) | |
452 | min_vruntime = vruntime; | |
453 | ||
454 | return min_vruntime; | |
455 | } | |
456 | ||
54fdc581 FC |
457 | static inline int entity_before(struct sched_entity *a, |
458 | struct sched_entity *b) | |
459 | { | |
460 | return (s64)(a->vruntime - b->vruntime) < 0; | |
461 | } | |
462 | ||
1af5f730 PZ |
463 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
464 | { | |
465 | u64 vruntime = cfs_rq->min_vruntime; | |
466 | ||
467 | if (cfs_rq->curr) | |
468 | vruntime = cfs_rq->curr->vruntime; | |
469 | ||
470 | if (cfs_rq->rb_leftmost) { | |
471 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
472 | struct sched_entity, | |
473 | run_node); | |
474 | ||
e17036da | 475 | if (!cfs_rq->curr) |
1af5f730 PZ |
476 | vruntime = se->vruntime; |
477 | else | |
478 | vruntime = min_vruntime(vruntime, se->vruntime); | |
479 | } | |
480 | ||
1bf08230 | 481 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 482 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
483 | #ifndef CONFIG_64BIT |
484 | smp_wmb(); | |
485 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
486 | #endif | |
1af5f730 PZ |
487 | } |
488 | ||
bf0f6f24 IM |
489 | /* |
490 | * Enqueue an entity into the rb-tree: | |
491 | */ | |
0702e3eb | 492 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
493 | { |
494 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
495 | struct rb_node *parent = NULL; | |
496 | struct sched_entity *entry; | |
bf0f6f24 IM |
497 | int leftmost = 1; |
498 | ||
499 | /* | |
500 | * Find the right place in the rbtree: | |
501 | */ | |
502 | while (*link) { | |
503 | parent = *link; | |
504 | entry = rb_entry(parent, struct sched_entity, run_node); | |
505 | /* | |
506 | * We dont care about collisions. Nodes with | |
507 | * the same key stay together. | |
508 | */ | |
2bd2d6f2 | 509 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
510 | link = &parent->rb_left; |
511 | } else { | |
512 | link = &parent->rb_right; | |
513 | leftmost = 0; | |
514 | } | |
515 | } | |
516 | ||
517 | /* | |
518 | * Maintain a cache of leftmost tree entries (it is frequently | |
519 | * used): | |
520 | */ | |
1af5f730 | 521 | if (leftmost) |
57cb499d | 522 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
523 | |
524 | rb_link_node(&se->run_node, parent, link); | |
525 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
526 | } |
527 | ||
0702e3eb | 528 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 529 | { |
3fe69747 PZ |
530 | if (cfs_rq->rb_leftmost == &se->run_node) { |
531 | struct rb_node *next_node; | |
3fe69747 PZ |
532 | |
533 | next_node = rb_next(&se->run_node); | |
534 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 535 | } |
e9acbff6 | 536 | |
bf0f6f24 | 537 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
538 | } |
539 | ||
029632fb | 540 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 541 | { |
f4b6755f PZ |
542 | struct rb_node *left = cfs_rq->rb_leftmost; |
543 | ||
544 | if (!left) | |
545 | return NULL; | |
546 | ||
547 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
548 | } |
549 | ||
ac53db59 RR |
550 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
551 | { | |
552 | struct rb_node *next = rb_next(&se->run_node); | |
553 | ||
554 | if (!next) | |
555 | return NULL; | |
556 | ||
557 | return rb_entry(next, struct sched_entity, run_node); | |
558 | } | |
559 | ||
560 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 561 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 562 | { |
7eee3e67 | 563 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 564 | |
70eee74b BS |
565 | if (!last) |
566 | return NULL; | |
7eee3e67 IM |
567 | |
568 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
569 | } |
570 | ||
bf0f6f24 IM |
571 | /************************************************************** |
572 | * Scheduling class statistics methods: | |
573 | */ | |
574 | ||
acb4a848 | 575 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 576 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
577 | loff_t *ppos) |
578 | { | |
8d65af78 | 579 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 580 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
581 | |
582 | if (ret || !write) | |
583 | return ret; | |
584 | ||
585 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
586 | sysctl_sched_min_granularity); | |
587 | ||
acb4a848 CE |
588 | #define WRT_SYSCTL(name) \ |
589 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
590 | WRT_SYSCTL(sched_min_granularity); | |
591 | WRT_SYSCTL(sched_latency); | |
592 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
593 | #undef WRT_SYSCTL |
594 | ||
b2be5e96 PZ |
595 | return 0; |
596 | } | |
597 | #endif | |
647e7cac | 598 | |
a7be37ac | 599 | /* |
f9c0b095 | 600 | * delta /= w |
a7be37ac | 601 | */ |
9dbdb155 | 602 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 603 | { |
f9c0b095 | 604 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 605 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
606 | |
607 | return delta; | |
608 | } | |
609 | ||
647e7cac IM |
610 | /* |
611 | * The idea is to set a period in which each task runs once. | |
612 | * | |
532b1858 | 613 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
614 | * this period because otherwise the slices get too small. |
615 | * | |
616 | * p = (nr <= nl) ? l : l*nr/nl | |
617 | */ | |
4d78e7b6 PZ |
618 | static u64 __sched_period(unsigned long nr_running) |
619 | { | |
8e2b0bf3 BF |
620 | if (unlikely(nr_running > sched_nr_latency)) |
621 | return nr_running * sysctl_sched_min_granularity; | |
622 | else | |
623 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
624 | } |
625 | ||
647e7cac IM |
626 | /* |
627 | * We calculate the wall-time slice from the period by taking a part | |
628 | * proportional to the weight. | |
629 | * | |
f9c0b095 | 630 | * s = p*P[w/rw] |
647e7cac | 631 | */ |
6d0f0ebd | 632 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 633 | { |
0a582440 | 634 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 635 | |
0a582440 | 636 | for_each_sched_entity(se) { |
6272d68c | 637 | struct load_weight *load; |
3104bf03 | 638 | struct load_weight lw; |
6272d68c LM |
639 | |
640 | cfs_rq = cfs_rq_of(se); | |
641 | load = &cfs_rq->load; | |
f9c0b095 | 642 | |
0a582440 | 643 | if (unlikely(!se->on_rq)) { |
3104bf03 | 644 | lw = cfs_rq->load; |
0a582440 MG |
645 | |
646 | update_load_add(&lw, se->load.weight); | |
647 | load = &lw; | |
648 | } | |
9dbdb155 | 649 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
650 | } |
651 | return slice; | |
bf0f6f24 IM |
652 | } |
653 | ||
647e7cac | 654 | /* |
660cc00f | 655 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 656 | * |
f9c0b095 | 657 | * vs = s/w |
647e7cac | 658 | */ |
f9c0b095 | 659 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 660 | { |
f9c0b095 | 661 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
662 | } |
663 | ||
a75cdaa9 | 664 | #ifdef CONFIG_SMP |
772bd008 | 665 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee MG |
666 | static unsigned long task_h_load(struct task_struct *p); |
667 | ||
9d89c257 YD |
668 | /* |
669 | * We choose a half-life close to 1 scheduling period. | |
84fb5a18 LY |
670 | * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are |
671 | * dependent on this value. | |
9d89c257 YD |
672 | */ |
673 | #define LOAD_AVG_PERIOD 32 | |
674 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
84fb5a18 | 675 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */ |
a75cdaa9 | 676 | |
540247fb YD |
677 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
678 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 679 | { |
540247fb | 680 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 681 | |
9d89c257 YD |
682 | sa->last_update_time = 0; |
683 | /* | |
684 | * sched_avg's period_contrib should be strictly less then 1024, so | |
685 | * we give it 1023 to make sure it is almost a period (1024us), and | |
686 | * will definitely be update (after enqueue). | |
687 | */ | |
688 | sa->period_contrib = 1023; | |
540247fb | 689 | sa->load_avg = scale_load_down(se->load.weight); |
9d89c257 | 690 | sa->load_sum = sa->load_avg * LOAD_AVG_MAX; |
2b8c41da YD |
691 | /* |
692 | * At this point, util_avg won't be used in select_task_rq_fair anyway | |
693 | */ | |
694 | sa->util_avg = 0; | |
695 | sa->util_sum = 0; | |
9d89c257 | 696 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 697 | } |
7ea241af | 698 | |
7dc603c9 PZ |
699 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
700 | static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq); | |
3d30544f | 701 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force); |
7dc603c9 PZ |
702 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se); |
703 | ||
2b8c41da YD |
704 | /* |
705 | * With new tasks being created, their initial util_avgs are extrapolated | |
706 | * based on the cfs_rq's current util_avg: | |
707 | * | |
708 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
709 | * | |
710 | * However, in many cases, the above util_avg does not give a desired | |
711 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
712 | * as when the series is a harmonic series. | |
713 | * | |
714 | * To solve this problem, we also cap the util_avg of successive tasks to | |
715 | * only 1/2 of the left utilization budget: | |
716 | * | |
717 | * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n | |
718 | * | |
719 | * where n denotes the nth task. | |
720 | * | |
721 | * For example, a simplest series from the beginning would be like: | |
722 | * | |
723 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
724 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
725 | * | |
726 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
727 | * if util_avg > util_avg_cap. | |
728 | */ | |
729 | void post_init_entity_util_avg(struct sched_entity *se) | |
730 | { | |
731 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
732 | struct sched_avg *sa = &se->avg; | |
172895e6 | 733 | long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2; |
7dc603c9 | 734 | u64 now = cfs_rq_clock_task(cfs_rq); |
2b8c41da YD |
735 | |
736 | if (cap > 0) { | |
737 | if (cfs_rq->avg.util_avg != 0) { | |
738 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
739 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
740 | ||
741 | if (sa->util_avg > cap) | |
742 | sa->util_avg = cap; | |
743 | } else { | |
744 | sa->util_avg = cap; | |
745 | } | |
746 | sa->util_sum = sa->util_avg * LOAD_AVG_MAX; | |
747 | } | |
7dc603c9 PZ |
748 | |
749 | if (entity_is_task(se)) { | |
750 | struct task_struct *p = task_of(se); | |
751 | if (p->sched_class != &fair_sched_class) { | |
752 | /* | |
753 | * For !fair tasks do: | |
754 | * | |
755 | update_cfs_rq_load_avg(now, cfs_rq, false); | |
756 | attach_entity_load_avg(cfs_rq, se); | |
757 | switched_from_fair(rq, p); | |
758 | * | |
759 | * such that the next switched_to_fair() has the | |
760 | * expected state. | |
761 | */ | |
762 | se->avg.last_update_time = now; | |
763 | return; | |
764 | } | |
765 | } | |
766 | ||
7c3edd2c | 767 | update_cfs_rq_load_avg(now, cfs_rq, false); |
7dc603c9 | 768 | attach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 769 | update_tg_load_avg(cfs_rq, false); |
2b8c41da YD |
770 | } |
771 | ||
7dc603c9 | 772 | #else /* !CONFIG_SMP */ |
540247fb | 773 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
774 | { |
775 | } | |
2b8c41da YD |
776 | void post_init_entity_util_avg(struct sched_entity *se) |
777 | { | |
778 | } | |
3d30544f PZ |
779 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
780 | { | |
781 | } | |
7dc603c9 | 782 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 783 | |
bf0f6f24 | 784 | /* |
9dbdb155 | 785 | * Update the current task's runtime statistics. |
bf0f6f24 | 786 | */ |
b7cc0896 | 787 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 788 | { |
429d43bc | 789 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 790 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 791 | u64 delta_exec; |
bf0f6f24 IM |
792 | |
793 | if (unlikely(!curr)) | |
794 | return; | |
795 | ||
9dbdb155 PZ |
796 | delta_exec = now - curr->exec_start; |
797 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 798 | return; |
bf0f6f24 | 799 | |
8ebc91d9 | 800 | curr->exec_start = now; |
d842de87 | 801 | |
9dbdb155 PZ |
802 | schedstat_set(curr->statistics.exec_max, |
803 | max(delta_exec, curr->statistics.exec_max)); | |
804 | ||
805 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 806 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
807 | |
808 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
809 | update_min_vruntime(cfs_rq); | |
810 | ||
d842de87 SV |
811 | if (entity_is_task(curr)) { |
812 | struct task_struct *curtask = task_of(curr); | |
813 | ||
f977bb49 | 814 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 815 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 816 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 817 | } |
ec12cb7f PT |
818 | |
819 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
820 | } |
821 | ||
6e998916 SG |
822 | static void update_curr_fair(struct rq *rq) |
823 | { | |
824 | update_curr(cfs_rq_of(&rq->curr->se)); | |
825 | } | |
826 | ||
bf0f6f24 | 827 | static inline void |
5870db5b | 828 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 829 | { |
4fa8d299 JP |
830 | u64 wait_start, prev_wait_start; |
831 | ||
832 | if (!schedstat_enabled()) | |
833 | return; | |
834 | ||
835 | wait_start = rq_clock(rq_of(cfs_rq)); | |
836 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
837 | |
838 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
839 | likely(wait_start > prev_wait_start)) |
840 | wait_start -= prev_wait_start; | |
3ea94de1 | 841 | |
4fa8d299 | 842 | schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
843 | } |
844 | ||
4fa8d299 | 845 | static inline void |
3ea94de1 JP |
846 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
847 | { | |
848 | struct task_struct *p; | |
cb251765 MG |
849 | u64 delta; |
850 | ||
4fa8d299 JP |
851 | if (!schedstat_enabled()) |
852 | return; | |
853 | ||
854 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
855 | |
856 | if (entity_is_task(se)) { | |
857 | p = task_of(se); | |
858 | if (task_on_rq_migrating(p)) { | |
859 | /* | |
860 | * Preserve migrating task's wait time so wait_start | |
861 | * time stamp can be adjusted to accumulate wait time | |
862 | * prior to migration. | |
863 | */ | |
4fa8d299 | 864 | schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
865 | return; |
866 | } | |
867 | trace_sched_stat_wait(p, delta); | |
868 | } | |
869 | ||
4fa8d299 JP |
870 | schedstat_set(se->statistics.wait_max, |
871 | max(schedstat_val(se->statistics.wait_max), delta)); | |
872 | schedstat_inc(se->statistics.wait_count); | |
873 | schedstat_add(se->statistics.wait_sum, delta); | |
874 | schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 875 | } |
3ea94de1 | 876 | |
4fa8d299 | 877 | static inline void |
1a3d027c JP |
878 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
879 | { | |
880 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
881 | u64 sleep_start, block_start; |
882 | ||
883 | if (!schedstat_enabled()) | |
884 | return; | |
885 | ||
886 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
887 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
888 | |
889 | if (entity_is_task(se)) | |
890 | tsk = task_of(se); | |
891 | ||
4fa8d299 JP |
892 | if (sleep_start) { |
893 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
894 | |
895 | if ((s64)delta < 0) | |
896 | delta = 0; | |
897 | ||
4fa8d299 JP |
898 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
899 | schedstat_set(se->statistics.sleep_max, delta); | |
1a3d027c | 900 | |
4fa8d299 JP |
901 | schedstat_set(se->statistics.sleep_start, 0); |
902 | schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
903 | |
904 | if (tsk) { | |
905 | account_scheduler_latency(tsk, delta >> 10, 1); | |
906 | trace_sched_stat_sleep(tsk, delta); | |
907 | } | |
908 | } | |
4fa8d299 JP |
909 | if (block_start) { |
910 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
911 | |
912 | if ((s64)delta < 0) | |
913 | delta = 0; | |
914 | ||
4fa8d299 JP |
915 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
916 | schedstat_set(se->statistics.block_max, delta); | |
1a3d027c | 917 | |
4fa8d299 JP |
918 | schedstat_set(se->statistics.block_start, 0); |
919 | schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
920 | |
921 | if (tsk) { | |
922 | if (tsk->in_iowait) { | |
4fa8d299 JP |
923 | schedstat_add(se->statistics.iowait_sum, delta); |
924 | schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
925 | trace_sched_stat_iowait(tsk, delta); |
926 | } | |
927 | ||
928 | trace_sched_stat_blocked(tsk, delta); | |
929 | ||
930 | /* | |
931 | * Blocking time is in units of nanosecs, so shift by | |
932 | * 20 to get a milliseconds-range estimation of the | |
933 | * amount of time that the task spent sleeping: | |
934 | */ | |
935 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
936 | profile_hits(SLEEP_PROFILING, | |
937 | (void *)get_wchan(tsk), | |
938 | delta >> 20); | |
939 | } | |
940 | account_scheduler_latency(tsk, delta >> 10, 0); | |
941 | } | |
942 | } | |
3ea94de1 | 943 | } |
3ea94de1 | 944 | |
bf0f6f24 IM |
945 | /* |
946 | * Task is being enqueued - update stats: | |
947 | */ | |
cb251765 | 948 | static inline void |
1a3d027c | 949 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 950 | { |
4fa8d299 JP |
951 | if (!schedstat_enabled()) |
952 | return; | |
953 | ||
bf0f6f24 IM |
954 | /* |
955 | * Are we enqueueing a waiting task? (for current tasks | |
956 | * a dequeue/enqueue event is a NOP) | |
957 | */ | |
429d43bc | 958 | if (se != cfs_rq->curr) |
5870db5b | 959 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
960 | |
961 | if (flags & ENQUEUE_WAKEUP) | |
962 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
963 | } |
964 | ||
bf0f6f24 | 965 | static inline void |
cb251765 | 966 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 967 | { |
4fa8d299 JP |
968 | |
969 | if (!schedstat_enabled()) | |
970 | return; | |
971 | ||
bf0f6f24 IM |
972 | /* |
973 | * Mark the end of the wait period if dequeueing a | |
974 | * waiting task: | |
975 | */ | |
429d43bc | 976 | if (se != cfs_rq->curr) |
9ef0a961 | 977 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 978 | |
4fa8d299 JP |
979 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
980 | struct task_struct *tsk = task_of(se); | |
cb251765 | 981 | |
4fa8d299 JP |
982 | if (tsk->state & TASK_INTERRUPTIBLE) |
983 | schedstat_set(se->statistics.sleep_start, | |
984 | rq_clock(rq_of(cfs_rq))); | |
985 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
986 | schedstat_set(se->statistics.block_start, | |
987 | rq_clock(rq_of(cfs_rq))); | |
cb251765 | 988 | } |
cb251765 MG |
989 | } |
990 | ||
bf0f6f24 IM |
991 | /* |
992 | * We are picking a new current task - update its stats: | |
993 | */ | |
994 | static inline void | |
79303e9e | 995 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
996 | { |
997 | /* | |
998 | * We are starting a new run period: | |
999 | */ | |
78becc27 | 1000 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1001 | } |
1002 | ||
bf0f6f24 IM |
1003 | /************************************************** |
1004 | * Scheduling class queueing methods: | |
1005 | */ | |
1006 | ||
cbee9f88 PZ |
1007 | #ifdef CONFIG_NUMA_BALANCING |
1008 | /* | |
598f0ec0 MG |
1009 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1010 | * calculated based on the tasks virtual memory size and | |
1011 | * numa_balancing_scan_size. | |
cbee9f88 | 1012 | */ |
598f0ec0 MG |
1013 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1014 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1015 | |
1016 | /* Portion of address space to scan in MB */ | |
1017 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1018 | |
4b96a29b PZ |
1019 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1020 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1021 | ||
598f0ec0 MG |
1022 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1023 | { | |
1024 | unsigned long rss = 0; | |
1025 | unsigned long nr_scan_pages; | |
1026 | ||
1027 | /* | |
1028 | * Calculations based on RSS as non-present and empty pages are skipped | |
1029 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1030 | * on resident pages | |
1031 | */ | |
1032 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1033 | rss = get_mm_rss(p->mm); | |
1034 | if (!rss) | |
1035 | rss = nr_scan_pages; | |
1036 | ||
1037 | rss = round_up(rss, nr_scan_pages); | |
1038 | return rss / nr_scan_pages; | |
1039 | } | |
1040 | ||
1041 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
1042 | #define MAX_SCAN_WINDOW 2560 | |
1043 | ||
1044 | static unsigned int task_scan_min(struct task_struct *p) | |
1045 | { | |
316c1608 | 1046 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1047 | unsigned int scan, floor; |
1048 | unsigned int windows = 1; | |
1049 | ||
64192658 KT |
1050 | if (scan_size < MAX_SCAN_WINDOW) |
1051 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1052 | floor = 1000 / windows; |
1053 | ||
1054 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1055 | return max_t(unsigned int, floor, scan); | |
1056 | } | |
1057 | ||
1058 | static unsigned int task_scan_max(struct task_struct *p) | |
1059 | { | |
1060 | unsigned int smin = task_scan_min(p); | |
1061 | unsigned int smax; | |
1062 | ||
1063 | /* Watch for min being lower than max due to floor calculations */ | |
1064 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
1065 | return max(smin, smax); | |
1066 | } | |
1067 | ||
0ec8aa00 PZ |
1068 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1069 | { | |
1070 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
1071 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
1072 | } | |
1073 | ||
1074 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1075 | { | |
1076 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
1077 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
1078 | } | |
1079 | ||
8c8a743c PZ |
1080 | struct numa_group { |
1081 | atomic_t refcount; | |
1082 | ||
1083 | spinlock_t lock; /* nr_tasks, tasks */ | |
1084 | int nr_tasks; | |
e29cf08b | 1085 | pid_t gid; |
4142c3eb | 1086 | int active_nodes; |
8c8a743c PZ |
1087 | |
1088 | struct rcu_head rcu; | |
989348b5 | 1089 | unsigned long total_faults; |
4142c3eb | 1090 | unsigned long max_faults_cpu; |
7e2703e6 RR |
1091 | /* |
1092 | * Faults_cpu is used to decide whether memory should move | |
1093 | * towards the CPU. As a consequence, these stats are weighted | |
1094 | * more by CPU use than by memory faults. | |
1095 | */ | |
50ec8a40 | 1096 | unsigned long *faults_cpu; |
989348b5 | 1097 | unsigned long faults[0]; |
8c8a743c PZ |
1098 | }; |
1099 | ||
be1e4e76 RR |
1100 | /* Shared or private faults. */ |
1101 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1102 | ||
1103 | /* Memory and CPU locality */ | |
1104 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1105 | ||
1106 | /* Averaged statistics, and temporary buffers. */ | |
1107 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1108 | ||
e29cf08b MG |
1109 | pid_t task_numa_group_id(struct task_struct *p) |
1110 | { | |
1111 | return p->numa_group ? p->numa_group->gid : 0; | |
1112 | } | |
1113 | ||
44dba3d5 IM |
1114 | /* |
1115 | * The averaged statistics, shared & private, memory & cpu, | |
1116 | * occupy the first half of the array. The second half of the | |
1117 | * array is for current counters, which are averaged into the | |
1118 | * first set by task_numa_placement. | |
1119 | */ | |
1120 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1121 | { |
44dba3d5 | 1122 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1123 | } |
1124 | ||
1125 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1126 | { | |
44dba3d5 | 1127 | if (!p->numa_faults) |
ac8e895b MG |
1128 | return 0; |
1129 | ||
44dba3d5 IM |
1130 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1131 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1132 | } |
1133 | ||
83e1d2cd MG |
1134 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1135 | { | |
1136 | if (!p->numa_group) | |
1137 | return 0; | |
1138 | ||
44dba3d5 IM |
1139 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1140 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1141 | } |
1142 | ||
20e07dea RR |
1143 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1144 | { | |
44dba3d5 IM |
1145 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1146 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1147 | } |
1148 | ||
4142c3eb RR |
1149 | /* |
1150 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1151 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1152 | * between these nodes are slowed down, to allow things to settle down. | |
1153 | */ | |
1154 | #define ACTIVE_NODE_FRACTION 3 | |
1155 | ||
1156 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1157 | { | |
1158 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1159 | } | |
1160 | ||
6c6b1193 RR |
1161 | /* Handle placement on systems where not all nodes are directly connected. */ |
1162 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1163 | int maxdist, bool task) | |
1164 | { | |
1165 | unsigned long score = 0; | |
1166 | int node; | |
1167 | ||
1168 | /* | |
1169 | * All nodes are directly connected, and the same distance | |
1170 | * from each other. No need for fancy placement algorithms. | |
1171 | */ | |
1172 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1173 | return 0; | |
1174 | ||
1175 | /* | |
1176 | * This code is called for each node, introducing N^2 complexity, | |
1177 | * which should be ok given the number of nodes rarely exceeds 8. | |
1178 | */ | |
1179 | for_each_online_node(node) { | |
1180 | unsigned long faults; | |
1181 | int dist = node_distance(nid, node); | |
1182 | ||
1183 | /* | |
1184 | * The furthest away nodes in the system are not interesting | |
1185 | * for placement; nid was already counted. | |
1186 | */ | |
1187 | if (dist == sched_max_numa_distance || node == nid) | |
1188 | continue; | |
1189 | ||
1190 | /* | |
1191 | * On systems with a backplane NUMA topology, compare groups | |
1192 | * of nodes, and move tasks towards the group with the most | |
1193 | * memory accesses. When comparing two nodes at distance | |
1194 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1195 | * of each group. Skip other nodes. | |
1196 | */ | |
1197 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
1198 | dist > maxdist) | |
1199 | continue; | |
1200 | ||
1201 | /* Add up the faults from nearby nodes. */ | |
1202 | if (task) | |
1203 | faults = task_faults(p, node); | |
1204 | else | |
1205 | faults = group_faults(p, node); | |
1206 | ||
1207 | /* | |
1208 | * On systems with a glueless mesh NUMA topology, there are | |
1209 | * no fixed "groups of nodes". Instead, nodes that are not | |
1210 | * directly connected bounce traffic through intermediate | |
1211 | * nodes; a numa_group can occupy any set of nodes. | |
1212 | * The further away a node is, the less the faults count. | |
1213 | * This seems to result in good task placement. | |
1214 | */ | |
1215 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1216 | faults *= (sched_max_numa_distance - dist); | |
1217 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1218 | } | |
1219 | ||
1220 | score += faults; | |
1221 | } | |
1222 | ||
1223 | return score; | |
1224 | } | |
1225 | ||
83e1d2cd MG |
1226 | /* |
1227 | * These return the fraction of accesses done by a particular task, or | |
1228 | * task group, on a particular numa node. The group weight is given a | |
1229 | * larger multiplier, in order to group tasks together that are almost | |
1230 | * evenly spread out between numa nodes. | |
1231 | */ | |
7bd95320 RR |
1232 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1233 | int dist) | |
83e1d2cd | 1234 | { |
7bd95320 | 1235 | unsigned long faults, total_faults; |
83e1d2cd | 1236 | |
44dba3d5 | 1237 | if (!p->numa_faults) |
83e1d2cd MG |
1238 | return 0; |
1239 | ||
1240 | total_faults = p->total_numa_faults; | |
1241 | ||
1242 | if (!total_faults) | |
1243 | return 0; | |
1244 | ||
7bd95320 | 1245 | faults = task_faults(p, nid); |
6c6b1193 RR |
1246 | faults += score_nearby_nodes(p, nid, dist, true); |
1247 | ||
7bd95320 | 1248 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1249 | } |
1250 | ||
7bd95320 RR |
1251 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1252 | int dist) | |
83e1d2cd | 1253 | { |
7bd95320 RR |
1254 | unsigned long faults, total_faults; |
1255 | ||
1256 | if (!p->numa_group) | |
1257 | return 0; | |
1258 | ||
1259 | total_faults = p->numa_group->total_faults; | |
1260 | ||
1261 | if (!total_faults) | |
83e1d2cd MG |
1262 | return 0; |
1263 | ||
7bd95320 | 1264 | faults = group_faults(p, nid); |
6c6b1193 RR |
1265 | faults += score_nearby_nodes(p, nid, dist, false); |
1266 | ||
7bd95320 | 1267 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1268 | } |
1269 | ||
10f39042 RR |
1270 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1271 | int src_nid, int dst_cpu) | |
1272 | { | |
1273 | struct numa_group *ng = p->numa_group; | |
1274 | int dst_nid = cpu_to_node(dst_cpu); | |
1275 | int last_cpupid, this_cpupid; | |
1276 | ||
1277 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
1278 | ||
1279 | /* | |
1280 | * Multi-stage node selection is used in conjunction with a periodic | |
1281 | * migration fault to build a temporal task<->page relation. By using | |
1282 | * a two-stage filter we remove short/unlikely relations. | |
1283 | * | |
1284 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1285 | * a task's usage of a particular page (n_p) per total usage of this | |
1286 | * page (n_t) (in a given time-span) to a probability. | |
1287 | * | |
1288 | * Our periodic faults will sample this probability and getting the | |
1289 | * same result twice in a row, given these samples are fully | |
1290 | * independent, is then given by P(n)^2, provided our sample period | |
1291 | * is sufficiently short compared to the usage pattern. | |
1292 | * | |
1293 | * This quadric squishes small probabilities, making it less likely we | |
1294 | * act on an unlikely task<->page relation. | |
1295 | */ | |
1296 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
1297 | if (!cpupid_pid_unset(last_cpupid) && | |
1298 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1299 | return false; | |
1300 | ||
1301 | /* Always allow migrate on private faults */ | |
1302 | if (cpupid_match_pid(p, last_cpupid)) | |
1303 | return true; | |
1304 | ||
1305 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1306 | if (!ng) | |
1307 | return true; | |
1308 | ||
1309 | /* | |
4142c3eb RR |
1310 | * Destination node is much more heavily used than the source |
1311 | * node? Allow migration. | |
10f39042 | 1312 | */ |
4142c3eb RR |
1313 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1314 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1315 | return true; |
1316 | ||
1317 | /* | |
4142c3eb RR |
1318 | * Distribute memory according to CPU & memory use on each node, |
1319 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1320 | * | |
1321 | * faults_cpu(dst) 3 faults_cpu(src) | |
1322 | * --------------- * - > --------------- | |
1323 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1324 | */ |
4142c3eb RR |
1325 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1326 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1327 | } |
1328 | ||
e6628d5b | 1329 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
1330 | static unsigned long source_load(int cpu, int type); |
1331 | static unsigned long target_load(int cpu, int type); | |
ced549fa | 1332 | static unsigned long capacity_of(int cpu); |
58d081b5 MG |
1333 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); |
1334 | ||
fb13c7ee | 1335 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1336 | struct numa_stats { |
fb13c7ee | 1337 | unsigned long nr_running; |
58d081b5 | 1338 | unsigned long load; |
fb13c7ee MG |
1339 | |
1340 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1341 | unsigned long compute_capacity; |
fb13c7ee MG |
1342 | |
1343 | /* Approximate capacity in terms of runnable tasks on a node */ | |
5ef20ca1 | 1344 | unsigned long task_capacity; |
1b6a7495 | 1345 | int has_free_capacity; |
58d081b5 | 1346 | }; |
e6628d5b | 1347 | |
fb13c7ee MG |
1348 | /* |
1349 | * XXX borrowed from update_sg_lb_stats | |
1350 | */ | |
1351 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1352 | { | |
83d7f242 RR |
1353 | int smt, cpu, cpus = 0; |
1354 | unsigned long capacity; | |
fb13c7ee MG |
1355 | |
1356 | memset(ns, 0, sizeof(*ns)); | |
1357 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1358 | struct rq *rq = cpu_rq(cpu); | |
1359 | ||
1360 | ns->nr_running += rq->nr_running; | |
1361 | ns->load += weighted_cpuload(cpu); | |
ced549fa | 1362 | ns->compute_capacity += capacity_of(cpu); |
5eca82a9 PZ |
1363 | |
1364 | cpus++; | |
fb13c7ee MG |
1365 | } |
1366 | ||
5eca82a9 PZ |
1367 | /* |
1368 | * If we raced with hotplug and there are no CPUs left in our mask | |
1369 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1370 | * not find this node attractive. | |
1371 | * | |
1b6a7495 NP |
1372 | * We'll either bail at !has_free_capacity, or we'll detect a huge |
1373 | * imbalance and bail there. | |
5eca82a9 PZ |
1374 | */ |
1375 | if (!cpus) | |
1376 | return; | |
1377 | ||
83d7f242 RR |
1378 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ |
1379 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); | |
1380 | capacity = cpus / smt; /* cores */ | |
1381 | ||
1382 | ns->task_capacity = min_t(unsigned, capacity, | |
1383 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); | |
1b6a7495 | 1384 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); |
fb13c7ee MG |
1385 | } |
1386 | ||
58d081b5 MG |
1387 | struct task_numa_env { |
1388 | struct task_struct *p; | |
e6628d5b | 1389 | |
58d081b5 MG |
1390 | int src_cpu, src_nid; |
1391 | int dst_cpu, dst_nid; | |
e6628d5b | 1392 | |
58d081b5 | 1393 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1394 | |
40ea2b42 | 1395 | int imbalance_pct; |
7bd95320 | 1396 | int dist; |
fb13c7ee MG |
1397 | |
1398 | struct task_struct *best_task; | |
1399 | long best_imp; | |
58d081b5 MG |
1400 | int best_cpu; |
1401 | }; | |
1402 | ||
fb13c7ee MG |
1403 | static void task_numa_assign(struct task_numa_env *env, |
1404 | struct task_struct *p, long imp) | |
1405 | { | |
1406 | if (env->best_task) | |
1407 | put_task_struct(env->best_task); | |
bac78573 ON |
1408 | if (p) |
1409 | get_task_struct(p); | |
fb13c7ee MG |
1410 | |
1411 | env->best_task = p; | |
1412 | env->best_imp = imp; | |
1413 | env->best_cpu = env->dst_cpu; | |
1414 | } | |
1415 | ||
28a21745 | 1416 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1417 | struct task_numa_env *env) |
1418 | { | |
e4991b24 RR |
1419 | long imb, old_imb; |
1420 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1421 | long src_capacity, dst_capacity; |
1422 | ||
1423 | /* | |
1424 | * The load is corrected for the CPU capacity available on each node. | |
1425 | * | |
1426 | * src_load dst_load | |
1427 | * ------------ vs --------- | |
1428 | * src_capacity dst_capacity | |
1429 | */ | |
1430 | src_capacity = env->src_stats.compute_capacity; | |
1431 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 RR |
1432 | |
1433 | /* We care about the slope of the imbalance, not the direction. */ | |
e4991b24 RR |
1434 | if (dst_load < src_load) |
1435 | swap(dst_load, src_load); | |
e63da036 RR |
1436 | |
1437 | /* Is the difference below the threshold? */ | |
e4991b24 RR |
1438 | imb = dst_load * src_capacity * 100 - |
1439 | src_load * dst_capacity * env->imbalance_pct; | |
e63da036 RR |
1440 | if (imb <= 0) |
1441 | return false; | |
1442 | ||
1443 | /* | |
1444 | * The imbalance is above the allowed threshold. | |
e4991b24 | 1445 | * Compare it with the old imbalance. |
e63da036 | 1446 | */ |
28a21745 | 1447 | orig_src_load = env->src_stats.load; |
e4991b24 | 1448 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1449 | |
e4991b24 RR |
1450 | if (orig_dst_load < orig_src_load) |
1451 | swap(orig_dst_load, orig_src_load); | |
e63da036 | 1452 | |
e4991b24 RR |
1453 | old_imb = orig_dst_load * src_capacity * 100 - |
1454 | orig_src_load * dst_capacity * env->imbalance_pct; | |
1455 | ||
1456 | /* Would this change make things worse? */ | |
1457 | return (imb > old_imb); | |
e63da036 RR |
1458 | } |
1459 | ||
fb13c7ee MG |
1460 | /* |
1461 | * This checks if the overall compute and NUMA accesses of the system would | |
1462 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1463 | * into account that it might be best if task running on the dst_cpu should | |
1464 | * be exchanged with the source task | |
1465 | */ | |
887c290e RR |
1466 | static void task_numa_compare(struct task_numa_env *env, |
1467 | long taskimp, long groupimp) | |
fb13c7ee MG |
1468 | { |
1469 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1470 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1471 | struct task_struct *cur; | |
28a21745 | 1472 | long src_load, dst_load; |
fb13c7ee | 1473 | long load; |
1c5d3eb3 | 1474 | long imp = env->p->numa_group ? groupimp : taskimp; |
0132c3e1 | 1475 | long moveimp = imp; |
7bd95320 | 1476 | int dist = env->dist; |
fb13c7ee MG |
1477 | |
1478 | rcu_read_lock(); | |
bac78573 ON |
1479 | cur = task_rcu_dereference(&dst_rq->curr); |
1480 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) | |
fb13c7ee MG |
1481 | cur = NULL; |
1482 | ||
7af68335 PZ |
1483 | /* |
1484 | * Because we have preemption enabled we can get migrated around and | |
1485 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1486 | */ | |
1487 | if (cur == env->p) | |
1488 | goto unlock; | |
1489 | ||
fb13c7ee MG |
1490 | /* |
1491 | * "imp" is the fault differential for the source task between the | |
1492 | * source and destination node. Calculate the total differential for | |
1493 | * the source task and potential destination task. The more negative | |
1494 | * the value is, the more rmeote accesses that would be expected to | |
1495 | * be incurred if the tasks were swapped. | |
1496 | */ | |
1497 | if (cur) { | |
1498 | /* Skip this swap candidate if cannot move to the source cpu */ | |
1499 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | |
1500 | goto unlock; | |
1501 | ||
887c290e RR |
1502 | /* |
1503 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1504 | * in any group then look only at task weights. |
887c290e | 1505 | */ |
ca28aa53 | 1506 | if (cur->numa_group == env->p->numa_group) { |
7bd95320 RR |
1507 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1508 | task_weight(cur, env->dst_nid, dist); | |
ca28aa53 RR |
1509 | /* |
1510 | * Add some hysteresis to prevent swapping the | |
1511 | * tasks within a group over tiny differences. | |
1512 | */ | |
1513 | if (cur->numa_group) | |
1514 | imp -= imp/16; | |
887c290e | 1515 | } else { |
ca28aa53 RR |
1516 | /* |
1517 | * Compare the group weights. If a task is all by | |
1518 | * itself (not part of a group), use the task weight | |
1519 | * instead. | |
1520 | */ | |
ca28aa53 | 1521 | if (cur->numa_group) |
7bd95320 RR |
1522 | imp += group_weight(cur, env->src_nid, dist) - |
1523 | group_weight(cur, env->dst_nid, dist); | |
ca28aa53 | 1524 | else |
7bd95320 RR |
1525 | imp += task_weight(cur, env->src_nid, dist) - |
1526 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1527 | } |
fb13c7ee MG |
1528 | } |
1529 | ||
0132c3e1 | 1530 | if (imp <= env->best_imp && moveimp <= env->best_imp) |
fb13c7ee MG |
1531 | goto unlock; |
1532 | ||
1533 | if (!cur) { | |
1534 | /* Is there capacity at our destination? */ | |
b932c03c | 1535 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && |
1b6a7495 | 1536 | !env->dst_stats.has_free_capacity) |
fb13c7ee MG |
1537 | goto unlock; |
1538 | ||
1539 | goto balance; | |
1540 | } | |
1541 | ||
1542 | /* Balance doesn't matter much if we're running a task per cpu */ | |
0132c3e1 RR |
1543 | if (imp > env->best_imp && src_rq->nr_running == 1 && |
1544 | dst_rq->nr_running == 1) | |
fb13c7ee MG |
1545 | goto assign; |
1546 | ||
1547 | /* | |
1548 | * In the overloaded case, try and keep the load balanced. | |
1549 | */ | |
1550 | balance: | |
e720fff6 PZ |
1551 | load = task_h_load(env->p); |
1552 | dst_load = env->dst_stats.load + load; | |
1553 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1554 | |
0132c3e1 RR |
1555 | if (moveimp > imp && moveimp > env->best_imp) { |
1556 | /* | |
1557 | * If the improvement from just moving env->p direction is | |
1558 | * better than swapping tasks around, check if a move is | |
1559 | * possible. Store a slightly smaller score than moveimp, | |
1560 | * so an actually idle CPU will win. | |
1561 | */ | |
1562 | if (!load_too_imbalanced(src_load, dst_load, env)) { | |
1563 | imp = moveimp - 1; | |
1564 | cur = NULL; | |
1565 | goto assign; | |
1566 | } | |
1567 | } | |
1568 | ||
1569 | if (imp <= env->best_imp) | |
1570 | goto unlock; | |
1571 | ||
fb13c7ee | 1572 | if (cur) { |
e720fff6 PZ |
1573 | load = task_h_load(cur); |
1574 | dst_load -= load; | |
1575 | src_load += load; | |
fb13c7ee MG |
1576 | } |
1577 | ||
28a21745 | 1578 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1579 | goto unlock; |
1580 | ||
ba7e5a27 RR |
1581 | /* |
1582 | * One idle CPU per node is evaluated for a task numa move. | |
1583 | * Call select_idle_sibling to maybe find a better one. | |
1584 | */ | |
1585 | if (!cur) | |
772bd008 MR |
1586 | env->dst_cpu = select_idle_sibling(env->p, env->src_cpu, |
1587 | env->dst_cpu); | |
ba7e5a27 | 1588 | |
fb13c7ee MG |
1589 | assign: |
1590 | task_numa_assign(env, cur, imp); | |
1591 | unlock: | |
1592 | rcu_read_unlock(); | |
1593 | } | |
1594 | ||
887c290e RR |
1595 | static void task_numa_find_cpu(struct task_numa_env *env, |
1596 | long taskimp, long groupimp) | |
2c8a50aa MG |
1597 | { |
1598 | int cpu; | |
1599 | ||
1600 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1601 | /* Skip this CPU if the source task cannot migrate */ | |
1602 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | |
1603 | continue; | |
1604 | ||
1605 | env->dst_cpu = cpu; | |
887c290e | 1606 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1607 | } |
1608 | } | |
1609 | ||
6f9aad0b RR |
1610 | /* Only move tasks to a NUMA node less busy than the current node. */ |
1611 | static bool numa_has_capacity(struct task_numa_env *env) | |
1612 | { | |
1613 | struct numa_stats *src = &env->src_stats; | |
1614 | struct numa_stats *dst = &env->dst_stats; | |
1615 | ||
1616 | if (src->has_free_capacity && !dst->has_free_capacity) | |
1617 | return false; | |
1618 | ||
1619 | /* | |
1620 | * Only consider a task move if the source has a higher load | |
1621 | * than the destination, corrected for CPU capacity on each node. | |
1622 | * | |
1623 | * src->load dst->load | |
1624 | * --------------------- vs --------------------- | |
1625 | * src->compute_capacity dst->compute_capacity | |
1626 | */ | |
44dcb04f SD |
1627 | if (src->load * dst->compute_capacity * env->imbalance_pct > |
1628 | ||
1629 | dst->load * src->compute_capacity * 100) | |
6f9aad0b RR |
1630 | return true; |
1631 | ||
1632 | return false; | |
1633 | } | |
1634 | ||
58d081b5 MG |
1635 | static int task_numa_migrate(struct task_struct *p) |
1636 | { | |
58d081b5 MG |
1637 | struct task_numa_env env = { |
1638 | .p = p, | |
fb13c7ee | 1639 | |
58d081b5 | 1640 | .src_cpu = task_cpu(p), |
b32e86b4 | 1641 | .src_nid = task_node(p), |
fb13c7ee MG |
1642 | |
1643 | .imbalance_pct = 112, | |
1644 | ||
1645 | .best_task = NULL, | |
1646 | .best_imp = 0, | |
4142c3eb | 1647 | .best_cpu = -1, |
58d081b5 MG |
1648 | }; |
1649 | struct sched_domain *sd; | |
887c290e | 1650 | unsigned long taskweight, groupweight; |
7bd95320 | 1651 | int nid, ret, dist; |
887c290e | 1652 | long taskimp, groupimp; |
e6628d5b | 1653 | |
58d081b5 | 1654 | /* |
fb13c7ee MG |
1655 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1656 | * imbalance and would be the first to start moving tasks about. | |
1657 | * | |
1658 | * And we want to avoid any moving of tasks about, as that would create | |
1659 | * random movement of tasks -- counter the numa conditions we're trying | |
1660 | * to satisfy here. | |
58d081b5 MG |
1661 | */ |
1662 | rcu_read_lock(); | |
fb13c7ee | 1663 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1664 | if (sd) |
1665 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1666 | rcu_read_unlock(); |
1667 | ||
46a73e8a RR |
1668 | /* |
1669 | * Cpusets can break the scheduler domain tree into smaller | |
1670 | * balance domains, some of which do not cross NUMA boundaries. | |
1671 | * Tasks that are "trapped" in such domains cannot be migrated | |
1672 | * elsewhere, so there is no point in (re)trying. | |
1673 | */ | |
1674 | if (unlikely(!sd)) { | |
de1b301a | 1675 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1676 | return -EINVAL; |
1677 | } | |
1678 | ||
2c8a50aa | 1679 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1680 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1681 | taskweight = task_weight(p, env.src_nid, dist); | |
1682 | groupweight = group_weight(p, env.src_nid, dist); | |
1683 | update_numa_stats(&env.src_stats, env.src_nid); | |
1684 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1685 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1686 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1687 | |
a43455a1 | 1688 | /* Try to find a spot on the preferred nid. */ |
6f9aad0b RR |
1689 | if (numa_has_capacity(&env)) |
1690 | task_numa_find_cpu(&env, taskimp, groupimp); | |
e1dda8a7 | 1691 | |
9de05d48 RR |
1692 | /* |
1693 | * Look at other nodes in these cases: | |
1694 | * - there is no space available on the preferred_nid | |
1695 | * - the task is part of a numa_group that is interleaved across | |
1696 | * multiple NUMA nodes; in order to better consolidate the group, | |
1697 | * we need to check other locations. | |
1698 | */ | |
4142c3eb | 1699 | if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) { |
2c8a50aa MG |
1700 | for_each_online_node(nid) { |
1701 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1702 | continue; | |
58d081b5 | 1703 | |
7bd95320 | 1704 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1705 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1706 | dist != env.dist) { | |
1707 | taskweight = task_weight(p, env.src_nid, dist); | |
1708 | groupweight = group_weight(p, env.src_nid, dist); | |
1709 | } | |
7bd95320 | 1710 | |
83e1d2cd | 1711 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1712 | taskimp = task_weight(p, nid, dist) - taskweight; |
1713 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1714 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1715 | continue; |
1716 | ||
7bd95320 | 1717 | env.dist = dist; |
2c8a50aa MG |
1718 | env.dst_nid = nid; |
1719 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
6f9aad0b RR |
1720 | if (numa_has_capacity(&env)) |
1721 | task_numa_find_cpu(&env, taskimp, groupimp); | |
58d081b5 MG |
1722 | } |
1723 | } | |
1724 | ||
68d1b02a RR |
1725 | /* |
1726 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1727 | * and is migrating into one of the workload's active nodes, remember | |
1728 | * this node as the task's preferred numa node, so the workload can | |
1729 | * settle down. | |
1730 | * A task that migrated to a second choice node will be better off | |
1731 | * trying for a better one later. Do not set the preferred node here. | |
1732 | */ | |
db015dae | 1733 | if (p->numa_group) { |
4142c3eb RR |
1734 | struct numa_group *ng = p->numa_group; |
1735 | ||
db015dae RR |
1736 | if (env.best_cpu == -1) |
1737 | nid = env.src_nid; | |
1738 | else | |
1739 | nid = env.dst_nid; | |
1740 | ||
4142c3eb | 1741 | if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng)) |
db015dae RR |
1742 | sched_setnuma(p, env.dst_nid); |
1743 | } | |
1744 | ||
1745 | /* No better CPU than the current one was found. */ | |
1746 | if (env.best_cpu == -1) | |
1747 | return -EAGAIN; | |
0ec8aa00 | 1748 | |
04bb2f94 RR |
1749 | /* |
1750 | * Reset the scan period if the task is being rescheduled on an | |
1751 | * alternative node to recheck if the tasks is now properly placed. | |
1752 | */ | |
1753 | p->numa_scan_period = task_scan_min(p); | |
1754 | ||
fb13c7ee | 1755 | if (env.best_task == NULL) { |
286549dc MG |
1756 | ret = migrate_task_to(p, env.best_cpu); |
1757 | if (ret != 0) | |
1758 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1759 | return ret; |
1760 | } | |
1761 | ||
1762 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1763 | if (ret != 0) |
1764 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1765 | put_task_struct(env.best_task); |
1766 | return ret; | |
e6628d5b MG |
1767 | } |
1768 | ||
6b9a7460 MG |
1769 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1770 | static void numa_migrate_preferred(struct task_struct *p) | |
1771 | { | |
5085e2a3 RR |
1772 | unsigned long interval = HZ; |
1773 | ||
2739d3ee | 1774 | /* This task has no NUMA fault statistics yet */ |
44dba3d5 | 1775 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
6b9a7460 MG |
1776 | return; |
1777 | ||
2739d3ee | 1778 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 RR |
1779 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1780 | p->numa_migrate_retry = jiffies + interval; | |
2739d3ee RR |
1781 | |
1782 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1783 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1784 | return; |
1785 | ||
1786 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1787 | task_numa_migrate(p); |
6b9a7460 MG |
1788 | } |
1789 | ||
20e07dea | 1790 | /* |
4142c3eb | 1791 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
1792 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
1793 | * be different from the set of nodes where the workload's memory is currently | |
1794 | * located. | |
20e07dea | 1795 | */ |
4142c3eb | 1796 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
1797 | { |
1798 | unsigned long faults, max_faults = 0; | |
4142c3eb | 1799 | int nid, active_nodes = 0; |
20e07dea RR |
1800 | |
1801 | for_each_online_node(nid) { | |
1802 | faults = group_faults_cpu(numa_group, nid); | |
1803 | if (faults > max_faults) | |
1804 | max_faults = faults; | |
1805 | } | |
1806 | ||
1807 | for_each_online_node(nid) { | |
1808 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
1809 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
1810 | active_nodes++; | |
20e07dea | 1811 | } |
4142c3eb RR |
1812 | |
1813 | numa_group->max_faults_cpu = max_faults; | |
1814 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
1815 | } |
1816 | ||
04bb2f94 RR |
1817 | /* |
1818 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1819 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1820 | * period will be for the next scan window. If local/(local+remote) ratio is |
1821 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1822 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1823 | */ |
1824 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1825 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1826 | |
1827 | /* | |
1828 | * Increase the scan period (slow down scanning) if the majority of | |
1829 | * our memory is already on our local node, or if the majority of | |
1830 | * the page accesses are shared with other processes. | |
1831 | * Otherwise, decrease the scan period. | |
1832 | */ | |
1833 | static void update_task_scan_period(struct task_struct *p, | |
1834 | unsigned long shared, unsigned long private) | |
1835 | { | |
1836 | unsigned int period_slot; | |
1837 | int ratio; | |
1838 | int diff; | |
1839 | ||
1840 | unsigned long remote = p->numa_faults_locality[0]; | |
1841 | unsigned long local = p->numa_faults_locality[1]; | |
1842 | ||
1843 | /* | |
1844 | * If there were no record hinting faults then either the task is | |
1845 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
1846 | * to automatic numa balancing. Related to that, if there were failed |
1847 | * migration then it implies we are migrating too quickly or the local | |
1848 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 1849 | */ |
074c2381 | 1850 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
1851 | p->numa_scan_period = min(p->numa_scan_period_max, |
1852 | p->numa_scan_period << 1); | |
1853 | ||
1854 | p->mm->numa_next_scan = jiffies + | |
1855 | msecs_to_jiffies(p->numa_scan_period); | |
1856 | ||
1857 | return; | |
1858 | } | |
1859 | ||
1860 | /* | |
1861 | * Prepare to scale scan period relative to the current period. | |
1862 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1863 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1864 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1865 | */ | |
1866 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1867 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1868 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1869 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1870 | if (!slot) | |
1871 | slot = 1; | |
1872 | diff = slot * period_slot; | |
1873 | } else { | |
1874 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1875 | ||
1876 | /* | |
1877 | * Scale scan rate increases based on sharing. There is an | |
1878 | * inverse relationship between the degree of sharing and | |
1879 | * the adjustment made to the scanning period. Broadly | |
1880 | * speaking the intent is that there is little point | |
1881 | * scanning faster if shared accesses dominate as it may | |
1882 | * simply bounce migrations uselessly | |
1883 | */ | |
2847c90e | 1884 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); |
04bb2f94 RR |
1885 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; |
1886 | } | |
1887 | ||
1888 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1889 | task_scan_min(p), task_scan_max(p)); | |
1890 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1891 | } | |
1892 | ||
7e2703e6 RR |
1893 | /* |
1894 | * Get the fraction of time the task has been running since the last | |
1895 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1896 | * decays those on a 32ms period, which is orders of magnitude off | |
1897 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1898 | * stats only if the task is so new there are no NUMA statistics yet. | |
1899 | */ | |
1900 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1901 | { | |
1902 | u64 runtime, delta, now; | |
1903 | /* Use the start of this time slice to avoid calculations. */ | |
1904 | now = p->se.exec_start; | |
1905 | runtime = p->se.sum_exec_runtime; | |
1906 | ||
1907 | if (p->last_task_numa_placement) { | |
1908 | delta = runtime - p->last_sum_exec_runtime; | |
1909 | *period = now - p->last_task_numa_placement; | |
1910 | } else { | |
9d89c257 YD |
1911 | delta = p->se.avg.load_sum / p->se.load.weight; |
1912 | *period = LOAD_AVG_MAX; | |
7e2703e6 RR |
1913 | } |
1914 | ||
1915 | p->last_sum_exec_runtime = runtime; | |
1916 | p->last_task_numa_placement = now; | |
1917 | ||
1918 | return delta; | |
1919 | } | |
1920 | ||
54009416 RR |
1921 | /* |
1922 | * Determine the preferred nid for a task in a numa_group. This needs to | |
1923 | * be done in a way that produces consistent results with group_weight, | |
1924 | * otherwise workloads might not converge. | |
1925 | */ | |
1926 | static int preferred_group_nid(struct task_struct *p, int nid) | |
1927 | { | |
1928 | nodemask_t nodes; | |
1929 | int dist; | |
1930 | ||
1931 | /* Direct connections between all NUMA nodes. */ | |
1932 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1933 | return nid; | |
1934 | ||
1935 | /* | |
1936 | * On a system with glueless mesh NUMA topology, group_weight | |
1937 | * scores nodes according to the number of NUMA hinting faults on | |
1938 | * both the node itself, and on nearby nodes. | |
1939 | */ | |
1940 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1941 | unsigned long score, max_score = 0; | |
1942 | int node, max_node = nid; | |
1943 | ||
1944 | dist = sched_max_numa_distance; | |
1945 | ||
1946 | for_each_online_node(node) { | |
1947 | score = group_weight(p, node, dist); | |
1948 | if (score > max_score) { | |
1949 | max_score = score; | |
1950 | max_node = node; | |
1951 | } | |
1952 | } | |
1953 | return max_node; | |
1954 | } | |
1955 | ||
1956 | /* | |
1957 | * Finding the preferred nid in a system with NUMA backplane | |
1958 | * interconnect topology is more involved. The goal is to locate | |
1959 | * tasks from numa_groups near each other in the system, and | |
1960 | * untangle workloads from different sides of the system. This requires | |
1961 | * searching down the hierarchy of node groups, recursively searching | |
1962 | * inside the highest scoring group of nodes. The nodemask tricks | |
1963 | * keep the complexity of the search down. | |
1964 | */ | |
1965 | nodes = node_online_map; | |
1966 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
1967 | unsigned long max_faults = 0; | |
81907478 | 1968 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
1969 | int a, b; |
1970 | ||
1971 | /* Are there nodes at this distance from each other? */ | |
1972 | if (!find_numa_distance(dist)) | |
1973 | continue; | |
1974 | ||
1975 | for_each_node_mask(a, nodes) { | |
1976 | unsigned long faults = 0; | |
1977 | nodemask_t this_group; | |
1978 | nodes_clear(this_group); | |
1979 | ||
1980 | /* Sum group's NUMA faults; includes a==b case. */ | |
1981 | for_each_node_mask(b, nodes) { | |
1982 | if (node_distance(a, b) < dist) { | |
1983 | faults += group_faults(p, b); | |
1984 | node_set(b, this_group); | |
1985 | node_clear(b, nodes); | |
1986 | } | |
1987 | } | |
1988 | ||
1989 | /* Remember the top group. */ | |
1990 | if (faults > max_faults) { | |
1991 | max_faults = faults; | |
1992 | max_group = this_group; | |
1993 | /* | |
1994 | * subtle: at the smallest distance there is | |
1995 | * just one node left in each "group", the | |
1996 | * winner is the preferred nid. | |
1997 | */ | |
1998 | nid = a; | |
1999 | } | |
2000 | } | |
2001 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2002 | if (!max_faults) |
2003 | break; | |
54009416 RR |
2004 | nodes = max_group; |
2005 | } | |
2006 | return nid; | |
2007 | } | |
2008 | ||
cbee9f88 PZ |
2009 | static void task_numa_placement(struct task_struct *p) |
2010 | { | |
83e1d2cd MG |
2011 | int seq, nid, max_nid = -1, max_group_nid = -1; |
2012 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 2013 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2014 | unsigned long total_faults; |
2015 | u64 runtime, period; | |
7dbd13ed | 2016 | spinlock_t *group_lock = NULL; |
cbee9f88 | 2017 | |
7e5a2c17 JL |
2018 | /* |
2019 | * The p->mm->numa_scan_seq field gets updated without | |
2020 | * exclusive access. Use READ_ONCE() here to ensure | |
2021 | * that the field is read in a single access: | |
2022 | */ | |
316c1608 | 2023 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2024 | if (p->numa_scan_seq == seq) |
2025 | return; | |
2026 | p->numa_scan_seq = seq; | |
598f0ec0 | 2027 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2028 | |
7e2703e6 RR |
2029 | total_faults = p->numa_faults_locality[0] + |
2030 | p->numa_faults_locality[1]; | |
2031 | runtime = numa_get_avg_runtime(p, &period); | |
2032 | ||
7dbd13ed MG |
2033 | /* If the task is part of a group prevent parallel updates to group stats */ |
2034 | if (p->numa_group) { | |
2035 | group_lock = &p->numa_group->lock; | |
60e69eed | 2036 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2037 | } |
2038 | ||
688b7585 MG |
2039 | /* Find the node with the highest number of faults */ |
2040 | for_each_online_node(nid) { | |
44dba3d5 IM |
2041 | /* Keep track of the offsets in numa_faults array */ |
2042 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2043 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2044 | int priv; |
745d6147 | 2045 | |
be1e4e76 | 2046 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2047 | long diff, f_diff, f_weight; |
8c8a743c | 2048 | |
44dba3d5 IM |
2049 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2050 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2051 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2052 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2053 | |
ac8e895b | 2054 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2055 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2056 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2057 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2058 | |
7e2703e6 RR |
2059 | /* |
2060 | * Normalize the faults_from, so all tasks in a group | |
2061 | * count according to CPU use, instead of by the raw | |
2062 | * number of faults. Tasks with little runtime have | |
2063 | * little over-all impact on throughput, and thus their | |
2064 | * faults are less important. | |
2065 | */ | |
2066 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2067 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2068 | (total_faults + 1); |
44dba3d5 IM |
2069 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2070 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2071 | |
44dba3d5 IM |
2072 | p->numa_faults[mem_idx] += diff; |
2073 | p->numa_faults[cpu_idx] += f_diff; | |
2074 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2075 | p->total_numa_faults += diff; |
8c8a743c | 2076 | if (p->numa_group) { |
44dba3d5 IM |
2077 | /* |
2078 | * safe because we can only change our own group | |
2079 | * | |
2080 | * mem_idx represents the offset for a given | |
2081 | * nid and priv in a specific region because it | |
2082 | * is at the beginning of the numa_faults array. | |
2083 | */ | |
2084 | p->numa_group->faults[mem_idx] += diff; | |
2085 | p->numa_group->faults_cpu[mem_idx] += f_diff; | |
989348b5 | 2086 | p->numa_group->total_faults += diff; |
44dba3d5 | 2087 | group_faults += p->numa_group->faults[mem_idx]; |
8c8a743c | 2088 | } |
ac8e895b MG |
2089 | } |
2090 | ||
688b7585 MG |
2091 | if (faults > max_faults) { |
2092 | max_faults = faults; | |
2093 | max_nid = nid; | |
2094 | } | |
83e1d2cd MG |
2095 | |
2096 | if (group_faults > max_group_faults) { | |
2097 | max_group_faults = group_faults; | |
2098 | max_group_nid = nid; | |
2099 | } | |
2100 | } | |
2101 | ||
04bb2f94 RR |
2102 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
2103 | ||
7dbd13ed | 2104 | if (p->numa_group) { |
4142c3eb | 2105 | numa_group_count_active_nodes(p->numa_group); |
60e69eed | 2106 | spin_unlock_irq(group_lock); |
54009416 | 2107 | max_nid = preferred_group_nid(p, max_group_nid); |
688b7585 MG |
2108 | } |
2109 | ||
bb97fc31 RR |
2110 | if (max_faults) { |
2111 | /* Set the new preferred node */ | |
2112 | if (max_nid != p->numa_preferred_nid) | |
2113 | sched_setnuma(p, max_nid); | |
2114 | ||
2115 | if (task_node(p) != p->numa_preferred_nid) | |
2116 | numa_migrate_preferred(p); | |
3a7053b3 | 2117 | } |
cbee9f88 PZ |
2118 | } |
2119 | ||
8c8a743c PZ |
2120 | static inline int get_numa_group(struct numa_group *grp) |
2121 | { | |
2122 | return atomic_inc_not_zero(&grp->refcount); | |
2123 | } | |
2124 | ||
2125 | static inline void put_numa_group(struct numa_group *grp) | |
2126 | { | |
2127 | if (atomic_dec_and_test(&grp->refcount)) | |
2128 | kfree_rcu(grp, rcu); | |
2129 | } | |
2130 | ||
3e6a9418 MG |
2131 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2132 | int *priv) | |
8c8a743c PZ |
2133 | { |
2134 | struct numa_group *grp, *my_grp; | |
2135 | struct task_struct *tsk; | |
2136 | bool join = false; | |
2137 | int cpu = cpupid_to_cpu(cpupid); | |
2138 | int i; | |
2139 | ||
2140 | if (unlikely(!p->numa_group)) { | |
2141 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 2142 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2143 | |
2144 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2145 | if (!grp) | |
2146 | return; | |
2147 | ||
2148 | atomic_set(&grp->refcount, 1); | |
4142c3eb RR |
2149 | grp->active_nodes = 1; |
2150 | grp->max_faults_cpu = 0; | |
8c8a743c | 2151 | spin_lock_init(&grp->lock); |
e29cf08b | 2152 | grp->gid = p->pid; |
50ec8a40 | 2153 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2154 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2155 | nr_node_ids; | |
8c8a743c | 2156 | |
be1e4e76 | 2157 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2158 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2159 | |
989348b5 | 2160 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2161 | |
8c8a743c PZ |
2162 | grp->nr_tasks++; |
2163 | rcu_assign_pointer(p->numa_group, grp); | |
2164 | } | |
2165 | ||
2166 | rcu_read_lock(); | |
316c1608 | 2167 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2168 | |
2169 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2170 | goto no_join; |
8c8a743c PZ |
2171 | |
2172 | grp = rcu_dereference(tsk->numa_group); | |
2173 | if (!grp) | |
3354781a | 2174 | goto no_join; |
8c8a743c PZ |
2175 | |
2176 | my_grp = p->numa_group; | |
2177 | if (grp == my_grp) | |
3354781a | 2178 | goto no_join; |
8c8a743c PZ |
2179 | |
2180 | /* | |
2181 | * Only join the other group if its bigger; if we're the bigger group, | |
2182 | * the other task will join us. | |
2183 | */ | |
2184 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2185 | goto no_join; |
8c8a743c PZ |
2186 | |
2187 | /* | |
2188 | * Tie-break on the grp address. | |
2189 | */ | |
2190 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2191 | goto no_join; |
8c8a743c | 2192 | |
dabe1d99 RR |
2193 | /* Always join threads in the same process. */ |
2194 | if (tsk->mm == current->mm) | |
2195 | join = true; | |
2196 | ||
2197 | /* Simple filter to avoid false positives due to PID collisions */ | |
2198 | if (flags & TNF_SHARED) | |
2199 | join = true; | |
8c8a743c | 2200 | |
3e6a9418 MG |
2201 | /* Update priv based on whether false sharing was detected */ |
2202 | *priv = !join; | |
2203 | ||
dabe1d99 | 2204 | if (join && !get_numa_group(grp)) |
3354781a | 2205 | goto no_join; |
8c8a743c | 2206 | |
8c8a743c PZ |
2207 | rcu_read_unlock(); |
2208 | ||
2209 | if (!join) | |
2210 | return; | |
2211 | ||
60e69eed MG |
2212 | BUG_ON(irqs_disabled()); |
2213 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2214 | |
be1e4e76 | 2215 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2216 | my_grp->faults[i] -= p->numa_faults[i]; |
2217 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2218 | } |
989348b5 MG |
2219 | my_grp->total_faults -= p->total_numa_faults; |
2220 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2221 | |
8c8a743c PZ |
2222 | my_grp->nr_tasks--; |
2223 | grp->nr_tasks++; | |
2224 | ||
2225 | spin_unlock(&my_grp->lock); | |
60e69eed | 2226 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2227 | |
2228 | rcu_assign_pointer(p->numa_group, grp); | |
2229 | ||
2230 | put_numa_group(my_grp); | |
3354781a PZ |
2231 | return; |
2232 | ||
2233 | no_join: | |
2234 | rcu_read_unlock(); | |
2235 | return; | |
8c8a743c PZ |
2236 | } |
2237 | ||
2238 | void task_numa_free(struct task_struct *p) | |
2239 | { | |
2240 | struct numa_group *grp = p->numa_group; | |
44dba3d5 | 2241 | void *numa_faults = p->numa_faults; |
e9dd685c SR |
2242 | unsigned long flags; |
2243 | int i; | |
8c8a743c PZ |
2244 | |
2245 | if (grp) { | |
e9dd685c | 2246 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2247 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2248 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2249 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2250 | |
8c8a743c | 2251 | grp->nr_tasks--; |
e9dd685c | 2252 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2253 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2254 | put_numa_group(grp); |
2255 | } | |
2256 | ||
44dba3d5 | 2257 | p->numa_faults = NULL; |
82727018 | 2258 | kfree(numa_faults); |
8c8a743c PZ |
2259 | } |
2260 | ||
cbee9f88 PZ |
2261 | /* |
2262 | * Got a PROT_NONE fault for a page on @node. | |
2263 | */ | |
58b46da3 | 2264 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2265 | { |
2266 | struct task_struct *p = current; | |
6688cc05 | 2267 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2268 | int cpu_node = task_node(current); |
792568ec | 2269 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2270 | struct numa_group *ng; |
ac8e895b | 2271 | int priv; |
cbee9f88 | 2272 | |
2a595721 | 2273 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2274 | return; |
2275 | ||
9ff1d9ff MG |
2276 | /* for example, ksmd faulting in a user's mm */ |
2277 | if (!p->mm) | |
2278 | return; | |
2279 | ||
f809ca9a | 2280 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2281 | if (unlikely(!p->numa_faults)) { |
2282 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2283 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2284 | |
44dba3d5 IM |
2285 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2286 | if (!p->numa_faults) | |
f809ca9a | 2287 | return; |
745d6147 | 2288 | |
83e1d2cd | 2289 | p->total_numa_faults = 0; |
04bb2f94 | 2290 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2291 | } |
cbee9f88 | 2292 | |
8c8a743c PZ |
2293 | /* |
2294 | * First accesses are treated as private, otherwise consider accesses | |
2295 | * to be private if the accessing pid has not changed | |
2296 | */ | |
2297 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2298 | priv = 1; | |
2299 | } else { | |
2300 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2301 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2302 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2303 | } |
2304 | ||
792568ec RR |
2305 | /* |
2306 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2307 | * occurs wholly within the set of nodes that the workload is | |
2308 | * actively using should be counted as local. This allows the | |
2309 | * scan rate to slow down when a workload has settled down. | |
2310 | */ | |
4142c3eb RR |
2311 | ng = p->numa_group; |
2312 | if (!priv && !local && ng && ng->active_nodes > 1 && | |
2313 | numa_is_active_node(cpu_node, ng) && | |
2314 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2315 | local = 1; |
2316 | ||
cbee9f88 | 2317 | task_numa_placement(p); |
f809ca9a | 2318 | |
2739d3ee RR |
2319 | /* |
2320 | * Retry task to preferred node migration periodically, in case it | |
2321 | * case it previously failed, or the scheduler moved us. | |
2322 | */ | |
2323 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
2324 | numa_migrate_preferred(p); |
2325 | ||
b32e86b4 IM |
2326 | if (migrated) |
2327 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2328 | if (flags & TNF_MIGRATE_FAIL) |
2329 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2330 | |
44dba3d5 IM |
2331 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2332 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2333 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2334 | } |
2335 | ||
6e5fb223 PZ |
2336 | static void reset_ptenuma_scan(struct task_struct *p) |
2337 | { | |
7e5a2c17 JL |
2338 | /* |
2339 | * We only did a read acquisition of the mmap sem, so | |
2340 | * p->mm->numa_scan_seq is written to without exclusive access | |
2341 | * and the update is not guaranteed to be atomic. That's not | |
2342 | * much of an issue though, since this is just used for | |
2343 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2344 | * expensive, to avoid any form of compiler optimizations: | |
2345 | */ | |
316c1608 | 2346 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2347 | p->mm->numa_scan_offset = 0; |
2348 | } | |
2349 | ||
cbee9f88 PZ |
2350 | /* |
2351 | * The expensive part of numa migration is done from task_work context. | |
2352 | * Triggered from task_tick_numa(). | |
2353 | */ | |
2354 | void task_numa_work(struct callback_head *work) | |
2355 | { | |
2356 | unsigned long migrate, next_scan, now = jiffies; | |
2357 | struct task_struct *p = current; | |
2358 | struct mm_struct *mm = p->mm; | |
51170840 | 2359 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2360 | struct vm_area_struct *vma; |
9f40604c | 2361 | unsigned long start, end; |
598f0ec0 | 2362 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2363 | long pages, virtpages; |
cbee9f88 PZ |
2364 | |
2365 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | |
2366 | ||
2367 | work->next = work; /* protect against double add */ | |
2368 | /* | |
2369 | * Who cares about NUMA placement when they're dying. | |
2370 | * | |
2371 | * NOTE: make sure not to dereference p->mm before this check, | |
2372 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2373 | * without p->mm even though we still had it when we enqueued this | |
2374 | * work. | |
2375 | */ | |
2376 | if (p->flags & PF_EXITING) | |
2377 | return; | |
2378 | ||
930aa174 | 2379 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2380 | mm->numa_next_scan = now + |
2381 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2382 | } |
2383 | ||
cbee9f88 PZ |
2384 | /* |
2385 | * Enforce maximal scan/migration frequency.. | |
2386 | */ | |
2387 | migrate = mm->numa_next_scan; | |
2388 | if (time_before(now, migrate)) | |
2389 | return; | |
2390 | ||
598f0ec0 MG |
2391 | if (p->numa_scan_period == 0) { |
2392 | p->numa_scan_period_max = task_scan_max(p); | |
2393 | p->numa_scan_period = task_scan_min(p); | |
2394 | } | |
cbee9f88 | 2395 | |
fb003b80 | 2396 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2397 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2398 | return; | |
2399 | ||
19a78d11 PZ |
2400 | /* |
2401 | * Delay this task enough that another task of this mm will likely win | |
2402 | * the next time around. | |
2403 | */ | |
2404 | p->node_stamp += 2 * TICK_NSEC; | |
2405 | ||
9f40604c MG |
2406 | start = mm->numa_scan_offset; |
2407 | pages = sysctl_numa_balancing_scan_size; | |
2408 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2409 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2410 | if (!pages) |
2411 | return; | |
cbee9f88 | 2412 | |
4620f8c1 | 2413 | |
6e5fb223 | 2414 | down_read(&mm->mmap_sem); |
9f40604c | 2415 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2416 | if (!vma) { |
2417 | reset_ptenuma_scan(p); | |
9f40604c | 2418 | start = 0; |
6e5fb223 PZ |
2419 | vma = mm->mmap; |
2420 | } | |
9f40604c | 2421 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2422 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2423 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2424 | continue; |
6b79c57b | 2425 | } |
6e5fb223 | 2426 | |
4591ce4f MG |
2427 | /* |
2428 | * Shared library pages mapped by multiple processes are not | |
2429 | * migrated as it is expected they are cache replicated. Avoid | |
2430 | * hinting faults in read-only file-backed mappings or the vdso | |
2431 | * as migrating the pages will be of marginal benefit. | |
2432 | */ | |
2433 | if (!vma->vm_mm || | |
2434 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2435 | continue; | |
2436 | ||
3c67f474 MG |
2437 | /* |
2438 | * Skip inaccessible VMAs to avoid any confusion between | |
2439 | * PROT_NONE and NUMA hinting ptes | |
2440 | */ | |
2441 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2442 | continue; | |
4591ce4f | 2443 | |
9f40604c MG |
2444 | do { |
2445 | start = max(start, vma->vm_start); | |
2446 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2447 | end = min(end, vma->vm_end); | |
4620f8c1 | 2448 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2449 | |
2450 | /* | |
4620f8c1 RR |
2451 | * Try to scan sysctl_numa_balancing_size worth of |
2452 | * hpages that have at least one present PTE that | |
2453 | * is not already pte-numa. If the VMA contains | |
2454 | * areas that are unused or already full of prot_numa | |
2455 | * PTEs, scan up to virtpages, to skip through those | |
2456 | * areas faster. | |
598f0ec0 MG |
2457 | */ |
2458 | if (nr_pte_updates) | |
2459 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2460 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2461 | |
9f40604c | 2462 | start = end; |
4620f8c1 | 2463 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2464 | goto out; |
3cf1962c RR |
2465 | |
2466 | cond_resched(); | |
9f40604c | 2467 | } while (end != vma->vm_end); |
cbee9f88 | 2468 | } |
6e5fb223 | 2469 | |
9f40604c | 2470 | out: |
6e5fb223 | 2471 | /* |
c69307d5 PZ |
2472 | * It is possible to reach the end of the VMA list but the last few |
2473 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2474 | * would find the !migratable VMA on the next scan but not reset the | |
2475 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2476 | */ |
2477 | if (vma) | |
9f40604c | 2478 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2479 | else |
2480 | reset_ptenuma_scan(p); | |
2481 | up_read(&mm->mmap_sem); | |
51170840 RR |
2482 | |
2483 | /* | |
2484 | * Make sure tasks use at least 32x as much time to run other code | |
2485 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2486 | * Usually update_task_scan_period slows down scanning enough; on an | |
2487 | * overloaded system we need to limit overhead on a per task basis. | |
2488 | */ | |
2489 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2490 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2491 | p->node_stamp += 32 * diff; | |
2492 | } | |
cbee9f88 PZ |
2493 | } |
2494 | ||
2495 | /* | |
2496 | * Drive the periodic memory faults.. | |
2497 | */ | |
2498 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2499 | { | |
2500 | struct callback_head *work = &curr->numa_work; | |
2501 | u64 period, now; | |
2502 | ||
2503 | /* | |
2504 | * We don't care about NUMA placement if we don't have memory. | |
2505 | */ | |
2506 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2507 | return; | |
2508 | ||
2509 | /* | |
2510 | * Using runtime rather than walltime has the dual advantage that | |
2511 | * we (mostly) drive the selection from busy threads and that the | |
2512 | * task needs to have done some actual work before we bother with | |
2513 | * NUMA placement. | |
2514 | */ | |
2515 | now = curr->se.sum_exec_runtime; | |
2516 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2517 | ||
25b3e5a3 | 2518 | if (now > curr->node_stamp + period) { |
4b96a29b | 2519 | if (!curr->node_stamp) |
598f0ec0 | 2520 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 2521 | curr->node_stamp += period; |
cbee9f88 PZ |
2522 | |
2523 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2524 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2525 | task_work_add(curr, work, true); | |
2526 | } | |
2527 | } | |
2528 | } | |
2529 | #else | |
2530 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2531 | { | |
2532 | } | |
0ec8aa00 PZ |
2533 | |
2534 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2535 | { | |
2536 | } | |
2537 | ||
2538 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2539 | { | |
2540 | } | |
cbee9f88 PZ |
2541 | #endif /* CONFIG_NUMA_BALANCING */ |
2542 | ||
30cfdcfc DA |
2543 | static void |
2544 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2545 | { | |
2546 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2547 | if (!parent_entity(se)) |
029632fb | 2548 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2549 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2550 | if (entity_is_task(se)) { |
2551 | struct rq *rq = rq_of(cfs_rq); | |
2552 | ||
2553 | account_numa_enqueue(rq, task_of(se)); | |
2554 | list_add(&se->group_node, &rq->cfs_tasks); | |
2555 | } | |
367456c7 | 2556 | #endif |
30cfdcfc | 2557 | cfs_rq->nr_running++; |
30cfdcfc DA |
2558 | } |
2559 | ||
2560 | static void | |
2561 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2562 | { | |
2563 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2564 | if (!parent_entity(se)) |
029632fb | 2565 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
bfdb198c | 2566 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2567 | if (entity_is_task(se)) { |
2568 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2569 | list_del_init(&se->group_node); |
0ec8aa00 | 2570 | } |
bfdb198c | 2571 | #endif |
30cfdcfc | 2572 | cfs_rq->nr_running--; |
30cfdcfc DA |
2573 | } |
2574 | ||
3ff6dcac YZ |
2575 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2576 | # ifdef CONFIG_SMP | |
ea1dc6fc | 2577 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
cf5f0acf | 2578 | { |
ea1dc6fc | 2579 | long tg_weight, load, shares; |
cf5f0acf PZ |
2580 | |
2581 | /* | |
ea1dc6fc PZ |
2582 | * This really should be: cfs_rq->avg.load_avg, but instead we use |
2583 | * cfs_rq->load.weight, which is its upper bound. This helps ramp up | |
2584 | * the shares for small weight interactive tasks. | |
cf5f0acf | 2585 | */ |
ea1dc6fc | 2586 | load = scale_load_down(cfs_rq->load.weight); |
cf5f0acf | 2587 | |
ea1dc6fc | 2588 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 2589 | |
ea1dc6fc PZ |
2590 | /* Ensure tg_weight >= load */ |
2591 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
2592 | tg_weight += load; | |
3ff6dcac | 2593 | |
3ff6dcac | 2594 | shares = (tg->shares * load); |
cf5f0acf PZ |
2595 | if (tg_weight) |
2596 | shares /= tg_weight; | |
3ff6dcac YZ |
2597 | |
2598 | if (shares < MIN_SHARES) | |
2599 | shares = MIN_SHARES; | |
2600 | if (shares > tg->shares) | |
2601 | shares = tg->shares; | |
2602 | ||
2603 | return shares; | |
2604 | } | |
3ff6dcac | 2605 | # else /* CONFIG_SMP */ |
6d5ab293 | 2606 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
2607 | { |
2608 | return tg->shares; | |
2609 | } | |
3ff6dcac | 2610 | # endif /* CONFIG_SMP */ |
ea1dc6fc | 2611 | |
2069dd75 PZ |
2612 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2613 | unsigned long weight) | |
2614 | { | |
19e5eebb PT |
2615 | if (se->on_rq) { |
2616 | /* commit outstanding execution time */ | |
2617 | if (cfs_rq->curr == se) | |
2618 | update_curr(cfs_rq); | |
2069dd75 | 2619 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 2620 | } |
2069dd75 PZ |
2621 | |
2622 | update_load_set(&se->load, weight); | |
2623 | ||
2624 | if (se->on_rq) | |
2625 | account_entity_enqueue(cfs_rq, se); | |
2626 | } | |
2627 | ||
82958366 PT |
2628 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2629 | ||
6d5ab293 | 2630 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2631 | { |
2632 | struct task_group *tg; | |
2633 | struct sched_entity *se; | |
3ff6dcac | 2634 | long shares; |
2069dd75 | 2635 | |
2069dd75 PZ |
2636 | tg = cfs_rq->tg; |
2637 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | |
64660c86 | 2638 | if (!se || throttled_hierarchy(cfs_rq)) |
2069dd75 | 2639 | return; |
3ff6dcac YZ |
2640 | #ifndef CONFIG_SMP |
2641 | if (likely(se->load.weight == tg->shares)) | |
2642 | return; | |
2643 | #endif | |
6d5ab293 | 2644 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
2645 | |
2646 | reweight_entity(cfs_rq_of(se), se, shares); | |
2647 | } | |
2648 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
6d5ab293 | 2649 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2069dd75 PZ |
2650 | { |
2651 | } | |
2652 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
2653 | ||
141965c7 | 2654 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
2655 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ |
2656 | static const u32 runnable_avg_yN_inv[] = { | |
2657 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
2658 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
2659 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
2660 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
2661 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
2662 | 0x85aac367, 0x82cd8698, | |
2663 | }; | |
2664 | ||
2665 | /* | |
2666 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | |
2667 | * over-estimates when re-combining. | |
2668 | */ | |
2669 | static const u32 runnable_avg_yN_sum[] = { | |
2670 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | |
2671 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | |
2672 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | |
2673 | }; | |
2674 | ||
7b20b916 YD |
2675 | /* |
2676 | * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to | |
2677 | * lower integers. See Documentation/scheduler/sched-avg.txt how these | |
2678 | * were generated: | |
2679 | */ | |
2680 | static const u32 __accumulated_sum_N32[] = { | |
2681 | 0, 23371, 35056, 40899, 43820, 45281, | |
2682 | 46011, 46376, 46559, 46650, 46696, 46719, | |
2683 | }; | |
2684 | ||
9d85f21c PT |
2685 | /* |
2686 | * Approximate: | |
2687 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
2688 | */ | |
2689 | static __always_inline u64 decay_load(u64 val, u64 n) | |
2690 | { | |
5b51f2f8 PT |
2691 | unsigned int local_n; |
2692 | ||
2693 | if (!n) | |
2694 | return val; | |
2695 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | |
2696 | return 0; | |
2697 | ||
2698 | /* after bounds checking we can collapse to 32-bit */ | |
2699 | local_n = n; | |
2700 | ||
2701 | /* | |
2702 | * As y^PERIOD = 1/2, we can combine | |
9c58c79a ZZ |
2703 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) |
2704 | * With a look-up table which covers y^n (n<PERIOD) | |
5b51f2f8 PT |
2705 | * |
2706 | * To achieve constant time decay_load. | |
2707 | */ | |
2708 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2709 | val >>= local_n / LOAD_AVG_PERIOD; | |
2710 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2711 | } |
2712 | ||
9d89c257 YD |
2713 | val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32); |
2714 | return val; | |
5b51f2f8 PT |
2715 | } |
2716 | ||
2717 | /* | |
2718 | * For updates fully spanning n periods, the contribution to runnable | |
2719 | * average will be: \Sum 1024*y^n | |
2720 | * | |
2721 | * We can compute this reasonably efficiently by combining: | |
2722 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | |
2723 | */ | |
2724 | static u32 __compute_runnable_contrib(u64 n) | |
2725 | { | |
2726 | u32 contrib = 0; | |
2727 | ||
2728 | if (likely(n <= LOAD_AVG_PERIOD)) | |
2729 | return runnable_avg_yN_sum[n]; | |
2730 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | |
2731 | return LOAD_AVG_MAX; | |
2732 | ||
7b20b916 YD |
2733 | /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */ |
2734 | contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD]; | |
2735 | n %= LOAD_AVG_PERIOD; | |
5b51f2f8 PT |
2736 | contrib = decay_load(contrib, n); |
2737 | return contrib + runnable_avg_yN_sum[n]; | |
9d85f21c PT |
2738 | } |
2739 | ||
54a21385 | 2740 | #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) |
e0f5f3af | 2741 | |
9d85f21c PT |
2742 | /* |
2743 | * We can represent the historical contribution to runnable average as the | |
2744 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2745 | * history into segments of approximately 1ms (1024us); label the segment that | |
2746 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2747 | * | |
2748 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2749 | * p0 p1 p2 | |
2750 | * (now) (~1ms ago) (~2ms ago) | |
2751 | * | |
2752 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2753 | * | |
2754 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2755 | * following representation of historical load: | |
2756 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2757 | * | |
2758 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2759 | * y^32 = 0.5 | |
2760 | * | |
2761 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2762 | * approximately half as much as the contribution to load within the last ms | |
2763 | * (u_0). | |
2764 | * | |
2765 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2766 | * sum again by y is sufficient to update: | |
2767 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2768 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2769 | */ | |
9d89c257 YD |
2770 | static __always_inline int |
2771 | __update_load_avg(u64 now, int cpu, struct sched_avg *sa, | |
13962234 | 2772 | unsigned long weight, int running, struct cfs_rq *cfs_rq) |
9d85f21c | 2773 | { |
e0f5f3af | 2774 | u64 delta, scaled_delta, periods; |
9d89c257 | 2775 | u32 contrib; |
6115c793 | 2776 | unsigned int delta_w, scaled_delta_w, decayed = 0; |
6f2b0452 | 2777 | unsigned long scale_freq, scale_cpu; |
9d85f21c | 2778 | |
9d89c257 | 2779 | delta = now - sa->last_update_time; |
9d85f21c PT |
2780 | /* |
2781 | * This should only happen when time goes backwards, which it | |
2782 | * unfortunately does during sched clock init when we swap over to TSC. | |
2783 | */ | |
2784 | if ((s64)delta < 0) { | |
9d89c257 | 2785 | sa->last_update_time = now; |
9d85f21c PT |
2786 | return 0; |
2787 | } | |
2788 | ||
2789 | /* | |
2790 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2791 | * approximation of 1us and fast to compute. | |
2792 | */ | |
2793 | delta >>= 10; | |
2794 | if (!delta) | |
2795 | return 0; | |
9d89c257 | 2796 | sa->last_update_time = now; |
9d85f21c | 2797 | |
6f2b0452 DE |
2798 | scale_freq = arch_scale_freq_capacity(NULL, cpu); |
2799 | scale_cpu = arch_scale_cpu_capacity(NULL, cpu); | |
2800 | ||
9d85f21c | 2801 | /* delta_w is the amount already accumulated against our next period */ |
9d89c257 | 2802 | delta_w = sa->period_contrib; |
9d85f21c | 2803 | if (delta + delta_w >= 1024) { |
9d85f21c PT |
2804 | decayed = 1; |
2805 | ||
9d89c257 YD |
2806 | /* how much left for next period will start over, we don't know yet */ |
2807 | sa->period_contrib = 0; | |
2808 | ||
9d85f21c PT |
2809 | /* |
2810 | * Now that we know we're crossing a period boundary, figure | |
2811 | * out how much from delta we need to complete the current | |
2812 | * period and accrue it. | |
2813 | */ | |
2814 | delta_w = 1024 - delta_w; | |
54a21385 | 2815 | scaled_delta_w = cap_scale(delta_w, scale_freq); |
13962234 | 2816 | if (weight) { |
e0f5f3af DE |
2817 | sa->load_sum += weight * scaled_delta_w; |
2818 | if (cfs_rq) { | |
2819 | cfs_rq->runnable_load_sum += | |
2820 | weight * scaled_delta_w; | |
2821 | } | |
13962234 | 2822 | } |
36ee28e4 | 2823 | if (running) |
006cdf02 | 2824 | sa->util_sum += scaled_delta_w * scale_cpu; |
5b51f2f8 PT |
2825 | |
2826 | delta -= delta_w; | |
2827 | ||
2828 | /* Figure out how many additional periods this update spans */ | |
2829 | periods = delta / 1024; | |
2830 | delta %= 1024; | |
2831 | ||
9d89c257 | 2832 | sa->load_sum = decay_load(sa->load_sum, periods + 1); |
13962234 YD |
2833 | if (cfs_rq) { |
2834 | cfs_rq->runnable_load_sum = | |
2835 | decay_load(cfs_rq->runnable_load_sum, periods + 1); | |
2836 | } | |
9d89c257 | 2837 | sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1); |
5b51f2f8 PT |
2838 | |
2839 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | |
9d89c257 | 2840 | contrib = __compute_runnable_contrib(periods); |
54a21385 | 2841 | contrib = cap_scale(contrib, scale_freq); |
13962234 | 2842 | if (weight) { |
9d89c257 | 2843 | sa->load_sum += weight * contrib; |
13962234 YD |
2844 | if (cfs_rq) |
2845 | cfs_rq->runnable_load_sum += weight * contrib; | |
2846 | } | |
36ee28e4 | 2847 | if (running) |
006cdf02 | 2848 | sa->util_sum += contrib * scale_cpu; |
9d85f21c PT |
2849 | } |
2850 | ||
2851 | /* Remainder of delta accrued against u_0` */ | |
54a21385 | 2852 | scaled_delta = cap_scale(delta, scale_freq); |
13962234 | 2853 | if (weight) { |
e0f5f3af | 2854 | sa->load_sum += weight * scaled_delta; |
13962234 | 2855 | if (cfs_rq) |
e0f5f3af | 2856 | cfs_rq->runnable_load_sum += weight * scaled_delta; |
13962234 | 2857 | } |
36ee28e4 | 2858 | if (running) |
006cdf02 | 2859 | sa->util_sum += scaled_delta * scale_cpu; |
9ee474f5 | 2860 | |
9d89c257 | 2861 | sa->period_contrib += delta; |
9ee474f5 | 2862 | |
9d89c257 YD |
2863 | if (decayed) { |
2864 | sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX); | |
13962234 YD |
2865 | if (cfs_rq) { |
2866 | cfs_rq->runnable_load_avg = | |
2867 | div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX); | |
2868 | } | |
006cdf02 | 2869 | sa->util_avg = sa->util_sum / LOAD_AVG_MAX; |
9d89c257 | 2870 | } |
aff3e498 | 2871 | |
9d89c257 | 2872 | return decayed; |
9ee474f5 PT |
2873 | } |
2874 | ||
c566e8e9 | 2875 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7c3edd2c PZ |
2876 | /** |
2877 | * update_tg_load_avg - update the tg's load avg | |
2878 | * @cfs_rq: the cfs_rq whose avg changed | |
2879 | * @force: update regardless of how small the difference | |
2880 | * | |
2881 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
2882 | * However, because tg->load_avg is a global value there are performance | |
2883 | * considerations. | |
2884 | * | |
2885 | * In order to avoid having to look at the other cfs_rq's, we use a | |
2886 | * differential update where we store the last value we propagated. This in | |
2887 | * turn allows skipping updates if the differential is 'small'. | |
2888 | * | |
2889 | * Updating tg's load_avg is necessary before update_cfs_share() (which is | |
2890 | * done) and effective_load() (which is not done because it is too costly). | |
bb17f655 | 2891 | */ |
9d89c257 | 2892 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 2893 | { |
9d89c257 | 2894 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 2895 | |
aa0b7ae0 WL |
2896 | /* |
2897 | * No need to update load_avg for root_task_group as it is not used. | |
2898 | */ | |
2899 | if (cfs_rq->tg == &root_task_group) | |
2900 | return; | |
2901 | ||
9d89c257 YD |
2902 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
2903 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
2904 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 2905 | } |
8165e145 | 2906 | } |
f5f9739d | 2907 | |
ad936d86 BP |
2908 | /* |
2909 | * Called within set_task_rq() right before setting a task's cpu. The | |
2910 | * caller only guarantees p->pi_lock is held; no other assumptions, | |
2911 | * including the state of rq->lock, should be made. | |
2912 | */ | |
2913 | void set_task_rq_fair(struct sched_entity *se, | |
2914 | struct cfs_rq *prev, struct cfs_rq *next) | |
2915 | { | |
2916 | if (!sched_feat(ATTACH_AGE_LOAD)) | |
2917 | return; | |
2918 | ||
2919 | /* | |
2920 | * We are supposed to update the task to "current" time, then its up to | |
2921 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
2922 | * getting what current time is, so simply throw away the out-of-date | |
2923 | * time. This will result in the wakee task is less decayed, but giving | |
2924 | * the wakee more load sounds not bad. | |
2925 | */ | |
2926 | if (se->avg.last_update_time && prev) { | |
2927 | u64 p_last_update_time; | |
2928 | u64 n_last_update_time; | |
2929 | ||
2930 | #ifndef CONFIG_64BIT | |
2931 | u64 p_last_update_time_copy; | |
2932 | u64 n_last_update_time_copy; | |
2933 | ||
2934 | do { | |
2935 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
2936 | n_last_update_time_copy = next->load_last_update_time_copy; | |
2937 | ||
2938 | smp_rmb(); | |
2939 | ||
2940 | p_last_update_time = prev->avg.last_update_time; | |
2941 | n_last_update_time = next->avg.last_update_time; | |
2942 | ||
2943 | } while (p_last_update_time != p_last_update_time_copy || | |
2944 | n_last_update_time != n_last_update_time_copy); | |
2945 | #else | |
2946 | p_last_update_time = prev->avg.last_update_time; | |
2947 | n_last_update_time = next->avg.last_update_time; | |
2948 | #endif | |
2949 | __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)), | |
2950 | &se->avg, 0, 0, NULL); | |
2951 | se->avg.last_update_time = n_last_update_time; | |
2952 | } | |
2953 | } | |
6e83125c | 2954 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
9d89c257 | 2955 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
6e83125c | 2956 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 2957 | |
a2c6c91f SM |
2958 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq) |
2959 | { | |
58919e83 | 2960 | if (&this_rq()->cfs == cfs_rq) { |
a2c6c91f SM |
2961 | /* |
2962 | * There are a few boundary cases this might miss but it should | |
2963 | * get called often enough that that should (hopefully) not be | |
2964 | * a real problem -- added to that it only calls on the local | |
2965 | * CPU, so if we enqueue remotely we'll miss an update, but | |
2966 | * the next tick/schedule should update. | |
2967 | * | |
2968 | * It will not get called when we go idle, because the idle | |
2969 | * thread is a different class (!fair), nor will the utilization | |
2970 | * number include things like RT tasks. | |
2971 | * | |
2972 | * As is, the util number is not freq-invariant (we'd have to | |
2973 | * implement arch_scale_freq_capacity() for that). | |
2974 | * | |
2975 | * See cpu_util(). | |
2976 | */ | |
12bde33d | 2977 | cpufreq_update_util(rq_of(cfs_rq), 0); |
a2c6c91f SM |
2978 | } |
2979 | } | |
2980 | ||
89741892 PZ |
2981 | /* |
2982 | * Unsigned subtract and clamp on underflow. | |
2983 | * | |
2984 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2985 | * memory. This allows lockless observations without ever seeing the negative | |
2986 | * values. | |
2987 | */ | |
2988 | #define sub_positive(_ptr, _val) do { \ | |
2989 | typeof(_ptr) ptr = (_ptr); \ | |
2990 | typeof(*ptr) val = (_val); \ | |
2991 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2992 | res = var - val; \ | |
2993 | if (res > var) \ | |
2994 | res = 0; \ | |
2995 | WRITE_ONCE(*ptr, res); \ | |
2996 | } while (0) | |
2997 | ||
3d30544f PZ |
2998 | /** |
2999 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
3000 | * @now: current time, as per cfs_rq_clock_task() | |
3001 | * @cfs_rq: cfs_rq to update | |
3002 | * @update_freq: should we call cfs_rq_util_change() or will the call do so | |
3003 | * | |
3004 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3005 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3006 | * post_init_entity_util_avg(). | |
3007 | * | |
3008 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3009 | * | |
7c3edd2c PZ |
3010 | * Returns true if the load decayed or we removed load. |
3011 | * | |
3012 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3013 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3014 | */ |
a2c6c91f SM |
3015 | static inline int |
3016 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) | |
2dac754e | 3017 | { |
9d89c257 | 3018 | struct sched_avg *sa = &cfs_rq->avg; |
41e0d37f | 3019 | int decayed, removed_load = 0, removed_util = 0; |
2dac754e | 3020 | |
9d89c257 | 3021 | if (atomic_long_read(&cfs_rq->removed_load_avg)) { |
9e0e83a1 | 3022 | s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); |
89741892 PZ |
3023 | sub_positive(&sa->load_avg, r); |
3024 | sub_positive(&sa->load_sum, r * LOAD_AVG_MAX); | |
41e0d37f | 3025 | removed_load = 1; |
8165e145 | 3026 | } |
2dac754e | 3027 | |
9d89c257 YD |
3028 | if (atomic_long_read(&cfs_rq->removed_util_avg)) { |
3029 | long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0); | |
89741892 PZ |
3030 | sub_positive(&sa->util_avg, r); |
3031 | sub_positive(&sa->util_sum, r * LOAD_AVG_MAX); | |
41e0d37f | 3032 | removed_util = 1; |
9d89c257 | 3033 | } |
36ee28e4 | 3034 | |
a2c6c91f | 3035 | decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa, |
13962234 | 3036 | scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq); |
36ee28e4 | 3037 | |
9d89c257 YD |
3038 | #ifndef CONFIG_64BIT |
3039 | smp_wmb(); | |
3040 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3041 | #endif | |
36ee28e4 | 3042 | |
a2c6c91f SM |
3043 | if (update_freq && (decayed || removed_util)) |
3044 | cfs_rq_util_change(cfs_rq); | |
21e96f88 | 3045 | |
41e0d37f | 3046 | return decayed || removed_load; |
21e96f88 SM |
3047 | } |
3048 | ||
3049 | /* Update task and its cfs_rq load average */ | |
3050 | static inline void update_load_avg(struct sched_entity *se, int update_tg) | |
3051 | { | |
3052 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3053 | u64 now = cfs_rq_clock_task(cfs_rq); | |
3054 | struct rq *rq = rq_of(cfs_rq); | |
3055 | int cpu = cpu_of(rq); | |
3056 | ||
3057 | /* | |
3058 | * Track task load average for carrying it to new CPU after migrated, and | |
3059 | * track group sched_entity load average for task_h_load calc in migration | |
3060 | */ | |
3061 | __update_load_avg(now, cpu, &se->avg, | |
3062 | se->on_rq * scale_load_down(se->load.weight), | |
3063 | cfs_rq->curr == se, NULL); | |
3064 | ||
a2c6c91f | 3065 | if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg) |
21e96f88 | 3066 | update_tg_load_avg(cfs_rq, 0); |
9ee474f5 PT |
3067 | } |
3068 | ||
3d30544f PZ |
3069 | /** |
3070 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3071 | * @cfs_rq: cfs_rq to attach to | |
3072 | * @se: sched_entity to attach | |
3073 | * | |
3074 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3075 | * cfs_rq->avg.last_update_time being current. | |
3076 | */ | |
a05e8c51 BP |
3077 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3078 | { | |
a9280514 PZ |
3079 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3080 | goto skip_aging; | |
3081 | ||
6efdb105 BP |
3082 | /* |
3083 | * If we got migrated (either between CPUs or between cgroups) we'll | |
3084 | * have aged the average right before clearing @last_update_time. | |
7dc603c9 PZ |
3085 | * |
3086 | * Or we're fresh through post_init_entity_util_avg(). | |
6efdb105 BP |
3087 | */ |
3088 | if (se->avg.last_update_time) { | |
3089 | __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)), | |
3090 | &se->avg, 0, 0, NULL); | |
3091 | ||
3092 | /* | |
3093 | * XXX: we could have just aged the entire load away if we've been | |
3094 | * absent from the fair class for too long. | |
3095 | */ | |
3096 | } | |
3097 | ||
a9280514 | 3098 | skip_aging: |
a05e8c51 BP |
3099 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
3100 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3101 | cfs_rq->avg.load_sum += se->avg.load_sum; | |
3102 | cfs_rq->avg.util_avg += se->avg.util_avg; | |
3103 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
a2c6c91f SM |
3104 | |
3105 | cfs_rq_util_change(cfs_rq); | |
a05e8c51 BP |
3106 | } |
3107 | ||
3d30544f PZ |
3108 | /** |
3109 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3110 | * @cfs_rq: cfs_rq to detach from | |
3111 | * @se: sched_entity to detach | |
3112 | * | |
3113 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3114 | * cfs_rq->avg.last_update_time being current. | |
3115 | */ | |
a05e8c51 BP |
3116 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3117 | { | |
3118 | __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)), | |
3119 | &se->avg, se->on_rq * scale_load_down(se->load.weight), | |
3120 | cfs_rq->curr == se, NULL); | |
3121 | ||
89741892 PZ |
3122 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); |
3123 | sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum); | |
3124 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); | |
3125 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
a2c6c91f SM |
3126 | |
3127 | cfs_rq_util_change(cfs_rq); | |
a05e8c51 BP |
3128 | } |
3129 | ||
9d89c257 YD |
3130 | /* Add the load generated by se into cfs_rq's load average */ |
3131 | static inline void | |
3132 | enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
9ee474f5 | 3133 | { |
9d89c257 YD |
3134 | struct sched_avg *sa = &se->avg; |
3135 | u64 now = cfs_rq_clock_task(cfs_rq); | |
a05e8c51 | 3136 | int migrated, decayed; |
9ee474f5 | 3137 | |
a05e8c51 BP |
3138 | migrated = !sa->last_update_time; |
3139 | if (!migrated) { | |
9d89c257 | 3140 | __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa, |
13962234 YD |
3141 | se->on_rq * scale_load_down(se->load.weight), |
3142 | cfs_rq->curr == se, NULL); | |
aff3e498 | 3143 | } |
c566e8e9 | 3144 | |
a2c6c91f | 3145 | decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated); |
18bf2805 | 3146 | |
13962234 YD |
3147 | cfs_rq->runnable_load_avg += sa->load_avg; |
3148 | cfs_rq->runnable_load_sum += sa->load_sum; | |
3149 | ||
a05e8c51 BP |
3150 | if (migrated) |
3151 | attach_entity_load_avg(cfs_rq, se); | |
9ee474f5 | 3152 | |
9d89c257 YD |
3153 | if (decayed || migrated) |
3154 | update_tg_load_avg(cfs_rq, 0); | |
2dac754e PT |
3155 | } |
3156 | ||
13962234 YD |
3157 | /* Remove the runnable load generated by se from cfs_rq's runnable load average */ |
3158 | static inline void | |
3159 | dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3160 | { | |
3161 | update_load_avg(se, 1); | |
3162 | ||
3163 | cfs_rq->runnable_load_avg = | |
3164 | max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0); | |
3165 | cfs_rq->runnable_load_sum = | |
a05e8c51 | 3166 | max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0); |
13962234 YD |
3167 | } |
3168 | ||
9d89c257 | 3169 | #ifndef CONFIG_64BIT |
0905f04e YD |
3170 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3171 | { | |
9d89c257 | 3172 | u64 last_update_time_copy; |
0905f04e | 3173 | u64 last_update_time; |
9ee474f5 | 3174 | |
9d89c257 YD |
3175 | do { |
3176 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3177 | smp_rmb(); | |
3178 | last_update_time = cfs_rq->avg.last_update_time; | |
3179 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3180 | |
3181 | return last_update_time; | |
3182 | } | |
9d89c257 | 3183 | #else |
0905f04e YD |
3184 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3185 | { | |
3186 | return cfs_rq->avg.last_update_time; | |
3187 | } | |
9d89c257 YD |
3188 | #endif |
3189 | ||
0905f04e YD |
3190 | /* |
3191 | * Task first catches up with cfs_rq, and then subtract | |
3192 | * itself from the cfs_rq (task must be off the queue now). | |
3193 | */ | |
3194 | void remove_entity_load_avg(struct sched_entity *se) | |
3195 | { | |
3196 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3197 | u64 last_update_time; | |
3198 | ||
3199 | /* | |
7dc603c9 PZ |
3200 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3201 | * post_init_entity_util_avg() which will have added things to the | |
3202 | * cfs_rq, so we can remove unconditionally. | |
3203 | * | |
3204 | * Similarly for groups, they will have passed through | |
3205 | * post_init_entity_util_avg() before unregister_sched_fair_group() | |
3206 | * calls this. | |
0905f04e | 3207 | */ |
0905f04e YD |
3208 | |
3209 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
3210 | ||
13962234 | 3211 | __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL); |
9d89c257 YD |
3212 | atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg); |
3213 | atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg); | |
2dac754e | 3214 | } |
642dbc39 | 3215 | |
7ea241af YD |
3216 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) |
3217 | { | |
3218 | return cfs_rq->runnable_load_avg; | |
3219 | } | |
3220 | ||
3221 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) | |
3222 | { | |
3223 | return cfs_rq->avg.load_avg; | |
3224 | } | |
3225 | ||
6e83125c PZ |
3226 | static int idle_balance(struct rq *this_rq); |
3227 | ||
38033c37 PZ |
3228 | #else /* CONFIG_SMP */ |
3229 | ||
01011473 PZ |
3230 | static inline int |
3231 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) | |
3232 | { | |
3233 | return 0; | |
3234 | } | |
3235 | ||
536bd00c RW |
3236 | static inline void update_load_avg(struct sched_entity *se, int not_used) |
3237 | { | |
12bde33d | 3238 | cpufreq_update_util(rq_of(cfs_rq_of(se)), 0); |
536bd00c RW |
3239 | } |
3240 | ||
9d89c257 YD |
3241 | static inline void |
3242 | enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
13962234 YD |
3243 | static inline void |
3244 | dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
9d89c257 | 3245 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 3246 | |
a05e8c51 BP |
3247 | static inline void |
3248 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3249 | static inline void | |
3250 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3251 | ||
6e83125c PZ |
3252 | static inline int idle_balance(struct rq *rq) |
3253 | { | |
3254 | return 0; | |
3255 | } | |
3256 | ||
38033c37 | 3257 | #endif /* CONFIG_SMP */ |
9d85f21c | 3258 | |
ddc97297 PZ |
3259 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3260 | { | |
3261 | #ifdef CONFIG_SCHED_DEBUG | |
3262 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
3263 | ||
3264 | if (d < 0) | |
3265 | d = -d; | |
3266 | ||
3267 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 3268 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
3269 | #endif |
3270 | } | |
3271 | ||
aeb73b04 PZ |
3272 | static void |
3273 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
3274 | { | |
1af5f730 | 3275 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 3276 | |
2cb8600e PZ |
3277 | /* |
3278 | * The 'current' period is already promised to the current tasks, | |
3279 | * however the extra weight of the new task will slow them down a | |
3280 | * little, place the new task so that it fits in the slot that | |
3281 | * stays open at the end. | |
3282 | */ | |
94dfb5e7 | 3283 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 3284 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 3285 | |
a2e7a7eb | 3286 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 3287 | if (!initial) { |
a2e7a7eb | 3288 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 3289 | |
a2e7a7eb MG |
3290 | /* |
3291 | * Halve their sleep time's effect, to allow | |
3292 | * for a gentler effect of sleepers: | |
3293 | */ | |
3294 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
3295 | thresh >>= 1; | |
51e0304c | 3296 | |
a2e7a7eb | 3297 | vruntime -= thresh; |
aeb73b04 PZ |
3298 | } |
3299 | ||
b5d9d734 | 3300 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3301 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3302 | } |
3303 | ||
d3d9dc33 PT |
3304 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3305 | ||
cb251765 MG |
3306 | static inline void check_schedstat_required(void) |
3307 | { | |
3308 | #ifdef CONFIG_SCHEDSTATS | |
3309 | if (schedstat_enabled()) | |
3310 | return; | |
3311 | ||
3312 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
3313 | if (trace_sched_stat_wait_enabled() || | |
3314 | trace_sched_stat_sleep_enabled() || | |
3315 | trace_sched_stat_iowait_enabled() || | |
3316 | trace_sched_stat_blocked_enabled() || | |
3317 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 3318 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 MG |
3319 | "stat_blocked and stat_runtime require the " |
3320 | "kernel parameter schedstats=enabled or " | |
3321 | "kernel.sched_schedstats=1\n"); | |
3322 | } | |
3323 | #endif | |
3324 | } | |
3325 | ||
b5179ac7 PZ |
3326 | |
3327 | /* | |
3328 | * MIGRATION | |
3329 | * | |
3330 | * dequeue | |
3331 | * update_curr() | |
3332 | * update_min_vruntime() | |
3333 | * vruntime -= min_vruntime | |
3334 | * | |
3335 | * enqueue | |
3336 | * update_curr() | |
3337 | * update_min_vruntime() | |
3338 | * vruntime += min_vruntime | |
3339 | * | |
3340 | * this way the vruntime transition between RQs is done when both | |
3341 | * min_vruntime are up-to-date. | |
3342 | * | |
3343 | * WAKEUP (remote) | |
3344 | * | |
59efa0ba | 3345 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
3346 | * vruntime -= min_vruntime |
3347 | * | |
3348 | * enqueue | |
3349 | * update_curr() | |
3350 | * update_min_vruntime() | |
3351 | * vruntime += min_vruntime | |
3352 | * | |
3353 | * this way we don't have the most up-to-date min_vruntime on the originating | |
3354 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
3355 | */ | |
3356 | ||
bf0f6f24 | 3357 | static void |
88ec22d3 | 3358 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3359 | { |
2f950354 PZ |
3360 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
3361 | bool curr = cfs_rq->curr == se; | |
3362 | ||
88ec22d3 | 3363 | /* |
2f950354 PZ |
3364 | * If we're the current task, we must renormalise before calling |
3365 | * update_curr(). | |
88ec22d3 | 3366 | */ |
2f950354 | 3367 | if (renorm && curr) |
88ec22d3 PZ |
3368 | se->vruntime += cfs_rq->min_vruntime; |
3369 | ||
2f950354 PZ |
3370 | update_curr(cfs_rq); |
3371 | ||
bf0f6f24 | 3372 | /* |
2f950354 PZ |
3373 | * Otherwise, renormalise after, such that we're placed at the current |
3374 | * moment in time, instead of some random moment in the past. Being | |
3375 | * placed in the past could significantly boost this task to the | |
3376 | * fairness detriment of existing tasks. | |
bf0f6f24 | 3377 | */ |
2f950354 PZ |
3378 | if (renorm && !curr) |
3379 | se->vruntime += cfs_rq->min_vruntime; | |
3380 | ||
9d89c257 | 3381 | enqueue_entity_load_avg(cfs_rq, se); |
17bc14b7 LT |
3382 | account_entity_enqueue(cfs_rq, se); |
3383 | update_cfs_shares(cfs_rq); | |
bf0f6f24 | 3384 | |
1a3d027c | 3385 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 3386 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 3387 | |
cb251765 | 3388 | check_schedstat_required(); |
4fa8d299 JP |
3389 | update_stats_enqueue(cfs_rq, se, flags); |
3390 | check_spread(cfs_rq, se); | |
2f950354 | 3391 | if (!curr) |
83b699ed | 3392 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 3393 | se->on_rq = 1; |
3d4b47b4 | 3394 | |
d3d9dc33 | 3395 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 3396 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
3397 | check_enqueue_throttle(cfs_rq); |
3398 | } | |
bf0f6f24 IM |
3399 | } |
3400 | ||
2c13c919 | 3401 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 3402 | { |
2c13c919 RR |
3403 | for_each_sched_entity(se) { |
3404 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3405 | if (cfs_rq->last != se) |
2c13c919 | 3406 | break; |
f1044799 PZ |
3407 | |
3408 | cfs_rq->last = NULL; | |
2c13c919 RR |
3409 | } |
3410 | } | |
2002c695 | 3411 | |
2c13c919 RR |
3412 | static void __clear_buddies_next(struct sched_entity *se) |
3413 | { | |
3414 | for_each_sched_entity(se) { | |
3415 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3416 | if (cfs_rq->next != se) |
2c13c919 | 3417 | break; |
f1044799 PZ |
3418 | |
3419 | cfs_rq->next = NULL; | |
2c13c919 | 3420 | } |
2002c695 PZ |
3421 | } |
3422 | ||
ac53db59 RR |
3423 | static void __clear_buddies_skip(struct sched_entity *se) |
3424 | { | |
3425 | for_each_sched_entity(se) { | |
3426 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3427 | if (cfs_rq->skip != se) |
ac53db59 | 3428 | break; |
f1044799 PZ |
3429 | |
3430 | cfs_rq->skip = NULL; | |
ac53db59 RR |
3431 | } |
3432 | } | |
3433 | ||
a571bbea PZ |
3434 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3435 | { | |
2c13c919 RR |
3436 | if (cfs_rq->last == se) |
3437 | __clear_buddies_last(se); | |
3438 | ||
3439 | if (cfs_rq->next == se) | |
3440 | __clear_buddies_next(se); | |
ac53db59 RR |
3441 | |
3442 | if (cfs_rq->skip == se) | |
3443 | __clear_buddies_skip(se); | |
a571bbea PZ |
3444 | } |
3445 | ||
6c16a6dc | 3446 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 3447 | |
bf0f6f24 | 3448 | static void |
371fd7e7 | 3449 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3450 | { |
a2a2d680 DA |
3451 | /* |
3452 | * Update run-time statistics of the 'current'. | |
3453 | */ | |
3454 | update_curr(cfs_rq); | |
13962234 | 3455 | dequeue_entity_load_avg(cfs_rq, se); |
a2a2d680 | 3456 | |
4fa8d299 | 3457 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 3458 | |
2002c695 | 3459 | clear_buddies(cfs_rq, se); |
4793241b | 3460 | |
83b699ed | 3461 | if (se != cfs_rq->curr) |
30cfdcfc | 3462 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 3463 | se->on_rq = 0; |
30cfdcfc | 3464 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
3465 | |
3466 | /* | |
3467 | * Normalize the entity after updating the min_vruntime because the | |
3468 | * update can refer to the ->curr item and we need to reflect this | |
3469 | * movement in our normalized position. | |
3470 | */ | |
371fd7e7 | 3471 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 3472 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 3473 | |
d8b4986d PT |
3474 | /* return excess runtime on last dequeue */ |
3475 | return_cfs_rq_runtime(cfs_rq); | |
3476 | ||
1e876231 | 3477 | update_min_vruntime(cfs_rq); |
17bc14b7 | 3478 | update_cfs_shares(cfs_rq); |
bf0f6f24 IM |
3479 | } |
3480 | ||
3481 | /* | |
3482 | * Preempt the current task with a newly woken task if needed: | |
3483 | */ | |
7c92e54f | 3484 | static void |
2e09bf55 | 3485 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 3486 | { |
11697830 | 3487 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
3488 | struct sched_entity *se; |
3489 | s64 delta; | |
11697830 | 3490 | |
6d0f0ebd | 3491 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 3492 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 3493 | if (delta_exec > ideal_runtime) { |
8875125e | 3494 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
3495 | /* |
3496 | * The current task ran long enough, ensure it doesn't get | |
3497 | * re-elected due to buddy favours. | |
3498 | */ | |
3499 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
3500 | return; |
3501 | } | |
3502 | ||
3503 | /* | |
3504 | * Ensure that a task that missed wakeup preemption by a | |
3505 | * narrow margin doesn't have to wait for a full slice. | |
3506 | * This also mitigates buddy induced latencies under load. | |
3507 | */ | |
f685ceac MG |
3508 | if (delta_exec < sysctl_sched_min_granularity) |
3509 | return; | |
3510 | ||
f4cfb33e WX |
3511 | se = __pick_first_entity(cfs_rq); |
3512 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 3513 | |
f4cfb33e WX |
3514 | if (delta < 0) |
3515 | return; | |
d7d82944 | 3516 | |
f4cfb33e | 3517 | if (delta > ideal_runtime) |
8875125e | 3518 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
3519 | } |
3520 | ||
83b699ed | 3521 | static void |
8494f412 | 3522 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 3523 | { |
83b699ed SV |
3524 | /* 'current' is not kept within the tree. */ |
3525 | if (se->on_rq) { | |
3526 | /* | |
3527 | * Any task has to be enqueued before it get to execute on | |
3528 | * a CPU. So account for the time it spent waiting on the | |
3529 | * runqueue. | |
3530 | */ | |
4fa8d299 | 3531 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 3532 | __dequeue_entity(cfs_rq, se); |
9d89c257 | 3533 | update_load_avg(se, 1); |
83b699ed SV |
3534 | } |
3535 | ||
79303e9e | 3536 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 3537 | cfs_rq->curr = se; |
4fa8d299 | 3538 | |
eba1ed4b IM |
3539 | /* |
3540 | * Track our maximum slice length, if the CPU's load is at | |
3541 | * least twice that of our own weight (i.e. dont track it | |
3542 | * when there are only lesser-weight tasks around): | |
3543 | */ | |
cb251765 | 3544 | if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
4fa8d299 JP |
3545 | schedstat_set(se->statistics.slice_max, |
3546 | max((u64)schedstat_val(se->statistics.slice_max), | |
3547 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 3548 | } |
4fa8d299 | 3549 | |
4a55b450 | 3550 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
3551 | } |
3552 | ||
3f3a4904 PZ |
3553 | static int |
3554 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
3555 | ||
ac53db59 RR |
3556 | /* |
3557 | * Pick the next process, keeping these things in mind, in this order: | |
3558 | * 1) keep things fair between processes/task groups | |
3559 | * 2) pick the "next" process, since someone really wants that to run | |
3560 | * 3) pick the "last" process, for cache locality | |
3561 | * 4) do not run the "skip" process, if something else is available | |
3562 | */ | |
678d5718 PZ |
3563 | static struct sched_entity * |
3564 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 3565 | { |
678d5718 PZ |
3566 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
3567 | struct sched_entity *se; | |
3568 | ||
3569 | /* | |
3570 | * If curr is set we have to see if its left of the leftmost entity | |
3571 | * still in the tree, provided there was anything in the tree at all. | |
3572 | */ | |
3573 | if (!left || (curr && entity_before(curr, left))) | |
3574 | left = curr; | |
3575 | ||
3576 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 3577 | |
ac53db59 RR |
3578 | /* |
3579 | * Avoid running the skip buddy, if running something else can | |
3580 | * be done without getting too unfair. | |
3581 | */ | |
3582 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
3583 | struct sched_entity *second; |
3584 | ||
3585 | if (se == curr) { | |
3586 | second = __pick_first_entity(cfs_rq); | |
3587 | } else { | |
3588 | second = __pick_next_entity(se); | |
3589 | if (!second || (curr && entity_before(curr, second))) | |
3590 | second = curr; | |
3591 | } | |
3592 | ||
ac53db59 RR |
3593 | if (second && wakeup_preempt_entity(second, left) < 1) |
3594 | se = second; | |
3595 | } | |
aa2ac252 | 3596 | |
f685ceac MG |
3597 | /* |
3598 | * Prefer last buddy, try to return the CPU to a preempted task. | |
3599 | */ | |
3600 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
3601 | se = cfs_rq->last; | |
3602 | ||
ac53db59 RR |
3603 | /* |
3604 | * Someone really wants this to run. If it's not unfair, run it. | |
3605 | */ | |
3606 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
3607 | se = cfs_rq->next; | |
3608 | ||
f685ceac | 3609 | clear_buddies(cfs_rq, se); |
4793241b PZ |
3610 | |
3611 | return se; | |
aa2ac252 PZ |
3612 | } |
3613 | ||
678d5718 | 3614 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 3615 | |
ab6cde26 | 3616 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
3617 | { |
3618 | /* | |
3619 | * If still on the runqueue then deactivate_task() | |
3620 | * was not called and update_curr() has to be done: | |
3621 | */ | |
3622 | if (prev->on_rq) | |
b7cc0896 | 3623 | update_curr(cfs_rq); |
bf0f6f24 | 3624 | |
d3d9dc33 PT |
3625 | /* throttle cfs_rqs exceeding runtime */ |
3626 | check_cfs_rq_runtime(cfs_rq); | |
3627 | ||
4fa8d299 | 3628 | check_spread(cfs_rq, prev); |
cb251765 | 3629 | |
30cfdcfc | 3630 | if (prev->on_rq) { |
4fa8d299 | 3631 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
3632 | /* Put 'current' back into the tree. */ |
3633 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 3634 | /* in !on_rq case, update occurred at dequeue */ |
9d89c257 | 3635 | update_load_avg(prev, 0); |
30cfdcfc | 3636 | } |
429d43bc | 3637 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
3638 | } |
3639 | ||
8f4d37ec PZ |
3640 | static void |
3641 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 3642 | { |
bf0f6f24 | 3643 | /* |
30cfdcfc | 3644 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3645 | */ |
30cfdcfc | 3646 | update_curr(cfs_rq); |
bf0f6f24 | 3647 | |
9d85f21c PT |
3648 | /* |
3649 | * Ensure that runnable average is periodically updated. | |
3650 | */ | |
9d89c257 | 3651 | update_load_avg(curr, 1); |
bf0bd948 | 3652 | update_cfs_shares(cfs_rq); |
9d85f21c | 3653 | |
8f4d37ec PZ |
3654 | #ifdef CONFIG_SCHED_HRTICK |
3655 | /* | |
3656 | * queued ticks are scheduled to match the slice, so don't bother | |
3657 | * validating it and just reschedule. | |
3658 | */ | |
983ed7a6 | 3659 | if (queued) { |
8875125e | 3660 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
3661 | return; |
3662 | } | |
8f4d37ec PZ |
3663 | /* |
3664 | * don't let the period tick interfere with the hrtick preemption | |
3665 | */ | |
3666 | if (!sched_feat(DOUBLE_TICK) && | |
3667 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
3668 | return; | |
3669 | #endif | |
3670 | ||
2c2efaed | 3671 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 3672 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
3673 | } |
3674 | ||
ab84d31e PT |
3675 | |
3676 | /************************************************** | |
3677 | * CFS bandwidth control machinery | |
3678 | */ | |
3679 | ||
3680 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
3681 | |
3682 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 3683 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
3684 | |
3685 | static inline bool cfs_bandwidth_used(void) | |
3686 | { | |
c5905afb | 3687 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
3688 | } |
3689 | ||
1ee14e6c | 3690 | void cfs_bandwidth_usage_inc(void) |
029632fb | 3691 | { |
1ee14e6c BS |
3692 | static_key_slow_inc(&__cfs_bandwidth_used); |
3693 | } | |
3694 | ||
3695 | void cfs_bandwidth_usage_dec(void) | |
3696 | { | |
3697 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
3698 | } |
3699 | #else /* HAVE_JUMP_LABEL */ | |
3700 | static bool cfs_bandwidth_used(void) | |
3701 | { | |
3702 | return true; | |
3703 | } | |
3704 | ||
1ee14e6c BS |
3705 | void cfs_bandwidth_usage_inc(void) {} |
3706 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
3707 | #endif /* HAVE_JUMP_LABEL */ |
3708 | ||
ab84d31e PT |
3709 | /* |
3710 | * default period for cfs group bandwidth. | |
3711 | * default: 0.1s, units: nanoseconds | |
3712 | */ | |
3713 | static inline u64 default_cfs_period(void) | |
3714 | { | |
3715 | return 100000000ULL; | |
3716 | } | |
ec12cb7f PT |
3717 | |
3718 | static inline u64 sched_cfs_bandwidth_slice(void) | |
3719 | { | |
3720 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
3721 | } | |
3722 | ||
a9cf55b2 PT |
3723 | /* |
3724 | * Replenish runtime according to assigned quota and update expiration time. | |
3725 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
3726 | * additional synchronization around rq->lock. | |
3727 | * | |
3728 | * requires cfs_b->lock | |
3729 | */ | |
029632fb | 3730 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
3731 | { |
3732 | u64 now; | |
3733 | ||
3734 | if (cfs_b->quota == RUNTIME_INF) | |
3735 | return; | |
3736 | ||
3737 | now = sched_clock_cpu(smp_processor_id()); | |
3738 | cfs_b->runtime = cfs_b->quota; | |
3739 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
3740 | } | |
3741 | ||
029632fb PZ |
3742 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3743 | { | |
3744 | return &tg->cfs_bandwidth; | |
3745 | } | |
3746 | ||
f1b17280 PT |
3747 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
3748 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
3749 | { | |
3750 | if (unlikely(cfs_rq->throttle_count)) | |
1a99ae3f | 3751 | return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time; |
f1b17280 | 3752 | |
78becc27 | 3753 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
3754 | } |
3755 | ||
85dac906 PT |
3756 | /* returns 0 on failure to allocate runtime */ |
3757 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
3758 | { |
3759 | struct task_group *tg = cfs_rq->tg; | |
3760 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 3761 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
3762 | |
3763 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
3764 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
3765 | ||
3766 | raw_spin_lock(&cfs_b->lock); | |
3767 | if (cfs_b->quota == RUNTIME_INF) | |
3768 | amount = min_amount; | |
58088ad0 | 3769 | else { |
77a4d1a1 | 3770 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
3771 | |
3772 | if (cfs_b->runtime > 0) { | |
3773 | amount = min(cfs_b->runtime, min_amount); | |
3774 | cfs_b->runtime -= amount; | |
3775 | cfs_b->idle = 0; | |
3776 | } | |
ec12cb7f | 3777 | } |
a9cf55b2 | 3778 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
3779 | raw_spin_unlock(&cfs_b->lock); |
3780 | ||
3781 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
3782 | /* |
3783 | * we may have advanced our local expiration to account for allowed | |
3784 | * spread between our sched_clock and the one on which runtime was | |
3785 | * issued. | |
3786 | */ | |
3787 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
3788 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
3789 | |
3790 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
3791 | } |
3792 | ||
a9cf55b2 PT |
3793 | /* |
3794 | * Note: This depends on the synchronization provided by sched_clock and the | |
3795 | * fact that rq->clock snapshots this value. | |
3796 | */ | |
3797 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 3798 | { |
a9cf55b2 | 3799 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
3800 | |
3801 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 3802 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
3803 | return; |
3804 | ||
a9cf55b2 PT |
3805 | if (cfs_rq->runtime_remaining < 0) |
3806 | return; | |
3807 | ||
3808 | /* | |
3809 | * If the local deadline has passed we have to consider the | |
3810 | * possibility that our sched_clock is 'fast' and the global deadline | |
3811 | * has not truly expired. | |
3812 | * | |
3813 | * Fortunately we can check determine whether this the case by checking | |
51f2176d BS |
3814 | * whether the global deadline has advanced. It is valid to compare |
3815 | * cfs_b->runtime_expires without any locks since we only care about | |
3816 | * exact equality, so a partial write will still work. | |
a9cf55b2 PT |
3817 | */ |
3818 | ||
51f2176d | 3819 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { |
a9cf55b2 PT |
3820 | /* extend local deadline, drift is bounded above by 2 ticks */ |
3821 | cfs_rq->runtime_expires += TICK_NSEC; | |
3822 | } else { | |
3823 | /* global deadline is ahead, expiration has passed */ | |
3824 | cfs_rq->runtime_remaining = 0; | |
3825 | } | |
3826 | } | |
3827 | ||
9dbdb155 | 3828 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
3829 | { |
3830 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 3831 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
3832 | expire_cfs_rq_runtime(cfs_rq); |
3833 | ||
3834 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
3835 | return; |
3836 | ||
85dac906 PT |
3837 | /* |
3838 | * if we're unable to extend our runtime we resched so that the active | |
3839 | * hierarchy can be throttled | |
3840 | */ | |
3841 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 3842 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
3843 | } |
3844 | ||
6c16a6dc | 3845 | static __always_inline |
9dbdb155 | 3846 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 3847 | { |
56f570e5 | 3848 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
3849 | return; |
3850 | ||
3851 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
3852 | } | |
3853 | ||
85dac906 PT |
3854 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3855 | { | |
56f570e5 | 3856 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
3857 | } |
3858 | ||
64660c86 PT |
3859 | /* check whether cfs_rq, or any parent, is throttled */ |
3860 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
3861 | { | |
56f570e5 | 3862 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
3863 | } |
3864 | ||
3865 | /* | |
3866 | * Ensure that neither of the group entities corresponding to src_cpu or | |
3867 | * dest_cpu are members of a throttled hierarchy when performing group | |
3868 | * load-balance operations. | |
3869 | */ | |
3870 | static inline int throttled_lb_pair(struct task_group *tg, | |
3871 | int src_cpu, int dest_cpu) | |
3872 | { | |
3873 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
3874 | ||
3875 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
3876 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
3877 | ||
3878 | return throttled_hierarchy(src_cfs_rq) || | |
3879 | throttled_hierarchy(dest_cfs_rq); | |
3880 | } | |
3881 | ||
3882 | /* updated child weight may affect parent so we have to do this bottom up */ | |
3883 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
3884 | { | |
3885 | struct rq *rq = data; | |
3886 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3887 | ||
3888 | cfs_rq->throttle_count--; | |
64660c86 | 3889 | if (!cfs_rq->throttle_count) { |
f1b17280 | 3890 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 3891 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 3892 | cfs_rq->throttled_clock_task; |
64660c86 | 3893 | } |
64660c86 PT |
3894 | |
3895 | return 0; | |
3896 | } | |
3897 | ||
3898 | static int tg_throttle_down(struct task_group *tg, void *data) | |
3899 | { | |
3900 | struct rq *rq = data; | |
3901 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
3902 | ||
82958366 PT |
3903 | /* group is entering throttled state, stop time */ |
3904 | if (!cfs_rq->throttle_count) | |
78becc27 | 3905 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
3906 | cfs_rq->throttle_count++; |
3907 | ||
3908 | return 0; | |
3909 | } | |
3910 | ||
d3d9dc33 | 3911 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
3912 | { |
3913 | struct rq *rq = rq_of(cfs_rq); | |
3914 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3915 | struct sched_entity *se; | |
3916 | long task_delta, dequeue = 1; | |
77a4d1a1 | 3917 | bool empty; |
85dac906 PT |
3918 | |
3919 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
3920 | ||
f1b17280 | 3921 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
3922 | rcu_read_lock(); |
3923 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
3924 | rcu_read_unlock(); | |
85dac906 PT |
3925 | |
3926 | task_delta = cfs_rq->h_nr_running; | |
3927 | for_each_sched_entity(se) { | |
3928 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
3929 | /* throttled entity or throttle-on-deactivate */ | |
3930 | if (!se->on_rq) | |
3931 | break; | |
3932 | ||
3933 | if (dequeue) | |
3934 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
3935 | qcfs_rq->h_nr_running -= task_delta; | |
3936 | ||
3937 | if (qcfs_rq->load.weight) | |
3938 | dequeue = 0; | |
3939 | } | |
3940 | ||
3941 | if (!se) | |
72465447 | 3942 | sub_nr_running(rq, task_delta); |
85dac906 PT |
3943 | |
3944 | cfs_rq->throttled = 1; | |
78becc27 | 3945 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 3946 | raw_spin_lock(&cfs_b->lock); |
d49db342 | 3947 | empty = list_empty(&cfs_b->throttled_cfs_rq); |
77a4d1a1 | 3948 | |
c06f04c7 BS |
3949 | /* |
3950 | * Add to the _head_ of the list, so that an already-started | |
3951 | * distribute_cfs_runtime will not see us | |
3952 | */ | |
3953 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
77a4d1a1 PZ |
3954 | |
3955 | /* | |
3956 | * If we're the first throttled task, make sure the bandwidth | |
3957 | * timer is running. | |
3958 | */ | |
3959 | if (empty) | |
3960 | start_cfs_bandwidth(cfs_b); | |
3961 | ||
85dac906 PT |
3962 | raw_spin_unlock(&cfs_b->lock); |
3963 | } | |
3964 | ||
029632fb | 3965 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
3966 | { |
3967 | struct rq *rq = rq_of(cfs_rq); | |
3968 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
3969 | struct sched_entity *se; | |
3970 | int enqueue = 1; | |
3971 | long task_delta; | |
3972 | ||
22b958d8 | 3973 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
3974 | |
3975 | cfs_rq->throttled = 0; | |
1a55af2e FW |
3976 | |
3977 | update_rq_clock(rq); | |
3978 | ||
671fd9da | 3979 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 3980 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
3981 | list_del_rcu(&cfs_rq->throttled_list); |
3982 | raw_spin_unlock(&cfs_b->lock); | |
3983 | ||
64660c86 PT |
3984 | /* update hierarchical throttle state */ |
3985 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
3986 | ||
671fd9da PT |
3987 | if (!cfs_rq->load.weight) |
3988 | return; | |
3989 | ||
3990 | task_delta = cfs_rq->h_nr_running; | |
3991 | for_each_sched_entity(se) { | |
3992 | if (se->on_rq) | |
3993 | enqueue = 0; | |
3994 | ||
3995 | cfs_rq = cfs_rq_of(se); | |
3996 | if (enqueue) | |
3997 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
3998 | cfs_rq->h_nr_running += task_delta; | |
3999 | ||
4000 | if (cfs_rq_throttled(cfs_rq)) | |
4001 | break; | |
4002 | } | |
4003 | ||
4004 | if (!se) | |
72465447 | 4005 | add_nr_running(rq, task_delta); |
671fd9da PT |
4006 | |
4007 | /* determine whether we need to wake up potentially idle cpu */ | |
4008 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
8875125e | 4009 | resched_curr(rq); |
671fd9da PT |
4010 | } |
4011 | ||
4012 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
4013 | u64 remaining, u64 expires) | |
4014 | { | |
4015 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
4016 | u64 runtime; |
4017 | u64 starting_runtime = remaining; | |
671fd9da PT |
4018 | |
4019 | rcu_read_lock(); | |
4020 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4021 | throttled_list) { | |
4022 | struct rq *rq = rq_of(cfs_rq); | |
4023 | ||
4024 | raw_spin_lock(&rq->lock); | |
4025 | if (!cfs_rq_throttled(cfs_rq)) | |
4026 | goto next; | |
4027 | ||
4028 | runtime = -cfs_rq->runtime_remaining + 1; | |
4029 | if (runtime > remaining) | |
4030 | runtime = remaining; | |
4031 | remaining -= runtime; | |
4032 | ||
4033 | cfs_rq->runtime_remaining += runtime; | |
4034 | cfs_rq->runtime_expires = expires; | |
4035 | ||
4036 | /* we check whether we're throttled above */ | |
4037 | if (cfs_rq->runtime_remaining > 0) | |
4038 | unthrottle_cfs_rq(cfs_rq); | |
4039 | ||
4040 | next: | |
4041 | raw_spin_unlock(&rq->lock); | |
4042 | ||
4043 | if (!remaining) | |
4044 | break; | |
4045 | } | |
4046 | rcu_read_unlock(); | |
4047 | ||
c06f04c7 | 4048 | return starting_runtime - remaining; |
671fd9da PT |
4049 | } |
4050 | ||
58088ad0 PT |
4051 | /* |
4052 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4053 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4054 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4055 | * used to track this state. | |
4056 | */ | |
4057 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
4058 | { | |
671fd9da | 4059 | u64 runtime, runtime_expires; |
51f2176d | 4060 | int throttled; |
58088ad0 | 4061 | |
58088ad0 PT |
4062 | /* no need to continue the timer with no bandwidth constraint */ |
4063 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4064 | goto out_deactivate; |
58088ad0 | 4065 | |
671fd9da | 4066 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4067 | cfs_b->nr_periods += overrun; |
671fd9da | 4068 | |
51f2176d BS |
4069 | /* |
4070 | * idle depends on !throttled (for the case of a large deficit), and if | |
4071 | * we're going inactive then everything else can be deferred | |
4072 | */ | |
4073 | if (cfs_b->idle && !throttled) | |
4074 | goto out_deactivate; | |
a9cf55b2 PT |
4075 | |
4076 | __refill_cfs_bandwidth_runtime(cfs_b); | |
4077 | ||
671fd9da PT |
4078 | if (!throttled) { |
4079 | /* mark as potentially idle for the upcoming period */ | |
4080 | cfs_b->idle = 1; | |
51f2176d | 4081 | return 0; |
671fd9da PT |
4082 | } |
4083 | ||
e8da1b18 NR |
4084 | /* account preceding periods in which throttling occurred */ |
4085 | cfs_b->nr_throttled += overrun; | |
4086 | ||
671fd9da | 4087 | runtime_expires = cfs_b->runtime_expires; |
671fd9da PT |
4088 | |
4089 | /* | |
c06f04c7 BS |
4090 | * This check is repeated as we are holding onto the new bandwidth while |
4091 | * we unthrottle. This can potentially race with an unthrottled group | |
4092 | * trying to acquire new bandwidth from the global pool. This can result | |
4093 | * in us over-using our runtime if it is all used during this loop, but | |
4094 | * only by limited amounts in that extreme case. | |
671fd9da | 4095 | */ |
c06f04c7 BS |
4096 | while (throttled && cfs_b->runtime > 0) { |
4097 | runtime = cfs_b->runtime; | |
671fd9da PT |
4098 | raw_spin_unlock(&cfs_b->lock); |
4099 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
4100 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
4101 | runtime_expires); | |
4102 | raw_spin_lock(&cfs_b->lock); | |
4103 | ||
4104 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
c06f04c7 BS |
4105 | |
4106 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | |
671fd9da | 4107 | } |
58088ad0 | 4108 | |
671fd9da PT |
4109 | /* |
4110 | * While we are ensured activity in the period following an | |
4111 | * unthrottle, this also covers the case in which the new bandwidth is | |
4112 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
4113 | * timer to remain active while there are any throttled entities.) | |
4114 | */ | |
4115 | cfs_b->idle = 0; | |
58088ad0 | 4116 | |
51f2176d BS |
4117 | return 0; |
4118 | ||
4119 | out_deactivate: | |
51f2176d | 4120 | return 1; |
58088ad0 | 4121 | } |
d3d9dc33 | 4122 | |
d8b4986d PT |
4123 | /* a cfs_rq won't donate quota below this amount */ |
4124 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
4125 | /* minimum remaining period time to redistribute slack quota */ | |
4126 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
4127 | /* how long we wait to gather additional slack before distributing */ | |
4128 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
4129 | ||
db06e78c BS |
4130 | /* |
4131 | * Are we near the end of the current quota period? | |
4132 | * | |
4133 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 4134 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
4135 | * migrate_hrtimers, base is never cleared, so we are fine. |
4136 | */ | |
d8b4986d PT |
4137 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
4138 | { | |
4139 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
4140 | u64 remaining; | |
4141 | ||
4142 | /* if the call-back is running a quota refresh is already occurring */ | |
4143 | if (hrtimer_callback_running(refresh_timer)) | |
4144 | return 1; | |
4145 | ||
4146 | /* is a quota refresh about to occur? */ | |
4147 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
4148 | if (remaining < min_expire) | |
4149 | return 1; | |
4150 | ||
4151 | return 0; | |
4152 | } | |
4153 | ||
4154 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
4155 | { | |
4156 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
4157 | ||
4158 | /* if there's a quota refresh soon don't bother with slack */ | |
4159 | if (runtime_refresh_within(cfs_b, min_left)) | |
4160 | return; | |
4161 | ||
4cfafd30 PZ |
4162 | hrtimer_start(&cfs_b->slack_timer, |
4163 | ns_to_ktime(cfs_bandwidth_slack_period), | |
4164 | HRTIMER_MODE_REL); | |
d8b4986d PT |
4165 | } |
4166 | ||
4167 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
4168 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4169 | { | |
4170 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4171 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
4172 | ||
4173 | if (slack_runtime <= 0) | |
4174 | return; | |
4175 | ||
4176 | raw_spin_lock(&cfs_b->lock); | |
4177 | if (cfs_b->quota != RUNTIME_INF && | |
4178 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
4179 | cfs_b->runtime += slack_runtime; | |
4180 | ||
4181 | /* we are under rq->lock, defer unthrottling using a timer */ | |
4182 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
4183 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
4184 | start_cfs_slack_bandwidth(cfs_b); | |
4185 | } | |
4186 | raw_spin_unlock(&cfs_b->lock); | |
4187 | ||
4188 | /* even if it's not valid for return we don't want to try again */ | |
4189 | cfs_rq->runtime_remaining -= slack_runtime; | |
4190 | } | |
4191 | ||
4192 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4193 | { | |
56f570e5 PT |
4194 | if (!cfs_bandwidth_used()) |
4195 | return; | |
4196 | ||
fccfdc6f | 4197 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
4198 | return; |
4199 | ||
4200 | __return_cfs_rq_runtime(cfs_rq); | |
4201 | } | |
4202 | ||
4203 | /* | |
4204 | * This is done with a timer (instead of inline with bandwidth return) since | |
4205 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
4206 | */ | |
4207 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
4208 | { | |
4209 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
4210 | u64 expires; | |
4211 | ||
4212 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
4213 | raw_spin_lock(&cfs_b->lock); |
4214 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
4215 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 4216 | return; |
db06e78c | 4217 | } |
d8b4986d | 4218 | |
c06f04c7 | 4219 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 4220 | runtime = cfs_b->runtime; |
c06f04c7 | 4221 | |
d8b4986d PT |
4222 | expires = cfs_b->runtime_expires; |
4223 | raw_spin_unlock(&cfs_b->lock); | |
4224 | ||
4225 | if (!runtime) | |
4226 | return; | |
4227 | ||
4228 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
4229 | ||
4230 | raw_spin_lock(&cfs_b->lock); | |
4231 | if (expires == cfs_b->runtime_expires) | |
c06f04c7 | 4232 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
d8b4986d PT |
4233 | raw_spin_unlock(&cfs_b->lock); |
4234 | } | |
4235 | ||
d3d9dc33 PT |
4236 | /* |
4237 | * When a group wakes up we want to make sure that its quota is not already | |
4238 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
4239 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
4240 | */ | |
4241 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
4242 | { | |
56f570e5 PT |
4243 | if (!cfs_bandwidth_used()) |
4244 | return; | |
4245 | ||
d3d9dc33 PT |
4246 | /* an active group must be handled by the update_curr()->put() path */ |
4247 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
4248 | return; | |
4249 | ||
4250 | /* ensure the group is not already throttled */ | |
4251 | if (cfs_rq_throttled(cfs_rq)) | |
4252 | return; | |
4253 | ||
4254 | /* update runtime allocation */ | |
4255 | account_cfs_rq_runtime(cfs_rq, 0); | |
4256 | if (cfs_rq->runtime_remaining <= 0) | |
4257 | throttle_cfs_rq(cfs_rq); | |
4258 | } | |
4259 | ||
55e16d30 PZ |
4260 | static void sync_throttle(struct task_group *tg, int cpu) |
4261 | { | |
4262 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
4263 | ||
4264 | if (!cfs_bandwidth_used()) | |
4265 | return; | |
4266 | ||
4267 | if (!tg->parent) | |
4268 | return; | |
4269 | ||
4270 | cfs_rq = tg->cfs_rq[cpu]; | |
4271 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
4272 | ||
4273 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 4274 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
4275 | } |
4276 | ||
d3d9dc33 | 4277 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 4278 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 4279 | { |
56f570e5 | 4280 | if (!cfs_bandwidth_used()) |
678d5718 | 4281 | return false; |
56f570e5 | 4282 | |
d3d9dc33 | 4283 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 4284 | return false; |
d3d9dc33 PT |
4285 | |
4286 | /* | |
4287 | * it's possible for a throttled entity to be forced into a running | |
4288 | * state (e.g. set_curr_task), in this case we're finished. | |
4289 | */ | |
4290 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 4291 | return true; |
d3d9dc33 PT |
4292 | |
4293 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 4294 | return true; |
d3d9dc33 | 4295 | } |
029632fb | 4296 | |
029632fb PZ |
4297 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
4298 | { | |
4299 | struct cfs_bandwidth *cfs_b = | |
4300 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 4301 | |
029632fb PZ |
4302 | do_sched_cfs_slack_timer(cfs_b); |
4303 | ||
4304 | return HRTIMER_NORESTART; | |
4305 | } | |
4306 | ||
4307 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
4308 | { | |
4309 | struct cfs_bandwidth *cfs_b = | |
4310 | container_of(timer, struct cfs_bandwidth, period_timer); | |
029632fb PZ |
4311 | int overrun; |
4312 | int idle = 0; | |
4313 | ||
51f2176d | 4314 | raw_spin_lock(&cfs_b->lock); |
029632fb | 4315 | for (;;) { |
77a4d1a1 | 4316 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
4317 | if (!overrun) |
4318 | break; | |
4319 | ||
4320 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
4321 | } | |
4cfafd30 PZ |
4322 | if (idle) |
4323 | cfs_b->period_active = 0; | |
51f2176d | 4324 | raw_spin_unlock(&cfs_b->lock); |
029632fb PZ |
4325 | |
4326 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
4327 | } | |
4328 | ||
4329 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4330 | { | |
4331 | raw_spin_lock_init(&cfs_b->lock); | |
4332 | cfs_b->runtime = 0; | |
4333 | cfs_b->quota = RUNTIME_INF; | |
4334 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
4335 | ||
4336 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 4337 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
4338 | cfs_b->period_timer.function = sched_cfs_period_timer; |
4339 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
4340 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
4341 | } | |
4342 | ||
4343 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4344 | { | |
4345 | cfs_rq->runtime_enabled = 0; | |
4346 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
4347 | } | |
4348 | ||
77a4d1a1 | 4349 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 4350 | { |
4cfafd30 | 4351 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 4352 | |
4cfafd30 PZ |
4353 | if (!cfs_b->period_active) { |
4354 | cfs_b->period_active = 1; | |
4355 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); | |
4356 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); | |
4357 | } | |
029632fb PZ |
4358 | } |
4359 | ||
4360 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4361 | { | |
7f1a169b TH |
4362 | /* init_cfs_bandwidth() was not called */ |
4363 | if (!cfs_b->throttled_cfs_rq.next) | |
4364 | return; | |
4365 | ||
029632fb PZ |
4366 | hrtimer_cancel(&cfs_b->period_timer); |
4367 | hrtimer_cancel(&cfs_b->slack_timer); | |
4368 | } | |
4369 | ||
0e59bdae KT |
4370 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
4371 | { | |
4372 | struct cfs_rq *cfs_rq; | |
4373 | ||
4374 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
4375 | struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; | |
4376 | ||
4377 | raw_spin_lock(&cfs_b->lock); | |
4378 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
4379 | raw_spin_unlock(&cfs_b->lock); | |
4380 | } | |
4381 | } | |
4382 | ||
38dc3348 | 4383 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
4384 | { |
4385 | struct cfs_rq *cfs_rq; | |
4386 | ||
4387 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
029632fb PZ |
4388 | if (!cfs_rq->runtime_enabled) |
4389 | continue; | |
4390 | ||
4391 | /* | |
4392 | * clock_task is not advancing so we just need to make sure | |
4393 | * there's some valid quota amount | |
4394 | */ | |
51f2176d | 4395 | cfs_rq->runtime_remaining = 1; |
0e59bdae KT |
4396 | /* |
4397 | * Offline rq is schedulable till cpu is completely disabled | |
4398 | * in take_cpu_down(), so we prevent new cfs throttling here. | |
4399 | */ | |
4400 | cfs_rq->runtime_enabled = 0; | |
4401 | ||
029632fb PZ |
4402 | if (cfs_rq_throttled(cfs_rq)) |
4403 | unthrottle_cfs_rq(cfs_rq); | |
4404 | } | |
4405 | } | |
4406 | ||
4407 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
4408 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
4409 | { | |
78becc27 | 4410 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
4411 | } |
4412 | ||
9dbdb155 | 4413 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 4414 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 4415 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 4416 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 4417 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
4418 | |
4419 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
4420 | { | |
4421 | return 0; | |
4422 | } | |
64660c86 PT |
4423 | |
4424 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4425 | { | |
4426 | return 0; | |
4427 | } | |
4428 | ||
4429 | static inline int throttled_lb_pair(struct task_group *tg, | |
4430 | int src_cpu, int dest_cpu) | |
4431 | { | |
4432 | return 0; | |
4433 | } | |
029632fb PZ |
4434 | |
4435 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
4436 | ||
4437 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
4438 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
4439 | #endif |
4440 | ||
029632fb PZ |
4441 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4442 | { | |
4443 | return NULL; | |
4444 | } | |
4445 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 4446 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 4447 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
4448 | |
4449 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
4450 | ||
bf0f6f24 IM |
4451 | /************************************************** |
4452 | * CFS operations on tasks: | |
4453 | */ | |
4454 | ||
8f4d37ec PZ |
4455 | #ifdef CONFIG_SCHED_HRTICK |
4456 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4457 | { | |
8f4d37ec PZ |
4458 | struct sched_entity *se = &p->se; |
4459 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4460 | ||
4461 | WARN_ON(task_rq(p) != rq); | |
4462 | ||
b39e66ea | 4463 | if (cfs_rq->nr_running > 1) { |
8f4d37ec PZ |
4464 | u64 slice = sched_slice(cfs_rq, se); |
4465 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
4466 | s64 delta = slice - ran; | |
4467 | ||
4468 | if (delta < 0) { | |
4469 | if (rq->curr == p) | |
8875125e | 4470 | resched_curr(rq); |
8f4d37ec PZ |
4471 | return; |
4472 | } | |
31656519 | 4473 | hrtick_start(rq, delta); |
8f4d37ec PZ |
4474 | } |
4475 | } | |
a4c2f00f PZ |
4476 | |
4477 | /* | |
4478 | * called from enqueue/dequeue and updates the hrtick when the | |
4479 | * current task is from our class and nr_running is low enough | |
4480 | * to matter. | |
4481 | */ | |
4482 | static void hrtick_update(struct rq *rq) | |
4483 | { | |
4484 | struct task_struct *curr = rq->curr; | |
4485 | ||
b39e66ea | 4486 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
4487 | return; |
4488 | ||
4489 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
4490 | hrtick_start_fair(rq, curr); | |
4491 | } | |
55e12e5e | 4492 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
4493 | static inline void |
4494 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4495 | { | |
4496 | } | |
a4c2f00f PZ |
4497 | |
4498 | static inline void hrtick_update(struct rq *rq) | |
4499 | { | |
4500 | } | |
8f4d37ec PZ |
4501 | #endif |
4502 | ||
bf0f6f24 IM |
4503 | /* |
4504 | * The enqueue_task method is called before nr_running is | |
4505 | * increased. Here we update the fair scheduling stats and | |
4506 | * then put the task into the rbtree: | |
4507 | */ | |
ea87bb78 | 4508 | static void |
371fd7e7 | 4509 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4510 | { |
4511 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4512 | struct sched_entity *se = &p->se; |
bf0f6f24 IM |
4513 | |
4514 | for_each_sched_entity(se) { | |
62fb1851 | 4515 | if (se->on_rq) |
bf0f6f24 IM |
4516 | break; |
4517 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 4518 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
4519 | |
4520 | /* | |
4521 | * end evaluation on encountering a throttled cfs_rq | |
4522 | * | |
4523 | * note: in the case of encountering a throttled cfs_rq we will | |
4524 | * post the final h_nr_running increment below. | |
e210bffd | 4525 | */ |
85dac906 PT |
4526 | if (cfs_rq_throttled(cfs_rq)) |
4527 | break; | |
953bfcd1 | 4528 | cfs_rq->h_nr_running++; |
85dac906 | 4529 | |
88ec22d3 | 4530 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 4531 | } |
8f4d37ec | 4532 | |
2069dd75 | 4533 | for_each_sched_entity(se) { |
0f317143 | 4534 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4535 | cfs_rq->h_nr_running++; |
2069dd75 | 4536 | |
85dac906 PT |
4537 | if (cfs_rq_throttled(cfs_rq)) |
4538 | break; | |
4539 | ||
9d89c257 | 4540 | update_load_avg(se, 1); |
17bc14b7 | 4541 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
4542 | } |
4543 | ||
cd126afe | 4544 | if (!se) |
72465447 | 4545 | add_nr_running(rq, 1); |
cd126afe | 4546 | |
a4c2f00f | 4547 | hrtick_update(rq); |
bf0f6f24 IM |
4548 | } |
4549 | ||
2f36825b VP |
4550 | static void set_next_buddy(struct sched_entity *se); |
4551 | ||
bf0f6f24 IM |
4552 | /* |
4553 | * The dequeue_task method is called before nr_running is | |
4554 | * decreased. We remove the task from the rbtree and | |
4555 | * update the fair scheduling stats: | |
4556 | */ | |
371fd7e7 | 4557 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4558 | { |
4559 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4560 | struct sched_entity *se = &p->se; |
2f36825b | 4561 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
4562 | |
4563 | for_each_sched_entity(se) { | |
4564 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 4565 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
4566 | |
4567 | /* | |
4568 | * end evaluation on encountering a throttled cfs_rq | |
4569 | * | |
4570 | * note: in the case of encountering a throttled cfs_rq we will | |
4571 | * post the final h_nr_running decrement below. | |
4572 | */ | |
4573 | if (cfs_rq_throttled(cfs_rq)) | |
4574 | break; | |
953bfcd1 | 4575 | cfs_rq->h_nr_running--; |
2069dd75 | 4576 | |
bf0f6f24 | 4577 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 4578 | if (cfs_rq->load.weight) { |
754bd598 KK |
4579 | /* Avoid re-evaluating load for this entity: */ |
4580 | se = parent_entity(se); | |
2f36825b VP |
4581 | /* |
4582 | * Bias pick_next to pick a task from this cfs_rq, as | |
4583 | * p is sleeping when it is within its sched_slice. | |
4584 | */ | |
754bd598 KK |
4585 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
4586 | set_next_buddy(se); | |
bf0f6f24 | 4587 | break; |
2f36825b | 4588 | } |
371fd7e7 | 4589 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 4590 | } |
8f4d37ec | 4591 | |
2069dd75 | 4592 | for_each_sched_entity(se) { |
0f317143 | 4593 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4594 | cfs_rq->h_nr_running--; |
2069dd75 | 4595 | |
85dac906 PT |
4596 | if (cfs_rq_throttled(cfs_rq)) |
4597 | break; | |
4598 | ||
9d89c257 | 4599 | update_load_avg(se, 1); |
17bc14b7 | 4600 | update_cfs_shares(cfs_rq); |
2069dd75 PZ |
4601 | } |
4602 | ||
cd126afe | 4603 | if (!se) |
72465447 | 4604 | sub_nr_running(rq, 1); |
cd126afe | 4605 | |
a4c2f00f | 4606 | hrtick_update(rq); |
bf0f6f24 IM |
4607 | } |
4608 | ||
e7693a36 | 4609 | #ifdef CONFIG_SMP |
9fd81dd5 | 4610 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 PZ |
4611 | /* |
4612 | * per rq 'load' arrray crap; XXX kill this. | |
4613 | */ | |
4614 | ||
4615 | /* | |
d937cdc5 | 4616 | * The exact cpuload calculated at every tick would be: |
3289bdb4 | 4617 | * |
d937cdc5 PZ |
4618 | * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load |
4619 | * | |
4620 | * If a cpu misses updates for n ticks (as it was idle) and update gets | |
4621 | * called on the n+1-th tick when cpu may be busy, then we have: | |
4622 | * | |
4623 | * load_n = (1 - 1/2^i)^n * load_0 | |
4624 | * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load | |
3289bdb4 PZ |
4625 | * |
4626 | * decay_load_missed() below does efficient calculation of | |
3289bdb4 | 4627 | * |
d937cdc5 PZ |
4628 | * load' = (1 - 1/2^i)^n * load |
4629 | * | |
4630 | * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors. | |
4631 | * This allows us to precompute the above in said factors, thereby allowing the | |
4632 | * reduction of an arbitrary n in O(log_2 n) steps. (See also | |
4633 | * fixed_power_int()) | |
3289bdb4 | 4634 | * |
d937cdc5 | 4635 | * The calculation is approximated on a 128 point scale. |
3289bdb4 PZ |
4636 | */ |
4637 | #define DEGRADE_SHIFT 7 | |
d937cdc5 PZ |
4638 | |
4639 | static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; | |
4640 | static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { | |
4641 | { 0, 0, 0, 0, 0, 0, 0, 0 }, | |
4642 | { 64, 32, 8, 0, 0, 0, 0, 0 }, | |
4643 | { 96, 72, 40, 12, 1, 0, 0, 0 }, | |
4644 | { 112, 98, 75, 43, 15, 1, 0, 0 }, | |
4645 | { 120, 112, 98, 76, 45, 16, 2, 0 } | |
4646 | }; | |
3289bdb4 PZ |
4647 | |
4648 | /* | |
4649 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog | |
4650 | * would be when CPU is idle and so we just decay the old load without | |
4651 | * adding any new load. | |
4652 | */ | |
4653 | static unsigned long | |
4654 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) | |
4655 | { | |
4656 | int j = 0; | |
4657 | ||
4658 | if (!missed_updates) | |
4659 | return load; | |
4660 | ||
4661 | if (missed_updates >= degrade_zero_ticks[idx]) | |
4662 | return 0; | |
4663 | ||
4664 | if (idx == 1) | |
4665 | return load >> missed_updates; | |
4666 | ||
4667 | while (missed_updates) { | |
4668 | if (missed_updates % 2) | |
4669 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; | |
4670 | ||
4671 | missed_updates >>= 1; | |
4672 | j++; | |
4673 | } | |
4674 | return load; | |
4675 | } | |
9fd81dd5 | 4676 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 4677 | |
59543275 | 4678 | /** |
cee1afce | 4679 | * __cpu_load_update - update the rq->cpu_load[] statistics |
59543275 BP |
4680 | * @this_rq: The rq to update statistics for |
4681 | * @this_load: The current load | |
4682 | * @pending_updates: The number of missed updates | |
59543275 | 4683 | * |
3289bdb4 | 4684 | * Update rq->cpu_load[] statistics. This function is usually called every |
59543275 BP |
4685 | * scheduler tick (TICK_NSEC). |
4686 | * | |
4687 | * This function computes a decaying average: | |
4688 | * | |
4689 | * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load | |
4690 | * | |
4691 | * Because of NOHZ it might not get called on every tick which gives need for | |
4692 | * the @pending_updates argument. | |
4693 | * | |
4694 | * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1 | |
4695 | * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load | |
4696 | * = A * (A * load[i]_n-2 + B) + B | |
4697 | * = A * (A * (A * load[i]_n-3 + B) + B) + B | |
4698 | * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B | |
4699 | * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B | |
4700 | * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B | |
4701 | * = (1 - 1/2^i)^n * (load[i]_0 - load) + load | |
4702 | * | |
4703 | * In the above we've assumed load_n := load, which is true for NOHZ_FULL as | |
4704 | * any change in load would have resulted in the tick being turned back on. | |
4705 | * | |
4706 | * For regular NOHZ, this reduces to: | |
4707 | * | |
4708 | * load[i]_n = (1 - 1/2^i)^n * load[i]_0 | |
4709 | * | |
4710 | * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra | |
1f41906a | 4711 | * term. |
3289bdb4 | 4712 | */ |
1f41906a FW |
4713 | static void cpu_load_update(struct rq *this_rq, unsigned long this_load, |
4714 | unsigned long pending_updates) | |
3289bdb4 | 4715 | { |
9fd81dd5 | 4716 | unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0]; |
3289bdb4 PZ |
4717 | int i, scale; |
4718 | ||
4719 | this_rq->nr_load_updates++; | |
4720 | ||
4721 | /* Update our load: */ | |
4722 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ | |
4723 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
4724 | unsigned long old_load, new_load; | |
4725 | ||
4726 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
4727 | ||
7400d3bb | 4728 | old_load = this_rq->cpu_load[i]; |
9fd81dd5 | 4729 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 | 4730 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
7400d3bb BP |
4731 | if (tickless_load) { |
4732 | old_load -= decay_load_missed(tickless_load, pending_updates - 1, i); | |
4733 | /* | |
4734 | * old_load can never be a negative value because a | |
4735 | * decayed tickless_load cannot be greater than the | |
4736 | * original tickless_load. | |
4737 | */ | |
4738 | old_load += tickless_load; | |
4739 | } | |
9fd81dd5 | 4740 | #endif |
3289bdb4 PZ |
4741 | new_load = this_load; |
4742 | /* | |
4743 | * Round up the averaging division if load is increasing. This | |
4744 | * prevents us from getting stuck on 9 if the load is 10, for | |
4745 | * example. | |
4746 | */ | |
4747 | if (new_load > old_load) | |
4748 | new_load += scale - 1; | |
4749 | ||
4750 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; | |
4751 | } | |
4752 | ||
4753 | sched_avg_update(this_rq); | |
4754 | } | |
4755 | ||
7ea241af YD |
4756 | /* Used instead of source_load when we know the type == 0 */ |
4757 | static unsigned long weighted_cpuload(const int cpu) | |
4758 | { | |
4759 | return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs); | |
4760 | } | |
4761 | ||
3289bdb4 | 4762 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
4763 | /* |
4764 | * There is no sane way to deal with nohz on smp when using jiffies because the | |
4765 | * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading | |
4766 | * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. | |
4767 | * | |
4768 | * Therefore we need to avoid the delta approach from the regular tick when | |
4769 | * possible since that would seriously skew the load calculation. This is why we | |
4770 | * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on | |
4771 | * jiffies deltas for updates happening while in nohz mode (idle ticks, idle | |
4772 | * loop exit, nohz_idle_balance, nohz full exit...) | |
4773 | * | |
4774 | * This means we might still be one tick off for nohz periods. | |
4775 | */ | |
4776 | ||
4777 | static void cpu_load_update_nohz(struct rq *this_rq, | |
4778 | unsigned long curr_jiffies, | |
4779 | unsigned long load) | |
be68a682 FW |
4780 | { |
4781 | unsigned long pending_updates; | |
4782 | ||
4783 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
4784 | if (pending_updates) { | |
4785 | this_rq->last_load_update_tick = curr_jiffies; | |
4786 | /* | |
4787 | * In the regular NOHZ case, we were idle, this means load 0. | |
4788 | * In the NOHZ_FULL case, we were non-idle, we should consider | |
4789 | * its weighted load. | |
4790 | */ | |
1f41906a | 4791 | cpu_load_update(this_rq, load, pending_updates); |
be68a682 FW |
4792 | } |
4793 | } | |
4794 | ||
3289bdb4 PZ |
4795 | /* |
4796 | * Called from nohz_idle_balance() to update the load ratings before doing the | |
4797 | * idle balance. | |
4798 | */ | |
cee1afce | 4799 | static void cpu_load_update_idle(struct rq *this_rq) |
3289bdb4 | 4800 | { |
3289bdb4 PZ |
4801 | /* |
4802 | * bail if there's load or we're actually up-to-date. | |
4803 | */ | |
be68a682 | 4804 | if (weighted_cpuload(cpu_of(this_rq))) |
3289bdb4 PZ |
4805 | return; |
4806 | ||
1f41906a | 4807 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0); |
3289bdb4 PZ |
4808 | } |
4809 | ||
4810 | /* | |
1f41906a FW |
4811 | * Record CPU load on nohz entry so we know the tickless load to account |
4812 | * on nohz exit. cpu_load[0] happens then to be updated more frequently | |
4813 | * than other cpu_load[idx] but it should be fine as cpu_load readers | |
4814 | * shouldn't rely into synchronized cpu_load[*] updates. | |
3289bdb4 | 4815 | */ |
1f41906a | 4816 | void cpu_load_update_nohz_start(void) |
3289bdb4 PZ |
4817 | { |
4818 | struct rq *this_rq = this_rq(); | |
1f41906a FW |
4819 | |
4820 | /* | |
4821 | * This is all lockless but should be fine. If weighted_cpuload changes | |
4822 | * concurrently we'll exit nohz. And cpu_load write can race with | |
4823 | * cpu_load_update_idle() but both updater would be writing the same. | |
4824 | */ | |
4825 | this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq)); | |
4826 | } | |
4827 | ||
4828 | /* | |
4829 | * Account the tickless load in the end of a nohz frame. | |
4830 | */ | |
4831 | void cpu_load_update_nohz_stop(void) | |
4832 | { | |
316c1608 | 4833 | unsigned long curr_jiffies = READ_ONCE(jiffies); |
1f41906a FW |
4834 | struct rq *this_rq = this_rq(); |
4835 | unsigned long load; | |
3289bdb4 PZ |
4836 | |
4837 | if (curr_jiffies == this_rq->last_load_update_tick) | |
4838 | return; | |
4839 | ||
1f41906a | 4840 | load = weighted_cpuload(cpu_of(this_rq)); |
3289bdb4 | 4841 | raw_spin_lock(&this_rq->lock); |
b52fad2d | 4842 | update_rq_clock(this_rq); |
1f41906a | 4843 | cpu_load_update_nohz(this_rq, curr_jiffies, load); |
3289bdb4 PZ |
4844 | raw_spin_unlock(&this_rq->lock); |
4845 | } | |
1f41906a FW |
4846 | #else /* !CONFIG_NO_HZ_COMMON */ |
4847 | static inline void cpu_load_update_nohz(struct rq *this_rq, | |
4848 | unsigned long curr_jiffies, | |
4849 | unsigned long load) { } | |
4850 | #endif /* CONFIG_NO_HZ_COMMON */ | |
4851 | ||
4852 | static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load) | |
4853 | { | |
9fd81dd5 | 4854 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
4855 | /* See the mess around cpu_load_update_nohz(). */ |
4856 | this_rq->last_load_update_tick = READ_ONCE(jiffies); | |
9fd81dd5 | 4857 | #endif |
1f41906a FW |
4858 | cpu_load_update(this_rq, load, 1); |
4859 | } | |
3289bdb4 PZ |
4860 | |
4861 | /* | |
4862 | * Called from scheduler_tick() | |
4863 | */ | |
cee1afce | 4864 | void cpu_load_update_active(struct rq *this_rq) |
3289bdb4 | 4865 | { |
7ea241af | 4866 | unsigned long load = weighted_cpuload(cpu_of(this_rq)); |
1f41906a FW |
4867 | |
4868 | if (tick_nohz_tick_stopped()) | |
4869 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load); | |
4870 | else | |
4871 | cpu_load_update_periodic(this_rq, load); | |
3289bdb4 PZ |
4872 | } |
4873 | ||
029632fb PZ |
4874 | /* |
4875 | * Return a low guess at the load of a migration-source cpu weighted | |
4876 | * according to the scheduling class and "nice" value. | |
4877 | * | |
4878 | * We want to under-estimate the load of migration sources, to | |
4879 | * balance conservatively. | |
4880 | */ | |
4881 | static unsigned long source_load(int cpu, int type) | |
4882 | { | |
4883 | struct rq *rq = cpu_rq(cpu); | |
4884 | unsigned long total = weighted_cpuload(cpu); | |
4885 | ||
4886 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4887 | return total; | |
4888 | ||
4889 | return min(rq->cpu_load[type-1], total); | |
4890 | } | |
4891 | ||
4892 | /* | |
4893 | * Return a high guess at the load of a migration-target cpu weighted | |
4894 | * according to the scheduling class and "nice" value. | |
4895 | */ | |
4896 | static unsigned long target_load(int cpu, int type) | |
4897 | { | |
4898 | struct rq *rq = cpu_rq(cpu); | |
4899 | unsigned long total = weighted_cpuload(cpu); | |
4900 | ||
4901 | if (type == 0 || !sched_feat(LB_BIAS)) | |
4902 | return total; | |
4903 | ||
4904 | return max(rq->cpu_load[type-1], total); | |
4905 | } | |
4906 | ||
ced549fa | 4907 | static unsigned long capacity_of(int cpu) |
029632fb | 4908 | { |
ced549fa | 4909 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
4910 | } |
4911 | ||
ca6d75e6 VG |
4912 | static unsigned long capacity_orig_of(int cpu) |
4913 | { | |
4914 | return cpu_rq(cpu)->cpu_capacity_orig; | |
4915 | } | |
4916 | ||
029632fb PZ |
4917 | static unsigned long cpu_avg_load_per_task(int cpu) |
4918 | { | |
4919 | struct rq *rq = cpu_rq(cpu); | |
316c1608 | 4920 | unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); |
7ea241af | 4921 | unsigned long load_avg = weighted_cpuload(cpu); |
029632fb PZ |
4922 | |
4923 | if (nr_running) | |
b92486cb | 4924 | return load_avg / nr_running; |
029632fb PZ |
4925 | |
4926 | return 0; | |
4927 | } | |
4928 | ||
bb3469ac | 4929 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
4930 | /* |
4931 | * effective_load() calculates the load change as seen from the root_task_group | |
4932 | * | |
4933 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
4934 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
4935 | * can calculate the shift in shares. | |
cf5f0acf PZ |
4936 | * |
4937 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
4938 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
4939 | * total group weight. | |
4940 | * | |
4941 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
4942 | * distribution (s_i) using: | |
4943 | * | |
4944 | * s_i = rw_i / \Sum rw_j (1) | |
4945 | * | |
4946 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
4947 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
4948 | * shares distribution (s_i): | |
4949 | * | |
4950 | * rw_i = { 2, 4, 1, 0 } | |
4951 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
4952 | * | |
4953 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
4954 | * task used to run on and the CPU the waker is running on), we need to | |
4955 | * compute the effect of waking a task on either CPU and, in case of a sync | |
4956 | * wakeup, compute the effect of the current task going to sleep. | |
4957 | * | |
4958 | * So for a change of @wl to the local @cpu with an overall group weight change | |
4959 | * of @wl we can compute the new shares distribution (s'_i) using: | |
4960 | * | |
4961 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
4962 | * | |
4963 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
4964 | * differences in waking a task to CPU 0. The additional task changes the | |
4965 | * weight and shares distributions like: | |
4966 | * | |
4967 | * rw'_i = { 3, 4, 1, 0 } | |
4968 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
4969 | * | |
4970 | * We can then compute the difference in effective weight by using: | |
4971 | * | |
4972 | * dw_i = S * (s'_i - s_i) (3) | |
4973 | * | |
4974 | * Where 'S' is the group weight as seen by its parent. | |
4975 | * | |
4976 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
4977 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
4978 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 4979 | */ |
2069dd75 | 4980 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 4981 | { |
4be9daaa | 4982 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 4983 | |
9722c2da | 4984 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
4985 | return wl; |
4986 | ||
4be9daaa | 4987 | for_each_sched_entity(se) { |
7dd49125 PZ |
4988 | struct cfs_rq *cfs_rq = se->my_q; |
4989 | long W, w = cfs_rq_load_avg(cfs_rq); | |
4be9daaa | 4990 | |
7dd49125 | 4991 | tg = cfs_rq->tg; |
bb3469ac | 4992 | |
cf5f0acf PZ |
4993 | /* |
4994 | * W = @wg + \Sum rw_j | |
4995 | */ | |
7dd49125 PZ |
4996 | W = wg + atomic_long_read(&tg->load_avg); |
4997 | ||
4998 | /* Ensure \Sum rw_j >= rw_i */ | |
4999 | W -= cfs_rq->tg_load_avg_contrib; | |
5000 | W += w; | |
4be9daaa | 5001 | |
cf5f0acf PZ |
5002 | /* |
5003 | * w = rw_i + @wl | |
5004 | */ | |
7dd49125 | 5005 | w += wl; |
940959e9 | 5006 | |
cf5f0acf PZ |
5007 | /* |
5008 | * wl = S * s'_i; see (2) | |
5009 | */ | |
5010 | if (W > 0 && w < W) | |
32a8df4e | 5011 | wl = (w * (long)tg->shares) / W; |
977dda7c PT |
5012 | else |
5013 | wl = tg->shares; | |
940959e9 | 5014 | |
cf5f0acf PZ |
5015 | /* |
5016 | * Per the above, wl is the new se->load.weight value; since | |
5017 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
5018 | * calc_cfs_shares(). | |
5019 | */ | |
977dda7c PT |
5020 | if (wl < MIN_SHARES) |
5021 | wl = MIN_SHARES; | |
cf5f0acf PZ |
5022 | |
5023 | /* | |
5024 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
5025 | */ | |
9d89c257 | 5026 | wl -= se->avg.load_avg; |
cf5f0acf PZ |
5027 | |
5028 | /* | |
5029 | * Recursively apply this logic to all parent groups to compute | |
5030 | * the final effective load change on the root group. Since | |
5031 | * only the @tg group gets extra weight, all parent groups can | |
5032 | * only redistribute existing shares. @wl is the shift in shares | |
5033 | * resulting from this level per the above. | |
5034 | */ | |
4be9daaa | 5035 | wg = 0; |
4be9daaa | 5036 | } |
bb3469ac | 5037 | |
4be9daaa | 5038 | return wl; |
bb3469ac PZ |
5039 | } |
5040 | #else | |
4be9daaa | 5041 | |
58d081b5 | 5042 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 5043 | { |
83378269 | 5044 | return wl; |
bb3469ac | 5045 | } |
4be9daaa | 5046 | |
bb3469ac PZ |
5047 | #endif |
5048 | ||
c58d25f3 PZ |
5049 | static void record_wakee(struct task_struct *p) |
5050 | { | |
5051 | /* | |
5052 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5053 | * jiffy will not have built up many flips. | |
5054 | */ | |
5055 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5056 | current->wakee_flips >>= 1; | |
5057 | current->wakee_flip_decay_ts = jiffies; | |
5058 | } | |
5059 | ||
5060 | if (current->last_wakee != p) { | |
5061 | current->last_wakee = p; | |
5062 | current->wakee_flips++; | |
5063 | } | |
5064 | } | |
5065 | ||
63b0e9ed MG |
5066 | /* |
5067 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5068 | * |
63b0e9ed | 5069 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5070 | * at a frequency roughly N times higher than one of its wakees. |
5071 | * | |
5072 | * In order to determine whether we should let the load spread vs consolidating | |
5073 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5074 | * partner, and a factor of lls_size higher frequency in the other. | |
5075 | * | |
5076 | * With both conditions met, we can be relatively sure that the relationship is | |
5077 | * non-monogamous, with partner count exceeding socket size. | |
5078 | * | |
5079 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5080 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5081 | * socket size. | |
63b0e9ed | 5082 | */ |
62470419 MW |
5083 | static int wake_wide(struct task_struct *p) |
5084 | { | |
63b0e9ed MG |
5085 | unsigned int master = current->wakee_flips; |
5086 | unsigned int slave = p->wakee_flips; | |
7d9ffa89 | 5087 | int factor = this_cpu_read(sd_llc_size); |
62470419 | 5088 | |
63b0e9ed MG |
5089 | if (master < slave) |
5090 | swap(master, slave); | |
5091 | if (slave < factor || master < slave * factor) | |
5092 | return 0; | |
5093 | return 1; | |
62470419 MW |
5094 | } |
5095 | ||
772bd008 MR |
5096 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
5097 | int prev_cpu, int sync) | |
098fb9db | 5098 | { |
e37b6a7b | 5099 | s64 this_load, load; |
bd61c98f | 5100 | s64 this_eff_load, prev_eff_load; |
772bd008 | 5101 | int idx, this_cpu; |
c88d5910 | 5102 | struct task_group *tg; |
83378269 | 5103 | unsigned long weight; |
b3137bc8 | 5104 | int balanced; |
098fb9db | 5105 | |
c88d5910 PZ |
5106 | idx = sd->wake_idx; |
5107 | this_cpu = smp_processor_id(); | |
c88d5910 PZ |
5108 | load = source_load(prev_cpu, idx); |
5109 | this_load = target_load(this_cpu, idx); | |
098fb9db | 5110 | |
b3137bc8 MG |
5111 | /* |
5112 | * If sync wakeup then subtract the (maximum possible) | |
5113 | * effect of the currently running task from the load | |
5114 | * of the current CPU: | |
5115 | */ | |
83378269 PZ |
5116 | if (sync) { |
5117 | tg = task_group(current); | |
9d89c257 | 5118 | weight = current->se.avg.load_avg; |
83378269 | 5119 | |
c88d5910 | 5120 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
5121 | load += effective_load(tg, prev_cpu, 0, -weight); |
5122 | } | |
b3137bc8 | 5123 | |
83378269 | 5124 | tg = task_group(p); |
9d89c257 | 5125 | weight = p->se.avg.load_avg; |
b3137bc8 | 5126 | |
71a29aa7 PZ |
5127 | /* |
5128 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
5129 | * due to the sync cause above having dropped this_load to 0, we'll |
5130 | * always have an imbalance, but there's really nothing you can do | |
5131 | * about that, so that's good too. | |
71a29aa7 PZ |
5132 | * |
5133 | * Otherwise check if either cpus are near enough in load to allow this | |
5134 | * task to be woken on this_cpu. | |
5135 | */ | |
bd61c98f VG |
5136 | this_eff_load = 100; |
5137 | this_eff_load *= capacity_of(prev_cpu); | |
e51fd5e2 | 5138 | |
bd61c98f VG |
5139 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; |
5140 | prev_eff_load *= capacity_of(this_cpu); | |
e51fd5e2 | 5141 | |
bd61c98f | 5142 | if (this_load > 0) { |
e51fd5e2 PZ |
5143 | this_eff_load *= this_load + |
5144 | effective_load(tg, this_cpu, weight, weight); | |
5145 | ||
e51fd5e2 | 5146 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); |
bd61c98f | 5147 | } |
e51fd5e2 | 5148 | |
bd61c98f | 5149 | balanced = this_eff_load <= prev_eff_load; |
098fb9db | 5150 | |
ae92882e | 5151 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
098fb9db | 5152 | |
05bfb65f VG |
5153 | if (!balanced) |
5154 | return 0; | |
098fb9db | 5155 | |
ae92882e JP |
5156 | schedstat_inc(sd->ttwu_move_affine); |
5157 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
05bfb65f VG |
5158 | |
5159 | return 1; | |
098fb9db IM |
5160 | } |
5161 | ||
aaee1203 PZ |
5162 | /* |
5163 | * find_idlest_group finds and returns the least busy CPU group within the | |
5164 | * domain. | |
5165 | */ | |
5166 | static struct sched_group * | |
78e7ed53 | 5167 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 5168 | int this_cpu, int sd_flag) |
e7693a36 | 5169 | { |
b3bd3de6 | 5170 | struct sched_group *idlest = NULL, *group = sd->groups; |
aaee1203 | 5171 | unsigned long min_load = ULONG_MAX, this_load = 0; |
c44f2a02 | 5172 | int load_idx = sd->forkexec_idx; |
aaee1203 | 5173 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
e7693a36 | 5174 | |
c44f2a02 VG |
5175 | if (sd_flag & SD_BALANCE_WAKE) |
5176 | load_idx = sd->wake_idx; | |
5177 | ||
aaee1203 PZ |
5178 | do { |
5179 | unsigned long load, avg_load; | |
5180 | int local_group; | |
5181 | int i; | |
e7693a36 | 5182 | |
aaee1203 PZ |
5183 | /* Skip over this group if it has no CPUs allowed */ |
5184 | if (!cpumask_intersects(sched_group_cpus(group), | |
fa17b507 | 5185 | tsk_cpus_allowed(p))) |
aaee1203 PZ |
5186 | continue; |
5187 | ||
5188 | local_group = cpumask_test_cpu(this_cpu, | |
5189 | sched_group_cpus(group)); | |
5190 | ||
5191 | /* Tally up the load of all CPUs in the group */ | |
5192 | avg_load = 0; | |
5193 | ||
5194 | for_each_cpu(i, sched_group_cpus(group)) { | |
5195 | /* Bias balancing toward cpus of our domain */ | |
5196 | if (local_group) | |
5197 | load = source_load(i, load_idx); | |
5198 | else | |
5199 | load = target_load(i, load_idx); | |
5200 | ||
5201 | avg_load += load; | |
5202 | } | |
5203 | ||
63b2ca30 | 5204 | /* Adjust by relative CPU capacity of the group */ |
ca8ce3d0 | 5205 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; |
aaee1203 PZ |
5206 | |
5207 | if (local_group) { | |
5208 | this_load = avg_load; | |
aaee1203 PZ |
5209 | } else if (avg_load < min_load) { |
5210 | min_load = avg_load; | |
5211 | idlest = group; | |
5212 | } | |
5213 | } while (group = group->next, group != sd->groups); | |
5214 | ||
5215 | if (!idlest || 100*this_load < imbalance*min_load) | |
5216 | return NULL; | |
5217 | return idlest; | |
5218 | } | |
5219 | ||
5220 | /* | |
5221 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
5222 | */ | |
5223 | static int | |
5224 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
5225 | { | |
5226 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5227 | unsigned int min_exit_latency = UINT_MAX; |
5228 | u64 latest_idle_timestamp = 0; | |
5229 | int least_loaded_cpu = this_cpu; | |
5230 | int shallowest_idle_cpu = -1; | |
aaee1203 PZ |
5231 | int i; |
5232 | ||
eaecf41f MR |
5233 | /* Check if we have any choice: */ |
5234 | if (group->group_weight == 1) | |
5235 | return cpumask_first(sched_group_cpus(group)); | |
5236 | ||
aaee1203 | 5237 | /* Traverse only the allowed CPUs */ |
fa17b507 | 5238 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
83a0a96a NP |
5239 | if (idle_cpu(i)) { |
5240 | struct rq *rq = cpu_rq(i); | |
5241 | struct cpuidle_state *idle = idle_get_state(rq); | |
5242 | if (idle && idle->exit_latency < min_exit_latency) { | |
5243 | /* | |
5244 | * We give priority to a CPU whose idle state | |
5245 | * has the smallest exit latency irrespective | |
5246 | * of any idle timestamp. | |
5247 | */ | |
5248 | min_exit_latency = idle->exit_latency; | |
5249 | latest_idle_timestamp = rq->idle_stamp; | |
5250 | shallowest_idle_cpu = i; | |
5251 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5252 | rq->idle_stamp > latest_idle_timestamp) { | |
5253 | /* | |
5254 | * If equal or no active idle state, then | |
5255 | * the most recently idled CPU might have | |
5256 | * a warmer cache. | |
5257 | */ | |
5258 | latest_idle_timestamp = rq->idle_stamp; | |
5259 | shallowest_idle_cpu = i; | |
5260 | } | |
9f96742a | 5261 | } else if (shallowest_idle_cpu == -1) { |
83a0a96a NP |
5262 | load = weighted_cpuload(i); |
5263 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
5264 | min_load = load; | |
5265 | least_loaded_cpu = i; | |
5266 | } | |
e7693a36 GH |
5267 | } |
5268 | } | |
5269 | ||
83a0a96a | 5270 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 5271 | } |
e7693a36 | 5272 | |
a50bde51 PZ |
5273 | /* |
5274 | * Try and locate an idle CPU in the sched_domain. | |
5275 | */ | |
772bd008 | 5276 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 5277 | { |
99bd5e2f | 5278 | struct sched_domain *sd; |
37407ea7 | 5279 | struct sched_group *sg; |
a50bde51 | 5280 | |
e0a79f52 MG |
5281 | if (idle_cpu(target)) |
5282 | return target; | |
99bd5e2f SS |
5283 | |
5284 | /* | |
e0a79f52 | 5285 | * If the prevous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 5286 | */ |
772bd008 MR |
5287 | if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev)) |
5288 | return prev; | |
a50bde51 PZ |
5289 | |
5290 | /* | |
d4335581 MF |
5291 | * Otherwise, iterate the domains and find an eligible idle cpu. |
5292 | * | |
5293 | * A completely idle sched group at higher domains is more | |
5294 | * desirable than an idle group at a lower level, because lower | |
5295 | * domains have smaller groups and usually share hardware | |
5296 | * resources which causes tasks to contend on them, e.g. x86 | |
5297 | * hyperthread siblings in the lowest domain (SMT) can contend | |
5298 | * on the shared cpu pipeline. | |
5299 | * | |
5300 | * However, while we prefer idle groups at higher domains | |
5301 | * finding an idle cpu at the lowest domain is still better than | |
5302 | * returning 'target', which we've already established, isn't | |
5303 | * idle. | |
a50bde51 | 5304 | */ |
518cd623 | 5305 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
970e1789 | 5306 | for_each_lower_domain(sd) { |
37407ea7 LT |
5307 | sg = sd->groups; |
5308 | do { | |
772bd008 MR |
5309 | int i; |
5310 | ||
37407ea7 LT |
5311 | if (!cpumask_intersects(sched_group_cpus(sg), |
5312 | tsk_cpus_allowed(p))) | |
5313 | goto next; | |
5314 | ||
d4335581 | 5315 | /* Ensure the entire group is idle */ |
37407ea7 | 5316 | for_each_cpu(i, sched_group_cpus(sg)) { |
e0a79f52 | 5317 | if (i == target || !idle_cpu(i)) |
37407ea7 LT |
5318 | goto next; |
5319 | } | |
970e1789 | 5320 | |
d4335581 MF |
5321 | /* |
5322 | * It doesn't matter which cpu we pick, the | |
5323 | * whole group is idle. | |
5324 | */ | |
37407ea7 LT |
5325 | target = cpumask_first_and(sched_group_cpus(sg), |
5326 | tsk_cpus_allowed(p)); | |
5327 | goto done; | |
5328 | next: | |
5329 | sg = sg->next; | |
5330 | } while (sg != sd->groups); | |
5331 | } | |
5332 | done: | |
a50bde51 PZ |
5333 | return target; |
5334 | } | |
231678b7 | 5335 | |
8bb5b00c | 5336 | /* |
9e91d61d | 5337 | * cpu_util returns the amount of capacity of a CPU that is used by CFS |
8bb5b00c | 5338 | * tasks. The unit of the return value must be the one of capacity so we can |
9e91d61d DE |
5339 | * compare the utilization with the capacity of the CPU that is available for |
5340 | * CFS task (ie cpu_capacity). | |
231678b7 DE |
5341 | * |
5342 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
5343 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
5344 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
5345 | * capacity_orig is the cpu_capacity available at the highest frequency | |
5346 | * (arch_scale_freq_capacity()). | |
5347 | * The utilization of a CPU converges towards a sum equal to or less than the | |
5348 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
5349 | * the running time on this CPU scaled by capacity_curr. | |
5350 | * | |
5351 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even | |
5352 | * higher than capacity_orig because of unfortunate rounding in | |
5353 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
5354 | * the average stabilizes with the new running time. We need to check that the | |
5355 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
5356 | * necessary. Without utilization capping, a group could be seen as overloaded | |
5357 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
5358 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
5359 | * capacity_orig) as it useful for predicting the capacity required after task | |
5360 | * migrations (scheduler-driven DVFS). | |
8bb5b00c | 5361 | */ |
9e91d61d | 5362 | static int cpu_util(int cpu) |
8bb5b00c | 5363 | { |
9e91d61d | 5364 | unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg; |
8bb5b00c VG |
5365 | unsigned long capacity = capacity_orig_of(cpu); |
5366 | ||
231678b7 | 5367 | return (util >= capacity) ? capacity : util; |
8bb5b00c | 5368 | } |
a50bde51 | 5369 | |
3273163c MR |
5370 | static inline int task_util(struct task_struct *p) |
5371 | { | |
5372 | return p->se.avg.util_avg; | |
5373 | } | |
5374 | ||
5375 | /* | |
5376 | * Disable WAKE_AFFINE in the case where task @p doesn't fit in the | |
5377 | * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu. | |
5378 | * | |
5379 | * In that case WAKE_AFFINE doesn't make sense and we'll let | |
5380 | * BALANCE_WAKE sort things out. | |
5381 | */ | |
5382 | static int wake_cap(struct task_struct *p, int cpu, int prev_cpu) | |
5383 | { | |
5384 | long min_cap, max_cap; | |
5385 | ||
5386 | min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu)); | |
5387 | max_cap = cpu_rq(cpu)->rd->max_cpu_capacity; | |
5388 | ||
5389 | /* Minimum capacity is close to max, no need to abort wake_affine */ | |
5390 | if (max_cap - min_cap < max_cap >> 3) | |
5391 | return 0; | |
5392 | ||
5393 | return min_cap * 1024 < task_util(p) * capacity_margin; | |
5394 | } | |
5395 | ||
aaee1203 | 5396 | /* |
de91b9cb MR |
5397 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
5398 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
5399 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 5400 | * |
de91b9cb MR |
5401 | * Balances load by selecting the idlest cpu in the idlest group, or under |
5402 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 5403 | * |
de91b9cb | 5404 | * Returns the target cpu number. |
aaee1203 PZ |
5405 | * |
5406 | * preempt must be disabled. | |
5407 | */ | |
0017d735 | 5408 | static int |
ac66f547 | 5409 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 5410 | { |
29cd8bae | 5411 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 5412 | int cpu = smp_processor_id(); |
63b0e9ed | 5413 | int new_cpu = prev_cpu; |
99bd5e2f | 5414 | int want_affine = 0; |
5158f4e4 | 5415 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 5416 | |
c58d25f3 PZ |
5417 | if (sd_flag & SD_BALANCE_WAKE) { |
5418 | record_wakee(p); | |
3273163c MR |
5419 | want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) |
5420 | && cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); | |
c58d25f3 | 5421 | } |
aaee1203 | 5422 | |
dce840a0 | 5423 | rcu_read_lock(); |
aaee1203 | 5424 | for_each_domain(cpu, tmp) { |
e4f42888 | 5425 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
63b0e9ed | 5426 | break; |
e4f42888 | 5427 | |
fe3bcfe1 | 5428 | /* |
99bd5e2f SS |
5429 | * If both cpu and prev_cpu are part of this domain, |
5430 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 5431 | */ |
99bd5e2f SS |
5432 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
5433 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
5434 | affine_sd = tmp; | |
29cd8bae | 5435 | break; |
f03542a7 | 5436 | } |
29cd8bae | 5437 | |
f03542a7 | 5438 | if (tmp->flags & sd_flag) |
29cd8bae | 5439 | sd = tmp; |
63b0e9ed MG |
5440 | else if (!want_affine) |
5441 | break; | |
29cd8bae PZ |
5442 | } |
5443 | ||
63b0e9ed MG |
5444 | if (affine_sd) { |
5445 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
772bd008 | 5446 | if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync)) |
63b0e9ed | 5447 | new_cpu = cpu; |
8b911acd | 5448 | } |
e7693a36 | 5449 | |
63b0e9ed MG |
5450 | if (!sd) { |
5451 | if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */ | |
772bd008 | 5452 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
63b0e9ed MG |
5453 | |
5454 | } else while (sd) { | |
aaee1203 | 5455 | struct sched_group *group; |
c88d5910 | 5456 | int weight; |
098fb9db | 5457 | |
0763a660 | 5458 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
5459 | sd = sd->child; |
5460 | continue; | |
5461 | } | |
098fb9db | 5462 | |
c44f2a02 | 5463 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
5464 | if (!group) { |
5465 | sd = sd->child; | |
5466 | continue; | |
5467 | } | |
4ae7d5ce | 5468 | |
d7c33c49 | 5469 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
5470 | if (new_cpu == -1 || new_cpu == cpu) { |
5471 | /* Now try balancing at a lower domain level of cpu */ | |
5472 | sd = sd->child; | |
5473 | continue; | |
e7693a36 | 5474 | } |
aaee1203 PZ |
5475 | |
5476 | /* Now try balancing at a lower domain level of new_cpu */ | |
5477 | cpu = new_cpu; | |
669c55e9 | 5478 | weight = sd->span_weight; |
aaee1203 PZ |
5479 | sd = NULL; |
5480 | for_each_domain(cpu, tmp) { | |
669c55e9 | 5481 | if (weight <= tmp->span_weight) |
aaee1203 | 5482 | break; |
0763a660 | 5483 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
5484 | sd = tmp; |
5485 | } | |
5486 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 5487 | } |
dce840a0 | 5488 | rcu_read_unlock(); |
e7693a36 | 5489 | |
c88d5910 | 5490 | return new_cpu; |
e7693a36 | 5491 | } |
0a74bef8 PT |
5492 | |
5493 | /* | |
5494 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
5495 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
525628c7 | 5496 | * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 5497 | */ |
5a4fd036 | 5498 | static void migrate_task_rq_fair(struct task_struct *p) |
0a74bef8 | 5499 | { |
59efa0ba PZ |
5500 | /* |
5501 | * As blocked tasks retain absolute vruntime the migration needs to | |
5502 | * deal with this by subtracting the old and adding the new | |
5503 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
5504 | * the task on the new runqueue. | |
5505 | */ | |
5506 | if (p->state == TASK_WAKING) { | |
5507 | struct sched_entity *se = &p->se; | |
5508 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5509 | u64 min_vruntime; | |
5510 | ||
5511 | #ifndef CONFIG_64BIT | |
5512 | u64 min_vruntime_copy; | |
5513 | ||
5514 | do { | |
5515 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
5516 | smp_rmb(); | |
5517 | min_vruntime = cfs_rq->min_vruntime; | |
5518 | } while (min_vruntime != min_vruntime_copy); | |
5519 | #else | |
5520 | min_vruntime = cfs_rq->min_vruntime; | |
5521 | #endif | |
5522 | ||
5523 | se->vruntime -= min_vruntime; | |
5524 | } | |
5525 | ||
aff3e498 | 5526 | /* |
9d89c257 YD |
5527 | * We are supposed to update the task to "current" time, then its up to date |
5528 | * and ready to go to new CPU/cfs_rq. But we have difficulty in getting | |
5529 | * what current time is, so simply throw away the out-of-date time. This | |
5530 | * will result in the wakee task is less decayed, but giving the wakee more | |
5531 | * load sounds not bad. | |
aff3e498 | 5532 | */ |
9d89c257 YD |
5533 | remove_entity_load_avg(&p->se); |
5534 | ||
5535 | /* Tell new CPU we are migrated */ | |
5536 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
5537 | |
5538 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 5539 | p->se.exec_start = 0; |
0a74bef8 | 5540 | } |
12695578 YD |
5541 | |
5542 | static void task_dead_fair(struct task_struct *p) | |
5543 | { | |
5544 | remove_entity_load_avg(&p->se); | |
5545 | } | |
e7693a36 GH |
5546 | #endif /* CONFIG_SMP */ |
5547 | ||
e52fb7c0 PZ |
5548 | static unsigned long |
5549 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
5550 | { |
5551 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
5552 | ||
5553 | /* | |
e52fb7c0 PZ |
5554 | * Since its curr running now, convert the gran from real-time |
5555 | * to virtual-time in his units. | |
13814d42 MG |
5556 | * |
5557 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
5558 | * they get preempted easier. That is, if 'se' < 'curr' then | |
5559 | * the resulting gran will be larger, therefore penalizing the | |
5560 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
5561 | * be smaller, again penalizing the lighter task. | |
5562 | * | |
5563 | * This is especially important for buddies when the leftmost | |
5564 | * task is higher priority than the buddy. | |
0bbd3336 | 5565 | */ |
f4ad9bd2 | 5566 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
5567 | } |
5568 | ||
464b7527 PZ |
5569 | /* |
5570 | * Should 'se' preempt 'curr'. | |
5571 | * | |
5572 | * |s1 | |
5573 | * |s2 | |
5574 | * |s3 | |
5575 | * g | |
5576 | * |<--->|c | |
5577 | * | |
5578 | * w(c, s1) = -1 | |
5579 | * w(c, s2) = 0 | |
5580 | * w(c, s3) = 1 | |
5581 | * | |
5582 | */ | |
5583 | static int | |
5584 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
5585 | { | |
5586 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
5587 | ||
5588 | if (vdiff <= 0) | |
5589 | return -1; | |
5590 | ||
e52fb7c0 | 5591 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
5592 | if (vdiff > gran) |
5593 | return 1; | |
5594 | ||
5595 | return 0; | |
5596 | } | |
5597 | ||
02479099 PZ |
5598 | static void set_last_buddy(struct sched_entity *se) |
5599 | { | |
69c80f3e VP |
5600 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
5601 | return; | |
5602 | ||
5603 | for_each_sched_entity(se) | |
5604 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
5605 | } |
5606 | ||
5607 | static void set_next_buddy(struct sched_entity *se) | |
5608 | { | |
69c80f3e VP |
5609 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
5610 | return; | |
5611 | ||
5612 | for_each_sched_entity(se) | |
5613 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
5614 | } |
5615 | ||
ac53db59 RR |
5616 | static void set_skip_buddy(struct sched_entity *se) |
5617 | { | |
69c80f3e VP |
5618 | for_each_sched_entity(se) |
5619 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
5620 | } |
5621 | ||
bf0f6f24 IM |
5622 | /* |
5623 | * Preempt the current task with a newly woken task if needed: | |
5624 | */ | |
5a9b86f6 | 5625 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
5626 | { |
5627 | struct task_struct *curr = rq->curr; | |
8651a86c | 5628 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 5629 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 5630 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 5631 | int next_buddy_marked = 0; |
bf0f6f24 | 5632 | |
4ae7d5ce IM |
5633 | if (unlikely(se == pse)) |
5634 | return; | |
5635 | ||
5238cdd3 | 5636 | /* |
163122b7 | 5637 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
5638 | * unconditionally check_prempt_curr() after an enqueue (which may have |
5639 | * lead to a throttle). This both saves work and prevents false | |
5640 | * next-buddy nomination below. | |
5641 | */ | |
5642 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
5643 | return; | |
5644 | ||
2f36825b | 5645 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 5646 | set_next_buddy(pse); |
2f36825b VP |
5647 | next_buddy_marked = 1; |
5648 | } | |
57fdc26d | 5649 | |
aec0a514 BR |
5650 | /* |
5651 | * We can come here with TIF_NEED_RESCHED already set from new task | |
5652 | * wake up path. | |
5238cdd3 PT |
5653 | * |
5654 | * Note: this also catches the edge-case of curr being in a throttled | |
5655 | * group (e.g. via set_curr_task), since update_curr() (in the | |
5656 | * enqueue of curr) will have resulted in resched being set. This | |
5657 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
5658 | * below. | |
aec0a514 BR |
5659 | */ |
5660 | if (test_tsk_need_resched(curr)) | |
5661 | return; | |
5662 | ||
a2f5c9ab DH |
5663 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
5664 | if (unlikely(curr->policy == SCHED_IDLE) && | |
5665 | likely(p->policy != SCHED_IDLE)) | |
5666 | goto preempt; | |
5667 | ||
91c234b4 | 5668 | /* |
a2f5c9ab DH |
5669 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
5670 | * is driven by the tick): | |
91c234b4 | 5671 | */ |
f14bf141 JD |
5672 | if (unlikely(p->policy == SCHED_BATCH || p->policy == SCHED_IDLE) || |
5673 | !sched_feat(WAKEUP_PREEMPTION)) | |
91c234b4 | 5674 | return; |
bf0f6f24 | 5675 | |
464b7527 | 5676 | find_matching_se(&se, &pse); |
9bbd7374 | 5677 | update_curr(cfs_rq_of(se)); |
002f128b | 5678 | BUG_ON(!pse); |
2f36825b VP |
5679 | if (wakeup_preempt_entity(se, pse) == 1) { |
5680 | /* | |
5681 | * Bias pick_next to pick the sched entity that is | |
5682 | * triggering this preemption. | |
5683 | */ | |
5684 | if (!next_buddy_marked) | |
5685 | set_next_buddy(pse); | |
3a7e73a2 | 5686 | goto preempt; |
2f36825b | 5687 | } |
464b7527 | 5688 | |
3a7e73a2 | 5689 | return; |
a65ac745 | 5690 | |
3a7e73a2 | 5691 | preempt: |
8875125e | 5692 | resched_curr(rq); |
3a7e73a2 PZ |
5693 | /* |
5694 | * Only set the backward buddy when the current task is still | |
5695 | * on the rq. This can happen when a wakeup gets interleaved | |
5696 | * with schedule on the ->pre_schedule() or idle_balance() | |
5697 | * point, either of which can * drop the rq lock. | |
5698 | * | |
5699 | * Also, during early boot the idle thread is in the fair class, | |
5700 | * for obvious reasons its a bad idea to schedule back to it. | |
5701 | */ | |
5702 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
5703 | return; | |
5704 | ||
5705 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
5706 | set_last_buddy(se); | |
bf0f6f24 IM |
5707 | } |
5708 | ||
606dba2e | 5709 | static struct task_struct * |
e7904a28 | 5710 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie) |
bf0f6f24 IM |
5711 | { |
5712 | struct cfs_rq *cfs_rq = &rq->cfs; | |
5713 | struct sched_entity *se; | |
678d5718 | 5714 | struct task_struct *p; |
37e117c0 | 5715 | int new_tasks; |
678d5718 | 5716 | |
6e83125c | 5717 | again: |
678d5718 PZ |
5718 | #ifdef CONFIG_FAIR_GROUP_SCHED |
5719 | if (!cfs_rq->nr_running) | |
38033c37 | 5720 | goto idle; |
678d5718 | 5721 | |
3f1d2a31 | 5722 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
5723 | goto simple; |
5724 | ||
5725 | /* | |
5726 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
5727 | * likely that a next task is from the same cgroup as the current. | |
5728 | * | |
5729 | * Therefore attempt to avoid putting and setting the entire cgroup | |
5730 | * hierarchy, only change the part that actually changes. | |
5731 | */ | |
5732 | ||
5733 | do { | |
5734 | struct sched_entity *curr = cfs_rq->curr; | |
5735 | ||
5736 | /* | |
5737 | * Since we got here without doing put_prev_entity() we also | |
5738 | * have to consider cfs_rq->curr. If it is still a runnable | |
5739 | * entity, update_curr() will update its vruntime, otherwise | |
5740 | * forget we've ever seen it. | |
5741 | */ | |
54d27365 BS |
5742 | if (curr) { |
5743 | if (curr->on_rq) | |
5744 | update_curr(cfs_rq); | |
5745 | else | |
5746 | curr = NULL; | |
678d5718 | 5747 | |
54d27365 BS |
5748 | /* |
5749 | * This call to check_cfs_rq_runtime() will do the | |
5750 | * throttle and dequeue its entity in the parent(s). | |
5751 | * Therefore the 'simple' nr_running test will indeed | |
5752 | * be correct. | |
5753 | */ | |
5754 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
5755 | goto simple; | |
5756 | } | |
678d5718 PZ |
5757 | |
5758 | se = pick_next_entity(cfs_rq, curr); | |
5759 | cfs_rq = group_cfs_rq(se); | |
5760 | } while (cfs_rq); | |
5761 | ||
5762 | p = task_of(se); | |
5763 | ||
5764 | /* | |
5765 | * Since we haven't yet done put_prev_entity and if the selected task | |
5766 | * is a different task than we started out with, try and touch the | |
5767 | * least amount of cfs_rqs. | |
5768 | */ | |
5769 | if (prev != p) { | |
5770 | struct sched_entity *pse = &prev->se; | |
5771 | ||
5772 | while (!(cfs_rq = is_same_group(se, pse))) { | |
5773 | int se_depth = se->depth; | |
5774 | int pse_depth = pse->depth; | |
5775 | ||
5776 | if (se_depth <= pse_depth) { | |
5777 | put_prev_entity(cfs_rq_of(pse), pse); | |
5778 | pse = parent_entity(pse); | |
5779 | } | |
5780 | if (se_depth >= pse_depth) { | |
5781 | set_next_entity(cfs_rq_of(se), se); | |
5782 | se = parent_entity(se); | |
5783 | } | |
5784 | } | |
5785 | ||
5786 | put_prev_entity(cfs_rq, pse); | |
5787 | set_next_entity(cfs_rq, se); | |
5788 | } | |
5789 | ||
5790 | if (hrtick_enabled(rq)) | |
5791 | hrtick_start_fair(rq, p); | |
5792 | ||
5793 | return p; | |
5794 | simple: | |
5795 | cfs_rq = &rq->cfs; | |
5796 | #endif | |
bf0f6f24 | 5797 | |
36ace27e | 5798 | if (!cfs_rq->nr_running) |
38033c37 | 5799 | goto idle; |
bf0f6f24 | 5800 | |
3f1d2a31 | 5801 | put_prev_task(rq, prev); |
606dba2e | 5802 | |
bf0f6f24 | 5803 | do { |
678d5718 | 5804 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 5805 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
5806 | cfs_rq = group_cfs_rq(se); |
5807 | } while (cfs_rq); | |
5808 | ||
8f4d37ec | 5809 | p = task_of(se); |
678d5718 | 5810 | |
b39e66ea MG |
5811 | if (hrtick_enabled(rq)) |
5812 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
5813 | |
5814 | return p; | |
38033c37 PZ |
5815 | |
5816 | idle: | |
cbce1a68 PZ |
5817 | /* |
5818 | * This is OK, because current is on_cpu, which avoids it being picked | |
5819 | * for load-balance and preemption/IRQs are still disabled avoiding | |
5820 | * further scheduler activity on it and we're being very careful to | |
5821 | * re-start the picking loop. | |
5822 | */ | |
e7904a28 | 5823 | lockdep_unpin_lock(&rq->lock, cookie); |
e4aa358b | 5824 | new_tasks = idle_balance(rq); |
e7904a28 | 5825 | lockdep_repin_lock(&rq->lock, cookie); |
37e117c0 PZ |
5826 | /* |
5827 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
5828 | * possible for any higher priority task to appear. In that case we | |
5829 | * must re-start the pick_next_entity() loop. | |
5830 | */ | |
e4aa358b | 5831 | if (new_tasks < 0) |
37e117c0 PZ |
5832 | return RETRY_TASK; |
5833 | ||
e4aa358b | 5834 | if (new_tasks > 0) |
38033c37 | 5835 | goto again; |
38033c37 PZ |
5836 | |
5837 | return NULL; | |
bf0f6f24 IM |
5838 | } |
5839 | ||
5840 | /* | |
5841 | * Account for a descheduled task: | |
5842 | */ | |
31ee529c | 5843 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
5844 | { |
5845 | struct sched_entity *se = &prev->se; | |
5846 | struct cfs_rq *cfs_rq; | |
5847 | ||
5848 | for_each_sched_entity(se) { | |
5849 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 5850 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
5851 | } |
5852 | } | |
5853 | ||
ac53db59 RR |
5854 | /* |
5855 | * sched_yield() is very simple | |
5856 | * | |
5857 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
5858 | */ | |
5859 | static void yield_task_fair(struct rq *rq) | |
5860 | { | |
5861 | struct task_struct *curr = rq->curr; | |
5862 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
5863 | struct sched_entity *se = &curr->se; | |
5864 | ||
5865 | /* | |
5866 | * Are we the only task in the tree? | |
5867 | */ | |
5868 | if (unlikely(rq->nr_running == 1)) | |
5869 | return; | |
5870 | ||
5871 | clear_buddies(cfs_rq, se); | |
5872 | ||
5873 | if (curr->policy != SCHED_BATCH) { | |
5874 | update_rq_clock(rq); | |
5875 | /* | |
5876 | * Update run-time statistics of the 'current'. | |
5877 | */ | |
5878 | update_curr(cfs_rq); | |
916671c0 MG |
5879 | /* |
5880 | * Tell update_rq_clock() that we've just updated, | |
5881 | * so we don't do microscopic update in schedule() | |
5882 | * and double the fastpath cost. | |
5883 | */ | |
9edfbfed | 5884 | rq_clock_skip_update(rq, true); |
ac53db59 RR |
5885 | } |
5886 | ||
5887 | set_skip_buddy(se); | |
5888 | } | |
5889 | ||
d95f4122 MG |
5890 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
5891 | { | |
5892 | struct sched_entity *se = &p->se; | |
5893 | ||
5238cdd3 PT |
5894 | /* throttled hierarchies are not runnable */ |
5895 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
5896 | return false; |
5897 | ||
5898 | /* Tell the scheduler that we'd really like pse to run next. */ | |
5899 | set_next_buddy(se); | |
5900 | ||
d95f4122 MG |
5901 | yield_task_fair(rq); |
5902 | ||
5903 | return true; | |
5904 | } | |
5905 | ||
681f3e68 | 5906 | #ifdef CONFIG_SMP |
bf0f6f24 | 5907 | /************************************************** |
e9c84cb8 PZ |
5908 | * Fair scheduling class load-balancing methods. |
5909 | * | |
5910 | * BASICS | |
5911 | * | |
5912 | * The purpose of load-balancing is to achieve the same basic fairness the | |
5913 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
5914 | * time to each task. This is expressed in the following equation: | |
5915 | * | |
5916 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
5917 | * | |
5918 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
5919 | * W_i,0 is defined as: | |
5920 | * | |
5921 | * W_i,0 = \Sum_j w_i,j (2) | |
5922 | * | |
5923 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
1c3de5e1 | 5924 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
5925 | * |
5926 | * The weight average is an exponential decay average of the instantaneous | |
5927 | * weight: | |
5928 | * | |
5929 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
5930 | * | |
ced549fa | 5931 | * C_i is the compute capacity of cpu i, typically it is the |
e9c84cb8 PZ |
5932 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
5933 | * can also include other factors [XXX]. | |
5934 | * | |
5935 | * To achieve this balance we define a measure of imbalance which follows | |
5936 | * directly from (1): | |
5937 | * | |
ced549fa | 5938 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
5939 | * |
5940 | * We them move tasks around to minimize the imbalance. In the continuous | |
5941 | * function space it is obvious this converges, in the discrete case we get | |
5942 | * a few fun cases generally called infeasible weight scenarios. | |
5943 | * | |
5944 | * [XXX expand on: | |
5945 | * - infeasible weights; | |
5946 | * - local vs global optima in the discrete case. ] | |
5947 | * | |
5948 | * | |
5949 | * SCHED DOMAINS | |
5950 | * | |
5951 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
5952 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
5953 | * topology where each level pairs two lower groups (or better). This results | |
5954 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
5955 | * tree to only the first of the previous level and we decrease the frequency | |
5956 | * of load-balance at each level inv. proportional to the number of cpus in | |
5957 | * the groups. | |
5958 | * | |
5959 | * This yields: | |
5960 | * | |
5961 | * log_2 n 1 n | |
5962 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
5963 | * i = 0 2^i 2^i | |
5964 | * `- size of each group | |
5965 | * | | `- number of cpus doing load-balance | |
5966 | * | `- freq | |
5967 | * `- sum over all levels | |
5968 | * | |
5969 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
5970 | * this makes (5) the runtime complexity of the balancer. | |
5971 | * | |
5972 | * An important property here is that each CPU is still (indirectly) connected | |
5973 | * to every other cpu in at most O(log n) steps: | |
5974 | * | |
5975 | * The adjacency matrix of the resulting graph is given by: | |
5976 | * | |
97a7142f | 5977 | * log_2 n |
e9c84cb8 PZ |
5978 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
5979 | * k = 0 | |
5980 | * | |
5981 | * And you'll find that: | |
5982 | * | |
5983 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
5984 | * | |
5985 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
5986 | * The task movement gives a factor of O(m), giving a convergence complexity | |
5987 | * of: | |
5988 | * | |
5989 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
5990 | * | |
5991 | * | |
5992 | * WORK CONSERVING | |
5993 | * | |
5994 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
5995 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
5996 | * tree itself instead of relying on other CPUs to bring it work. | |
5997 | * | |
5998 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
5999 | * time. | |
6000 | * | |
6001 | * [XXX more?] | |
6002 | * | |
6003 | * | |
6004 | * CGROUPS | |
6005 | * | |
6006 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
6007 | * | |
6008 | * s_k,i | |
6009 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
6010 | * S_k | |
6011 | * | |
6012 | * Where | |
6013 | * | |
6014 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
6015 | * | |
6016 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
6017 | * | |
6018 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
6019 | * property. | |
6020 | * | |
6021 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
6022 | * rewrite all of this once again.] | |
97a7142f | 6023 | */ |
bf0f6f24 | 6024 | |
ed387b78 HS |
6025 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
6026 | ||
0ec8aa00 PZ |
6027 | enum fbq_type { regular, remote, all }; |
6028 | ||
ddcdf6e7 | 6029 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 6030 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
6031 | #define LBF_DST_PINNED 0x04 |
6032 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
6033 | |
6034 | struct lb_env { | |
6035 | struct sched_domain *sd; | |
6036 | ||
ddcdf6e7 | 6037 | struct rq *src_rq; |
85c1e7da | 6038 | int src_cpu; |
ddcdf6e7 PZ |
6039 | |
6040 | int dst_cpu; | |
6041 | struct rq *dst_rq; | |
6042 | ||
88b8dac0 SV |
6043 | struct cpumask *dst_grpmask; |
6044 | int new_dst_cpu; | |
ddcdf6e7 | 6045 | enum cpu_idle_type idle; |
bd939f45 | 6046 | long imbalance; |
b9403130 MW |
6047 | /* The set of CPUs under consideration for load-balancing */ |
6048 | struct cpumask *cpus; | |
6049 | ||
ddcdf6e7 | 6050 | unsigned int flags; |
367456c7 PZ |
6051 | |
6052 | unsigned int loop; | |
6053 | unsigned int loop_break; | |
6054 | unsigned int loop_max; | |
0ec8aa00 PZ |
6055 | |
6056 | enum fbq_type fbq_type; | |
163122b7 | 6057 | struct list_head tasks; |
ddcdf6e7 PZ |
6058 | }; |
6059 | ||
029632fb PZ |
6060 | /* |
6061 | * Is this task likely cache-hot: | |
6062 | */ | |
5d5e2b1b | 6063 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
6064 | { |
6065 | s64 delta; | |
6066 | ||
e5673f28 KT |
6067 | lockdep_assert_held(&env->src_rq->lock); |
6068 | ||
029632fb PZ |
6069 | if (p->sched_class != &fair_sched_class) |
6070 | return 0; | |
6071 | ||
6072 | if (unlikely(p->policy == SCHED_IDLE)) | |
6073 | return 0; | |
6074 | ||
6075 | /* | |
6076 | * Buddy candidates are cache hot: | |
6077 | */ | |
5d5e2b1b | 6078 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
6079 | (&p->se == cfs_rq_of(&p->se)->next || |
6080 | &p->se == cfs_rq_of(&p->se)->last)) | |
6081 | return 1; | |
6082 | ||
6083 | if (sysctl_sched_migration_cost == -1) | |
6084 | return 1; | |
6085 | if (sysctl_sched_migration_cost == 0) | |
6086 | return 0; | |
6087 | ||
5d5e2b1b | 6088 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
6089 | |
6090 | return delta < (s64)sysctl_sched_migration_cost; | |
6091 | } | |
6092 | ||
3a7053b3 | 6093 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 6094 | /* |
2a1ed24c SD |
6095 | * Returns 1, if task migration degrades locality |
6096 | * Returns 0, if task migration improves locality i.e migration preferred. | |
6097 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 6098 | */ |
2a1ed24c | 6099 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 6100 | { |
b1ad065e | 6101 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
c1ceac62 | 6102 | unsigned long src_faults, dst_faults; |
3a7053b3 MG |
6103 | int src_nid, dst_nid; |
6104 | ||
2a595721 | 6105 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
6106 | return -1; |
6107 | ||
c3b9bc5b | 6108 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 6109 | return -1; |
7a0f3083 MG |
6110 | |
6111 | src_nid = cpu_to_node(env->src_cpu); | |
6112 | dst_nid = cpu_to_node(env->dst_cpu); | |
6113 | ||
83e1d2cd | 6114 | if (src_nid == dst_nid) |
2a1ed24c | 6115 | return -1; |
7a0f3083 | 6116 | |
2a1ed24c SD |
6117 | /* Migrating away from the preferred node is always bad. */ |
6118 | if (src_nid == p->numa_preferred_nid) { | |
6119 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
6120 | return 1; | |
6121 | else | |
6122 | return -1; | |
6123 | } | |
b1ad065e | 6124 | |
c1ceac62 RR |
6125 | /* Encourage migration to the preferred node. */ |
6126 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 6127 | return 0; |
b1ad065e | 6128 | |
c1ceac62 RR |
6129 | if (numa_group) { |
6130 | src_faults = group_faults(p, src_nid); | |
6131 | dst_faults = group_faults(p, dst_nid); | |
6132 | } else { | |
6133 | src_faults = task_faults(p, src_nid); | |
6134 | dst_faults = task_faults(p, dst_nid); | |
b1ad065e RR |
6135 | } |
6136 | ||
c1ceac62 | 6137 | return dst_faults < src_faults; |
7a0f3083 MG |
6138 | } |
6139 | ||
3a7053b3 | 6140 | #else |
2a1ed24c | 6141 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
6142 | struct lb_env *env) |
6143 | { | |
2a1ed24c | 6144 | return -1; |
7a0f3083 | 6145 | } |
3a7053b3 MG |
6146 | #endif |
6147 | ||
1e3c88bd PZ |
6148 | /* |
6149 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
6150 | */ | |
6151 | static | |
8e45cb54 | 6152 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 6153 | { |
2a1ed24c | 6154 | int tsk_cache_hot; |
e5673f28 KT |
6155 | |
6156 | lockdep_assert_held(&env->src_rq->lock); | |
6157 | ||
1e3c88bd PZ |
6158 | /* |
6159 | * We do not migrate tasks that are: | |
d3198084 | 6160 | * 1) throttled_lb_pair, or |
1e3c88bd | 6161 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
6162 | * 3) running (obviously), or |
6163 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 6164 | */ |
d3198084 JK |
6165 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
6166 | return 0; | |
6167 | ||
ddcdf6e7 | 6168 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
e02e60c1 | 6169 | int cpu; |
88b8dac0 | 6170 | |
ae92882e | 6171 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 6172 | |
6263322c PZ |
6173 | env->flags |= LBF_SOME_PINNED; |
6174 | ||
88b8dac0 SV |
6175 | /* |
6176 | * Remember if this task can be migrated to any other cpu in | |
6177 | * our sched_group. We may want to revisit it if we couldn't | |
6178 | * meet load balance goals by pulling other tasks on src_cpu. | |
6179 | * | |
6180 | * Also avoid computing new_dst_cpu if we have already computed | |
6181 | * one in current iteration. | |
6182 | */ | |
6263322c | 6183 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
6184 | return 0; |
6185 | ||
e02e60c1 JK |
6186 | /* Prevent to re-select dst_cpu via env's cpus */ |
6187 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
6188 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | |
6263322c | 6189 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
6190 | env->new_dst_cpu = cpu; |
6191 | break; | |
6192 | } | |
88b8dac0 | 6193 | } |
e02e60c1 | 6194 | |
1e3c88bd PZ |
6195 | return 0; |
6196 | } | |
88b8dac0 SV |
6197 | |
6198 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 6199 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 6200 | |
ddcdf6e7 | 6201 | if (task_running(env->src_rq, p)) { |
ae92882e | 6202 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
6203 | return 0; |
6204 | } | |
6205 | ||
6206 | /* | |
6207 | * Aggressive migration if: | |
3a7053b3 MG |
6208 | * 1) destination numa is preferred |
6209 | * 2) task is cache cold, or | |
6210 | * 3) too many balance attempts have failed. | |
1e3c88bd | 6211 | */ |
2a1ed24c SD |
6212 | tsk_cache_hot = migrate_degrades_locality(p, env); |
6213 | if (tsk_cache_hot == -1) | |
6214 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 6215 | |
2a1ed24c | 6216 | if (tsk_cache_hot <= 0 || |
7a96c231 | 6217 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 6218 | if (tsk_cache_hot == 1) { |
ae92882e JP |
6219 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
6220 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 6221 | } |
1e3c88bd PZ |
6222 | return 1; |
6223 | } | |
6224 | ||
ae92882e | 6225 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 6226 | return 0; |
1e3c88bd PZ |
6227 | } |
6228 | ||
897c395f | 6229 | /* |
163122b7 KT |
6230 | * detach_task() -- detach the task for the migration specified in env |
6231 | */ | |
6232 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
6233 | { | |
6234 | lockdep_assert_held(&env->src_rq->lock); | |
6235 | ||
163122b7 | 6236 | p->on_rq = TASK_ON_RQ_MIGRATING; |
3ea94de1 | 6237 | deactivate_task(env->src_rq, p, 0); |
163122b7 KT |
6238 | set_task_cpu(p, env->dst_cpu); |
6239 | } | |
6240 | ||
897c395f | 6241 | /* |
e5673f28 | 6242 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 6243 | * part of active balancing operations within "domain". |
897c395f | 6244 | * |
e5673f28 | 6245 | * Returns a task if successful and NULL otherwise. |
897c395f | 6246 | */ |
e5673f28 | 6247 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f PZ |
6248 | { |
6249 | struct task_struct *p, *n; | |
897c395f | 6250 | |
e5673f28 KT |
6251 | lockdep_assert_held(&env->src_rq->lock); |
6252 | ||
367456c7 | 6253 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
6254 | if (!can_migrate_task(p, env)) |
6255 | continue; | |
897c395f | 6256 | |
163122b7 | 6257 | detach_task(p, env); |
e5673f28 | 6258 | |
367456c7 | 6259 | /* |
e5673f28 | 6260 | * Right now, this is only the second place where |
163122b7 | 6261 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 6262 | * so we can safely collect stats here rather than |
163122b7 | 6263 | * inside detach_tasks(). |
367456c7 | 6264 | */ |
ae92882e | 6265 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 6266 | return p; |
897c395f | 6267 | } |
e5673f28 | 6268 | return NULL; |
897c395f PZ |
6269 | } |
6270 | ||
eb95308e PZ |
6271 | static const unsigned int sched_nr_migrate_break = 32; |
6272 | ||
5d6523eb | 6273 | /* |
163122b7 KT |
6274 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
6275 | * busiest_rq, as part of a balancing operation within domain "sd". | |
5d6523eb | 6276 | * |
163122b7 | 6277 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 6278 | */ |
163122b7 | 6279 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 6280 | { |
5d6523eb PZ |
6281 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
6282 | struct task_struct *p; | |
367456c7 | 6283 | unsigned long load; |
163122b7 KT |
6284 | int detached = 0; |
6285 | ||
6286 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 6287 | |
bd939f45 | 6288 | if (env->imbalance <= 0) |
5d6523eb | 6289 | return 0; |
1e3c88bd | 6290 | |
5d6523eb | 6291 | while (!list_empty(tasks)) { |
985d3a4c YD |
6292 | /* |
6293 | * We don't want to steal all, otherwise we may be treated likewise, | |
6294 | * which could at worst lead to a livelock crash. | |
6295 | */ | |
6296 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
6297 | break; | |
6298 | ||
5d6523eb | 6299 | p = list_first_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 6300 | |
367456c7 PZ |
6301 | env->loop++; |
6302 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 6303 | if (env->loop > env->loop_max) |
367456c7 | 6304 | break; |
5d6523eb PZ |
6305 | |
6306 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 6307 | if (env->loop > env->loop_break) { |
eb95308e | 6308 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 6309 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 6310 | break; |
a195f004 | 6311 | } |
1e3c88bd | 6312 | |
d3198084 | 6313 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
6314 | goto next; |
6315 | ||
6316 | load = task_h_load(p); | |
5d6523eb | 6317 | |
eb95308e | 6318 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
6319 | goto next; |
6320 | ||
bd939f45 | 6321 | if ((load / 2) > env->imbalance) |
367456c7 | 6322 | goto next; |
1e3c88bd | 6323 | |
163122b7 KT |
6324 | detach_task(p, env); |
6325 | list_add(&p->se.group_node, &env->tasks); | |
6326 | ||
6327 | detached++; | |
bd939f45 | 6328 | env->imbalance -= load; |
1e3c88bd PZ |
6329 | |
6330 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
6331 | /* |
6332 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 6333 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
6334 | * the critical section. |
6335 | */ | |
5d6523eb | 6336 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 6337 | break; |
1e3c88bd PZ |
6338 | #endif |
6339 | ||
ee00e66f PZ |
6340 | /* |
6341 | * We only want to steal up to the prescribed amount of | |
6342 | * weighted load. | |
6343 | */ | |
bd939f45 | 6344 | if (env->imbalance <= 0) |
ee00e66f | 6345 | break; |
367456c7 PZ |
6346 | |
6347 | continue; | |
6348 | next: | |
5d6523eb | 6349 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 6350 | } |
5d6523eb | 6351 | |
1e3c88bd | 6352 | /* |
163122b7 KT |
6353 | * Right now, this is one of only two places we collect this stat |
6354 | * so we can safely collect detach_one_task() stats here rather | |
6355 | * than inside detach_one_task(). | |
1e3c88bd | 6356 | */ |
ae92882e | 6357 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 6358 | |
163122b7 KT |
6359 | return detached; |
6360 | } | |
6361 | ||
6362 | /* | |
6363 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
6364 | */ | |
6365 | static void attach_task(struct rq *rq, struct task_struct *p) | |
6366 | { | |
6367 | lockdep_assert_held(&rq->lock); | |
6368 | ||
6369 | BUG_ON(task_rq(p) != rq); | |
163122b7 | 6370 | activate_task(rq, p, 0); |
3ea94de1 | 6371 | p->on_rq = TASK_ON_RQ_QUEUED; |
163122b7 KT |
6372 | check_preempt_curr(rq, p, 0); |
6373 | } | |
6374 | ||
6375 | /* | |
6376 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
6377 | * its new rq. | |
6378 | */ | |
6379 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
6380 | { | |
6381 | raw_spin_lock(&rq->lock); | |
6382 | attach_task(rq, p); | |
6383 | raw_spin_unlock(&rq->lock); | |
6384 | } | |
6385 | ||
6386 | /* | |
6387 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
6388 | * new rq. | |
6389 | */ | |
6390 | static void attach_tasks(struct lb_env *env) | |
6391 | { | |
6392 | struct list_head *tasks = &env->tasks; | |
6393 | struct task_struct *p; | |
6394 | ||
6395 | raw_spin_lock(&env->dst_rq->lock); | |
6396 | ||
6397 | while (!list_empty(tasks)) { | |
6398 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
6399 | list_del_init(&p->se.group_node); | |
1e3c88bd | 6400 | |
163122b7 KT |
6401 | attach_task(env->dst_rq, p); |
6402 | } | |
6403 | ||
6404 | raw_spin_unlock(&env->dst_rq->lock); | |
1e3c88bd PZ |
6405 | } |
6406 | ||
230059de | 6407 | #ifdef CONFIG_FAIR_GROUP_SCHED |
48a16753 | 6408 | static void update_blocked_averages(int cpu) |
9e3081ca | 6409 | { |
9e3081ca | 6410 | struct rq *rq = cpu_rq(cpu); |
48a16753 PT |
6411 | struct cfs_rq *cfs_rq; |
6412 | unsigned long flags; | |
9e3081ca | 6413 | |
48a16753 PT |
6414 | raw_spin_lock_irqsave(&rq->lock, flags); |
6415 | update_rq_clock(rq); | |
9d89c257 | 6416 | |
9763b67f PZ |
6417 | /* |
6418 | * Iterates the task_group tree in a bottom up fashion, see | |
6419 | * list_add_leaf_cfs_rq() for details. | |
6420 | */ | |
64660c86 | 6421 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
9d89c257 YD |
6422 | /* throttled entities do not contribute to load */ |
6423 | if (throttled_hierarchy(cfs_rq)) | |
6424 | continue; | |
48a16753 | 6425 | |
a2c6c91f | 6426 | if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true)) |
9d89c257 YD |
6427 | update_tg_load_avg(cfs_rq, 0); |
6428 | } | |
48a16753 | 6429 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
9e3081ca PZ |
6430 | } |
6431 | ||
9763b67f | 6432 | /* |
68520796 | 6433 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
6434 | * This needs to be done in a top-down fashion because the load of a child |
6435 | * group is a fraction of its parents load. | |
6436 | */ | |
68520796 | 6437 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 6438 | { |
68520796 VD |
6439 | struct rq *rq = rq_of(cfs_rq); |
6440 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 6441 | unsigned long now = jiffies; |
68520796 | 6442 | unsigned long load; |
a35b6466 | 6443 | |
68520796 | 6444 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
6445 | return; |
6446 | ||
68520796 VD |
6447 | cfs_rq->h_load_next = NULL; |
6448 | for_each_sched_entity(se) { | |
6449 | cfs_rq = cfs_rq_of(se); | |
6450 | cfs_rq->h_load_next = se; | |
6451 | if (cfs_rq->last_h_load_update == now) | |
6452 | break; | |
6453 | } | |
a35b6466 | 6454 | |
68520796 | 6455 | if (!se) { |
7ea241af | 6456 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
6457 | cfs_rq->last_h_load_update = now; |
6458 | } | |
6459 | ||
6460 | while ((se = cfs_rq->h_load_next) != NULL) { | |
6461 | load = cfs_rq->h_load; | |
7ea241af YD |
6462 | load = div64_ul(load * se->avg.load_avg, |
6463 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
6464 | cfs_rq = group_cfs_rq(se); |
6465 | cfs_rq->h_load = load; | |
6466 | cfs_rq->last_h_load_update = now; | |
6467 | } | |
9763b67f PZ |
6468 | } |
6469 | ||
367456c7 | 6470 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 6471 | { |
367456c7 | 6472 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 6473 | |
68520796 | 6474 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 6475 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 6476 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
6477 | } |
6478 | #else | |
48a16753 | 6479 | static inline void update_blocked_averages(int cpu) |
9e3081ca | 6480 | { |
6c1d47c0 VG |
6481 | struct rq *rq = cpu_rq(cpu); |
6482 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6483 | unsigned long flags; | |
6484 | ||
6485 | raw_spin_lock_irqsave(&rq->lock, flags); | |
6486 | update_rq_clock(rq); | |
a2c6c91f | 6487 | update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true); |
6c1d47c0 | 6488 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
9e3081ca PZ |
6489 | } |
6490 | ||
367456c7 | 6491 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 6492 | { |
9d89c257 | 6493 | return p->se.avg.load_avg; |
1e3c88bd | 6494 | } |
230059de | 6495 | #endif |
1e3c88bd | 6496 | |
1e3c88bd | 6497 | /********** Helpers for find_busiest_group ************************/ |
caeb178c RR |
6498 | |
6499 | enum group_type { | |
6500 | group_other = 0, | |
6501 | group_imbalanced, | |
6502 | group_overloaded, | |
6503 | }; | |
6504 | ||
1e3c88bd PZ |
6505 | /* |
6506 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
6507 | */ | |
6508 | struct sg_lb_stats { | |
6509 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
6510 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 6511 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 6512 | unsigned long load_per_task; |
63b2ca30 | 6513 | unsigned long group_capacity; |
9e91d61d | 6514 | unsigned long group_util; /* Total utilization of the group */ |
147c5fc2 | 6515 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
147c5fc2 PZ |
6516 | unsigned int idle_cpus; |
6517 | unsigned int group_weight; | |
caeb178c | 6518 | enum group_type group_type; |
ea67821b | 6519 | int group_no_capacity; |
0ec8aa00 PZ |
6520 | #ifdef CONFIG_NUMA_BALANCING |
6521 | unsigned int nr_numa_running; | |
6522 | unsigned int nr_preferred_running; | |
6523 | #endif | |
1e3c88bd PZ |
6524 | }; |
6525 | ||
56cf515b JK |
6526 | /* |
6527 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
6528 | * during load balancing. | |
6529 | */ | |
6530 | struct sd_lb_stats { | |
6531 | struct sched_group *busiest; /* Busiest group in this sd */ | |
6532 | struct sched_group *local; /* Local group in this sd */ | |
6533 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 6534 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
6535 | unsigned long avg_load; /* Average load across all groups in sd */ |
6536 | ||
56cf515b | 6537 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 6538 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
6539 | }; |
6540 | ||
147c5fc2 PZ |
6541 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
6542 | { | |
6543 | /* | |
6544 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
6545 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
6546 | * We must however clear busiest_stat::avg_load because | |
6547 | * update_sd_pick_busiest() reads this before assignment. | |
6548 | */ | |
6549 | *sds = (struct sd_lb_stats){ | |
6550 | .busiest = NULL, | |
6551 | .local = NULL, | |
6552 | .total_load = 0UL, | |
63b2ca30 | 6553 | .total_capacity = 0UL, |
147c5fc2 PZ |
6554 | .busiest_stat = { |
6555 | .avg_load = 0UL, | |
caeb178c RR |
6556 | .sum_nr_running = 0, |
6557 | .group_type = group_other, | |
147c5fc2 PZ |
6558 | }, |
6559 | }; | |
6560 | } | |
6561 | ||
1e3c88bd PZ |
6562 | /** |
6563 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
6564 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 6565 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
6566 | * |
6567 | * Return: The load index. | |
1e3c88bd PZ |
6568 | */ |
6569 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
6570 | enum cpu_idle_type idle) | |
6571 | { | |
6572 | int load_idx; | |
6573 | ||
6574 | switch (idle) { | |
6575 | case CPU_NOT_IDLE: | |
6576 | load_idx = sd->busy_idx; | |
6577 | break; | |
6578 | ||
6579 | case CPU_NEWLY_IDLE: | |
6580 | load_idx = sd->newidle_idx; | |
6581 | break; | |
6582 | default: | |
6583 | load_idx = sd->idle_idx; | |
6584 | break; | |
6585 | } | |
6586 | ||
6587 | return load_idx; | |
6588 | } | |
6589 | ||
ced549fa | 6590 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
6591 | { |
6592 | struct rq *rq = cpu_rq(cpu); | |
b5b4860d | 6593 | u64 total, used, age_stamp, avg; |
cadefd3d | 6594 | s64 delta; |
1e3c88bd | 6595 | |
b654f7de PZ |
6596 | /* |
6597 | * Since we're reading these variables without serialization make sure | |
6598 | * we read them once before doing sanity checks on them. | |
6599 | */ | |
316c1608 JL |
6600 | age_stamp = READ_ONCE(rq->age_stamp); |
6601 | avg = READ_ONCE(rq->rt_avg); | |
cebde6d6 | 6602 | delta = __rq_clock_broken(rq) - age_stamp; |
b654f7de | 6603 | |
cadefd3d PZ |
6604 | if (unlikely(delta < 0)) |
6605 | delta = 0; | |
6606 | ||
6607 | total = sched_avg_period() + delta; | |
aa483808 | 6608 | |
b5b4860d | 6609 | used = div_u64(avg, total); |
1e3c88bd | 6610 | |
b5b4860d VG |
6611 | if (likely(used < SCHED_CAPACITY_SCALE)) |
6612 | return SCHED_CAPACITY_SCALE - used; | |
1e3c88bd | 6613 | |
b5b4860d | 6614 | return 1; |
1e3c88bd PZ |
6615 | } |
6616 | ||
ced549fa | 6617 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 6618 | { |
8cd5601c | 6619 | unsigned long capacity = arch_scale_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
6620 | struct sched_group *sdg = sd->groups; |
6621 | ||
ca6d75e6 | 6622 | cpu_rq(cpu)->cpu_capacity_orig = capacity; |
9d5efe05 | 6623 | |
ced549fa | 6624 | capacity *= scale_rt_capacity(cpu); |
ca8ce3d0 | 6625 | capacity >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd | 6626 | |
ced549fa NP |
6627 | if (!capacity) |
6628 | capacity = 1; | |
1e3c88bd | 6629 | |
ced549fa NP |
6630 | cpu_rq(cpu)->cpu_capacity = capacity; |
6631 | sdg->sgc->capacity = capacity; | |
1e3c88bd PZ |
6632 | } |
6633 | ||
63b2ca30 | 6634 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
6635 | { |
6636 | struct sched_domain *child = sd->child; | |
6637 | struct sched_group *group, *sdg = sd->groups; | |
dc7ff76e | 6638 | unsigned long capacity; |
4ec4412e VG |
6639 | unsigned long interval; |
6640 | ||
6641 | interval = msecs_to_jiffies(sd->balance_interval); | |
6642 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 6643 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
6644 | |
6645 | if (!child) { | |
ced549fa | 6646 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
6647 | return; |
6648 | } | |
6649 | ||
dc7ff76e | 6650 | capacity = 0; |
1e3c88bd | 6651 | |
74a5ce20 PZ |
6652 | if (child->flags & SD_OVERLAP) { |
6653 | /* | |
6654 | * SD_OVERLAP domains cannot assume that child groups | |
6655 | * span the current group. | |
6656 | */ | |
6657 | ||
863bffc8 | 6658 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
63b2ca30 | 6659 | struct sched_group_capacity *sgc; |
9abf24d4 | 6660 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 6661 | |
9abf24d4 | 6662 | /* |
63b2ca30 | 6663 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
6664 | * gets here before we've attached the domains to the |
6665 | * runqueues. | |
6666 | * | |
ced549fa NP |
6667 | * Use capacity_of(), which is set irrespective of domains |
6668 | * in update_cpu_capacity(). | |
9abf24d4 | 6669 | * |
dc7ff76e | 6670 | * This avoids capacity from being 0 and |
9abf24d4 | 6671 | * causing divide-by-zero issues on boot. |
9abf24d4 SD |
6672 | */ |
6673 | if (unlikely(!rq->sd)) { | |
ced549fa | 6674 | capacity += capacity_of(cpu); |
9abf24d4 SD |
6675 | continue; |
6676 | } | |
863bffc8 | 6677 | |
63b2ca30 | 6678 | sgc = rq->sd->groups->sgc; |
63b2ca30 | 6679 | capacity += sgc->capacity; |
863bffc8 | 6680 | } |
74a5ce20 PZ |
6681 | } else { |
6682 | /* | |
6683 | * !SD_OVERLAP domains can assume that child groups | |
6684 | * span the current group. | |
97a7142f | 6685 | */ |
74a5ce20 PZ |
6686 | |
6687 | group = child->groups; | |
6688 | do { | |
63b2ca30 | 6689 | capacity += group->sgc->capacity; |
74a5ce20 PZ |
6690 | group = group->next; |
6691 | } while (group != child->groups); | |
6692 | } | |
1e3c88bd | 6693 | |
63b2ca30 | 6694 | sdg->sgc->capacity = capacity; |
1e3c88bd PZ |
6695 | } |
6696 | ||
9d5efe05 | 6697 | /* |
ea67821b VG |
6698 | * Check whether the capacity of the rq has been noticeably reduced by side |
6699 | * activity. The imbalance_pct is used for the threshold. | |
6700 | * Return true is the capacity is reduced | |
9d5efe05 SV |
6701 | */ |
6702 | static inline int | |
ea67821b | 6703 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 6704 | { |
ea67821b VG |
6705 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
6706 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
6707 | } |
6708 | ||
30ce5dab PZ |
6709 | /* |
6710 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
6711 | * groups is inadequate due to tsk_cpus_allowed() constraints. | |
6712 | * | |
6713 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
6714 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
6715 | * Something like: | |
6716 | * | |
6717 | * { 0 1 2 3 } { 4 5 6 7 } | |
6718 | * * * * * | |
6719 | * | |
6720 | * If we were to balance group-wise we'd place two tasks in the first group and | |
6721 | * two tasks in the second group. Clearly this is undesired as it will overload | |
6722 | * cpu 3 and leave one of the cpus in the second group unused. | |
6723 | * | |
6724 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
6725 | * by noticing the lower domain failed to reach balance and had difficulty |
6726 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
6727 | * |
6728 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 6729 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 6730 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
6731 | * to create an effective group imbalance. |
6732 | * | |
6733 | * This is a somewhat tricky proposition since the next run might not find the | |
6734 | * group imbalance and decide the groups need to be balanced again. A most | |
6735 | * subtle and fragile situation. | |
6736 | */ | |
6737 | ||
6263322c | 6738 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 6739 | { |
63b2ca30 | 6740 | return group->sgc->imbalance; |
30ce5dab PZ |
6741 | } |
6742 | ||
b37d9316 | 6743 | /* |
ea67821b VG |
6744 | * group_has_capacity returns true if the group has spare capacity that could |
6745 | * be used by some tasks. | |
6746 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
6747 | * smaller than the number of CPUs or if the utilization is lower than the |
6748 | * available capacity for CFS tasks. | |
ea67821b VG |
6749 | * For the latter, we use a threshold to stabilize the state, to take into |
6750 | * account the variance of the tasks' load and to return true if the available | |
6751 | * capacity in meaningful for the load balancer. | |
6752 | * As an example, an available capacity of 1% can appear but it doesn't make | |
6753 | * any benefit for the load balance. | |
b37d9316 | 6754 | */ |
ea67821b VG |
6755 | static inline bool |
6756 | group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) | |
b37d9316 | 6757 | { |
ea67821b VG |
6758 | if (sgs->sum_nr_running < sgs->group_weight) |
6759 | return true; | |
c61037e9 | 6760 | |
ea67821b | 6761 | if ((sgs->group_capacity * 100) > |
9e91d61d | 6762 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 6763 | return true; |
b37d9316 | 6764 | |
ea67821b VG |
6765 | return false; |
6766 | } | |
6767 | ||
6768 | /* | |
6769 | * group_is_overloaded returns true if the group has more tasks than it can | |
6770 | * handle. | |
6771 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
6772 | * with the exact right number of tasks, has no more spare capacity but is not | |
6773 | * overloaded so both group_has_capacity and group_is_overloaded return | |
6774 | * false. | |
6775 | */ | |
6776 | static inline bool | |
6777 | group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) | |
6778 | { | |
6779 | if (sgs->sum_nr_running <= sgs->group_weight) | |
6780 | return false; | |
b37d9316 | 6781 | |
ea67821b | 6782 | if ((sgs->group_capacity * 100) < |
9e91d61d | 6783 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 6784 | return true; |
b37d9316 | 6785 | |
ea67821b | 6786 | return false; |
b37d9316 PZ |
6787 | } |
6788 | ||
79a89f92 LY |
6789 | static inline enum |
6790 | group_type group_classify(struct sched_group *group, | |
6791 | struct sg_lb_stats *sgs) | |
caeb178c | 6792 | { |
ea67821b | 6793 | if (sgs->group_no_capacity) |
caeb178c RR |
6794 | return group_overloaded; |
6795 | ||
6796 | if (sg_imbalanced(group)) | |
6797 | return group_imbalanced; | |
6798 | ||
6799 | return group_other; | |
6800 | } | |
6801 | ||
1e3c88bd PZ |
6802 | /** |
6803 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 6804 | * @env: The load balancing environment. |
1e3c88bd | 6805 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 6806 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 6807 | * @local_group: Does group contain this_cpu. |
1e3c88bd | 6808 | * @sgs: variable to hold the statistics for this group. |
cd3bd4e6 | 6809 | * @overload: Indicate more than one runnable task for any CPU. |
1e3c88bd | 6810 | */ |
bd939f45 PZ |
6811 | static inline void update_sg_lb_stats(struct lb_env *env, |
6812 | struct sched_group *group, int load_idx, | |
4486edd1 TC |
6813 | int local_group, struct sg_lb_stats *sgs, |
6814 | bool *overload) | |
1e3c88bd | 6815 | { |
30ce5dab | 6816 | unsigned long load; |
a426f99c | 6817 | int i, nr_running; |
1e3c88bd | 6818 | |
b72ff13c PZ |
6819 | memset(sgs, 0, sizeof(*sgs)); |
6820 | ||
b9403130 | 6821 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
6822 | struct rq *rq = cpu_rq(i); |
6823 | ||
1e3c88bd | 6824 | /* Bias balancing toward cpus of our domain */ |
6263322c | 6825 | if (local_group) |
04f733b4 | 6826 | load = target_load(i, load_idx); |
6263322c | 6827 | else |
1e3c88bd | 6828 | load = source_load(i, load_idx); |
1e3c88bd PZ |
6829 | |
6830 | sgs->group_load += load; | |
9e91d61d | 6831 | sgs->group_util += cpu_util(i); |
65fdac08 | 6832 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 6833 | |
a426f99c WL |
6834 | nr_running = rq->nr_running; |
6835 | if (nr_running > 1) | |
4486edd1 TC |
6836 | *overload = true; |
6837 | ||
0ec8aa00 PZ |
6838 | #ifdef CONFIG_NUMA_BALANCING |
6839 | sgs->nr_numa_running += rq->nr_numa_running; | |
6840 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
6841 | #endif | |
1e3c88bd | 6842 | sgs->sum_weighted_load += weighted_cpuload(i); |
a426f99c WL |
6843 | /* |
6844 | * No need to call idle_cpu() if nr_running is not 0 | |
6845 | */ | |
6846 | if (!nr_running && idle_cpu(i)) | |
aae6d3dd | 6847 | sgs->idle_cpus++; |
1e3c88bd PZ |
6848 | } |
6849 | ||
63b2ca30 NP |
6850 | /* Adjust by relative CPU capacity of the group */ |
6851 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 6852 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 6853 | |
dd5feea1 | 6854 | if (sgs->sum_nr_running) |
38d0f770 | 6855 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 6856 | |
aae6d3dd | 6857 | sgs->group_weight = group->group_weight; |
b37d9316 | 6858 | |
ea67821b | 6859 | sgs->group_no_capacity = group_is_overloaded(env, sgs); |
79a89f92 | 6860 | sgs->group_type = group_classify(group, sgs); |
1e3c88bd PZ |
6861 | } |
6862 | ||
532cb4c4 MN |
6863 | /** |
6864 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 6865 | * @env: The load balancing environment. |
532cb4c4 MN |
6866 | * @sds: sched_domain statistics |
6867 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 6868 | * @sgs: sched_group statistics |
532cb4c4 MN |
6869 | * |
6870 | * Determine if @sg is a busier group than the previously selected | |
6871 | * busiest group. | |
e69f6186 YB |
6872 | * |
6873 | * Return: %true if @sg is a busier group than the previously selected | |
6874 | * busiest group. %false otherwise. | |
532cb4c4 | 6875 | */ |
bd939f45 | 6876 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
6877 | struct sd_lb_stats *sds, |
6878 | struct sched_group *sg, | |
bd939f45 | 6879 | struct sg_lb_stats *sgs) |
532cb4c4 | 6880 | { |
caeb178c | 6881 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 6882 | |
caeb178c | 6883 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
6884 | return true; |
6885 | ||
caeb178c RR |
6886 | if (sgs->group_type < busiest->group_type) |
6887 | return false; | |
6888 | ||
6889 | if (sgs->avg_load <= busiest->avg_load) | |
6890 | return false; | |
6891 | ||
6892 | /* This is the busiest node in its class. */ | |
6893 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
6894 | return true; |
6895 | ||
1f621e02 SD |
6896 | /* No ASYM_PACKING if target cpu is already busy */ |
6897 | if (env->idle == CPU_NOT_IDLE) | |
6898 | return true; | |
532cb4c4 MN |
6899 | /* |
6900 | * ASYM_PACKING needs to move all the work to the lowest | |
6901 | * numbered CPUs in the group, therefore mark all groups | |
6902 | * higher than ourself as busy. | |
6903 | */ | |
caeb178c | 6904 | if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { |
532cb4c4 MN |
6905 | if (!sds->busiest) |
6906 | return true; | |
6907 | ||
1f621e02 SD |
6908 | /* Prefer to move from highest possible cpu's work */ |
6909 | if (group_first_cpu(sds->busiest) < group_first_cpu(sg)) | |
532cb4c4 MN |
6910 | return true; |
6911 | } | |
6912 | ||
6913 | return false; | |
6914 | } | |
6915 | ||
0ec8aa00 PZ |
6916 | #ifdef CONFIG_NUMA_BALANCING |
6917 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
6918 | { | |
6919 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
6920 | return regular; | |
6921 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
6922 | return remote; | |
6923 | return all; | |
6924 | } | |
6925 | ||
6926 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
6927 | { | |
6928 | if (rq->nr_running > rq->nr_numa_running) | |
6929 | return regular; | |
6930 | if (rq->nr_running > rq->nr_preferred_running) | |
6931 | return remote; | |
6932 | return all; | |
6933 | } | |
6934 | #else | |
6935 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
6936 | { | |
6937 | return all; | |
6938 | } | |
6939 | ||
6940 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
6941 | { | |
6942 | return regular; | |
6943 | } | |
6944 | #endif /* CONFIG_NUMA_BALANCING */ | |
6945 | ||
1e3c88bd | 6946 | /** |
461819ac | 6947 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 6948 | * @env: The load balancing environment. |
1e3c88bd PZ |
6949 | * @sds: variable to hold the statistics for this sched_domain. |
6950 | */ | |
0ec8aa00 | 6951 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 6952 | { |
bd939f45 PZ |
6953 | struct sched_domain *child = env->sd->child; |
6954 | struct sched_group *sg = env->sd->groups; | |
56cf515b | 6955 | struct sg_lb_stats tmp_sgs; |
1e3c88bd | 6956 | int load_idx, prefer_sibling = 0; |
4486edd1 | 6957 | bool overload = false; |
1e3c88bd PZ |
6958 | |
6959 | if (child && child->flags & SD_PREFER_SIBLING) | |
6960 | prefer_sibling = 1; | |
6961 | ||
bd939f45 | 6962 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
6963 | |
6964 | do { | |
56cf515b | 6965 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
6966 | int local_group; |
6967 | ||
bd939f45 | 6968 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
6969 | if (local_group) { |
6970 | sds->local = sg; | |
6971 | sgs = &sds->local_stat; | |
b72ff13c PZ |
6972 | |
6973 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
6974 | time_after_eq(jiffies, sg->sgc->next_update)) |
6975 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 6976 | } |
1e3c88bd | 6977 | |
4486edd1 TC |
6978 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, |
6979 | &overload); | |
1e3c88bd | 6980 | |
b72ff13c PZ |
6981 | if (local_group) |
6982 | goto next_group; | |
6983 | ||
1e3c88bd PZ |
6984 | /* |
6985 | * In case the child domain prefers tasks go to siblings | |
ea67821b | 6986 | * first, lower the sg capacity so that we'll try |
75dd321d NR |
6987 | * and move all the excess tasks away. We lower the capacity |
6988 | * of a group only if the local group has the capacity to fit | |
ea67821b VG |
6989 | * these excess tasks. The extra check prevents the case where |
6990 | * you always pull from the heaviest group when it is already | |
6991 | * under-utilized (possible with a large weight task outweighs | |
6992 | * the tasks on the system). | |
1e3c88bd | 6993 | */ |
b72ff13c | 6994 | if (prefer_sibling && sds->local && |
ea67821b VG |
6995 | group_has_capacity(env, &sds->local_stat) && |
6996 | (sgs->sum_nr_running > 1)) { | |
6997 | sgs->group_no_capacity = 1; | |
79a89f92 | 6998 | sgs->group_type = group_classify(sg, sgs); |
cb0b9f24 | 6999 | } |
1e3c88bd | 7000 | |
b72ff13c | 7001 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 7002 | sds->busiest = sg; |
56cf515b | 7003 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
7004 | } |
7005 | ||
b72ff13c PZ |
7006 | next_group: |
7007 | /* Now, start updating sd_lb_stats */ | |
7008 | sds->total_load += sgs->group_load; | |
63b2ca30 | 7009 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 7010 | |
532cb4c4 | 7011 | sg = sg->next; |
bd939f45 | 7012 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
7013 | |
7014 | if (env->sd->flags & SD_NUMA) | |
7015 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
7016 | |
7017 | if (!env->sd->parent) { | |
7018 | /* update overload indicator if we are at root domain */ | |
7019 | if (env->dst_rq->rd->overload != overload) | |
7020 | env->dst_rq->rd->overload = overload; | |
7021 | } | |
7022 | ||
532cb4c4 MN |
7023 | } |
7024 | ||
532cb4c4 MN |
7025 | /** |
7026 | * check_asym_packing - Check to see if the group is packed into the | |
7027 | * sched doman. | |
7028 | * | |
7029 | * This is primarily intended to used at the sibling level. Some | |
7030 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
7031 | * case of POWER7, it can move to lower SMT modes only when higher | |
7032 | * threads are idle. When in lower SMT modes, the threads will | |
7033 | * perform better since they share less core resources. Hence when we | |
7034 | * have idle threads, we want them to be the higher ones. | |
7035 | * | |
7036 | * This packing function is run on idle threads. It checks to see if | |
7037 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
7038 | * CPU number than the packing function is being run on. Here we are | |
7039 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
7040 | * number. | |
7041 | * | |
e69f6186 | 7042 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
7043 | * this CPU. The amount of the imbalance is returned in *imbalance. |
7044 | * | |
cd96891d | 7045 | * @env: The load balancing environment. |
532cb4c4 | 7046 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 7047 | */ |
bd939f45 | 7048 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
7049 | { |
7050 | int busiest_cpu; | |
7051 | ||
bd939f45 | 7052 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
7053 | return 0; |
7054 | ||
1f621e02 SD |
7055 | if (env->idle == CPU_NOT_IDLE) |
7056 | return 0; | |
7057 | ||
532cb4c4 MN |
7058 | if (!sds->busiest) |
7059 | return 0; | |
7060 | ||
7061 | busiest_cpu = group_first_cpu(sds->busiest); | |
bd939f45 | 7062 | if (env->dst_cpu > busiest_cpu) |
532cb4c4 MN |
7063 | return 0; |
7064 | ||
bd939f45 | 7065 | env->imbalance = DIV_ROUND_CLOSEST( |
63b2ca30 | 7066 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, |
ca8ce3d0 | 7067 | SCHED_CAPACITY_SCALE); |
bd939f45 | 7068 | |
532cb4c4 | 7069 | return 1; |
1e3c88bd PZ |
7070 | } |
7071 | ||
7072 | /** | |
7073 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
7074 | * amongst the groups of a sched_domain, during | |
7075 | * load balancing. | |
cd96891d | 7076 | * @env: The load balancing environment. |
1e3c88bd | 7077 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 7078 | */ |
bd939f45 PZ |
7079 | static inline |
7080 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 7081 | { |
63b2ca30 | 7082 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 7083 | unsigned int imbn = 2; |
dd5feea1 | 7084 | unsigned long scaled_busy_load_per_task; |
56cf515b | 7085 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 7086 | |
56cf515b JK |
7087 | local = &sds->local_stat; |
7088 | busiest = &sds->busiest_stat; | |
1e3c88bd | 7089 | |
56cf515b JK |
7090 | if (!local->sum_nr_running) |
7091 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
7092 | else if (busiest->load_per_task > local->load_per_task) | |
7093 | imbn = 1; | |
dd5feea1 | 7094 | |
56cf515b | 7095 | scaled_busy_load_per_task = |
ca8ce3d0 | 7096 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 7097 | busiest->group_capacity; |
56cf515b | 7098 | |
3029ede3 VD |
7099 | if (busiest->avg_load + scaled_busy_load_per_task >= |
7100 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 7101 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
7102 | return; |
7103 | } | |
7104 | ||
7105 | /* | |
7106 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 7107 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
7108 | * moving them. |
7109 | */ | |
7110 | ||
63b2ca30 | 7111 | capa_now += busiest->group_capacity * |
56cf515b | 7112 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 7113 | capa_now += local->group_capacity * |
56cf515b | 7114 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 7115 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
7116 | |
7117 | /* Amount of load we'd subtract */ | |
a2cd4260 | 7118 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 7119 | capa_move += busiest->group_capacity * |
56cf515b | 7120 | min(busiest->load_per_task, |
a2cd4260 | 7121 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 7122 | } |
1e3c88bd PZ |
7123 | |
7124 | /* Amount of load we'd add */ | |
63b2ca30 | 7125 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 7126 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
7127 | tmp = (busiest->avg_load * busiest->group_capacity) / |
7128 | local->group_capacity; | |
56cf515b | 7129 | } else { |
ca8ce3d0 | 7130 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 7131 | local->group_capacity; |
56cf515b | 7132 | } |
63b2ca30 | 7133 | capa_move += local->group_capacity * |
3ae11c90 | 7134 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 7135 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
7136 | |
7137 | /* Move if we gain throughput */ | |
63b2ca30 | 7138 | if (capa_move > capa_now) |
56cf515b | 7139 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
7140 | } |
7141 | ||
7142 | /** | |
7143 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
7144 | * groups of a given sched_domain during load balance. | |
bd939f45 | 7145 | * @env: load balance environment |
1e3c88bd | 7146 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 7147 | */ |
bd939f45 | 7148 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 7149 | { |
dd5feea1 | 7150 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
7151 | struct sg_lb_stats *local, *busiest; |
7152 | ||
7153 | local = &sds->local_stat; | |
56cf515b | 7154 | busiest = &sds->busiest_stat; |
dd5feea1 | 7155 | |
caeb178c | 7156 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
7157 | /* |
7158 | * In the group_imb case we cannot rely on group-wide averages | |
7159 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
7160 | */ | |
56cf515b JK |
7161 | busiest->load_per_task = |
7162 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
7163 | } |
7164 | ||
1e3c88bd | 7165 | /* |
885e542c DE |
7166 | * Avg load of busiest sg can be less and avg load of local sg can |
7167 | * be greater than avg load across all sgs of sd because avg load | |
7168 | * factors in sg capacity and sgs with smaller group_type are | |
7169 | * skipped when updating the busiest sg: | |
1e3c88bd | 7170 | */ |
b1885550 VD |
7171 | if (busiest->avg_load <= sds->avg_load || |
7172 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
7173 | env->imbalance = 0; |
7174 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
7175 | } |
7176 | ||
9a5d9ba6 PZ |
7177 | /* |
7178 | * If there aren't any idle cpus, avoid creating some. | |
7179 | */ | |
7180 | if (busiest->group_type == group_overloaded && | |
7181 | local->group_type == group_overloaded) { | |
1be0eb2a | 7182 | load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE; |
cfa10334 | 7183 | if (load_above_capacity > busiest->group_capacity) { |
ea67821b | 7184 | load_above_capacity -= busiest->group_capacity; |
26656215 | 7185 | load_above_capacity *= scale_load_down(NICE_0_LOAD); |
cfa10334 MR |
7186 | load_above_capacity /= busiest->group_capacity; |
7187 | } else | |
ea67821b | 7188 | load_above_capacity = ~0UL; |
dd5feea1 SS |
7189 | } |
7190 | ||
7191 | /* | |
7192 | * We're trying to get all the cpus to the average_load, so we don't | |
7193 | * want to push ourselves above the average load, nor do we wish to | |
7194 | * reduce the max loaded cpu below the average load. At the same time, | |
0a9b23ce DE |
7195 | * we also don't want to reduce the group load below the group |
7196 | * capacity. Thus we look for the minimum possible imbalance. | |
dd5feea1 | 7197 | */ |
30ce5dab | 7198 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
7199 | |
7200 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 7201 | env->imbalance = min( |
63b2ca30 NP |
7202 | max_pull * busiest->group_capacity, |
7203 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 7204 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
7205 | |
7206 | /* | |
7207 | * if *imbalance is less than the average load per runnable task | |
25985edc | 7208 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
7209 | * a think about bumping its value to force at least one task to be |
7210 | * moved | |
7211 | */ | |
56cf515b | 7212 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 7213 | return fix_small_imbalance(env, sds); |
1e3c88bd | 7214 | } |
fab47622 | 7215 | |
1e3c88bd PZ |
7216 | /******* find_busiest_group() helpers end here *********************/ |
7217 | ||
7218 | /** | |
7219 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 7220 | * if there is an imbalance. |
1e3c88bd PZ |
7221 | * |
7222 | * Also calculates the amount of weighted load which should be moved | |
7223 | * to restore balance. | |
7224 | * | |
cd96891d | 7225 | * @env: The load balancing environment. |
1e3c88bd | 7226 | * |
e69f6186 | 7227 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 7228 | */ |
56cf515b | 7229 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 7230 | { |
56cf515b | 7231 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
7232 | struct sd_lb_stats sds; |
7233 | ||
147c5fc2 | 7234 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
7235 | |
7236 | /* | |
7237 | * Compute the various statistics relavent for load balancing at | |
7238 | * this level. | |
7239 | */ | |
23f0d209 | 7240 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
7241 | local = &sds.local_stat; |
7242 | busiest = &sds.busiest_stat; | |
1e3c88bd | 7243 | |
ea67821b | 7244 | /* ASYM feature bypasses nice load balance check */ |
1f621e02 | 7245 | if (check_asym_packing(env, &sds)) |
532cb4c4 MN |
7246 | return sds.busiest; |
7247 | ||
cc57aa8f | 7248 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 7249 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
7250 | goto out_balanced; |
7251 | ||
ca8ce3d0 NP |
7252 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
7253 | / sds.total_capacity; | |
b0432d8f | 7254 | |
866ab43e PZ |
7255 | /* |
7256 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 7257 | * work because they assume all things are equal, which typically |
866ab43e PZ |
7258 | * isn't true due to cpus_allowed constraints and the like. |
7259 | */ | |
caeb178c | 7260 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
7261 | goto force_balance; |
7262 | ||
cc57aa8f | 7263 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
ea67821b VG |
7264 | if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) && |
7265 | busiest->group_no_capacity) | |
fab47622 NR |
7266 | goto force_balance; |
7267 | ||
cc57aa8f | 7268 | /* |
9c58c79a | 7269 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
7270 | * don't try and pull any tasks. |
7271 | */ | |
56cf515b | 7272 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
7273 | goto out_balanced; |
7274 | ||
cc57aa8f PZ |
7275 | /* |
7276 | * Don't pull any tasks if this group is already above the domain | |
7277 | * average load. | |
7278 | */ | |
56cf515b | 7279 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
7280 | goto out_balanced; |
7281 | ||
bd939f45 | 7282 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 7283 | /* |
43f4d666 VG |
7284 | * This cpu is idle. If the busiest group is not overloaded |
7285 | * and there is no imbalance between this and busiest group | |
7286 | * wrt idle cpus, it is balanced. The imbalance becomes | |
7287 | * significant if the diff is greater than 1 otherwise we | |
7288 | * might end up to just move the imbalance on another group | |
aae6d3dd | 7289 | */ |
43f4d666 VG |
7290 | if ((busiest->group_type != group_overloaded) && |
7291 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 7292 | goto out_balanced; |
c186fafe PZ |
7293 | } else { |
7294 | /* | |
7295 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
7296 | * imbalance_pct to be conservative. | |
7297 | */ | |
56cf515b JK |
7298 | if (100 * busiest->avg_load <= |
7299 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 7300 | goto out_balanced; |
aae6d3dd | 7301 | } |
1e3c88bd | 7302 | |
fab47622 | 7303 | force_balance: |
1e3c88bd | 7304 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 7305 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
7306 | return sds.busiest; |
7307 | ||
7308 | out_balanced: | |
bd939f45 | 7309 | env->imbalance = 0; |
1e3c88bd PZ |
7310 | return NULL; |
7311 | } | |
7312 | ||
7313 | /* | |
7314 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
7315 | */ | |
bd939f45 | 7316 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 7317 | struct sched_group *group) |
1e3c88bd PZ |
7318 | { |
7319 | struct rq *busiest = NULL, *rq; | |
ced549fa | 7320 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
7321 | int i; |
7322 | ||
6906a408 | 7323 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
ea67821b | 7324 | unsigned long capacity, wl; |
0ec8aa00 PZ |
7325 | enum fbq_type rt; |
7326 | ||
7327 | rq = cpu_rq(i); | |
7328 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 7329 | |
0ec8aa00 PZ |
7330 | /* |
7331 | * We classify groups/runqueues into three groups: | |
7332 | * - regular: there are !numa tasks | |
7333 | * - remote: there are numa tasks that run on the 'wrong' node | |
7334 | * - all: there is no distinction | |
7335 | * | |
7336 | * In order to avoid migrating ideally placed numa tasks, | |
7337 | * ignore those when there's better options. | |
7338 | * | |
7339 | * If we ignore the actual busiest queue to migrate another | |
7340 | * task, the next balance pass can still reduce the busiest | |
7341 | * queue by moving tasks around inside the node. | |
7342 | * | |
7343 | * If we cannot move enough load due to this classification | |
7344 | * the next pass will adjust the group classification and | |
7345 | * allow migration of more tasks. | |
7346 | * | |
7347 | * Both cases only affect the total convergence complexity. | |
7348 | */ | |
7349 | if (rt > env->fbq_type) | |
7350 | continue; | |
7351 | ||
ced549fa | 7352 | capacity = capacity_of(i); |
9d5efe05 | 7353 | |
6e40f5bb | 7354 | wl = weighted_cpuload(i); |
1e3c88bd | 7355 | |
6e40f5bb TG |
7356 | /* |
7357 | * When comparing with imbalance, use weighted_cpuload() | |
ced549fa | 7358 | * which is not scaled with the cpu capacity. |
6e40f5bb | 7359 | */ |
ea67821b VG |
7360 | |
7361 | if (rq->nr_running == 1 && wl > env->imbalance && | |
7362 | !check_cpu_capacity(rq, env->sd)) | |
1e3c88bd PZ |
7363 | continue; |
7364 | ||
6e40f5bb TG |
7365 | /* |
7366 | * For the load comparisons with the other cpu's, consider | |
ced549fa NP |
7367 | * the weighted_cpuload() scaled with the cpu capacity, so |
7368 | * that the load can be moved away from the cpu that is | |
7369 | * potentially running at a lower capacity. | |
95a79b80 | 7370 | * |
ced549fa | 7371 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
95a79b80 | 7372 | * multiplication to rid ourselves of the division works out |
ced549fa NP |
7373 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
7374 | * our previous maximum. | |
6e40f5bb | 7375 | */ |
ced549fa | 7376 | if (wl * busiest_capacity > busiest_load * capacity) { |
95a79b80 | 7377 | busiest_load = wl; |
ced549fa | 7378 | busiest_capacity = capacity; |
1e3c88bd PZ |
7379 | busiest = rq; |
7380 | } | |
7381 | } | |
7382 | ||
7383 | return busiest; | |
7384 | } | |
7385 | ||
7386 | /* | |
7387 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
7388 | * so long as it is large enough. | |
7389 | */ | |
7390 | #define MAX_PINNED_INTERVAL 512 | |
7391 | ||
7392 | /* Working cpumask for load_balance and load_balance_newidle. */ | |
e6252c3e | 7393 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
1e3c88bd | 7394 | |
bd939f45 | 7395 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 7396 | { |
bd939f45 PZ |
7397 | struct sched_domain *sd = env->sd; |
7398 | ||
7399 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
7400 | |
7401 | /* | |
7402 | * ASYM_PACKING needs to force migrate tasks from busy but | |
7403 | * higher numbered CPUs in order to pack all tasks in the | |
7404 | * lowest numbered CPUs. | |
7405 | */ | |
bd939f45 | 7406 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
532cb4c4 | 7407 | return 1; |
1af3ed3d PZ |
7408 | } |
7409 | ||
1aaf90a4 VG |
7410 | /* |
7411 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
7412 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
7413 | * because of other sched_class or IRQs if more capacity stays | |
7414 | * available on dst_cpu. | |
7415 | */ | |
7416 | if ((env->idle != CPU_NOT_IDLE) && | |
7417 | (env->src_rq->cfs.h_nr_running == 1)) { | |
7418 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
7419 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
7420 | return 1; | |
7421 | } | |
7422 | ||
1af3ed3d PZ |
7423 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
7424 | } | |
7425 | ||
969c7921 TH |
7426 | static int active_load_balance_cpu_stop(void *data); |
7427 | ||
23f0d209 JK |
7428 | static int should_we_balance(struct lb_env *env) |
7429 | { | |
7430 | struct sched_group *sg = env->sd->groups; | |
7431 | struct cpumask *sg_cpus, *sg_mask; | |
7432 | int cpu, balance_cpu = -1; | |
7433 | ||
7434 | /* | |
7435 | * In the newly idle case, we will allow all the cpu's | |
7436 | * to do the newly idle load balance. | |
7437 | */ | |
7438 | if (env->idle == CPU_NEWLY_IDLE) | |
7439 | return 1; | |
7440 | ||
7441 | sg_cpus = sched_group_cpus(sg); | |
7442 | sg_mask = sched_group_mask(sg); | |
7443 | /* Try to find first idle cpu */ | |
7444 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
7445 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
7446 | continue; | |
7447 | ||
7448 | balance_cpu = cpu; | |
7449 | break; | |
7450 | } | |
7451 | ||
7452 | if (balance_cpu == -1) | |
7453 | balance_cpu = group_balance_cpu(sg); | |
7454 | ||
7455 | /* | |
7456 | * First idle cpu or the first cpu(busiest) in this sched group | |
7457 | * is eligible for doing load balancing at this and above domains. | |
7458 | */ | |
b0cff9d8 | 7459 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
7460 | } |
7461 | ||
1e3c88bd PZ |
7462 | /* |
7463 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
7464 | * tasks if there is an imbalance. | |
7465 | */ | |
7466 | static int load_balance(int this_cpu, struct rq *this_rq, | |
7467 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 7468 | int *continue_balancing) |
1e3c88bd | 7469 | { |
88b8dac0 | 7470 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 7471 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 7472 | struct sched_group *group; |
1e3c88bd PZ |
7473 | struct rq *busiest; |
7474 | unsigned long flags; | |
4ba29684 | 7475 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 7476 | |
8e45cb54 PZ |
7477 | struct lb_env env = { |
7478 | .sd = sd, | |
ddcdf6e7 PZ |
7479 | .dst_cpu = this_cpu, |
7480 | .dst_rq = this_rq, | |
88b8dac0 | 7481 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 7482 | .idle = idle, |
eb95308e | 7483 | .loop_break = sched_nr_migrate_break, |
b9403130 | 7484 | .cpus = cpus, |
0ec8aa00 | 7485 | .fbq_type = all, |
163122b7 | 7486 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
7487 | }; |
7488 | ||
cfc03118 JK |
7489 | /* |
7490 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
7491 | * other cpus in our group | |
7492 | */ | |
e02e60c1 | 7493 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 7494 | env.dst_grpmask = NULL; |
cfc03118 | 7495 | |
1e3c88bd PZ |
7496 | cpumask_copy(cpus, cpu_active_mask); |
7497 | ||
ae92882e | 7498 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
7499 | |
7500 | redo: | |
23f0d209 JK |
7501 | if (!should_we_balance(&env)) { |
7502 | *continue_balancing = 0; | |
1e3c88bd | 7503 | goto out_balanced; |
23f0d209 | 7504 | } |
1e3c88bd | 7505 | |
23f0d209 | 7506 | group = find_busiest_group(&env); |
1e3c88bd | 7507 | if (!group) { |
ae92882e | 7508 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
7509 | goto out_balanced; |
7510 | } | |
7511 | ||
b9403130 | 7512 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 7513 | if (!busiest) { |
ae92882e | 7514 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
7515 | goto out_balanced; |
7516 | } | |
7517 | ||
78feefc5 | 7518 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 7519 | |
ae92882e | 7520 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 7521 | |
1aaf90a4 VG |
7522 | env.src_cpu = busiest->cpu; |
7523 | env.src_rq = busiest; | |
7524 | ||
1e3c88bd PZ |
7525 | ld_moved = 0; |
7526 | if (busiest->nr_running > 1) { | |
7527 | /* | |
7528 | * Attempt to move tasks. If find_busiest_group has found | |
7529 | * an imbalance but busiest->nr_running <= 1, the group is | |
7530 | * still unbalanced. ld_moved simply stays zero, so it is | |
7531 | * correctly treated as an imbalance. | |
7532 | */ | |
8e45cb54 | 7533 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 7534 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 7535 | |
5d6523eb | 7536 | more_balance: |
163122b7 | 7537 | raw_spin_lock_irqsave(&busiest->lock, flags); |
88b8dac0 SV |
7538 | |
7539 | /* | |
7540 | * cur_ld_moved - load moved in current iteration | |
7541 | * ld_moved - cumulative load moved across iterations | |
7542 | */ | |
163122b7 | 7543 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
7544 | |
7545 | /* | |
163122b7 KT |
7546 | * We've detached some tasks from busiest_rq. Every |
7547 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
7548 | * unlock busiest->lock, and we are able to be sure | |
7549 | * that nobody can manipulate the tasks in parallel. | |
7550 | * See task_rq_lock() family for the details. | |
1e3c88bd | 7551 | */ |
163122b7 KT |
7552 | |
7553 | raw_spin_unlock(&busiest->lock); | |
7554 | ||
7555 | if (cur_ld_moved) { | |
7556 | attach_tasks(&env); | |
7557 | ld_moved += cur_ld_moved; | |
7558 | } | |
7559 | ||
1e3c88bd | 7560 | local_irq_restore(flags); |
88b8dac0 | 7561 | |
f1cd0858 JK |
7562 | if (env.flags & LBF_NEED_BREAK) { |
7563 | env.flags &= ~LBF_NEED_BREAK; | |
7564 | goto more_balance; | |
7565 | } | |
7566 | ||
88b8dac0 SV |
7567 | /* |
7568 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
7569 | * us and move them to an alternate dst_cpu in our sched_group | |
7570 | * where they can run. The upper limit on how many times we | |
7571 | * iterate on same src_cpu is dependent on number of cpus in our | |
7572 | * sched_group. | |
7573 | * | |
7574 | * This changes load balance semantics a bit on who can move | |
7575 | * load to a given_cpu. In addition to the given_cpu itself | |
7576 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
7577 | * nohz-idle), we now have balance_cpu in a position to move | |
7578 | * load to given_cpu. In rare situations, this may cause | |
7579 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
7580 | * _independently_ and at _same_ time to move some load to | |
7581 | * given_cpu) causing exceess load to be moved to given_cpu. | |
7582 | * This however should not happen so much in practice and | |
7583 | * moreover subsequent load balance cycles should correct the | |
7584 | * excess load moved. | |
7585 | */ | |
6263322c | 7586 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 7587 | |
7aff2e3a VD |
7588 | /* Prevent to re-select dst_cpu via env's cpus */ |
7589 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
7590 | ||
78feefc5 | 7591 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 7592 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 7593 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
7594 | env.loop = 0; |
7595 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 7596 | |
88b8dac0 SV |
7597 | /* |
7598 | * Go back to "more_balance" rather than "redo" since we | |
7599 | * need to continue with same src_cpu. | |
7600 | */ | |
7601 | goto more_balance; | |
7602 | } | |
1e3c88bd | 7603 | |
6263322c PZ |
7604 | /* |
7605 | * We failed to reach balance because of affinity. | |
7606 | */ | |
7607 | if (sd_parent) { | |
63b2ca30 | 7608 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 7609 | |
afdeee05 | 7610 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 7611 | *group_imbalance = 1; |
6263322c PZ |
7612 | } |
7613 | ||
1e3c88bd | 7614 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 7615 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 7616 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
7617 | if (!cpumask_empty(cpus)) { |
7618 | env.loop = 0; | |
7619 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 7620 | goto redo; |
bbf18b19 | 7621 | } |
afdeee05 | 7622 | goto out_all_pinned; |
1e3c88bd PZ |
7623 | } |
7624 | } | |
7625 | ||
7626 | if (!ld_moved) { | |
ae92882e | 7627 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
7628 | /* |
7629 | * Increment the failure counter only on periodic balance. | |
7630 | * We do not want newidle balance, which can be very | |
7631 | * frequent, pollute the failure counter causing | |
7632 | * excessive cache_hot migrations and active balances. | |
7633 | */ | |
7634 | if (idle != CPU_NEWLY_IDLE) | |
7635 | sd->nr_balance_failed++; | |
1e3c88bd | 7636 | |
bd939f45 | 7637 | if (need_active_balance(&env)) { |
1e3c88bd PZ |
7638 | raw_spin_lock_irqsave(&busiest->lock, flags); |
7639 | ||
969c7921 TH |
7640 | /* don't kick the active_load_balance_cpu_stop, |
7641 | * if the curr task on busiest cpu can't be | |
7642 | * moved to this_cpu | |
1e3c88bd PZ |
7643 | */ |
7644 | if (!cpumask_test_cpu(this_cpu, | |
fa17b507 | 7645 | tsk_cpus_allowed(busiest->curr))) { |
1e3c88bd PZ |
7646 | raw_spin_unlock_irqrestore(&busiest->lock, |
7647 | flags); | |
8e45cb54 | 7648 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
7649 | goto out_one_pinned; |
7650 | } | |
7651 | ||
969c7921 TH |
7652 | /* |
7653 | * ->active_balance synchronizes accesses to | |
7654 | * ->active_balance_work. Once set, it's cleared | |
7655 | * only after active load balance is finished. | |
7656 | */ | |
1e3c88bd PZ |
7657 | if (!busiest->active_balance) { |
7658 | busiest->active_balance = 1; | |
7659 | busiest->push_cpu = this_cpu; | |
7660 | active_balance = 1; | |
7661 | } | |
7662 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 7663 | |
bd939f45 | 7664 | if (active_balance) { |
969c7921 TH |
7665 | stop_one_cpu_nowait(cpu_of(busiest), |
7666 | active_load_balance_cpu_stop, busiest, | |
7667 | &busiest->active_balance_work); | |
bd939f45 | 7668 | } |
1e3c88bd | 7669 | |
d02c0711 | 7670 | /* We've kicked active balancing, force task migration. */ |
1e3c88bd PZ |
7671 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
7672 | } | |
7673 | } else | |
7674 | sd->nr_balance_failed = 0; | |
7675 | ||
7676 | if (likely(!active_balance)) { | |
7677 | /* We were unbalanced, so reset the balancing interval */ | |
7678 | sd->balance_interval = sd->min_interval; | |
7679 | } else { | |
7680 | /* | |
7681 | * If we've begun active balancing, start to back off. This | |
7682 | * case may not be covered by the all_pinned logic if there | |
7683 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 7684 | * detach_tasks). |
1e3c88bd PZ |
7685 | */ |
7686 | if (sd->balance_interval < sd->max_interval) | |
7687 | sd->balance_interval *= 2; | |
7688 | } | |
7689 | ||
1e3c88bd PZ |
7690 | goto out; |
7691 | ||
7692 | out_balanced: | |
afdeee05 VG |
7693 | /* |
7694 | * We reach balance although we may have faced some affinity | |
7695 | * constraints. Clear the imbalance flag if it was set. | |
7696 | */ | |
7697 | if (sd_parent) { | |
7698 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | |
7699 | ||
7700 | if (*group_imbalance) | |
7701 | *group_imbalance = 0; | |
7702 | } | |
7703 | ||
7704 | out_all_pinned: | |
7705 | /* | |
7706 | * We reach balance because all tasks are pinned at this level so | |
7707 | * we can't migrate them. Let the imbalance flag set so parent level | |
7708 | * can try to migrate them. | |
7709 | */ | |
ae92882e | 7710 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
7711 | |
7712 | sd->nr_balance_failed = 0; | |
7713 | ||
7714 | out_one_pinned: | |
7715 | /* tune up the balancing interval */ | |
8e45cb54 | 7716 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 7717 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
7718 | (sd->balance_interval < sd->max_interval)) |
7719 | sd->balance_interval *= 2; | |
7720 | ||
46e49b38 | 7721 | ld_moved = 0; |
1e3c88bd | 7722 | out: |
1e3c88bd PZ |
7723 | return ld_moved; |
7724 | } | |
7725 | ||
52a08ef1 JL |
7726 | static inline unsigned long |
7727 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
7728 | { | |
7729 | unsigned long interval = sd->balance_interval; | |
7730 | ||
7731 | if (cpu_busy) | |
7732 | interval *= sd->busy_factor; | |
7733 | ||
7734 | /* scale ms to jiffies */ | |
7735 | interval = msecs_to_jiffies(interval); | |
7736 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
7737 | ||
7738 | return interval; | |
7739 | } | |
7740 | ||
7741 | static inline void | |
31851a98 | 7742 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
7743 | { |
7744 | unsigned long interval, next; | |
7745 | ||
31851a98 LY |
7746 | /* used by idle balance, so cpu_busy = 0 */ |
7747 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
7748 | next = sd->last_balance + interval; |
7749 | ||
7750 | if (time_after(*next_balance, next)) | |
7751 | *next_balance = next; | |
7752 | } | |
7753 | ||
1e3c88bd PZ |
7754 | /* |
7755 | * idle_balance is called by schedule() if this_cpu is about to become | |
7756 | * idle. Attempts to pull tasks from other CPUs. | |
7757 | */ | |
6e83125c | 7758 | static int idle_balance(struct rq *this_rq) |
1e3c88bd | 7759 | { |
52a08ef1 JL |
7760 | unsigned long next_balance = jiffies + HZ; |
7761 | int this_cpu = this_rq->cpu; | |
1e3c88bd PZ |
7762 | struct sched_domain *sd; |
7763 | int pulled_task = 0; | |
9bd721c5 | 7764 | u64 curr_cost = 0; |
1e3c88bd | 7765 | |
6e83125c PZ |
7766 | /* |
7767 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
7768 | * measure the duration of idle_balance() as idle time. | |
7769 | */ | |
7770 | this_rq->idle_stamp = rq_clock(this_rq); | |
7771 | ||
4486edd1 TC |
7772 | if (this_rq->avg_idle < sysctl_sched_migration_cost || |
7773 | !this_rq->rd->overload) { | |
52a08ef1 JL |
7774 | rcu_read_lock(); |
7775 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
7776 | if (sd) | |
31851a98 | 7777 | update_next_balance(sd, &next_balance); |
52a08ef1 JL |
7778 | rcu_read_unlock(); |
7779 | ||
6e83125c | 7780 | goto out; |
52a08ef1 | 7781 | } |
1e3c88bd | 7782 | |
f492e12e PZ |
7783 | raw_spin_unlock(&this_rq->lock); |
7784 | ||
48a16753 | 7785 | update_blocked_averages(this_cpu); |
dce840a0 | 7786 | rcu_read_lock(); |
1e3c88bd | 7787 | for_each_domain(this_cpu, sd) { |
23f0d209 | 7788 | int continue_balancing = 1; |
9bd721c5 | 7789 | u64 t0, domain_cost; |
1e3c88bd PZ |
7790 | |
7791 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
7792 | continue; | |
7793 | ||
52a08ef1 | 7794 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
31851a98 | 7795 | update_next_balance(sd, &next_balance); |
9bd721c5 | 7796 | break; |
52a08ef1 | 7797 | } |
9bd721c5 | 7798 | |
f492e12e | 7799 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
7800 | t0 = sched_clock_cpu(this_cpu); |
7801 | ||
f492e12e | 7802 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
7803 | sd, CPU_NEWLY_IDLE, |
7804 | &continue_balancing); | |
9bd721c5 JL |
7805 | |
7806 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
7807 | if (domain_cost > sd->max_newidle_lb_cost) | |
7808 | sd->max_newidle_lb_cost = domain_cost; | |
7809 | ||
7810 | curr_cost += domain_cost; | |
f492e12e | 7811 | } |
1e3c88bd | 7812 | |
31851a98 | 7813 | update_next_balance(sd, &next_balance); |
39a4d9ca JL |
7814 | |
7815 | /* | |
7816 | * Stop searching for tasks to pull if there are | |
7817 | * now runnable tasks on this rq. | |
7818 | */ | |
7819 | if (pulled_task || this_rq->nr_running > 0) | |
1e3c88bd | 7820 | break; |
1e3c88bd | 7821 | } |
dce840a0 | 7822 | rcu_read_unlock(); |
f492e12e PZ |
7823 | |
7824 | raw_spin_lock(&this_rq->lock); | |
7825 | ||
0e5b5337 JL |
7826 | if (curr_cost > this_rq->max_idle_balance_cost) |
7827 | this_rq->max_idle_balance_cost = curr_cost; | |
7828 | ||
e5fc6611 | 7829 | /* |
0e5b5337 JL |
7830 | * While browsing the domains, we released the rq lock, a task could |
7831 | * have been enqueued in the meantime. Since we're not going idle, | |
7832 | * pretend we pulled a task. | |
e5fc6611 | 7833 | */ |
0e5b5337 | 7834 | if (this_rq->cfs.h_nr_running && !pulled_task) |
6e83125c | 7835 | pulled_task = 1; |
e5fc6611 | 7836 | |
52a08ef1 JL |
7837 | out: |
7838 | /* Move the next balance forward */ | |
7839 | if (time_after(this_rq->next_balance, next_balance)) | |
1e3c88bd | 7840 | this_rq->next_balance = next_balance; |
9bd721c5 | 7841 | |
e4aa358b | 7842 | /* Is there a task of a high priority class? */ |
46383648 | 7843 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
e4aa358b KT |
7844 | pulled_task = -1; |
7845 | ||
38c6ade2 | 7846 | if (pulled_task) |
6e83125c PZ |
7847 | this_rq->idle_stamp = 0; |
7848 | ||
3c4017c1 | 7849 | return pulled_task; |
1e3c88bd PZ |
7850 | } |
7851 | ||
7852 | /* | |
969c7921 TH |
7853 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
7854 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
7855 | * least 1 task to be running on each physical CPU where possible, and | |
7856 | * avoids physical / logical imbalances. | |
1e3c88bd | 7857 | */ |
969c7921 | 7858 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 7859 | { |
969c7921 TH |
7860 | struct rq *busiest_rq = data; |
7861 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 7862 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 7863 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 7864 | struct sched_domain *sd; |
e5673f28 | 7865 | struct task_struct *p = NULL; |
969c7921 TH |
7866 | |
7867 | raw_spin_lock_irq(&busiest_rq->lock); | |
7868 | ||
7869 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
7870 | if (unlikely(busiest_cpu != smp_processor_id() || | |
7871 | !busiest_rq->active_balance)) | |
7872 | goto out_unlock; | |
1e3c88bd PZ |
7873 | |
7874 | /* Is there any task to move? */ | |
7875 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 7876 | goto out_unlock; |
1e3c88bd PZ |
7877 | |
7878 | /* | |
7879 | * This condition is "impossible", if it occurs | |
7880 | * we need to fix it. Originally reported by | |
7881 | * Bjorn Helgaas on a 128-cpu setup. | |
7882 | */ | |
7883 | BUG_ON(busiest_rq == target_rq); | |
7884 | ||
1e3c88bd | 7885 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 7886 | rcu_read_lock(); |
1e3c88bd PZ |
7887 | for_each_domain(target_cpu, sd) { |
7888 | if ((sd->flags & SD_LOAD_BALANCE) && | |
7889 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
7890 | break; | |
7891 | } | |
7892 | ||
7893 | if (likely(sd)) { | |
8e45cb54 PZ |
7894 | struct lb_env env = { |
7895 | .sd = sd, | |
ddcdf6e7 PZ |
7896 | .dst_cpu = target_cpu, |
7897 | .dst_rq = target_rq, | |
7898 | .src_cpu = busiest_rq->cpu, | |
7899 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
7900 | .idle = CPU_IDLE, |
7901 | }; | |
7902 | ||
ae92882e | 7903 | schedstat_inc(sd->alb_count); |
1e3c88bd | 7904 | |
e5673f28 | 7905 | p = detach_one_task(&env); |
d02c0711 | 7906 | if (p) { |
ae92882e | 7907 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
7908 | /* Active balancing done, reset the failure counter. */ |
7909 | sd->nr_balance_failed = 0; | |
7910 | } else { | |
ae92882e | 7911 | schedstat_inc(sd->alb_failed); |
d02c0711 | 7912 | } |
1e3c88bd | 7913 | } |
dce840a0 | 7914 | rcu_read_unlock(); |
969c7921 TH |
7915 | out_unlock: |
7916 | busiest_rq->active_balance = 0; | |
e5673f28 KT |
7917 | raw_spin_unlock(&busiest_rq->lock); |
7918 | ||
7919 | if (p) | |
7920 | attach_one_task(target_rq, p); | |
7921 | ||
7922 | local_irq_enable(); | |
7923 | ||
969c7921 | 7924 | return 0; |
1e3c88bd PZ |
7925 | } |
7926 | ||
d987fc7f MG |
7927 | static inline int on_null_domain(struct rq *rq) |
7928 | { | |
7929 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
7930 | } | |
7931 | ||
3451d024 | 7932 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
7933 | /* |
7934 | * idle load balancing details | |
83cd4fe2 VP |
7935 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
7936 | * needed, they will kick the idle load balancer, which then does idle | |
7937 | * load balancing for all the idle CPUs. | |
7938 | */ | |
1e3c88bd | 7939 | static struct { |
83cd4fe2 | 7940 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 7941 | atomic_t nr_cpus; |
83cd4fe2 VP |
7942 | unsigned long next_balance; /* in jiffy units */ |
7943 | } nohz ____cacheline_aligned; | |
1e3c88bd | 7944 | |
3dd0337d | 7945 | static inline int find_new_ilb(void) |
1e3c88bd | 7946 | { |
0b005cf5 | 7947 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 7948 | |
786d6dc7 SS |
7949 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
7950 | return ilb; | |
7951 | ||
7952 | return nr_cpu_ids; | |
1e3c88bd | 7953 | } |
1e3c88bd | 7954 | |
83cd4fe2 VP |
7955 | /* |
7956 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
7957 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
7958 | * CPU (if there is one). | |
7959 | */ | |
0aeeeeba | 7960 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
7961 | { |
7962 | int ilb_cpu; | |
7963 | ||
7964 | nohz.next_balance++; | |
7965 | ||
3dd0337d | 7966 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 7967 | |
0b005cf5 SS |
7968 | if (ilb_cpu >= nr_cpu_ids) |
7969 | return; | |
83cd4fe2 | 7970 | |
cd490c5b | 7971 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
7972 | return; |
7973 | /* | |
7974 | * Use smp_send_reschedule() instead of resched_cpu(). | |
7975 | * This way we generate a sched IPI on the target cpu which | |
7976 | * is idle. And the softirq performing nohz idle load balance | |
7977 | * will be run before returning from the IPI. | |
7978 | */ | |
7979 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
7980 | return; |
7981 | } | |
7982 | ||
20a5c8cc | 7983 | void nohz_balance_exit_idle(unsigned int cpu) |
71325960 SS |
7984 | { |
7985 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
d987fc7f MG |
7986 | /* |
7987 | * Completely isolated CPUs don't ever set, so we must test. | |
7988 | */ | |
7989 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | |
7990 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
7991 | atomic_dec(&nohz.nr_cpus); | |
7992 | } | |
71325960 SS |
7993 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
7994 | } | |
7995 | } | |
7996 | ||
69e1e811 SS |
7997 | static inline void set_cpu_sd_state_busy(void) |
7998 | { | |
7999 | struct sched_domain *sd; | |
37dc6b50 | 8000 | int cpu = smp_processor_id(); |
69e1e811 | 8001 | |
69e1e811 | 8002 | rcu_read_lock(); |
37dc6b50 | 8003 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
8004 | |
8005 | if (!sd || !sd->nohz_idle) | |
8006 | goto unlock; | |
8007 | sd->nohz_idle = 0; | |
8008 | ||
63b2ca30 | 8009 | atomic_inc(&sd->groups->sgc->nr_busy_cpus); |
25f55d9d | 8010 | unlock: |
69e1e811 SS |
8011 | rcu_read_unlock(); |
8012 | } | |
8013 | ||
8014 | void set_cpu_sd_state_idle(void) | |
8015 | { | |
8016 | struct sched_domain *sd; | |
37dc6b50 | 8017 | int cpu = smp_processor_id(); |
69e1e811 | 8018 | |
69e1e811 | 8019 | rcu_read_lock(); |
37dc6b50 | 8020 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
25f55d9d VG |
8021 | |
8022 | if (!sd || sd->nohz_idle) | |
8023 | goto unlock; | |
8024 | sd->nohz_idle = 1; | |
8025 | ||
63b2ca30 | 8026 | atomic_dec(&sd->groups->sgc->nr_busy_cpus); |
25f55d9d | 8027 | unlock: |
69e1e811 SS |
8028 | rcu_read_unlock(); |
8029 | } | |
8030 | ||
1e3c88bd | 8031 | /* |
c1cc017c | 8032 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 8033 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 8034 | */ |
c1cc017c | 8035 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 8036 | { |
71325960 SS |
8037 | /* |
8038 | * If this cpu is going down, then nothing needs to be done. | |
8039 | */ | |
8040 | if (!cpu_active(cpu)) | |
8041 | return; | |
8042 | ||
c1cc017c AS |
8043 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
8044 | return; | |
1e3c88bd | 8045 | |
d987fc7f MG |
8046 | /* |
8047 | * If we're a completely isolated CPU, we don't play. | |
8048 | */ | |
8049 | if (on_null_domain(cpu_rq(cpu))) | |
8050 | return; | |
8051 | ||
c1cc017c AS |
8052 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
8053 | atomic_inc(&nohz.nr_cpus); | |
8054 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd PZ |
8055 | } |
8056 | #endif | |
8057 | ||
8058 | static DEFINE_SPINLOCK(balancing); | |
8059 | ||
49c022e6 PZ |
8060 | /* |
8061 | * Scale the max load_balance interval with the number of CPUs in the system. | |
8062 | * This trades load-balance latency on larger machines for less cross talk. | |
8063 | */ | |
029632fb | 8064 | void update_max_interval(void) |
49c022e6 PZ |
8065 | { |
8066 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
8067 | } | |
8068 | ||
1e3c88bd PZ |
8069 | /* |
8070 | * It checks each scheduling domain to see if it is due to be balanced, | |
8071 | * and initiates a balancing operation if so. | |
8072 | * | |
b9b0853a | 8073 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 8074 | */ |
f7ed0a89 | 8075 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 8076 | { |
23f0d209 | 8077 | int continue_balancing = 1; |
f7ed0a89 | 8078 | int cpu = rq->cpu; |
1e3c88bd | 8079 | unsigned long interval; |
04f733b4 | 8080 | struct sched_domain *sd; |
1e3c88bd PZ |
8081 | /* Earliest time when we have to do rebalance again */ |
8082 | unsigned long next_balance = jiffies + 60*HZ; | |
8083 | int update_next_balance = 0; | |
f48627e6 JL |
8084 | int need_serialize, need_decay = 0; |
8085 | u64 max_cost = 0; | |
1e3c88bd | 8086 | |
48a16753 | 8087 | update_blocked_averages(cpu); |
2069dd75 | 8088 | |
dce840a0 | 8089 | rcu_read_lock(); |
1e3c88bd | 8090 | for_each_domain(cpu, sd) { |
f48627e6 JL |
8091 | /* |
8092 | * Decay the newidle max times here because this is a regular | |
8093 | * visit to all the domains. Decay ~1% per second. | |
8094 | */ | |
8095 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
8096 | sd->max_newidle_lb_cost = | |
8097 | (sd->max_newidle_lb_cost * 253) / 256; | |
8098 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
8099 | need_decay = 1; | |
8100 | } | |
8101 | max_cost += sd->max_newidle_lb_cost; | |
8102 | ||
1e3c88bd PZ |
8103 | if (!(sd->flags & SD_LOAD_BALANCE)) |
8104 | continue; | |
8105 | ||
f48627e6 JL |
8106 | /* |
8107 | * Stop the load balance at this level. There is another | |
8108 | * CPU in our sched group which is doing load balancing more | |
8109 | * actively. | |
8110 | */ | |
8111 | if (!continue_balancing) { | |
8112 | if (need_decay) | |
8113 | continue; | |
8114 | break; | |
8115 | } | |
8116 | ||
52a08ef1 | 8117 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
8118 | |
8119 | need_serialize = sd->flags & SD_SERIALIZE; | |
1e3c88bd PZ |
8120 | if (need_serialize) { |
8121 | if (!spin_trylock(&balancing)) | |
8122 | goto out; | |
8123 | } | |
8124 | ||
8125 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 8126 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 8127 | /* |
6263322c | 8128 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
8129 | * env->dst_cpu, so we can't know our idle |
8130 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 8131 | */ |
de5eb2dd | 8132 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
8133 | } |
8134 | sd->last_balance = jiffies; | |
52a08ef1 | 8135 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
8136 | } |
8137 | if (need_serialize) | |
8138 | spin_unlock(&balancing); | |
8139 | out: | |
8140 | if (time_after(next_balance, sd->last_balance + interval)) { | |
8141 | next_balance = sd->last_balance + interval; | |
8142 | update_next_balance = 1; | |
8143 | } | |
f48627e6 JL |
8144 | } |
8145 | if (need_decay) { | |
1e3c88bd | 8146 | /* |
f48627e6 JL |
8147 | * Ensure the rq-wide value also decays but keep it at a |
8148 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 8149 | */ |
f48627e6 JL |
8150 | rq->max_idle_balance_cost = |
8151 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 8152 | } |
dce840a0 | 8153 | rcu_read_unlock(); |
1e3c88bd PZ |
8154 | |
8155 | /* | |
8156 | * next_balance will be updated only when there is a need. | |
8157 | * When the cpu is attached to null domain for ex, it will not be | |
8158 | * updated. | |
8159 | */ | |
c5afb6a8 | 8160 | if (likely(update_next_balance)) { |
1e3c88bd | 8161 | rq->next_balance = next_balance; |
c5afb6a8 VG |
8162 | |
8163 | #ifdef CONFIG_NO_HZ_COMMON | |
8164 | /* | |
8165 | * If this CPU has been elected to perform the nohz idle | |
8166 | * balance. Other idle CPUs have already rebalanced with | |
8167 | * nohz_idle_balance() and nohz.next_balance has been | |
8168 | * updated accordingly. This CPU is now running the idle load | |
8169 | * balance for itself and we need to update the | |
8170 | * nohz.next_balance accordingly. | |
8171 | */ | |
8172 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
8173 | nohz.next_balance = rq->next_balance; | |
8174 | #endif | |
8175 | } | |
1e3c88bd PZ |
8176 | } |
8177 | ||
3451d024 | 8178 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 8179 | /* |
3451d024 | 8180 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
8181 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
8182 | */ | |
208cb16b | 8183 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 8184 | { |
208cb16b | 8185 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
8186 | struct rq *rq; |
8187 | int balance_cpu; | |
c5afb6a8 VG |
8188 | /* Earliest time when we have to do rebalance again */ |
8189 | unsigned long next_balance = jiffies + 60*HZ; | |
8190 | int update_next_balance = 0; | |
83cd4fe2 | 8191 | |
1c792db7 SS |
8192 | if (idle != CPU_IDLE || |
8193 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
8194 | goto end; | |
83cd4fe2 VP |
8195 | |
8196 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 8197 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
8198 | continue; |
8199 | ||
8200 | /* | |
8201 | * If this cpu gets work to do, stop the load balancing | |
8202 | * work being done for other cpus. Next load | |
8203 | * balancing owner will pick it up. | |
8204 | */ | |
1c792db7 | 8205 | if (need_resched()) |
83cd4fe2 | 8206 | break; |
83cd4fe2 | 8207 | |
5ed4f1d9 VG |
8208 | rq = cpu_rq(balance_cpu); |
8209 | ||
ed61bbc6 TC |
8210 | /* |
8211 | * If time for next balance is due, | |
8212 | * do the balance. | |
8213 | */ | |
8214 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8215 | raw_spin_lock_irq(&rq->lock); | |
8216 | update_rq_clock(rq); | |
cee1afce | 8217 | cpu_load_update_idle(rq); |
ed61bbc6 TC |
8218 | raw_spin_unlock_irq(&rq->lock); |
8219 | rebalance_domains(rq, CPU_IDLE); | |
8220 | } | |
83cd4fe2 | 8221 | |
c5afb6a8 VG |
8222 | if (time_after(next_balance, rq->next_balance)) { |
8223 | next_balance = rq->next_balance; | |
8224 | update_next_balance = 1; | |
8225 | } | |
83cd4fe2 | 8226 | } |
c5afb6a8 VG |
8227 | |
8228 | /* | |
8229 | * next_balance will be updated only when there is a need. | |
8230 | * When the CPU is attached to null domain for ex, it will not be | |
8231 | * updated. | |
8232 | */ | |
8233 | if (likely(update_next_balance)) | |
8234 | nohz.next_balance = next_balance; | |
1c792db7 SS |
8235 | end: |
8236 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
8237 | } |
8238 | ||
8239 | /* | |
0b005cf5 | 8240 | * Current heuristic for kicking the idle load balancer in the presence |
1aaf90a4 | 8241 | * of an idle cpu in the system. |
0b005cf5 | 8242 | * - This rq has more than one task. |
1aaf90a4 VG |
8243 | * - This rq has at least one CFS task and the capacity of the CPU is |
8244 | * significantly reduced because of RT tasks or IRQs. | |
8245 | * - At parent of LLC scheduler domain level, this cpu's scheduler group has | |
8246 | * multiple busy cpu. | |
0b005cf5 SS |
8247 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler |
8248 | * domain span are idle. | |
83cd4fe2 | 8249 | */ |
1aaf90a4 | 8250 | static inline bool nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
8251 | { |
8252 | unsigned long now = jiffies; | |
0b005cf5 | 8253 | struct sched_domain *sd; |
63b2ca30 | 8254 | struct sched_group_capacity *sgc; |
4a725627 | 8255 | int nr_busy, cpu = rq->cpu; |
1aaf90a4 | 8256 | bool kick = false; |
83cd4fe2 | 8257 | |
4a725627 | 8258 | if (unlikely(rq->idle_balance)) |
1aaf90a4 | 8259 | return false; |
83cd4fe2 | 8260 | |
1c792db7 SS |
8261 | /* |
8262 | * We may be recently in ticked or tickless idle mode. At the first | |
8263 | * busy tick after returning from idle, we will update the busy stats. | |
8264 | */ | |
69e1e811 | 8265 | set_cpu_sd_state_busy(); |
c1cc017c | 8266 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
8267 | |
8268 | /* | |
8269 | * None are in tickless mode and hence no need for NOHZ idle load | |
8270 | * balancing. | |
8271 | */ | |
8272 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
1aaf90a4 | 8273 | return false; |
1c792db7 SS |
8274 | |
8275 | if (time_before(now, nohz.next_balance)) | |
1aaf90a4 | 8276 | return false; |
83cd4fe2 | 8277 | |
0b005cf5 | 8278 | if (rq->nr_running >= 2) |
1aaf90a4 | 8279 | return true; |
83cd4fe2 | 8280 | |
067491b7 | 8281 | rcu_read_lock(); |
37dc6b50 | 8282 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
37dc6b50 | 8283 | if (sd) { |
63b2ca30 NP |
8284 | sgc = sd->groups->sgc; |
8285 | nr_busy = atomic_read(&sgc->nr_busy_cpus); | |
0b005cf5 | 8286 | |
1aaf90a4 VG |
8287 | if (nr_busy > 1) { |
8288 | kick = true; | |
8289 | goto unlock; | |
8290 | } | |
8291 | ||
83cd4fe2 | 8292 | } |
37dc6b50 | 8293 | |
1aaf90a4 VG |
8294 | sd = rcu_dereference(rq->sd); |
8295 | if (sd) { | |
8296 | if ((rq->cfs.h_nr_running >= 1) && | |
8297 | check_cpu_capacity(rq, sd)) { | |
8298 | kick = true; | |
8299 | goto unlock; | |
8300 | } | |
8301 | } | |
37dc6b50 | 8302 | |
1aaf90a4 | 8303 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); |
37dc6b50 | 8304 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, |
1aaf90a4 VG |
8305 | sched_domain_span(sd)) < cpu)) { |
8306 | kick = true; | |
8307 | goto unlock; | |
8308 | } | |
067491b7 | 8309 | |
1aaf90a4 | 8310 | unlock: |
067491b7 | 8311 | rcu_read_unlock(); |
1aaf90a4 | 8312 | return kick; |
83cd4fe2 VP |
8313 | } |
8314 | #else | |
208cb16b | 8315 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
8316 | #endif |
8317 | ||
8318 | /* | |
8319 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
8320 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
8321 | */ | |
09dd109d | 8322 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 8323 | { |
208cb16b | 8324 | struct rq *this_rq = this_rq(); |
6eb57e0d | 8325 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
8326 | CPU_IDLE : CPU_NOT_IDLE; |
8327 | ||
1e3c88bd | 8328 | /* |
83cd4fe2 | 8329 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd | 8330 | * balancing on behalf of the other idle cpus whose ticks are |
d4573c3e PM |
8331 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
8332 | * give the idle cpus a chance to load balance. Else we may | |
8333 | * load balance only within the local sched_domain hierarchy | |
8334 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 8335 | */ |
208cb16b | 8336 | nohz_idle_balance(this_rq, idle); |
d4573c3e | 8337 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
8338 | } |
8339 | ||
1e3c88bd PZ |
8340 | /* |
8341 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 8342 | */ |
7caff66f | 8343 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 8344 | { |
1e3c88bd | 8345 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
8346 | if (unlikely(on_null_domain(rq))) |
8347 | return; | |
8348 | ||
8349 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 8350 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 8351 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 8352 | if (nohz_kick_needed(rq)) |
0aeeeeba | 8353 | nohz_balancer_kick(); |
83cd4fe2 | 8354 | #endif |
1e3c88bd PZ |
8355 | } |
8356 | ||
0bcdcf28 CE |
8357 | static void rq_online_fair(struct rq *rq) |
8358 | { | |
8359 | update_sysctl(); | |
0e59bdae KT |
8360 | |
8361 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
8362 | } |
8363 | ||
8364 | static void rq_offline_fair(struct rq *rq) | |
8365 | { | |
8366 | update_sysctl(); | |
a4c96ae3 PB |
8367 | |
8368 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
8369 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
8370 | } |
8371 | ||
55e12e5e | 8372 | #endif /* CONFIG_SMP */ |
e1d1484f | 8373 | |
bf0f6f24 IM |
8374 | /* |
8375 | * scheduler tick hitting a task of our scheduling class: | |
8376 | */ | |
8f4d37ec | 8377 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
8378 | { |
8379 | struct cfs_rq *cfs_rq; | |
8380 | struct sched_entity *se = &curr->se; | |
8381 | ||
8382 | for_each_sched_entity(se) { | |
8383 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 8384 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 8385 | } |
18bf2805 | 8386 | |
b52da86e | 8387 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 8388 | task_tick_numa(rq, curr); |
bf0f6f24 IM |
8389 | } |
8390 | ||
8391 | /* | |
cd29fe6f PZ |
8392 | * called on fork with the child task as argument from the parent's context |
8393 | * - child not yet on the tasklist | |
8394 | * - preemption disabled | |
bf0f6f24 | 8395 | */ |
cd29fe6f | 8396 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 8397 | { |
4fc420c9 DN |
8398 | struct cfs_rq *cfs_rq; |
8399 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 8400 | struct rq *rq = this_rq(); |
bf0f6f24 | 8401 | |
e210bffd | 8402 | raw_spin_lock(&rq->lock); |
861d034e PZ |
8403 | update_rq_clock(rq); |
8404 | ||
4fc420c9 DN |
8405 | cfs_rq = task_cfs_rq(current); |
8406 | curr = cfs_rq->curr; | |
e210bffd PZ |
8407 | if (curr) { |
8408 | update_curr(cfs_rq); | |
b5d9d734 | 8409 | se->vruntime = curr->vruntime; |
e210bffd | 8410 | } |
aeb73b04 | 8411 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 8412 | |
cd29fe6f | 8413 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 8414 | /* |
edcb60a3 IM |
8415 | * Upon rescheduling, sched_class::put_prev_task() will place |
8416 | * 'current' within the tree based on its new key value. | |
8417 | */ | |
4d78e7b6 | 8418 | swap(curr->vruntime, se->vruntime); |
8875125e | 8419 | resched_curr(rq); |
4d78e7b6 | 8420 | } |
bf0f6f24 | 8421 | |
88ec22d3 | 8422 | se->vruntime -= cfs_rq->min_vruntime; |
e210bffd | 8423 | raw_spin_unlock(&rq->lock); |
bf0f6f24 IM |
8424 | } |
8425 | ||
cb469845 SR |
8426 | /* |
8427 | * Priority of the task has changed. Check to see if we preempt | |
8428 | * the current task. | |
8429 | */ | |
da7a735e PZ |
8430 | static void |
8431 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 8432 | { |
da0c1e65 | 8433 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
8434 | return; |
8435 | ||
cb469845 SR |
8436 | /* |
8437 | * Reschedule if we are currently running on this runqueue and | |
8438 | * our priority decreased, or if we are not currently running on | |
8439 | * this runqueue and our priority is higher than the current's | |
8440 | */ | |
da7a735e | 8441 | if (rq->curr == p) { |
cb469845 | 8442 | if (p->prio > oldprio) |
8875125e | 8443 | resched_curr(rq); |
cb469845 | 8444 | } else |
15afe09b | 8445 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
8446 | } |
8447 | ||
daa59407 | 8448 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
8449 | { |
8450 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
8451 | |
8452 | /* | |
daa59407 BP |
8453 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
8454 | * the dequeue_entity(.flags=0) will already have normalized the | |
8455 | * vruntime. | |
8456 | */ | |
8457 | if (p->on_rq) | |
8458 | return true; | |
8459 | ||
8460 | /* | |
8461 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
8462 | * But there are some cases where it has already been normalized: | |
da7a735e | 8463 | * |
daa59407 BP |
8464 | * - A forked child which is waiting for being woken up by |
8465 | * wake_up_new_task(). | |
8466 | * - A task which has been woken up by try_to_wake_up() and | |
8467 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 8468 | */ |
daa59407 BP |
8469 | if (!se->sum_exec_runtime || p->state == TASK_WAKING) |
8470 | return true; | |
8471 | ||
8472 | return false; | |
8473 | } | |
8474 | ||
8475 | static void detach_task_cfs_rq(struct task_struct *p) | |
8476 | { | |
8477 | struct sched_entity *se = &p->se; | |
8478 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
01011473 | 8479 | u64 now = cfs_rq_clock_task(cfs_rq); |
daa59407 BP |
8480 | |
8481 | if (!vruntime_normalized(p)) { | |
da7a735e PZ |
8482 | /* |
8483 | * Fix up our vruntime so that the current sleep doesn't | |
8484 | * cause 'unlimited' sleep bonus. | |
8485 | */ | |
8486 | place_entity(cfs_rq, se, 0); | |
8487 | se->vruntime -= cfs_rq->min_vruntime; | |
8488 | } | |
9ee474f5 | 8489 | |
9d89c257 | 8490 | /* Catch up with the cfs_rq and remove our load when we leave */ |
7c3edd2c | 8491 | update_cfs_rq_load_avg(now, cfs_rq, false); |
a05e8c51 | 8492 | detach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 8493 | update_tg_load_avg(cfs_rq, false); |
da7a735e PZ |
8494 | } |
8495 | ||
daa59407 | 8496 | static void attach_task_cfs_rq(struct task_struct *p) |
cb469845 | 8497 | { |
f36c019c | 8498 | struct sched_entity *se = &p->se; |
daa59407 | 8499 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
01011473 | 8500 | u64 now = cfs_rq_clock_task(cfs_rq); |
7855a35a BP |
8501 | |
8502 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
8503 | /* |
8504 | * Since the real-depth could have been changed (only FAIR | |
8505 | * class maintain depth value), reset depth properly. | |
8506 | */ | |
8507 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
8508 | #endif | |
7855a35a | 8509 | |
6efdb105 | 8510 | /* Synchronize task with its cfs_rq */ |
7c3edd2c | 8511 | update_cfs_rq_load_avg(now, cfs_rq, false); |
daa59407 | 8512 | attach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 8513 | update_tg_load_avg(cfs_rq, false); |
daa59407 BP |
8514 | |
8515 | if (!vruntime_normalized(p)) | |
8516 | se->vruntime += cfs_rq->min_vruntime; | |
8517 | } | |
6efdb105 | 8518 | |
daa59407 BP |
8519 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
8520 | { | |
8521 | detach_task_cfs_rq(p); | |
8522 | } | |
8523 | ||
8524 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
8525 | { | |
8526 | attach_task_cfs_rq(p); | |
7855a35a | 8527 | |
daa59407 | 8528 | if (task_on_rq_queued(p)) { |
7855a35a | 8529 | /* |
daa59407 BP |
8530 | * We were most likely switched from sched_rt, so |
8531 | * kick off the schedule if running, otherwise just see | |
8532 | * if we can still preempt the current task. | |
7855a35a | 8533 | */ |
daa59407 BP |
8534 | if (rq->curr == p) |
8535 | resched_curr(rq); | |
8536 | else | |
8537 | check_preempt_curr(rq, p, 0); | |
7855a35a | 8538 | } |
cb469845 SR |
8539 | } |
8540 | ||
83b699ed SV |
8541 | /* Account for a task changing its policy or group. |
8542 | * | |
8543 | * This routine is mostly called to set cfs_rq->curr field when a task | |
8544 | * migrates between groups/classes. | |
8545 | */ | |
8546 | static void set_curr_task_fair(struct rq *rq) | |
8547 | { | |
8548 | struct sched_entity *se = &rq->curr->se; | |
8549 | ||
ec12cb7f PT |
8550 | for_each_sched_entity(se) { |
8551 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
8552 | ||
8553 | set_next_entity(cfs_rq, se); | |
8554 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
8555 | account_cfs_rq_runtime(cfs_rq, 0); | |
8556 | } | |
83b699ed SV |
8557 | } |
8558 | ||
029632fb PZ |
8559 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
8560 | { | |
8561 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
8562 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
8563 | #ifndef CONFIG_64BIT | |
8564 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
8565 | #endif | |
141965c7 | 8566 | #ifdef CONFIG_SMP |
9d89c257 YD |
8567 | atomic_long_set(&cfs_rq->removed_load_avg, 0); |
8568 | atomic_long_set(&cfs_rq->removed_util_avg, 0); | |
9ee474f5 | 8569 | #endif |
029632fb PZ |
8570 | } |
8571 | ||
810b3817 | 8572 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
8573 | static void task_set_group_fair(struct task_struct *p) |
8574 | { | |
8575 | struct sched_entity *se = &p->se; | |
8576 | ||
8577 | set_task_rq(p, task_cpu(p)); | |
8578 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
8579 | } | |
8580 | ||
bc54da21 | 8581 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 8582 | { |
daa59407 | 8583 | detach_task_cfs_rq(p); |
b2b5ce02 | 8584 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
8585 | |
8586 | #ifdef CONFIG_SMP | |
8587 | /* Tell se's cfs_rq has been changed -- migrated */ | |
8588 | p->se.avg.last_update_time = 0; | |
8589 | #endif | |
daa59407 | 8590 | attach_task_cfs_rq(p); |
810b3817 | 8591 | } |
029632fb | 8592 | |
ea86cb4b VG |
8593 | static void task_change_group_fair(struct task_struct *p, int type) |
8594 | { | |
8595 | switch (type) { | |
8596 | case TASK_SET_GROUP: | |
8597 | task_set_group_fair(p); | |
8598 | break; | |
8599 | ||
8600 | case TASK_MOVE_GROUP: | |
8601 | task_move_group_fair(p); | |
8602 | break; | |
8603 | } | |
8604 | } | |
8605 | ||
029632fb PZ |
8606 | void free_fair_sched_group(struct task_group *tg) |
8607 | { | |
8608 | int i; | |
8609 | ||
8610 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
8611 | ||
8612 | for_each_possible_cpu(i) { | |
8613 | if (tg->cfs_rq) | |
8614 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 8615 | if (tg->se) |
029632fb PZ |
8616 | kfree(tg->se[i]); |
8617 | } | |
8618 | ||
8619 | kfree(tg->cfs_rq); | |
8620 | kfree(tg->se); | |
8621 | } | |
8622 | ||
8623 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8624 | { | |
029632fb | 8625 | struct sched_entity *se; |
b7fa30c9 PZ |
8626 | struct cfs_rq *cfs_rq; |
8627 | struct rq *rq; | |
029632fb PZ |
8628 | int i; |
8629 | ||
8630 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
8631 | if (!tg->cfs_rq) | |
8632 | goto err; | |
8633 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
8634 | if (!tg->se) | |
8635 | goto err; | |
8636 | ||
8637 | tg->shares = NICE_0_LOAD; | |
8638 | ||
8639 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
8640 | ||
8641 | for_each_possible_cpu(i) { | |
b7fa30c9 PZ |
8642 | rq = cpu_rq(i); |
8643 | ||
029632fb PZ |
8644 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), |
8645 | GFP_KERNEL, cpu_to_node(i)); | |
8646 | if (!cfs_rq) | |
8647 | goto err; | |
8648 | ||
8649 | se = kzalloc_node(sizeof(struct sched_entity), | |
8650 | GFP_KERNEL, cpu_to_node(i)); | |
8651 | if (!se) | |
8652 | goto err_free_rq; | |
8653 | ||
8654 | init_cfs_rq(cfs_rq); | |
8655 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 8656 | init_entity_runnable_average(se); |
029632fb PZ |
8657 | } |
8658 | ||
8659 | return 1; | |
8660 | ||
8661 | err_free_rq: | |
8662 | kfree(cfs_rq); | |
8663 | err: | |
8664 | return 0; | |
8665 | } | |
8666 | ||
8663e24d PZ |
8667 | void online_fair_sched_group(struct task_group *tg) |
8668 | { | |
8669 | struct sched_entity *se; | |
8670 | struct rq *rq; | |
8671 | int i; | |
8672 | ||
8673 | for_each_possible_cpu(i) { | |
8674 | rq = cpu_rq(i); | |
8675 | se = tg->se[i]; | |
8676 | ||
8677 | raw_spin_lock_irq(&rq->lock); | |
8678 | post_init_entity_util_avg(se); | |
55e16d30 | 8679 | sync_throttle(tg, i); |
8663e24d PZ |
8680 | raw_spin_unlock_irq(&rq->lock); |
8681 | } | |
8682 | } | |
8683 | ||
6fe1f348 | 8684 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 8685 | { |
029632fb | 8686 | unsigned long flags; |
6fe1f348 PZ |
8687 | struct rq *rq; |
8688 | int cpu; | |
029632fb | 8689 | |
6fe1f348 PZ |
8690 | for_each_possible_cpu(cpu) { |
8691 | if (tg->se[cpu]) | |
8692 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 8693 | |
6fe1f348 PZ |
8694 | /* |
8695 | * Only empty task groups can be destroyed; so we can speculatively | |
8696 | * check on_list without danger of it being re-added. | |
8697 | */ | |
8698 | if (!tg->cfs_rq[cpu]->on_list) | |
8699 | continue; | |
8700 | ||
8701 | rq = cpu_rq(cpu); | |
8702 | ||
8703 | raw_spin_lock_irqsave(&rq->lock, flags); | |
8704 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
8705 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8706 | } | |
029632fb PZ |
8707 | } |
8708 | ||
8709 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
8710 | struct sched_entity *se, int cpu, | |
8711 | struct sched_entity *parent) | |
8712 | { | |
8713 | struct rq *rq = cpu_rq(cpu); | |
8714 | ||
8715 | cfs_rq->tg = tg; | |
8716 | cfs_rq->rq = rq; | |
029632fb PZ |
8717 | init_cfs_rq_runtime(cfs_rq); |
8718 | ||
8719 | tg->cfs_rq[cpu] = cfs_rq; | |
8720 | tg->se[cpu] = se; | |
8721 | ||
8722 | /* se could be NULL for root_task_group */ | |
8723 | if (!se) | |
8724 | return; | |
8725 | ||
fed14d45 | 8726 | if (!parent) { |
029632fb | 8727 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
8728 | se->depth = 0; |
8729 | } else { | |
029632fb | 8730 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
8731 | se->depth = parent->depth + 1; |
8732 | } | |
029632fb PZ |
8733 | |
8734 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
8735 | /* guarantee group entities always have weight */ |
8736 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
8737 | se->parent = parent; |
8738 | } | |
8739 | ||
8740 | static DEFINE_MUTEX(shares_mutex); | |
8741 | ||
8742 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
8743 | { | |
8744 | int i; | |
8745 | unsigned long flags; | |
8746 | ||
8747 | /* | |
8748 | * We can't change the weight of the root cgroup. | |
8749 | */ | |
8750 | if (!tg->se[0]) | |
8751 | return -EINVAL; | |
8752 | ||
8753 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
8754 | ||
8755 | mutex_lock(&shares_mutex); | |
8756 | if (tg->shares == shares) | |
8757 | goto done; | |
8758 | ||
8759 | tg->shares = shares; | |
8760 | for_each_possible_cpu(i) { | |
8761 | struct rq *rq = cpu_rq(i); | |
8762 | struct sched_entity *se; | |
8763 | ||
8764 | se = tg->se[i]; | |
8765 | /* Propagate contribution to hierarchy */ | |
8766 | raw_spin_lock_irqsave(&rq->lock, flags); | |
71b1da46 FW |
8767 | |
8768 | /* Possible calls to update_curr() need rq clock */ | |
8769 | update_rq_clock(rq); | |
17bc14b7 | 8770 | for_each_sched_entity(se) |
029632fb PZ |
8771 | update_cfs_shares(group_cfs_rq(se)); |
8772 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
8773 | } | |
8774 | ||
8775 | done: | |
8776 | mutex_unlock(&shares_mutex); | |
8777 | return 0; | |
8778 | } | |
8779 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
8780 | ||
8781 | void free_fair_sched_group(struct task_group *tg) { } | |
8782 | ||
8783 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
8784 | { | |
8785 | return 1; | |
8786 | } | |
8787 | ||
8663e24d PZ |
8788 | void online_fair_sched_group(struct task_group *tg) { } |
8789 | ||
6fe1f348 | 8790 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
8791 | |
8792 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
8793 | ||
810b3817 | 8794 | |
6d686f45 | 8795 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
8796 | { |
8797 | struct sched_entity *se = &task->se; | |
0d721cea PW |
8798 | unsigned int rr_interval = 0; |
8799 | ||
8800 | /* | |
8801 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
8802 | * idle runqueue: | |
8803 | */ | |
0d721cea | 8804 | if (rq->cfs.load.weight) |
a59f4e07 | 8805 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
8806 | |
8807 | return rr_interval; | |
8808 | } | |
8809 | ||
bf0f6f24 IM |
8810 | /* |
8811 | * All the scheduling class methods: | |
8812 | */ | |
029632fb | 8813 | const struct sched_class fair_sched_class = { |
5522d5d5 | 8814 | .next = &idle_sched_class, |
bf0f6f24 IM |
8815 | .enqueue_task = enqueue_task_fair, |
8816 | .dequeue_task = dequeue_task_fair, | |
8817 | .yield_task = yield_task_fair, | |
d95f4122 | 8818 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 8819 | |
2e09bf55 | 8820 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
8821 | |
8822 | .pick_next_task = pick_next_task_fair, | |
8823 | .put_prev_task = put_prev_task_fair, | |
8824 | ||
681f3e68 | 8825 | #ifdef CONFIG_SMP |
4ce72a2c | 8826 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 8827 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 8828 | |
0bcdcf28 CE |
8829 | .rq_online = rq_online_fair, |
8830 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 8831 | |
12695578 | 8832 | .task_dead = task_dead_fair, |
c5b28038 | 8833 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 8834 | #endif |
bf0f6f24 | 8835 | |
83b699ed | 8836 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 8837 | .task_tick = task_tick_fair, |
cd29fe6f | 8838 | .task_fork = task_fork_fair, |
cb469845 SR |
8839 | |
8840 | .prio_changed = prio_changed_fair, | |
da7a735e | 8841 | .switched_from = switched_from_fair, |
cb469845 | 8842 | .switched_to = switched_to_fair, |
810b3817 | 8843 | |
0d721cea PW |
8844 | .get_rr_interval = get_rr_interval_fair, |
8845 | ||
6e998916 SG |
8846 | .update_curr = update_curr_fair, |
8847 | ||
810b3817 | 8848 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 8849 | .task_change_group = task_change_group_fair, |
810b3817 | 8850 | #endif |
bf0f6f24 IM |
8851 | }; |
8852 | ||
8853 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 8854 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 8855 | { |
bf0f6f24 IM |
8856 | struct cfs_rq *cfs_rq; |
8857 | ||
5973e5b9 | 8858 | rcu_read_lock(); |
c3b64f1e | 8859 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 8860 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 8861 | rcu_read_unlock(); |
bf0f6f24 | 8862 | } |
397f2378 SD |
8863 | |
8864 | #ifdef CONFIG_NUMA_BALANCING | |
8865 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
8866 | { | |
8867 | int node; | |
8868 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
8869 | ||
8870 | for_each_online_node(node) { | |
8871 | if (p->numa_faults) { | |
8872 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
8873 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
8874 | } | |
8875 | if (p->numa_group) { | |
8876 | gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
8877 | gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
8878 | } | |
8879 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
8880 | } | |
8881 | } | |
8882 | #endif /* CONFIG_NUMA_BALANCING */ | |
8883 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
8884 | |
8885 | __init void init_sched_fair_class(void) | |
8886 | { | |
8887 | #ifdef CONFIG_SMP | |
8888 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
8889 | ||
3451d024 | 8890 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 8891 | nohz.next_balance = jiffies; |
029632fb | 8892 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
8893 | #endif |
8894 | #endif /* SMP */ | |
8895 | ||
8896 | } |