Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
bf0f6f24 IM |
2 | /* |
3 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
4 | * | |
5 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
6 | * | |
7 | * Interactivity improvements by Mike Galbraith | |
8 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
9 | * | |
10 | * Various enhancements by Dmitry Adamushko. | |
11 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
12 | * | |
13 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
14 | * Copyright IBM Corporation, 2007 | |
15 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
16 | * | |
17 | * Scaled math optimizations by Thomas Gleixner | |
18 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
19 | * |
20 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
90eec103 | 21 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra |
bf0f6f24 | 22 | */ |
325ea10c | 23 | #include "sched.h" |
029632fb | 24 | |
bf0f6f24 | 25 | /* |
21805085 | 26 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 27 | * |
21805085 | 28 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
29 | * 'timeslice length' - timeslices in CFS are of variable length |
30 | * and have no persistent notion like in traditional, time-slice | |
31 | * based scheduling concepts. | |
bf0f6f24 | 32 | * |
d274a4ce IM |
33 | * (to see the precise effective timeslice length of your workload, |
34 | * run vmstat and monitor the context-switches (cs) field) | |
2b4d5b25 IM |
35 | * |
36 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 37 | */ |
2b4d5b25 | 38 | unsigned int sysctl_sched_latency = 6000000ULL; |
ed8885a1 | 39 | static unsigned int normalized_sysctl_sched_latency = 6000000ULL; |
2bd8e6d4 | 40 | |
1983a922 CE |
41 | /* |
42 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
43 | * |
44 | * Options are: | |
2b4d5b25 IM |
45 | * |
46 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
47 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
48 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
49 | * | |
50 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 51 | */ |
8a99b683 | 52 | unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 53 | |
2bd8e6d4 | 54 | /* |
b2be5e96 | 55 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 56 | * |
864616ee | 57 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 58 | */ |
ed8885a1 MS |
59 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
60 | static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
61 | |
62 | /* | |
2b4d5b25 | 63 | * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity |
b2be5e96 | 64 | */ |
0bf377bb | 65 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
66 | |
67 | /* | |
2bba22c5 | 68 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 69 | * parent will (try to) run first. |
21805085 | 70 | */ |
2bba22c5 | 71 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 72 | |
bf0f6f24 IM |
73 | /* |
74 | * SCHED_OTHER wake-up granularity. | |
bf0f6f24 IM |
75 | * |
76 | * This option delays the preemption effects of decoupled workloads | |
77 | * and reduces their over-scheduling. Synchronous workloads will still | |
78 | * have immediate wakeup/sleep latencies. | |
2b4d5b25 IM |
79 | * |
80 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 81 | */ |
ed8885a1 MS |
82 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
83 | static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
bf0f6f24 | 84 | |
2b4d5b25 | 85 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 86 | |
05289b90 TG |
87 | int sched_thermal_decay_shift; |
88 | static int __init setup_sched_thermal_decay_shift(char *str) | |
89 | { | |
90 | int _shift = 0; | |
91 | ||
92 | if (kstrtoint(str, 0, &_shift)) | |
93 | pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n"); | |
94 | ||
95 | sched_thermal_decay_shift = clamp(_shift, 0, 10); | |
96 | return 1; | |
97 | } | |
98 | __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift); | |
99 | ||
afe06efd TC |
100 | #ifdef CONFIG_SMP |
101 | /* | |
97fb7a0a | 102 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
103 | */ |
104 | int __weak arch_asym_cpu_priority(int cpu) | |
105 | { | |
106 | return -cpu; | |
107 | } | |
6d101ba6 OJ |
108 | |
109 | /* | |
60e17f5c | 110 | * The margin used when comparing utilization with CPU capacity. |
6d101ba6 OJ |
111 | * |
112 | * (default: ~20%) | |
113 | */ | |
60e17f5c VK |
114 | #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) |
115 | ||
4aed8aa4 VS |
116 | /* |
117 | * The margin used when comparing CPU capacities. | |
118 | * is 'cap1' noticeably greater than 'cap2' | |
119 | * | |
120 | * (default: ~5%) | |
121 | */ | |
122 | #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078) | |
afe06efd TC |
123 | #endif |
124 | ||
ec12cb7f PT |
125 | #ifdef CONFIG_CFS_BANDWIDTH |
126 | /* | |
127 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
128 | * each time a cfs_rq requests quota. | |
129 | * | |
130 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
131 | * to consumption or the quota being specified to be smaller than the slice) | |
132 | * we will always only issue the remaining available time. | |
133 | * | |
2b4d5b25 IM |
134 | * (default: 5 msec, units: microseconds) |
135 | */ | |
136 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
ec12cb7f PT |
137 | #endif |
138 | ||
8527632d PG |
139 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
140 | { | |
141 | lw->weight += inc; | |
142 | lw->inv_weight = 0; | |
143 | } | |
144 | ||
145 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
146 | { | |
147 | lw->weight -= dec; | |
148 | lw->inv_weight = 0; | |
149 | } | |
150 | ||
151 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
152 | { | |
153 | lw->weight = w; | |
154 | lw->inv_weight = 0; | |
155 | } | |
156 | ||
029632fb PZ |
157 | /* |
158 | * Increase the granularity value when there are more CPUs, | |
159 | * because with more CPUs the 'effective latency' as visible | |
160 | * to users decreases. But the relationship is not linear, | |
161 | * so pick a second-best guess by going with the log2 of the | |
162 | * number of CPUs. | |
163 | * | |
164 | * This idea comes from the SD scheduler of Con Kolivas: | |
165 | */ | |
58ac93e4 | 166 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 167 | { |
58ac93e4 | 168 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
169 | unsigned int factor; |
170 | ||
171 | switch (sysctl_sched_tunable_scaling) { | |
172 | case SCHED_TUNABLESCALING_NONE: | |
173 | factor = 1; | |
174 | break; | |
175 | case SCHED_TUNABLESCALING_LINEAR: | |
176 | factor = cpus; | |
177 | break; | |
178 | case SCHED_TUNABLESCALING_LOG: | |
179 | default: | |
180 | factor = 1 + ilog2(cpus); | |
181 | break; | |
182 | } | |
183 | ||
184 | return factor; | |
185 | } | |
186 | ||
187 | static void update_sysctl(void) | |
188 | { | |
189 | unsigned int factor = get_update_sysctl_factor(); | |
190 | ||
191 | #define SET_SYSCTL(name) \ | |
192 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
193 | SET_SYSCTL(sched_min_granularity); | |
194 | SET_SYSCTL(sched_latency); | |
195 | SET_SYSCTL(sched_wakeup_granularity); | |
196 | #undef SET_SYSCTL | |
197 | } | |
198 | ||
f38f12d1 | 199 | void __init sched_init_granularity(void) |
029632fb PZ |
200 | { |
201 | update_sysctl(); | |
202 | } | |
203 | ||
9dbdb155 | 204 | #define WMULT_CONST (~0U) |
029632fb PZ |
205 | #define WMULT_SHIFT 32 |
206 | ||
9dbdb155 PZ |
207 | static void __update_inv_weight(struct load_weight *lw) |
208 | { | |
209 | unsigned long w; | |
210 | ||
211 | if (likely(lw->inv_weight)) | |
212 | return; | |
213 | ||
214 | w = scale_load_down(lw->weight); | |
215 | ||
216 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
217 | lw->inv_weight = 1; | |
218 | else if (unlikely(!w)) | |
219 | lw->inv_weight = WMULT_CONST; | |
220 | else | |
221 | lw->inv_weight = WMULT_CONST / w; | |
222 | } | |
029632fb PZ |
223 | |
224 | /* | |
9dbdb155 PZ |
225 | * delta_exec * weight / lw.weight |
226 | * OR | |
227 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
228 | * | |
1c3de5e1 | 229 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
230 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
231 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
232 | * | |
233 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
234 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 235 | */ |
9dbdb155 | 236 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 237 | { |
9dbdb155 | 238 | u64 fact = scale_load_down(weight); |
1e17fb8e | 239 | u32 fact_hi = (u32)(fact >> 32); |
9dbdb155 | 240 | int shift = WMULT_SHIFT; |
1e17fb8e | 241 | int fs; |
029632fb | 242 | |
9dbdb155 | 243 | __update_inv_weight(lw); |
029632fb | 244 | |
1e17fb8e CC |
245 | if (unlikely(fact_hi)) { |
246 | fs = fls(fact_hi); | |
247 | shift -= fs; | |
248 | fact >>= fs; | |
029632fb PZ |
249 | } |
250 | ||
2eeb01a2 | 251 | fact = mul_u32_u32(fact, lw->inv_weight); |
029632fb | 252 | |
1e17fb8e CC |
253 | fact_hi = (u32)(fact >> 32); |
254 | if (fact_hi) { | |
255 | fs = fls(fact_hi); | |
256 | shift -= fs; | |
257 | fact >>= fs; | |
9dbdb155 | 258 | } |
029632fb | 259 | |
9dbdb155 | 260 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
261 | } |
262 | ||
263 | ||
264 | const struct sched_class fair_sched_class; | |
a4c2f00f | 265 | |
bf0f6f24 IM |
266 | /************************************************************** |
267 | * CFS operations on generic schedulable entities: | |
268 | */ | |
269 | ||
62160e3f | 270 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8f48894f PZ |
271 | static inline struct task_struct *task_of(struct sched_entity *se) |
272 | { | |
9148a3a1 | 273 | SCHED_WARN_ON(!entity_is_task(se)); |
8f48894f PZ |
274 | return container_of(se, struct task_struct, se); |
275 | } | |
276 | ||
b758149c PZ |
277 | /* Walk up scheduling entities hierarchy */ |
278 | #define for_each_sched_entity(se) \ | |
279 | for (; se; se = se->parent) | |
280 | ||
281 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
282 | { | |
283 | return p->se.cfs_rq; | |
284 | } | |
285 | ||
286 | /* runqueue on which this entity is (to be) queued */ | |
287 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
288 | { | |
289 | return se->cfs_rq; | |
290 | } | |
291 | ||
292 | /* runqueue "owned" by this group */ | |
293 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
294 | { | |
295 | return grp->my_q; | |
296 | } | |
297 | ||
3c93a0c0 QY |
298 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
299 | { | |
300 | if (!path) | |
301 | return; | |
302 | ||
303 | if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) | |
304 | autogroup_path(cfs_rq->tg, path, len); | |
305 | else if (cfs_rq && cfs_rq->tg->css.cgroup) | |
306 | cgroup_path(cfs_rq->tg->css.cgroup, path, len); | |
307 | else | |
308 | strlcpy(path, "(null)", len); | |
309 | } | |
310 | ||
f6783319 | 311 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 312 | { |
5d299eab PZ |
313 | struct rq *rq = rq_of(cfs_rq); |
314 | int cpu = cpu_of(rq); | |
315 | ||
316 | if (cfs_rq->on_list) | |
f6783319 | 317 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
318 | |
319 | cfs_rq->on_list = 1; | |
320 | ||
321 | /* | |
322 | * Ensure we either appear before our parent (if already | |
323 | * enqueued) or force our parent to appear after us when it is | |
324 | * enqueued. The fact that we always enqueue bottom-up | |
325 | * reduces this to two cases and a special case for the root | |
326 | * cfs_rq. Furthermore, it also means that we will always reset | |
327 | * tmp_alone_branch either when the branch is connected | |
328 | * to a tree or when we reach the top of the tree | |
329 | */ | |
330 | if (cfs_rq->tg->parent && | |
331 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 332 | /* |
5d299eab PZ |
333 | * If parent is already on the list, we add the child |
334 | * just before. Thanks to circular linked property of | |
335 | * the list, this means to put the child at the tail | |
336 | * of the list that starts by parent. | |
67e86250 | 337 | */ |
5d299eab PZ |
338 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
339 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
340 | /* | |
341 | * The branch is now connected to its tree so we can | |
342 | * reset tmp_alone_branch to the beginning of the | |
343 | * list. | |
344 | */ | |
345 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 346 | return true; |
5d299eab | 347 | } |
3d4b47b4 | 348 | |
5d299eab PZ |
349 | if (!cfs_rq->tg->parent) { |
350 | /* | |
351 | * cfs rq without parent should be put | |
352 | * at the tail of the list. | |
353 | */ | |
354 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
355 | &rq->leaf_cfs_rq_list); | |
356 | /* | |
357 | * We have reach the top of a tree so we can reset | |
358 | * tmp_alone_branch to the beginning of the list. | |
359 | */ | |
360 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 361 | return true; |
3d4b47b4 | 362 | } |
5d299eab PZ |
363 | |
364 | /* | |
365 | * The parent has not already been added so we want to | |
366 | * make sure that it will be put after us. | |
367 | * tmp_alone_branch points to the begin of the branch | |
368 | * where we will add parent. | |
369 | */ | |
370 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
371 | /* | |
372 | * update tmp_alone_branch to points to the new begin | |
373 | * of the branch | |
374 | */ | |
375 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 376 | return false; |
3d4b47b4 PZ |
377 | } |
378 | ||
379 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
380 | { | |
381 | if (cfs_rq->on_list) { | |
31bc6aea VG |
382 | struct rq *rq = rq_of(cfs_rq); |
383 | ||
384 | /* | |
385 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
386 | * it can happen the tmp_alone_branch points the a leaf that | |
387 | * we finally want to del. In this case, tmp_alone_branch moves | |
388 | * to the prev element but it will point to rq->leaf_cfs_rq_list | |
389 | * at the end of the enqueue. | |
390 | */ | |
391 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
392 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
393 | ||
3d4b47b4 PZ |
394 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
395 | cfs_rq->on_list = 0; | |
396 | } | |
397 | } | |
398 | ||
5d299eab PZ |
399 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
400 | { | |
401 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
402 | } | |
403 | ||
039ae8bc VG |
404 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
405 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ | |
406 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
407 | leaf_cfs_rq_list) | |
b758149c PZ |
408 | |
409 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 410 | static inline struct cfs_rq * |
b758149c PZ |
411 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
412 | { | |
413 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 414 | return se->cfs_rq; |
b758149c | 415 | |
fed14d45 | 416 | return NULL; |
b758149c PZ |
417 | } |
418 | ||
419 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
420 | { | |
421 | return se->parent; | |
422 | } | |
423 | ||
464b7527 PZ |
424 | static void |
425 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
426 | { | |
427 | int se_depth, pse_depth; | |
428 | ||
429 | /* | |
430 | * preemption test can be made between sibling entities who are in the | |
431 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
432 | * both tasks until we find their ancestors who are siblings of common | |
433 | * parent. | |
434 | */ | |
435 | ||
436 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
437 | se_depth = (*se)->depth; |
438 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
439 | |
440 | while (se_depth > pse_depth) { | |
441 | se_depth--; | |
442 | *se = parent_entity(*se); | |
443 | } | |
444 | ||
445 | while (pse_depth > se_depth) { | |
446 | pse_depth--; | |
447 | *pse = parent_entity(*pse); | |
448 | } | |
449 | ||
450 | while (!is_same_group(*se, *pse)) { | |
451 | *se = parent_entity(*se); | |
452 | *pse = parent_entity(*pse); | |
453 | } | |
454 | } | |
455 | ||
8f48894f PZ |
456 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
457 | ||
458 | static inline struct task_struct *task_of(struct sched_entity *se) | |
459 | { | |
460 | return container_of(se, struct task_struct, se); | |
461 | } | |
bf0f6f24 | 462 | |
b758149c PZ |
463 | #define for_each_sched_entity(se) \ |
464 | for (; se; se = NULL) | |
bf0f6f24 | 465 | |
b758149c | 466 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 467 | { |
b758149c | 468 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
469 | } |
470 | ||
b758149c PZ |
471 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
472 | { | |
473 | struct task_struct *p = task_of(se); | |
474 | struct rq *rq = task_rq(p); | |
475 | ||
476 | return &rq->cfs; | |
477 | } | |
478 | ||
479 | /* runqueue "owned" by this group */ | |
480 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
481 | { | |
482 | return NULL; | |
483 | } | |
484 | ||
3c93a0c0 QY |
485 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
486 | { | |
487 | if (path) | |
488 | strlcpy(path, "(null)", len); | |
489 | } | |
490 | ||
f6783319 | 491 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 492 | { |
f6783319 | 493 | return true; |
3d4b47b4 PZ |
494 | } |
495 | ||
496 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
497 | { | |
498 | } | |
499 | ||
5d299eab PZ |
500 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
501 | { | |
502 | } | |
503 | ||
039ae8bc VG |
504 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
505 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 506 | |
b758149c PZ |
507 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
508 | { | |
509 | return NULL; | |
510 | } | |
511 | ||
464b7527 PZ |
512 | static inline void |
513 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
514 | { | |
515 | } | |
516 | ||
b758149c PZ |
517 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
518 | ||
6c16a6dc | 519 | static __always_inline |
9dbdb155 | 520 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
521 | |
522 | /************************************************************** | |
523 | * Scheduling class tree data structure manipulation methods: | |
524 | */ | |
525 | ||
1bf08230 | 526 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 527 | { |
1bf08230 | 528 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 529 | if (delta > 0) |
1bf08230 | 530 | max_vruntime = vruntime; |
02e0431a | 531 | |
1bf08230 | 532 | return max_vruntime; |
02e0431a PZ |
533 | } |
534 | ||
0702e3eb | 535 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
536 | { |
537 | s64 delta = (s64)(vruntime - min_vruntime); | |
538 | if (delta < 0) | |
539 | min_vruntime = vruntime; | |
540 | ||
541 | return min_vruntime; | |
542 | } | |
543 | ||
bf9be9a1 | 544 | static inline bool entity_before(struct sched_entity *a, |
54fdc581 FC |
545 | struct sched_entity *b) |
546 | { | |
547 | return (s64)(a->vruntime - b->vruntime) < 0; | |
548 | } | |
549 | ||
bf9be9a1 PZ |
550 | #define __node_2_se(node) \ |
551 | rb_entry((node), struct sched_entity, run_node) | |
552 | ||
1af5f730 PZ |
553 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
554 | { | |
b60205c7 | 555 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 556 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 557 | |
1af5f730 PZ |
558 | u64 vruntime = cfs_rq->min_vruntime; |
559 | ||
b60205c7 PZ |
560 | if (curr) { |
561 | if (curr->on_rq) | |
562 | vruntime = curr->vruntime; | |
563 | else | |
564 | curr = NULL; | |
565 | } | |
1af5f730 | 566 | |
bfb06889 | 567 | if (leftmost) { /* non-empty tree */ |
bf9be9a1 | 568 | struct sched_entity *se = __node_2_se(leftmost); |
1af5f730 | 569 | |
b60205c7 | 570 | if (!curr) |
1af5f730 PZ |
571 | vruntime = se->vruntime; |
572 | else | |
573 | vruntime = min_vruntime(vruntime, se->vruntime); | |
574 | } | |
575 | ||
1bf08230 | 576 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 577 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
578 | #ifndef CONFIG_64BIT |
579 | smp_wmb(); | |
580 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
581 | #endif | |
1af5f730 PZ |
582 | } |
583 | ||
bf9be9a1 PZ |
584 | static inline bool __entity_less(struct rb_node *a, const struct rb_node *b) |
585 | { | |
586 | return entity_before(__node_2_se(a), __node_2_se(b)); | |
587 | } | |
588 | ||
bf0f6f24 IM |
589 | /* |
590 | * Enqueue an entity into the rb-tree: | |
591 | */ | |
0702e3eb | 592 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 593 | { |
bf9be9a1 | 594 | rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less); |
bf0f6f24 IM |
595 | } |
596 | ||
0702e3eb | 597 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 598 | { |
bfb06889 | 599 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
600 | } |
601 | ||
029632fb | 602 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 603 | { |
bfb06889 | 604 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
605 | |
606 | if (!left) | |
607 | return NULL; | |
608 | ||
bf9be9a1 | 609 | return __node_2_se(left); |
bf0f6f24 IM |
610 | } |
611 | ||
ac53db59 RR |
612 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
613 | { | |
614 | struct rb_node *next = rb_next(&se->run_node); | |
615 | ||
616 | if (!next) | |
617 | return NULL; | |
618 | ||
bf9be9a1 | 619 | return __node_2_se(next); |
ac53db59 RR |
620 | } |
621 | ||
622 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 623 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 624 | { |
bfb06889 | 625 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 626 | |
70eee74b BS |
627 | if (!last) |
628 | return NULL; | |
7eee3e67 | 629 | |
bf9be9a1 | 630 | return __node_2_se(last); |
aeb73b04 PZ |
631 | } |
632 | ||
bf0f6f24 IM |
633 | /************************************************************** |
634 | * Scheduling class statistics methods: | |
635 | */ | |
636 | ||
8a99b683 | 637 | int sched_update_scaling(void) |
b2be5e96 | 638 | { |
58ac93e4 | 639 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 | 640 | |
b2be5e96 PZ |
641 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
642 | sysctl_sched_min_granularity); | |
643 | ||
acb4a848 CE |
644 | #define WRT_SYSCTL(name) \ |
645 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
646 | WRT_SYSCTL(sched_min_granularity); | |
647 | WRT_SYSCTL(sched_latency); | |
648 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
649 | #undef WRT_SYSCTL |
650 | ||
b2be5e96 PZ |
651 | return 0; |
652 | } | |
653 | #endif | |
647e7cac | 654 | |
a7be37ac | 655 | /* |
f9c0b095 | 656 | * delta /= w |
a7be37ac | 657 | */ |
9dbdb155 | 658 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 659 | { |
f9c0b095 | 660 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 661 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
662 | |
663 | return delta; | |
664 | } | |
665 | ||
647e7cac IM |
666 | /* |
667 | * The idea is to set a period in which each task runs once. | |
668 | * | |
532b1858 | 669 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
670 | * this period because otherwise the slices get too small. |
671 | * | |
672 | * p = (nr <= nl) ? l : l*nr/nl | |
673 | */ | |
4d78e7b6 PZ |
674 | static u64 __sched_period(unsigned long nr_running) |
675 | { | |
8e2b0bf3 BF |
676 | if (unlikely(nr_running > sched_nr_latency)) |
677 | return nr_running * sysctl_sched_min_granularity; | |
678 | else | |
679 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
680 | } |
681 | ||
647e7cac IM |
682 | /* |
683 | * We calculate the wall-time slice from the period by taking a part | |
684 | * proportional to the weight. | |
685 | * | |
f9c0b095 | 686 | * s = p*P[w/rw] |
647e7cac | 687 | */ |
6d0f0ebd | 688 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 689 | { |
0c2de3f0 PZ |
690 | unsigned int nr_running = cfs_rq->nr_running; |
691 | u64 slice; | |
692 | ||
693 | if (sched_feat(ALT_PERIOD)) | |
694 | nr_running = rq_of(cfs_rq)->cfs.h_nr_running; | |
695 | ||
696 | slice = __sched_period(nr_running + !se->on_rq); | |
f9c0b095 | 697 | |
0a582440 | 698 | for_each_sched_entity(se) { |
6272d68c | 699 | struct load_weight *load; |
3104bf03 | 700 | struct load_weight lw; |
6272d68c LM |
701 | |
702 | cfs_rq = cfs_rq_of(se); | |
703 | load = &cfs_rq->load; | |
f9c0b095 | 704 | |
0a582440 | 705 | if (unlikely(!se->on_rq)) { |
3104bf03 | 706 | lw = cfs_rq->load; |
0a582440 MG |
707 | |
708 | update_load_add(&lw, se->load.weight); | |
709 | load = &lw; | |
710 | } | |
9dbdb155 | 711 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 | 712 | } |
0c2de3f0 PZ |
713 | |
714 | if (sched_feat(BASE_SLICE)) | |
715 | slice = max(slice, (u64)sysctl_sched_min_granularity); | |
716 | ||
0a582440 | 717 | return slice; |
bf0f6f24 IM |
718 | } |
719 | ||
647e7cac | 720 | /* |
660cc00f | 721 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 722 | * |
f9c0b095 | 723 | * vs = s/w |
647e7cac | 724 | */ |
f9c0b095 | 725 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 726 | { |
f9c0b095 | 727 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
728 | } |
729 | ||
c0796298 | 730 | #include "pelt.h" |
23127296 | 731 | #ifdef CONFIG_SMP |
283e2ed3 | 732 | |
772bd008 | 733 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 734 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 735 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 736 | |
540247fb YD |
737 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
738 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 739 | { |
540247fb | 740 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 741 | |
f207934f PZ |
742 | memset(sa, 0, sizeof(*sa)); |
743 | ||
b5a9b340 | 744 | /* |
dfcb245e | 745 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 746 | * they get a chance to stabilize to their real load level. |
dfcb245e | 747 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
748 | * nothing has been attached to the task group yet. |
749 | */ | |
750 | if (entity_is_task(se)) | |
0dacee1b | 751 | sa->load_avg = scale_load_down(se->load.weight); |
f207934f | 752 | |
9d89c257 | 753 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 754 | } |
7ea241af | 755 | |
df217913 | 756 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 757 | |
2b8c41da YD |
758 | /* |
759 | * With new tasks being created, their initial util_avgs are extrapolated | |
760 | * based on the cfs_rq's current util_avg: | |
761 | * | |
762 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
763 | * | |
764 | * However, in many cases, the above util_avg does not give a desired | |
765 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
766 | * as when the series is a harmonic series. | |
767 | * | |
768 | * To solve this problem, we also cap the util_avg of successive tasks to | |
769 | * only 1/2 of the left utilization budget: | |
770 | * | |
8fe5c5a9 | 771 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 772 | * |
8fe5c5a9 | 773 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 774 | * |
8fe5c5a9 QP |
775 | * For example, for a CPU with 1024 of capacity, a simplest series from |
776 | * the beginning would be like: | |
2b8c41da YD |
777 | * |
778 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
779 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
780 | * | |
781 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
782 | * if util_avg > util_avg_cap. | |
783 | */ | |
d0fe0b9c | 784 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 785 | { |
d0fe0b9c | 786 | struct sched_entity *se = &p->se; |
2b8c41da YD |
787 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
788 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 789 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 790 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
791 | |
792 | if (cap > 0) { | |
793 | if (cfs_rq->avg.util_avg != 0) { | |
794 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
795 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
796 | ||
797 | if (sa->util_avg > cap) | |
798 | sa->util_avg = cap; | |
799 | } else { | |
800 | sa->util_avg = cap; | |
801 | } | |
2b8c41da | 802 | } |
7dc603c9 | 803 | |
e21cf434 | 804 | sa->runnable_avg = sa->util_avg; |
9f683953 | 805 | |
d0fe0b9c DE |
806 | if (p->sched_class != &fair_sched_class) { |
807 | /* | |
808 | * For !fair tasks do: | |
809 | * | |
810 | update_cfs_rq_load_avg(now, cfs_rq); | |
a4f9a0e5 | 811 | attach_entity_load_avg(cfs_rq, se); |
d0fe0b9c DE |
812 | switched_from_fair(rq, p); |
813 | * | |
814 | * such that the next switched_to_fair() has the | |
815 | * expected state. | |
816 | */ | |
817 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
818 | return; | |
7dc603c9 PZ |
819 | } |
820 | ||
df217913 | 821 | attach_entity_cfs_rq(se); |
2b8c41da YD |
822 | } |
823 | ||
7dc603c9 | 824 | #else /* !CONFIG_SMP */ |
540247fb | 825 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
826 | { |
827 | } | |
d0fe0b9c | 828 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
829 | { |
830 | } | |
fe749158 | 831 | static void update_tg_load_avg(struct cfs_rq *cfs_rq) |
3d30544f PZ |
832 | { |
833 | } | |
7dc603c9 | 834 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 835 | |
bf0f6f24 | 836 | /* |
9dbdb155 | 837 | * Update the current task's runtime statistics. |
bf0f6f24 | 838 | */ |
b7cc0896 | 839 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 840 | { |
429d43bc | 841 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 842 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 843 | u64 delta_exec; |
bf0f6f24 IM |
844 | |
845 | if (unlikely(!curr)) | |
846 | return; | |
847 | ||
9dbdb155 PZ |
848 | delta_exec = now - curr->exec_start; |
849 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 850 | return; |
bf0f6f24 | 851 | |
8ebc91d9 | 852 | curr->exec_start = now; |
d842de87 | 853 | |
9dbdb155 PZ |
854 | schedstat_set(curr->statistics.exec_max, |
855 | max(delta_exec, curr->statistics.exec_max)); | |
856 | ||
857 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 858 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
859 | |
860 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
861 | update_min_vruntime(cfs_rq); | |
862 | ||
d842de87 SV |
863 | if (entity_is_task(curr)) { |
864 | struct task_struct *curtask = task_of(curr); | |
865 | ||
f977bb49 | 866 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 867 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 868 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 869 | } |
ec12cb7f PT |
870 | |
871 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
872 | } |
873 | ||
6e998916 SG |
874 | static void update_curr_fair(struct rq *rq) |
875 | { | |
876 | update_curr(cfs_rq_of(&rq->curr->se)); | |
877 | } | |
878 | ||
bf0f6f24 | 879 | static inline void |
5870db5b | 880 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 881 | { |
4fa8d299 JP |
882 | u64 wait_start, prev_wait_start; |
883 | ||
884 | if (!schedstat_enabled()) | |
885 | return; | |
886 | ||
887 | wait_start = rq_clock(rq_of(cfs_rq)); | |
888 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
889 | |
890 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
891 | likely(wait_start > prev_wait_start)) |
892 | wait_start -= prev_wait_start; | |
3ea94de1 | 893 | |
2ed41a55 | 894 | __schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
895 | } |
896 | ||
4fa8d299 | 897 | static inline void |
3ea94de1 JP |
898 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
899 | { | |
900 | struct task_struct *p; | |
cb251765 MG |
901 | u64 delta; |
902 | ||
4fa8d299 JP |
903 | if (!schedstat_enabled()) |
904 | return; | |
905 | ||
b9c88f75 | 906 | /* |
907 | * When the sched_schedstat changes from 0 to 1, some sched se | |
908 | * maybe already in the runqueue, the se->statistics.wait_start | |
909 | * will be 0.So it will let the delta wrong. We need to avoid this | |
910 | * scenario. | |
911 | */ | |
912 | if (unlikely(!schedstat_val(se->statistics.wait_start))) | |
913 | return; | |
914 | ||
4fa8d299 | 915 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); |
3ea94de1 JP |
916 | |
917 | if (entity_is_task(se)) { | |
918 | p = task_of(se); | |
919 | if (task_on_rq_migrating(p)) { | |
920 | /* | |
921 | * Preserve migrating task's wait time so wait_start | |
922 | * time stamp can be adjusted to accumulate wait time | |
923 | * prior to migration. | |
924 | */ | |
2ed41a55 | 925 | __schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
926 | return; |
927 | } | |
928 | trace_sched_stat_wait(p, delta); | |
929 | } | |
930 | ||
2ed41a55 | 931 | __schedstat_set(se->statistics.wait_max, |
4fa8d299 | 932 | max(schedstat_val(se->statistics.wait_max), delta)); |
2ed41a55 PZ |
933 | __schedstat_inc(se->statistics.wait_count); |
934 | __schedstat_add(se->statistics.wait_sum, delta); | |
935 | __schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 936 | } |
3ea94de1 | 937 | |
4fa8d299 | 938 | static inline void |
1a3d027c JP |
939 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
940 | { | |
941 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
942 | u64 sleep_start, block_start; |
943 | ||
944 | if (!schedstat_enabled()) | |
945 | return; | |
946 | ||
947 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
948 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
949 | |
950 | if (entity_is_task(se)) | |
951 | tsk = task_of(se); | |
952 | ||
4fa8d299 JP |
953 | if (sleep_start) { |
954 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
955 | |
956 | if ((s64)delta < 0) | |
957 | delta = 0; | |
958 | ||
4fa8d299 | 959 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
2ed41a55 | 960 | __schedstat_set(se->statistics.sleep_max, delta); |
1a3d027c | 961 | |
2ed41a55 PZ |
962 | __schedstat_set(se->statistics.sleep_start, 0); |
963 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
964 | |
965 | if (tsk) { | |
966 | account_scheduler_latency(tsk, delta >> 10, 1); | |
967 | trace_sched_stat_sleep(tsk, delta); | |
968 | } | |
969 | } | |
4fa8d299 JP |
970 | if (block_start) { |
971 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
972 | |
973 | if ((s64)delta < 0) | |
974 | delta = 0; | |
975 | ||
4fa8d299 | 976 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
2ed41a55 | 977 | __schedstat_set(se->statistics.block_max, delta); |
1a3d027c | 978 | |
2ed41a55 PZ |
979 | __schedstat_set(se->statistics.block_start, 0); |
980 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
981 | |
982 | if (tsk) { | |
983 | if (tsk->in_iowait) { | |
2ed41a55 PZ |
984 | __schedstat_add(se->statistics.iowait_sum, delta); |
985 | __schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
986 | trace_sched_stat_iowait(tsk, delta); |
987 | } | |
988 | ||
989 | trace_sched_stat_blocked(tsk, delta); | |
990 | ||
991 | /* | |
992 | * Blocking time is in units of nanosecs, so shift by | |
993 | * 20 to get a milliseconds-range estimation of the | |
994 | * amount of time that the task spent sleeping: | |
995 | */ | |
996 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
997 | profile_hits(SLEEP_PROFILING, | |
998 | (void *)get_wchan(tsk), | |
999 | delta >> 20); | |
1000 | } | |
1001 | account_scheduler_latency(tsk, delta >> 10, 0); | |
1002 | } | |
1003 | } | |
3ea94de1 | 1004 | } |
3ea94de1 | 1005 | |
bf0f6f24 IM |
1006 | /* |
1007 | * Task is being enqueued - update stats: | |
1008 | */ | |
cb251765 | 1009 | static inline void |
1a3d027c | 1010 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1011 | { |
4fa8d299 JP |
1012 | if (!schedstat_enabled()) |
1013 | return; | |
1014 | ||
bf0f6f24 IM |
1015 | /* |
1016 | * Are we enqueueing a waiting task? (for current tasks | |
1017 | * a dequeue/enqueue event is a NOP) | |
1018 | */ | |
429d43bc | 1019 | if (se != cfs_rq->curr) |
5870db5b | 1020 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
1021 | |
1022 | if (flags & ENQUEUE_WAKEUP) | |
1023 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
1024 | } |
1025 | ||
bf0f6f24 | 1026 | static inline void |
cb251765 | 1027 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1028 | { |
4fa8d299 JP |
1029 | |
1030 | if (!schedstat_enabled()) | |
1031 | return; | |
1032 | ||
bf0f6f24 IM |
1033 | /* |
1034 | * Mark the end of the wait period if dequeueing a | |
1035 | * waiting task: | |
1036 | */ | |
429d43bc | 1037 | if (se != cfs_rq->curr) |
9ef0a961 | 1038 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 1039 | |
4fa8d299 JP |
1040 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1041 | struct task_struct *tsk = task_of(se); | |
cb251765 | 1042 | |
4fa8d299 | 1043 | if (tsk->state & TASK_INTERRUPTIBLE) |
2ed41a55 | 1044 | __schedstat_set(se->statistics.sleep_start, |
4fa8d299 JP |
1045 | rq_clock(rq_of(cfs_rq))); |
1046 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
2ed41a55 | 1047 | __schedstat_set(se->statistics.block_start, |
4fa8d299 | 1048 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1049 | } |
cb251765 MG |
1050 | } |
1051 | ||
bf0f6f24 IM |
1052 | /* |
1053 | * We are picking a new current task - update its stats: | |
1054 | */ | |
1055 | static inline void | |
79303e9e | 1056 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1057 | { |
1058 | /* | |
1059 | * We are starting a new run period: | |
1060 | */ | |
78becc27 | 1061 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1062 | } |
1063 | ||
bf0f6f24 IM |
1064 | /************************************************** |
1065 | * Scheduling class queueing methods: | |
1066 | */ | |
1067 | ||
cbee9f88 PZ |
1068 | #ifdef CONFIG_NUMA_BALANCING |
1069 | /* | |
598f0ec0 MG |
1070 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1071 | * calculated based on the tasks virtual memory size and | |
1072 | * numa_balancing_scan_size. | |
cbee9f88 | 1073 | */ |
598f0ec0 MG |
1074 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1075 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1076 | |
1077 | /* Portion of address space to scan in MB */ | |
1078 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1079 | |
4b96a29b PZ |
1080 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1081 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1082 | ||
b5dd77c8 | 1083 | struct numa_group { |
c45a7795 | 1084 | refcount_t refcount; |
b5dd77c8 RR |
1085 | |
1086 | spinlock_t lock; /* nr_tasks, tasks */ | |
1087 | int nr_tasks; | |
1088 | pid_t gid; | |
1089 | int active_nodes; | |
1090 | ||
1091 | struct rcu_head rcu; | |
1092 | unsigned long total_faults; | |
1093 | unsigned long max_faults_cpu; | |
1094 | /* | |
1095 | * Faults_cpu is used to decide whether memory should move | |
1096 | * towards the CPU. As a consequence, these stats are weighted | |
1097 | * more by CPU use than by memory faults. | |
1098 | */ | |
1099 | unsigned long *faults_cpu; | |
04f5c362 | 1100 | unsigned long faults[]; |
b5dd77c8 RR |
1101 | }; |
1102 | ||
cb361d8c JH |
1103 | /* |
1104 | * For functions that can be called in multiple contexts that permit reading | |
1105 | * ->numa_group (see struct task_struct for locking rules). | |
1106 | */ | |
1107 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1108 | { | |
1109 | return rcu_dereference_check(p->numa_group, p == current || | |
1110 | (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu))); | |
1111 | } | |
1112 | ||
1113 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1114 | { | |
1115 | return rcu_dereference_protected(p->numa_group, p == current); | |
1116 | } | |
1117 | ||
b5dd77c8 RR |
1118 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1119 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1120 | ||
598f0ec0 MG |
1121 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1122 | { | |
1123 | unsigned long rss = 0; | |
1124 | unsigned long nr_scan_pages; | |
1125 | ||
1126 | /* | |
1127 | * Calculations based on RSS as non-present and empty pages are skipped | |
1128 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1129 | * on resident pages | |
1130 | */ | |
1131 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1132 | rss = get_mm_rss(p->mm); | |
1133 | if (!rss) | |
1134 | rss = nr_scan_pages; | |
1135 | ||
1136 | rss = round_up(rss, nr_scan_pages); | |
1137 | return rss / nr_scan_pages; | |
1138 | } | |
1139 | ||
3b03706f | 1140 | /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ |
598f0ec0 MG |
1141 | #define MAX_SCAN_WINDOW 2560 |
1142 | ||
1143 | static unsigned int task_scan_min(struct task_struct *p) | |
1144 | { | |
316c1608 | 1145 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1146 | unsigned int scan, floor; |
1147 | unsigned int windows = 1; | |
1148 | ||
64192658 KT |
1149 | if (scan_size < MAX_SCAN_WINDOW) |
1150 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1151 | floor = 1000 / windows; |
1152 | ||
1153 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1154 | return max_t(unsigned int, floor, scan); | |
1155 | } | |
1156 | ||
b5dd77c8 RR |
1157 | static unsigned int task_scan_start(struct task_struct *p) |
1158 | { | |
1159 | unsigned long smin = task_scan_min(p); | |
1160 | unsigned long period = smin; | |
cb361d8c | 1161 | struct numa_group *ng; |
b5dd77c8 RR |
1162 | |
1163 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1164 | rcu_read_lock(); |
1165 | ng = rcu_dereference(p->numa_group); | |
1166 | if (ng) { | |
b5dd77c8 RR |
1167 | unsigned long shared = group_faults_shared(ng); |
1168 | unsigned long private = group_faults_priv(ng); | |
1169 | ||
c45a7795 | 1170 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1171 | period *= shared + 1; |
1172 | period /= private + shared + 1; | |
1173 | } | |
cb361d8c | 1174 | rcu_read_unlock(); |
b5dd77c8 RR |
1175 | |
1176 | return max(smin, period); | |
1177 | } | |
1178 | ||
598f0ec0 MG |
1179 | static unsigned int task_scan_max(struct task_struct *p) |
1180 | { | |
b5dd77c8 RR |
1181 | unsigned long smin = task_scan_min(p); |
1182 | unsigned long smax; | |
cb361d8c | 1183 | struct numa_group *ng; |
598f0ec0 MG |
1184 | |
1185 | /* Watch for min being lower than max due to floor calculations */ | |
1186 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1187 | |
1188 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1189 | ng = deref_curr_numa_group(p); |
1190 | if (ng) { | |
b5dd77c8 RR |
1191 | unsigned long shared = group_faults_shared(ng); |
1192 | unsigned long private = group_faults_priv(ng); | |
1193 | unsigned long period = smax; | |
1194 | ||
c45a7795 | 1195 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1196 | period *= shared + 1; |
1197 | period /= private + shared + 1; | |
1198 | ||
1199 | smax = max(smax, period); | |
1200 | } | |
1201 | ||
598f0ec0 MG |
1202 | return max(smin, smax); |
1203 | } | |
1204 | ||
0ec8aa00 PZ |
1205 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1206 | { | |
98fa15f3 | 1207 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1208 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1209 | } | |
1210 | ||
1211 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1212 | { | |
98fa15f3 | 1213 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1214 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1215 | } | |
1216 | ||
be1e4e76 RR |
1217 | /* Shared or private faults. */ |
1218 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1219 | ||
1220 | /* Memory and CPU locality */ | |
1221 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1222 | ||
1223 | /* Averaged statistics, and temporary buffers. */ | |
1224 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1225 | ||
e29cf08b MG |
1226 | pid_t task_numa_group_id(struct task_struct *p) |
1227 | { | |
cb361d8c JH |
1228 | struct numa_group *ng; |
1229 | pid_t gid = 0; | |
1230 | ||
1231 | rcu_read_lock(); | |
1232 | ng = rcu_dereference(p->numa_group); | |
1233 | if (ng) | |
1234 | gid = ng->gid; | |
1235 | rcu_read_unlock(); | |
1236 | ||
1237 | return gid; | |
e29cf08b MG |
1238 | } |
1239 | ||
44dba3d5 | 1240 | /* |
97fb7a0a | 1241 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1242 | * occupy the first half of the array. The second half of the |
1243 | * array is for current counters, which are averaged into the | |
1244 | * first set by task_numa_placement. | |
1245 | */ | |
1246 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1247 | { |
44dba3d5 | 1248 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1249 | } |
1250 | ||
1251 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1252 | { | |
44dba3d5 | 1253 | if (!p->numa_faults) |
ac8e895b MG |
1254 | return 0; |
1255 | ||
44dba3d5 IM |
1256 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1257 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1258 | } |
1259 | ||
83e1d2cd MG |
1260 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1261 | { | |
cb361d8c JH |
1262 | struct numa_group *ng = deref_task_numa_group(p); |
1263 | ||
1264 | if (!ng) | |
83e1d2cd MG |
1265 | return 0; |
1266 | ||
cb361d8c JH |
1267 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1268 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1269 | } |
1270 | ||
20e07dea RR |
1271 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1272 | { | |
44dba3d5 IM |
1273 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1274 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1275 | } |
1276 | ||
b5dd77c8 RR |
1277 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1278 | { | |
1279 | unsigned long faults = 0; | |
1280 | int node; | |
1281 | ||
1282 | for_each_online_node(node) { | |
1283 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1284 | } | |
1285 | ||
1286 | return faults; | |
1287 | } | |
1288 | ||
1289 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1290 | { | |
1291 | unsigned long faults = 0; | |
1292 | int node; | |
1293 | ||
1294 | for_each_online_node(node) { | |
1295 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1296 | } | |
1297 | ||
1298 | return faults; | |
1299 | } | |
1300 | ||
4142c3eb RR |
1301 | /* |
1302 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1303 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1304 | * between these nodes are slowed down, to allow things to settle down. | |
1305 | */ | |
1306 | #define ACTIVE_NODE_FRACTION 3 | |
1307 | ||
1308 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1309 | { | |
1310 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1311 | } | |
1312 | ||
6c6b1193 RR |
1313 | /* Handle placement on systems where not all nodes are directly connected. */ |
1314 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1315 | int maxdist, bool task) | |
1316 | { | |
1317 | unsigned long score = 0; | |
1318 | int node; | |
1319 | ||
1320 | /* | |
1321 | * All nodes are directly connected, and the same distance | |
1322 | * from each other. No need for fancy placement algorithms. | |
1323 | */ | |
1324 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1325 | return 0; | |
1326 | ||
1327 | /* | |
1328 | * This code is called for each node, introducing N^2 complexity, | |
1329 | * which should be ok given the number of nodes rarely exceeds 8. | |
1330 | */ | |
1331 | for_each_online_node(node) { | |
1332 | unsigned long faults; | |
1333 | int dist = node_distance(nid, node); | |
1334 | ||
1335 | /* | |
1336 | * The furthest away nodes in the system are not interesting | |
1337 | * for placement; nid was already counted. | |
1338 | */ | |
1339 | if (dist == sched_max_numa_distance || node == nid) | |
1340 | continue; | |
1341 | ||
1342 | /* | |
1343 | * On systems with a backplane NUMA topology, compare groups | |
1344 | * of nodes, and move tasks towards the group with the most | |
1345 | * memory accesses. When comparing two nodes at distance | |
1346 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1347 | * of each group. Skip other nodes. | |
1348 | */ | |
1349 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
0ee7e74d | 1350 | dist >= maxdist) |
6c6b1193 RR |
1351 | continue; |
1352 | ||
1353 | /* Add up the faults from nearby nodes. */ | |
1354 | if (task) | |
1355 | faults = task_faults(p, node); | |
1356 | else | |
1357 | faults = group_faults(p, node); | |
1358 | ||
1359 | /* | |
1360 | * On systems with a glueless mesh NUMA topology, there are | |
1361 | * no fixed "groups of nodes". Instead, nodes that are not | |
1362 | * directly connected bounce traffic through intermediate | |
1363 | * nodes; a numa_group can occupy any set of nodes. | |
1364 | * The further away a node is, the less the faults count. | |
1365 | * This seems to result in good task placement. | |
1366 | */ | |
1367 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1368 | faults *= (sched_max_numa_distance - dist); | |
1369 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1370 | } | |
1371 | ||
1372 | score += faults; | |
1373 | } | |
1374 | ||
1375 | return score; | |
1376 | } | |
1377 | ||
83e1d2cd MG |
1378 | /* |
1379 | * These return the fraction of accesses done by a particular task, or | |
1380 | * task group, on a particular numa node. The group weight is given a | |
1381 | * larger multiplier, in order to group tasks together that are almost | |
1382 | * evenly spread out between numa nodes. | |
1383 | */ | |
7bd95320 RR |
1384 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1385 | int dist) | |
83e1d2cd | 1386 | { |
7bd95320 | 1387 | unsigned long faults, total_faults; |
83e1d2cd | 1388 | |
44dba3d5 | 1389 | if (!p->numa_faults) |
83e1d2cd MG |
1390 | return 0; |
1391 | ||
1392 | total_faults = p->total_numa_faults; | |
1393 | ||
1394 | if (!total_faults) | |
1395 | return 0; | |
1396 | ||
7bd95320 | 1397 | faults = task_faults(p, nid); |
6c6b1193 RR |
1398 | faults += score_nearby_nodes(p, nid, dist, true); |
1399 | ||
7bd95320 | 1400 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1401 | } |
1402 | ||
7bd95320 RR |
1403 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1404 | int dist) | |
83e1d2cd | 1405 | { |
cb361d8c | 1406 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1407 | unsigned long faults, total_faults; |
1408 | ||
cb361d8c | 1409 | if (!ng) |
7bd95320 RR |
1410 | return 0; |
1411 | ||
cb361d8c | 1412 | total_faults = ng->total_faults; |
7bd95320 RR |
1413 | |
1414 | if (!total_faults) | |
83e1d2cd MG |
1415 | return 0; |
1416 | ||
7bd95320 | 1417 | faults = group_faults(p, nid); |
6c6b1193 RR |
1418 | faults += score_nearby_nodes(p, nid, dist, false); |
1419 | ||
7bd95320 | 1420 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1421 | } |
1422 | ||
10f39042 RR |
1423 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1424 | int src_nid, int dst_cpu) | |
1425 | { | |
cb361d8c | 1426 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1427 | int dst_nid = cpu_to_node(dst_cpu); |
1428 | int last_cpupid, this_cpupid; | |
1429 | ||
1430 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
37355bdc MG |
1431 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1432 | ||
1433 | /* | |
1434 | * Allow first faults or private faults to migrate immediately early in | |
1435 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1436 | * two full passes of the "multi-stage node selection" test that is | |
1437 | * executed below. | |
1438 | */ | |
98fa15f3 | 1439 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1440 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1441 | return true; | |
10f39042 RR |
1442 | |
1443 | /* | |
1444 | * Multi-stage node selection is used in conjunction with a periodic | |
1445 | * migration fault to build a temporal task<->page relation. By using | |
1446 | * a two-stage filter we remove short/unlikely relations. | |
1447 | * | |
1448 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1449 | * a task's usage of a particular page (n_p) per total usage of this | |
1450 | * page (n_t) (in a given time-span) to a probability. | |
1451 | * | |
1452 | * Our periodic faults will sample this probability and getting the | |
1453 | * same result twice in a row, given these samples are fully | |
1454 | * independent, is then given by P(n)^2, provided our sample period | |
1455 | * is sufficiently short compared to the usage pattern. | |
1456 | * | |
1457 | * This quadric squishes small probabilities, making it less likely we | |
1458 | * act on an unlikely task<->page relation. | |
1459 | */ | |
10f39042 RR |
1460 | if (!cpupid_pid_unset(last_cpupid) && |
1461 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1462 | return false; | |
1463 | ||
1464 | /* Always allow migrate on private faults */ | |
1465 | if (cpupid_match_pid(p, last_cpupid)) | |
1466 | return true; | |
1467 | ||
1468 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1469 | if (!ng) | |
1470 | return true; | |
1471 | ||
1472 | /* | |
4142c3eb RR |
1473 | * Destination node is much more heavily used than the source |
1474 | * node? Allow migration. | |
10f39042 | 1475 | */ |
4142c3eb RR |
1476 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1477 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1478 | return true; |
1479 | ||
1480 | /* | |
4142c3eb RR |
1481 | * Distribute memory according to CPU & memory use on each node, |
1482 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1483 | * | |
1484 | * faults_cpu(dst) 3 faults_cpu(src) | |
1485 | * --------------- * - > --------------- | |
1486 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1487 | */ |
4142c3eb RR |
1488 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1489 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1490 | } |
1491 | ||
6499b1b2 VG |
1492 | /* |
1493 | * 'numa_type' describes the node at the moment of load balancing. | |
1494 | */ | |
1495 | enum numa_type { | |
1496 | /* The node has spare capacity that can be used to run more tasks. */ | |
1497 | node_has_spare = 0, | |
1498 | /* | |
1499 | * The node is fully used and the tasks don't compete for more CPU | |
1500 | * cycles. Nevertheless, some tasks might wait before running. | |
1501 | */ | |
1502 | node_fully_busy, | |
1503 | /* | |
1504 | * The node is overloaded and can't provide expected CPU cycles to all | |
1505 | * tasks. | |
1506 | */ | |
1507 | node_overloaded | |
1508 | }; | |
58d081b5 | 1509 | |
fb13c7ee | 1510 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1511 | struct numa_stats { |
1512 | unsigned long load; | |
8e0e0eda | 1513 | unsigned long runnable; |
6499b1b2 | 1514 | unsigned long util; |
fb13c7ee | 1515 | /* Total compute capacity of CPUs on a node */ |
5ef20ca1 | 1516 | unsigned long compute_capacity; |
6499b1b2 VG |
1517 | unsigned int nr_running; |
1518 | unsigned int weight; | |
1519 | enum numa_type node_type; | |
ff7db0bf | 1520 | int idle_cpu; |
58d081b5 | 1521 | }; |
e6628d5b | 1522 | |
ff7db0bf MG |
1523 | static inline bool is_core_idle(int cpu) |
1524 | { | |
1525 | #ifdef CONFIG_SCHED_SMT | |
1526 | int sibling; | |
1527 | ||
1528 | for_each_cpu(sibling, cpu_smt_mask(cpu)) { | |
1529 | if (cpu == sibling) | |
1530 | continue; | |
1531 | ||
1532 | if (!idle_cpu(cpu)) | |
1533 | return false; | |
1534 | } | |
1535 | #endif | |
1536 | ||
1537 | return true; | |
1538 | } | |
1539 | ||
58d081b5 MG |
1540 | struct task_numa_env { |
1541 | struct task_struct *p; | |
e6628d5b | 1542 | |
58d081b5 MG |
1543 | int src_cpu, src_nid; |
1544 | int dst_cpu, dst_nid; | |
e6628d5b | 1545 | |
58d081b5 | 1546 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1547 | |
40ea2b42 | 1548 | int imbalance_pct; |
7bd95320 | 1549 | int dist; |
fb13c7ee MG |
1550 | |
1551 | struct task_struct *best_task; | |
1552 | long best_imp; | |
58d081b5 MG |
1553 | int best_cpu; |
1554 | }; | |
1555 | ||
6499b1b2 | 1556 | static unsigned long cpu_load(struct rq *rq); |
8e0e0eda | 1557 | static unsigned long cpu_runnable(struct rq *rq); |
6499b1b2 | 1558 | static unsigned long cpu_util(int cpu); |
7d2b5dd0 MG |
1559 | static inline long adjust_numa_imbalance(int imbalance, |
1560 | int dst_running, int dst_weight); | |
6499b1b2 VG |
1561 | |
1562 | static inline enum | |
1563 | numa_type numa_classify(unsigned int imbalance_pct, | |
1564 | struct numa_stats *ns) | |
1565 | { | |
1566 | if ((ns->nr_running > ns->weight) && | |
8e0e0eda VG |
1567 | (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || |
1568 | ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) | |
6499b1b2 VG |
1569 | return node_overloaded; |
1570 | ||
1571 | if ((ns->nr_running < ns->weight) || | |
8e0e0eda VG |
1572 | (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && |
1573 | ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) | |
6499b1b2 VG |
1574 | return node_has_spare; |
1575 | ||
1576 | return node_fully_busy; | |
1577 | } | |
1578 | ||
76c389ab VS |
1579 | #ifdef CONFIG_SCHED_SMT |
1580 | /* Forward declarations of select_idle_sibling helpers */ | |
1581 | static inline bool test_idle_cores(int cpu, bool def); | |
ff7db0bf MG |
1582 | static inline int numa_idle_core(int idle_core, int cpu) |
1583 | { | |
ff7db0bf MG |
1584 | if (!static_branch_likely(&sched_smt_present) || |
1585 | idle_core >= 0 || !test_idle_cores(cpu, false)) | |
1586 | return idle_core; | |
1587 | ||
1588 | /* | |
1589 | * Prefer cores instead of packing HT siblings | |
1590 | * and triggering future load balancing. | |
1591 | */ | |
1592 | if (is_core_idle(cpu)) | |
1593 | idle_core = cpu; | |
ff7db0bf MG |
1594 | |
1595 | return idle_core; | |
1596 | } | |
76c389ab VS |
1597 | #else |
1598 | static inline int numa_idle_core(int idle_core, int cpu) | |
1599 | { | |
1600 | return idle_core; | |
1601 | } | |
1602 | #endif | |
ff7db0bf | 1603 | |
6499b1b2 | 1604 | /* |
ff7db0bf MG |
1605 | * Gather all necessary information to make NUMA balancing placement |
1606 | * decisions that are compatible with standard load balancer. This | |
1607 | * borrows code and logic from update_sg_lb_stats but sharing a | |
1608 | * common implementation is impractical. | |
6499b1b2 VG |
1609 | */ |
1610 | static void update_numa_stats(struct task_numa_env *env, | |
ff7db0bf MG |
1611 | struct numa_stats *ns, int nid, |
1612 | bool find_idle) | |
6499b1b2 | 1613 | { |
ff7db0bf | 1614 | int cpu, idle_core = -1; |
6499b1b2 VG |
1615 | |
1616 | memset(ns, 0, sizeof(*ns)); | |
ff7db0bf MG |
1617 | ns->idle_cpu = -1; |
1618 | ||
0621df31 | 1619 | rcu_read_lock(); |
6499b1b2 VG |
1620 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
1621 | struct rq *rq = cpu_rq(cpu); | |
1622 | ||
1623 | ns->load += cpu_load(rq); | |
8e0e0eda | 1624 | ns->runnable += cpu_runnable(rq); |
6499b1b2 VG |
1625 | ns->util += cpu_util(cpu); |
1626 | ns->nr_running += rq->cfs.h_nr_running; | |
1627 | ns->compute_capacity += capacity_of(cpu); | |
ff7db0bf MG |
1628 | |
1629 | if (find_idle && !rq->nr_running && idle_cpu(cpu)) { | |
1630 | if (READ_ONCE(rq->numa_migrate_on) || | |
1631 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) | |
1632 | continue; | |
1633 | ||
1634 | if (ns->idle_cpu == -1) | |
1635 | ns->idle_cpu = cpu; | |
1636 | ||
1637 | idle_core = numa_idle_core(idle_core, cpu); | |
1638 | } | |
6499b1b2 | 1639 | } |
0621df31 | 1640 | rcu_read_unlock(); |
6499b1b2 VG |
1641 | |
1642 | ns->weight = cpumask_weight(cpumask_of_node(nid)); | |
1643 | ||
1644 | ns->node_type = numa_classify(env->imbalance_pct, ns); | |
ff7db0bf MG |
1645 | |
1646 | if (idle_core >= 0) | |
1647 | ns->idle_cpu = idle_core; | |
6499b1b2 VG |
1648 | } |
1649 | ||
fb13c7ee MG |
1650 | static void task_numa_assign(struct task_numa_env *env, |
1651 | struct task_struct *p, long imp) | |
1652 | { | |
a4739eca SD |
1653 | struct rq *rq = cpu_rq(env->dst_cpu); |
1654 | ||
5fb52dd9 MG |
1655 | /* Check if run-queue part of active NUMA balance. */ |
1656 | if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { | |
1657 | int cpu; | |
1658 | int start = env->dst_cpu; | |
1659 | ||
1660 | /* Find alternative idle CPU. */ | |
1661 | for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { | |
1662 | if (cpu == env->best_cpu || !idle_cpu(cpu) || | |
1663 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { | |
1664 | continue; | |
1665 | } | |
1666 | ||
1667 | env->dst_cpu = cpu; | |
1668 | rq = cpu_rq(env->dst_cpu); | |
1669 | if (!xchg(&rq->numa_migrate_on, 1)) | |
1670 | goto assign; | |
1671 | } | |
1672 | ||
1673 | /* Failed to find an alternative idle CPU */ | |
a4739eca | 1674 | return; |
5fb52dd9 | 1675 | } |
a4739eca | 1676 | |
5fb52dd9 | 1677 | assign: |
a4739eca SD |
1678 | /* |
1679 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1680 | * found a better CPU to move/swap. | |
1681 | */ | |
5fb52dd9 | 1682 | if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { |
a4739eca SD |
1683 | rq = cpu_rq(env->best_cpu); |
1684 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1685 | } | |
1686 | ||
fb13c7ee MG |
1687 | if (env->best_task) |
1688 | put_task_struct(env->best_task); | |
bac78573 ON |
1689 | if (p) |
1690 | get_task_struct(p); | |
fb13c7ee MG |
1691 | |
1692 | env->best_task = p; | |
1693 | env->best_imp = imp; | |
1694 | env->best_cpu = env->dst_cpu; | |
1695 | } | |
1696 | ||
28a21745 | 1697 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1698 | struct task_numa_env *env) |
1699 | { | |
e4991b24 RR |
1700 | long imb, old_imb; |
1701 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1702 | long src_capacity, dst_capacity; |
1703 | ||
1704 | /* | |
1705 | * The load is corrected for the CPU capacity available on each node. | |
1706 | * | |
1707 | * src_load dst_load | |
1708 | * ------------ vs --------- | |
1709 | * src_capacity dst_capacity | |
1710 | */ | |
1711 | src_capacity = env->src_stats.compute_capacity; | |
1712 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1713 | |
5f95ba7a | 1714 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1715 | |
28a21745 | 1716 | orig_src_load = env->src_stats.load; |
e4991b24 | 1717 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1718 | |
5f95ba7a | 1719 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1720 | |
1721 | /* Would this change make things worse? */ | |
1722 | return (imb > old_imb); | |
e63da036 RR |
1723 | } |
1724 | ||
6fd98e77 SD |
1725 | /* |
1726 | * Maximum NUMA importance can be 1998 (2*999); | |
1727 | * SMALLIMP @ 30 would be close to 1998/64. | |
1728 | * Used to deter task migration. | |
1729 | */ | |
1730 | #define SMALLIMP 30 | |
1731 | ||
fb13c7ee MG |
1732 | /* |
1733 | * This checks if the overall compute and NUMA accesses of the system would | |
1734 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1735 | * into account that it might be best if task running on the dst_cpu should | |
1736 | * be exchanged with the source task | |
1737 | */ | |
a0f03b61 | 1738 | static bool task_numa_compare(struct task_numa_env *env, |
305c1fac | 1739 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1740 | { |
cb361d8c | 1741 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 1742 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 1743 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 1744 | struct task_struct *cur; |
28a21745 | 1745 | long src_load, dst_load; |
7bd95320 | 1746 | int dist = env->dist; |
cb361d8c JH |
1747 | long moveimp = imp; |
1748 | long load; | |
a0f03b61 | 1749 | bool stopsearch = false; |
fb13c7ee | 1750 | |
a4739eca | 1751 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
a0f03b61 | 1752 | return false; |
a4739eca | 1753 | |
fb13c7ee | 1754 | rcu_read_lock(); |
154abafc | 1755 | cur = rcu_dereference(dst_rq->curr); |
bac78573 | 1756 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) |
fb13c7ee MG |
1757 | cur = NULL; |
1758 | ||
7af68335 PZ |
1759 | /* |
1760 | * Because we have preemption enabled we can get migrated around and | |
1761 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1762 | */ | |
a0f03b61 MG |
1763 | if (cur == env->p) { |
1764 | stopsearch = true; | |
7af68335 | 1765 | goto unlock; |
a0f03b61 | 1766 | } |
7af68335 | 1767 | |
305c1fac | 1768 | if (!cur) { |
6fd98e77 | 1769 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1770 | goto assign; |
1771 | else | |
1772 | goto unlock; | |
1773 | } | |
1774 | ||
88cca72c MG |
1775 | /* Skip this swap candidate if cannot move to the source cpu. */ |
1776 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) | |
1777 | goto unlock; | |
1778 | ||
1779 | /* | |
1780 | * Skip this swap candidate if it is not moving to its preferred | |
1781 | * node and the best task is. | |
1782 | */ | |
1783 | if (env->best_task && | |
1784 | env->best_task->numa_preferred_nid == env->src_nid && | |
1785 | cur->numa_preferred_nid != env->src_nid) { | |
1786 | goto unlock; | |
1787 | } | |
1788 | ||
fb13c7ee MG |
1789 | /* |
1790 | * "imp" is the fault differential for the source task between the | |
1791 | * source and destination node. Calculate the total differential for | |
1792 | * the source task and potential destination task. The more negative | |
305c1fac | 1793 | * the value is, the more remote accesses that would be expected to |
fb13c7ee | 1794 | * be incurred if the tasks were swapped. |
88cca72c | 1795 | * |
305c1fac SD |
1796 | * If dst and source tasks are in the same NUMA group, or not |
1797 | * in any group then look only at task weights. | |
1798 | */ | |
cb361d8c JH |
1799 | cur_ng = rcu_dereference(cur->numa_group); |
1800 | if (cur_ng == p_ng) { | |
305c1fac SD |
1801 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1802 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1803 | /* |
305c1fac SD |
1804 | * Add some hysteresis to prevent swapping the |
1805 | * tasks within a group over tiny differences. | |
887c290e | 1806 | */ |
cb361d8c | 1807 | if (cur_ng) |
305c1fac SD |
1808 | imp -= imp / 16; |
1809 | } else { | |
1810 | /* | |
1811 | * Compare the group weights. If a task is all by itself | |
1812 | * (not part of a group), use the task weight instead. | |
1813 | */ | |
cb361d8c | 1814 | if (cur_ng && p_ng) |
305c1fac SD |
1815 | imp += group_weight(cur, env->src_nid, dist) - |
1816 | group_weight(cur, env->dst_nid, dist); | |
1817 | else | |
1818 | imp += task_weight(cur, env->src_nid, dist) - | |
1819 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
1820 | } |
1821 | ||
88cca72c MG |
1822 | /* Discourage picking a task already on its preferred node */ |
1823 | if (cur->numa_preferred_nid == env->dst_nid) | |
1824 | imp -= imp / 16; | |
1825 | ||
1826 | /* | |
1827 | * Encourage picking a task that moves to its preferred node. | |
1828 | * This potentially makes imp larger than it's maximum of | |
1829 | * 1998 (see SMALLIMP and task_weight for why) but in this | |
1830 | * case, it does not matter. | |
1831 | */ | |
1832 | if (cur->numa_preferred_nid == env->src_nid) | |
1833 | imp += imp / 8; | |
1834 | ||
305c1fac | 1835 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 1836 | imp = moveimp; |
305c1fac | 1837 | cur = NULL; |
fb13c7ee | 1838 | goto assign; |
305c1fac | 1839 | } |
fb13c7ee | 1840 | |
88cca72c MG |
1841 | /* |
1842 | * Prefer swapping with a task moving to its preferred node over a | |
1843 | * task that is not. | |
1844 | */ | |
1845 | if (env->best_task && cur->numa_preferred_nid == env->src_nid && | |
1846 | env->best_task->numa_preferred_nid != env->src_nid) { | |
1847 | goto assign; | |
1848 | } | |
1849 | ||
6fd98e77 SD |
1850 | /* |
1851 | * If the NUMA importance is less than SMALLIMP, | |
1852 | * task migration might only result in ping pong | |
1853 | * of tasks and also hurt performance due to cache | |
1854 | * misses. | |
1855 | */ | |
1856 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
1857 | goto unlock; | |
1858 | ||
fb13c7ee MG |
1859 | /* |
1860 | * In the overloaded case, try and keep the load balanced. | |
1861 | */ | |
305c1fac SD |
1862 | load = task_h_load(env->p) - task_h_load(cur); |
1863 | if (!load) | |
1864 | goto assign; | |
1865 | ||
e720fff6 PZ |
1866 | dst_load = env->dst_stats.load + load; |
1867 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1868 | |
28a21745 | 1869 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1870 | goto unlock; |
1871 | ||
305c1fac | 1872 | assign: |
ff7db0bf | 1873 | /* Evaluate an idle CPU for a task numa move. */ |
10e2f1ac | 1874 | if (!cur) { |
ff7db0bf MG |
1875 | int cpu = env->dst_stats.idle_cpu; |
1876 | ||
1877 | /* Nothing cached so current CPU went idle since the search. */ | |
1878 | if (cpu < 0) | |
1879 | cpu = env->dst_cpu; | |
1880 | ||
10e2f1ac | 1881 | /* |
ff7db0bf MG |
1882 | * If the CPU is no longer truly idle and the previous best CPU |
1883 | * is, keep using it. | |
10e2f1ac | 1884 | */ |
ff7db0bf MG |
1885 | if (!idle_cpu(cpu) && env->best_cpu >= 0 && |
1886 | idle_cpu(env->best_cpu)) { | |
1887 | cpu = env->best_cpu; | |
1888 | } | |
1889 | ||
ff7db0bf | 1890 | env->dst_cpu = cpu; |
10e2f1ac | 1891 | } |
ba7e5a27 | 1892 | |
fb13c7ee | 1893 | task_numa_assign(env, cur, imp); |
a0f03b61 MG |
1894 | |
1895 | /* | |
1896 | * If a move to idle is allowed because there is capacity or load | |
1897 | * balance improves then stop the search. While a better swap | |
1898 | * candidate may exist, a search is not free. | |
1899 | */ | |
1900 | if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) | |
1901 | stopsearch = true; | |
1902 | ||
1903 | /* | |
1904 | * If a swap candidate must be identified and the current best task | |
1905 | * moves its preferred node then stop the search. | |
1906 | */ | |
1907 | if (!maymove && env->best_task && | |
1908 | env->best_task->numa_preferred_nid == env->src_nid) { | |
1909 | stopsearch = true; | |
1910 | } | |
fb13c7ee MG |
1911 | unlock: |
1912 | rcu_read_unlock(); | |
a0f03b61 MG |
1913 | |
1914 | return stopsearch; | |
fb13c7ee MG |
1915 | } |
1916 | ||
887c290e RR |
1917 | static void task_numa_find_cpu(struct task_numa_env *env, |
1918 | long taskimp, long groupimp) | |
2c8a50aa | 1919 | { |
305c1fac | 1920 | bool maymove = false; |
2c8a50aa MG |
1921 | int cpu; |
1922 | ||
305c1fac | 1923 | /* |
fb86f5b2 MG |
1924 | * If dst node has spare capacity, then check if there is an |
1925 | * imbalance that would be overruled by the load balancer. | |
305c1fac | 1926 | */ |
fb86f5b2 MG |
1927 | if (env->dst_stats.node_type == node_has_spare) { |
1928 | unsigned int imbalance; | |
1929 | int src_running, dst_running; | |
1930 | ||
1931 | /* | |
1932 | * Would movement cause an imbalance? Note that if src has | |
1933 | * more running tasks that the imbalance is ignored as the | |
1934 | * move improves the imbalance from the perspective of the | |
1935 | * CPU load balancer. | |
1936 | * */ | |
1937 | src_running = env->src_stats.nr_running - 1; | |
1938 | dst_running = env->dst_stats.nr_running + 1; | |
1939 | imbalance = max(0, dst_running - src_running); | |
7d2b5dd0 MG |
1940 | imbalance = adjust_numa_imbalance(imbalance, dst_running, |
1941 | env->dst_stats.weight); | |
fb86f5b2 MG |
1942 | |
1943 | /* Use idle CPU if there is no imbalance */ | |
ff7db0bf | 1944 | if (!imbalance) { |
fb86f5b2 | 1945 | maymove = true; |
ff7db0bf MG |
1946 | if (env->dst_stats.idle_cpu >= 0) { |
1947 | env->dst_cpu = env->dst_stats.idle_cpu; | |
1948 | task_numa_assign(env, NULL, 0); | |
1949 | return; | |
1950 | } | |
1951 | } | |
fb86f5b2 MG |
1952 | } else { |
1953 | long src_load, dst_load, load; | |
1954 | /* | |
1955 | * If the improvement from just moving env->p direction is better | |
1956 | * than swapping tasks around, check if a move is possible. | |
1957 | */ | |
1958 | load = task_h_load(env->p); | |
1959 | dst_load = env->dst_stats.load + load; | |
1960 | src_load = env->src_stats.load - load; | |
1961 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
1962 | } | |
305c1fac | 1963 | |
2c8a50aa MG |
1964 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1965 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 1966 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
1967 | continue; |
1968 | ||
1969 | env->dst_cpu = cpu; | |
a0f03b61 MG |
1970 | if (task_numa_compare(env, taskimp, groupimp, maymove)) |
1971 | break; | |
2c8a50aa MG |
1972 | } |
1973 | } | |
1974 | ||
58d081b5 MG |
1975 | static int task_numa_migrate(struct task_struct *p) |
1976 | { | |
58d081b5 MG |
1977 | struct task_numa_env env = { |
1978 | .p = p, | |
fb13c7ee | 1979 | |
58d081b5 | 1980 | .src_cpu = task_cpu(p), |
b32e86b4 | 1981 | .src_nid = task_node(p), |
fb13c7ee MG |
1982 | |
1983 | .imbalance_pct = 112, | |
1984 | ||
1985 | .best_task = NULL, | |
1986 | .best_imp = 0, | |
4142c3eb | 1987 | .best_cpu = -1, |
58d081b5 | 1988 | }; |
cb361d8c | 1989 | unsigned long taskweight, groupweight; |
58d081b5 | 1990 | struct sched_domain *sd; |
cb361d8c JH |
1991 | long taskimp, groupimp; |
1992 | struct numa_group *ng; | |
a4739eca | 1993 | struct rq *best_rq; |
7bd95320 | 1994 | int nid, ret, dist; |
e6628d5b | 1995 | |
58d081b5 | 1996 | /* |
fb13c7ee MG |
1997 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1998 | * imbalance and would be the first to start moving tasks about. | |
1999 | * | |
2000 | * And we want to avoid any moving of tasks about, as that would create | |
2001 | * random movement of tasks -- counter the numa conditions we're trying | |
2002 | * to satisfy here. | |
58d081b5 MG |
2003 | */ |
2004 | rcu_read_lock(); | |
fb13c7ee | 2005 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
2006 | if (sd) |
2007 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
2008 | rcu_read_unlock(); |
2009 | ||
46a73e8a RR |
2010 | /* |
2011 | * Cpusets can break the scheduler domain tree into smaller | |
2012 | * balance domains, some of which do not cross NUMA boundaries. | |
2013 | * Tasks that are "trapped" in such domains cannot be migrated | |
2014 | * elsewhere, so there is no point in (re)trying. | |
2015 | */ | |
2016 | if (unlikely(!sd)) { | |
8cd45eee | 2017 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
2018 | return -EINVAL; |
2019 | } | |
2020 | ||
2c8a50aa | 2021 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
2022 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
2023 | taskweight = task_weight(p, env.src_nid, dist); | |
2024 | groupweight = group_weight(p, env.src_nid, dist); | |
ff7db0bf | 2025 | update_numa_stats(&env, &env.src_stats, env.src_nid, false); |
7bd95320 RR |
2026 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
2027 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
ff7db0bf | 2028 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
58d081b5 | 2029 | |
a43455a1 | 2030 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 2031 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 2032 | |
9de05d48 RR |
2033 | /* |
2034 | * Look at other nodes in these cases: | |
2035 | * - there is no space available on the preferred_nid | |
2036 | * - the task is part of a numa_group that is interleaved across | |
2037 | * multiple NUMA nodes; in order to better consolidate the group, | |
2038 | * we need to check other locations. | |
2039 | */ | |
cb361d8c JH |
2040 | ng = deref_curr_numa_group(p); |
2041 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
2c8a50aa MG |
2042 | for_each_online_node(nid) { |
2043 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
2044 | continue; | |
58d081b5 | 2045 | |
7bd95320 | 2046 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
2047 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
2048 | dist != env.dist) { | |
2049 | taskweight = task_weight(p, env.src_nid, dist); | |
2050 | groupweight = group_weight(p, env.src_nid, dist); | |
2051 | } | |
7bd95320 | 2052 | |
83e1d2cd | 2053 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
2054 | taskimp = task_weight(p, nid, dist) - taskweight; |
2055 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 2056 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
2057 | continue; |
2058 | ||
7bd95320 | 2059 | env.dist = dist; |
2c8a50aa | 2060 | env.dst_nid = nid; |
ff7db0bf | 2061 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
2d4056fa | 2062 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
2063 | } |
2064 | } | |
2065 | ||
68d1b02a RR |
2066 | /* |
2067 | * If the task is part of a workload that spans multiple NUMA nodes, | |
2068 | * and is migrating into one of the workload's active nodes, remember | |
2069 | * this node as the task's preferred numa node, so the workload can | |
2070 | * settle down. | |
2071 | * A task that migrated to a second choice node will be better off | |
2072 | * trying for a better one later. Do not set the preferred node here. | |
2073 | */ | |
cb361d8c | 2074 | if (ng) { |
db015dae RR |
2075 | if (env.best_cpu == -1) |
2076 | nid = env.src_nid; | |
2077 | else | |
8cd45eee | 2078 | nid = cpu_to_node(env.best_cpu); |
db015dae | 2079 | |
8cd45eee SD |
2080 | if (nid != p->numa_preferred_nid) |
2081 | sched_setnuma(p, nid); | |
db015dae RR |
2082 | } |
2083 | ||
2084 | /* No better CPU than the current one was found. */ | |
f22aef4a | 2085 | if (env.best_cpu == -1) { |
b2b2042b | 2086 | trace_sched_stick_numa(p, env.src_cpu, NULL, -1); |
db015dae | 2087 | return -EAGAIN; |
f22aef4a | 2088 | } |
0ec8aa00 | 2089 | |
a4739eca | 2090 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 2091 | if (env.best_task == NULL) { |
286549dc | 2092 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 2093 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc | 2094 | if (ret != 0) |
b2b2042b | 2095 | trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); |
fb13c7ee MG |
2096 | return ret; |
2097 | } | |
2098 | ||
0ad4e3df | 2099 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 2100 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 2101 | |
286549dc | 2102 | if (ret != 0) |
b2b2042b | 2103 | trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); |
fb13c7ee MG |
2104 | put_task_struct(env.best_task); |
2105 | return ret; | |
e6628d5b MG |
2106 | } |
2107 | ||
6b9a7460 MG |
2108 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
2109 | static void numa_migrate_preferred(struct task_struct *p) | |
2110 | { | |
5085e2a3 RR |
2111 | unsigned long interval = HZ; |
2112 | ||
2739d3ee | 2113 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 2114 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
2115 | return; |
2116 | ||
2739d3ee | 2117 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 2118 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 2119 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
2120 | |
2121 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 2122 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
2123 | return; |
2124 | ||
2125 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 2126 | task_numa_migrate(p); |
6b9a7460 MG |
2127 | } |
2128 | ||
20e07dea | 2129 | /* |
4142c3eb | 2130 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
2131 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
2132 | * be different from the set of nodes where the workload's memory is currently | |
2133 | * located. | |
20e07dea | 2134 | */ |
4142c3eb | 2135 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
2136 | { |
2137 | unsigned long faults, max_faults = 0; | |
4142c3eb | 2138 | int nid, active_nodes = 0; |
20e07dea RR |
2139 | |
2140 | for_each_online_node(nid) { | |
2141 | faults = group_faults_cpu(numa_group, nid); | |
2142 | if (faults > max_faults) | |
2143 | max_faults = faults; | |
2144 | } | |
2145 | ||
2146 | for_each_online_node(nid) { | |
2147 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
2148 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
2149 | active_nodes++; | |
20e07dea | 2150 | } |
4142c3eb RR |
2151 | |
2152 | numa_group->max_faults_cpu = max_faults; | |
2153 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
2154 | } |
2155 | ||
04bb2f94 RR |
2156 | /* |
2157 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
2158 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
2159 | * period will be for the next scan window. If local/(local+remote) ratio is |
2160 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
2161 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
2162 | */ |
2163 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 2164 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
2165 | |
2166 | /* | |
2167 | * Increase the scan period (slow down scanning) if the majority of | |
2168 | * our memory is already on our local node, or if the majority of | |
2169 | * the page accesses are shared with other processes. | |
2170 | * Otherwise, decrease the scan period. | |
2171 | */ | |
2172 | static void update_task_scan_period(struct task_struct *p, | |
2173 | unsigned long shared, unsigned long private) | |
2174 | { | |
2175 | unsigned int period_slot; | |
37ec97de | 2176 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
2177 | int diff; |
2178 | ||
2179 | unsigned long remote = p->numa_faults_locality[0]; | |
2180 | unsigned long local = p->numa_faults_locality[1]; | |
2181 | ||
2182 | /* | |
2183 | * If there were no record hinting faults then either the task is | |
2184 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
2185 | * to automatic numa balancing. Related to that, if there were failed |
2186 | * migration then it implies we are migrating too quickly or the local | |
2187 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 2188 | */ |
074c2381 | 2189 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
2190 | p->numa_scan_period = min(p->numa_scan_period_max, |
2191 | p->numa_scan_period << 1); | |
2192 | ||
2193 | p->mm->numa_next_scan = jiffies + | |
2194 | msecs_to_jiffies(p->numa_scan_period); | |
2195 | ||
2196 | return; | |
2197 | } | |
2198 | ||
2199 | /* | |
2200 | * Prepare to scale scan period relative to the current period. | |
2201 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
2202 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
2203 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
2204 | */ | |
2205 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
2206 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
2207 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
2208 | ||
2209 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2210 | /* | |
2211 | * Most memory accesses are local. There is no need to | |
2212 | * do fast NUMA scanning, since memory is already local. | |
2213 | */ | |
2214 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
2215 | if (!slot) | |
2216 | slot = 1; | |
2217 | diff = slot * period_slot; | |
2218 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2219 | /* | |
2220 | * Most memory accesses are shared with other tasks. | |
2221 | * There is no point in continuing fast NUMA scanning, | |
2222 | * since other tasks may just move the memory elsewhere. | |
2223 | */ | |
2224 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
2225 | if (!slot) |
2226 | slot = 1; | |
2227 | diff = slot * period_slot; | |
2228 | } else { | |
04bb2f94 | 2229 | /* |
37ec97de RR |
2230 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
2231 | * yet they are not on the local NUMA node. Speed up | |
2232 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 2233 | */ |
37ec97de RR |
2234 | int ratio = max(lr_ratio, ps_ratio); |
2235 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
2236 | } |
2237 | ||
2238 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2239 | task_scan_min(p), task_scan_max(p)); | |
2240 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2241 | } | |
2242 | ||
7e2703e6 RR |
2243 | /* |
2244 | * Get the fraction of time the task has been running since the last | |
2245 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2246 | * decays those on a 32ms period, which is orders of magnitude off | |
2247 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2248 | * stats only if the task is so new there are no NUMA statistics yet. | |
2249 | */ | |
2250 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2251 | { | |
2252 | u64 runtime, delta, now; | |
2253 | /* Use the start of this time slice to avoid calculations. */ | |
2254 | now = p->se.exec_start; | |
2255 | runtime = p->se.sum_exec_runtime; | |
2256 | ||
2257 | if (p->last_task_numa_placement) { | |
2258 | delta = runtime - p->last_sum_exec_runtime; | |
2259 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2260 | |
2261 | /* Avoid time going backwards, prevent potential divide error: */ | |
2262 | if (unlikely((s64)*period < 0)) | |
2263 | *period = 0; | |
7e2703e6 | 2264 | } else { |
c7b50216 | 2265 | delta = p->se.avg.load_sum; |
9d89c257 | 2266 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2267 | } |
2268 | ||
2269 | p->last_sum_exec_runtime = runtime; | |
2270 | p->last_task_numa_placement = now; | |
2271 | ||
2272 | return delta; | |
2273 | } | |
2274 | ||
54009416 RR |
2275 | /* |
2276 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2277 | * be done in a way that produces consistent results with group_weight, | |
2278 | * otherwise workloads might not converge. | |
2279 | */ | |
2280 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2281 | { | |
2282 | nodemask_t nodes; | |
2283 | int dist; | |
2284 | ||
2285 | /* Direct connections between all NUMA nodes. */ | |
2286 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2287 | return nid; | |
2288 | ||
2289 | /* | |
2290 | * On a system with glueless mesh NUMA topology, group_weight | |
2291 | * scores nodes according to the number of NUMA hinting faults on | |
2292 | * both the node itself, and on nearby nodes. | |
2293 | */ | |
2294 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2295 | unsigned long score, max_score = 0; | |
2296 | int node, max_node = nid; | |
2297 | ||
2298 | dist = sched_max_numa_distance; | |
2299 | ||
2300 | for_each_online_node(node) { | |
2301 | score = group_weight(p, node, dist); | |
2302 | if (score > max_score) { | |
2303 | max_score = score; | |
2304 | max_node = node; | |
2305 | } | |
2306 | } | |
2307 | return max_node; | |
2308 | } | |
2309 | ||
2310 | /* | |
2311 | * Finding the preferred nid in a system with NUMA backplane | |
2312 | * interconnect topology is more involved. The goal is to locate | |
2313 | * tasks from numa_groups near each other in the system, and | |
2314 | * untangle workloads from different sides of the system. This requires | |
2315 | * searching down the hierarchy of node groups, recursively searching | |
2316 | * inside the highest scoring group of nodes. The nodemask tricks | |
2317 | * keep the complexity of the search down. | |
2318 | */ | |
2319 | nodes = node_online_map; | |
2320 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2321 | unsigned long max_faults = 0; | |
81907478 | 2322 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2323 | int a, b; |
2324 | ||
2325 | /* Are there nodes at this distance from each other? */ | |
2326 | if (!find_numa_distance(dist)) | |
2327 | continue; | |
2328 | ||
2329 | for_each_node_mask(a, nodes) { | |
2330 | unsigned long faults = 0; | |
2331 | nodemask_t this_group; | |
2332 | nodes_clear(this_group); | |
2333 | ||
2334 | /* Sum group's NUMA faults; includes a==b case. */ | |
2335 | for_each_node_mask(b, nodes) { | |
2336 | if (node_distance(a, b) < dist) { | |
2337 | faults += group_faults(p, b); | |
2338 | node_set(b, this_group); | |
2339 | node_clear(b, nodes); | |
2340 | } | |
2341 | } | |
2342 | ||
2343 | /* Remember the top group. */ | |
2344 | if (faults > max_faults) { | |
2345 | max_faults = faults; | |
2346 | max_group = this_group; | |
2347 | /* | |
2348 | * subtle: at the smallest distance there is | |
2349 | * just one node left in each "group", the | |
2350 | * winner is the preferred nid. | |
2351 | */ | |
2352 | nid = a; | |
2353 | } | |
2354 | } | |
2355 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2356 | if (!max_faults) |
2357 | break; | |
54009416 RR |
2358 | nodes = max_group; |
2359 | } | |
2360 | return nid; | |
2361 | } | |
2362 | ||
cbee9f88 PZ |
2363 | static void task_numa_placement(struct task_struct *p) |
2364 | { | |
98fa15f3 | 2365 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2366 | unsigned long max_faults = 0; |
04bb2f94 | 2367 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2368 | unsigned long total_faults; |
2369 | u64 runtime, period; | |
7dbd13ed | 2370 | spinlock_t *group_lock = NULL; |
cb361d8c | 2371 | struct numa_group *ng; |
cbee9f88 | 2372 | |
7e5a2c17 JL |
2373 | /* |
2374 | * The p->mm->numa_scan_seq field gets updated without | |
2375 | * exclusive access. Use READ_ONCE() here to ensure | |
2376 | * that the field is read in a single access: | |
2377 | */ | |
316c1608 | 2378 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2379 | if (p->numa_scan_seq == seq) |
2380 | return; | |
2381 | p->numa_scan_seq = seq; | |
598f0ec0 | 2382 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2383 | |
7e2703e6 RR |
2384 | total_faults = p->numa_faults_locality[0] + |
2385 | p->numa_faults_locality[1]; | |
2386 | runtime = numa_get_avg_runtime(p, &period); | |
2387 | ||
7dbd13ed | 2388 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2389 | ng = deref_curr_numa_group(p); |
2390 | if (ng) { | |
2391 | group_lock = &ng->lock; | |
60e69eed | 2392 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2393 | } |
2394 | ||
688b7585 MG |
2395 | /* Find the node with the highest number of faults */ |
2396 | for_each_online_node(nid) { | |
44dba3d5 IM |
2397 | /* Keep track of the offsets in numa_faults array */ |
2398 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2399 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2400 | int priv; |
745d6147 | 2401 | |
be1e4e76 | 2402 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2403 | long diff, f_diff, f_weight; |
8c8a743c | 2404 | |
44dba3d5 IM |
2405 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2406 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2407 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2408 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2409 | |
ac8e895b | 2410 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2411 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2412 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2413 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2414 | |
7e2703e6 RR |
2415 | /* |
2416 | * Normalize the faults_from, so all tasks in a group | |
2417 | * count according to CPU use, instead of by the raw | |
2418 | * number of faults. Tasks with little runtime have | |
2419 | * little over-all impact on throughput, and thus their | |
2420 | * faults are less important. | |
2421 | */ | |
2422 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2423 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2424 | (total_faults + 1); |
44dba3d5 IM |
2425 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2426 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2427 | |
44dba3d5 IM |
2428 | p->numa_faults[mem_idx] += diff; |
2429 | p->numa_faults[cpu_idx] += f_diff; | |
2430 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2431 | p->total_numa_faults += diff; |
cb361d8c | 2432 | if (ng) { |
44dba3d5 IM |
2433 | /* |
2434 | * safe because we can only change our own group | |
2435 | * | |
2436 | * mem_idx represents the offset for a given | |
2437 | * nid and priv in a specific region because it | |
2438 | * is at the beginning of the numa_faults array. | |
2439 | */ | |
cb361d8c JH |
2440 | ng->faults[mem_idx] += diff; |
2441 | ng->faults_cpu[mem_idx] += f_diff; | |
2442 | ng->total_faults += diff; | |
2443 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2444 | } |
ac8e895b MG |
2445 | } |
2446 | ||
cb361d8c | 2447 | if (!ng) { |
f03bb676 SD |
2448 | if (faults > max_faults) { |
2449 | max_faults = faults; | |
2450 | max_nid = nid; | |
2451 | } | |
2452 | } else if (group_faults > max_faults) { | |
2453 | max_faults = group_faults; | |
688b7585 MG |
2454 | max_nid = nid; |
2455 | } | |
83e1d2cd MG |
2456 | } |
2457 | ||
cb361d8c JH |
2458 | if (ng) { |
2459 | numa_group_count_active_nodes(ng); | |
60e69eed | 2460 | spin_unlock_irq(group_lock); |
f03bb676 | 2461 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2462 | } |
2463 | ||
bb97fc31 RR |
2464 | if (max_faults) { |
2465 | /* Set the new preferred node */ | |
2466 | if (max_nid != p->numa_preferred_nid) | |
2467 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2468 | } |
30619c89 SD |
2469 | |
2470 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2471 | } |
2472 | ||
8c8a743c PZ |
2473 | static inline int get_numa_group(struct numa_group *grp) |
2474 | { | |
c45a7795 | 2475 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2476 | } |
2477 | ||
2478 | static inline void put_numa_group(struct numa_group *grp) | |
2479 | { | |
c45a7795 | 2480 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2481 | kfree_rcu(grp, rcu); |
2482 | } | |
2483 | ||
3e6a9418 MG |
2484 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2485 | int *priv) | |
8c8a743c PZ |
2486 | { |
2487 | struct numa_group *grp, *my_grp; | |
2488 | struct task_struct *tsk; | |
2489 | bool join = false; | |
2490 | int cpu = cpupid_to_cpu(cpupid); | |
2491 | int i; | |
2492 | ||
cb361d8c | 2493 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2494 | unsigned int size = sizeof(struct numa_group) + |
50ec8a40 | 2495 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2496 | |
2497 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2498 | if (!grp) | |
2499 | return; | |
2500 | ||
c45a7795 | 2501 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2502 | grp->active_nodes = 1; |
2503 | grp->max_faults_cpu = 0; | |
8c8a743c | 2504 | spin_lock_init(&grp->lock); |
e29cf08b | 2505 | grp->gid = p->pid; |
50ec8a40 | 2506 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2507 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2508 | nr_node_ids; | |
8c8a743c | 2509 | |
be1e4e76 | 2510 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2511 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2512 | |
989348b5 | 2513 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2514 | |
8c8a743c PZ |
2515 | grp->nr_tasks++; |
2516 | rcu_assign_pointer(p->numa_group, grp); | |
2517 | } | |
2518 | ||
2519 | rcu_read_lock(); | |
316c1608 | 2520 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2521 | |
2522 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2523 | goto no_join; |
8c8a743c PZ |
2524 | |
2525 | grp = rcu_dereference(tsk->numa_group); | |
2526 | if (!grp) | |
3354781a | 2527 | goto no_join; |
8c8a743c | 2528 | |
cb361d8c | 2529 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2530 | if (grp == my_grp) |
3354781a | 2531 | goto no_join; |
8c8a743c PZ |
2532 | |
2533 | /* | |
2534 | * Only join the other group if its bigger; if we're the bigger group, | |
2535 | * the other task will join us. | |
2536 | */ | |
2537 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2538 | goto no_join; |
8c8a743c PZ |
2539 | |
2540 | /* | |
2541 | * Tie-break on the grp address. | |
2542 | */ | |
2543 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2544 | goto no_join; |
8c8a743c | 2545 | |
dabe1d99 RR |
2546 | /* Always join threads in the same process. */ |
2547 | if (tsk->mm == current->mm) | |
2548 | join = true; | |
2549 | ||
2550 | /* Simple filter to avoid false positives due to PID collisions */ | |
2551 | if (flags & TNF_SHARED) | |
2552 | join = true; | |
8c8a743c | 2553 | |
3e6a9418 MG |
2554 | /* Update priv based on whether false sharing was detected */ |
2555 | *priv = !join; | |
2556 | ||
dabe1d99 | 2557 | if (join && !get_numa_group(grp)) |
3354781a | 2558 | goto no_join; |
8c8a743c | 2559 | |
8c8a743c PZ |
2560 | rcu_read_unlock(); |
2561 | ||
2562 | if (!join) | |
2563 | return; | |
2564 | ||
60e69eed MG |
2565 | BUG_ON(irqs_disabled()); |
2566 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2567 | |
be1e4e76 | 2568 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2569 | my_grp->faults[i] -= p->numa_faults[i]; |
2570 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2571 | } |
989348b5 MG |
2572 | my_grp->total_faults -= p->total_numa_faults; |
2573 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2574 | |
8c8a743c PZ |
2575 | my_grp->nr_tasks--; |
2576 | grp->nr_tasks++; | |
2577 | ||
2578 | spin_unlock(&my_grp->lock); | |
60e69eed | 2579 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2580 | |
2581 | rcu_assign_pointer(p->numa_group, grp); | |
2582 | ||
2583 | put_numa_group(my_grp); | |
3354781a PZ |
2584 | return; |
2585 | ||
2586 | no_join: | |
2587 | rcu_read_unlock(); | |
2588 | return; | |
8c8a743c PZ |
2589 | } |
2590 | ||
16d51a59 | 2591 | /* |
3b03706f | 2592 | * Get rid of NUMA statistics associated with a task (either current or dead). |
16d51a59 JH |
2593 | * If @final is set, the task is dead and has reached refcount zero, so we can |
2594 | * safely free all relevant data structures. Otherwise, there might be | |
2595 | * concurrent reads from places like load balancing and procfs, and we should | |
2596 | * reset the data back to default state without freeing ->numa_faults. | |
2597 | */ | |
2598 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 2599 | { |
cb361d8c JH |
2600 | /* safe: p either is current or is being freed by current */ |
2601 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 2602 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
2603 | unsigned long flags; |
2604 | int i; | |
8c8a743c | 2605 | |
16d51a59 JH |
2606 | if (!numa_faults) |
2607 | return; | |
2608 | ||
8c8a743c | 2609 | if (grp) { |
e9dd685c | 2610 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2611 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2612 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2613 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2614 | |
8c8a743c | 2615 | grp->nr_tasks--; |
e9dd685c | 2616 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2617 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2618 | put_numa_group(grp); |
2619 | } | |
2620 | ||
16d51a59 JH |
2621 | if (final) { |
2622 | p->numa_faults = NULL; | |
2623 | kfree(numa_faults); | |
2624 | } else { | |
2625 | p->total_numa_faults = 0; | |
2626 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
2627 | numa_faults[i] = 0; | |
2628 | } | |
8c8a743c PZ |
2629 | } |
2630 | ||
cbee9f88 PZ |
2631 | /* |
2632 | * Got a PROT_NONE fault for a page on @node. | |
2633 | */ | |
58b46da3 | 2634 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2635 | { |
2636 | struct task_struct *p = current; | |
6688cc05 | 2637 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2638 | int cpu_node = task_node(current); |
792568ec | 2639 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2640 | struct numa_group *ng; |
ac8e895b | 2641 | int priv; |
cbee9f88 | 2642 | |
2a595721 | 2643 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2644 | return; |
2645 | ||
9ff1d9ff MG |
2646 | /* for example, ksmd faulting in a user's mm */ |
2647 | if (!p->mm) | |
2648 | return; | |
2649 | ||
f809ca9a | 2650 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2651 | if (unlikely(!p->numa_faults)) { |
2652 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2653 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2654 | |
44dba3d5 IM |
2655 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2656 | if (!p->numa_faults) | |
f809ca9a | 2657 | return; |
745d6147 | 2658 | |
83e1d2cd | 2659 | p->total_numa_faults = 0; |
04bb2f94 | 2660 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2661 | } |
cbee9f88 | 2662 | |
8c8a743c PZ |
2663 | /* |
2664 | * First accesses are treated as private, otherwise consider accesses | |
2665 | * to be private if the accessing pid has not changed | |
2666 | */ | |
2667 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2668 | priv = 1; | |
2669 | } else { | |
2670 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2671 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2672 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2673 | } |
2674 | ||
792568ec RR |
2675 | /* |
2676 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2677 | * occurs wholly within the set of nodes that the workload is | |
2678 | * actively using should be counted as local. This allows the | |
2679 | * scan rate to slow down when a workload has settled down. | |
2680 | */ | |
cb361d8c | 2681 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
2682 | if (!priv && !local && ng && ng->active_nodes > 1 && |
2683 | numa_is_active_node(cpu_node, ng) && | |
2684 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2685 | local = 1; |
2686 | ||
2739d3ee | 2687 | /* |
e1ff516a YW |
2688 | * Retry to migrate task to preferred node periodically, in case it |
2689 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2690 | */ |
b6a60cf3 SD |
2691 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2692 | task_numa_placement(p); | |
6b9a7460 | 2693 | numa_migrate_preferred(p); |
b6a60cf3 | 2694 | } |
6b9a7460 | 2695 | |
b32e86b4 IM |
2696 | if (migrated) |
2697 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2698 | if (flags & TNF_MIGRATE_FAIL) |
2699 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2700 | |
44dba3d5 IM |
2701 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2702 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2703 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2704 | } |
2705 | ||
6e5fb223 PZ |
2706 | static void reset_ptenuma_scan(struct task_struct *p) |
2707 | { | |
7e5a2c17 JL |
2708 | /* |
2709 | * We only did a read acquisition of the mmap sem, so | |
2710 | * p->mm->numa_scan_seq is written to without exclusive access | |
2711 | * and the update is not guaranteed to be atomic. That's not | |
2712 | * much of an issue though, since this is just used for | |
2713 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2714 | * expensive, to avoid any form of compiler optimizations: | |
2715 | */ | |
316c1608 | 2716 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2717 | p->mm->numa_scan_offset = 0; |
2718 | } | |
2719 | ||
cbee9f88 PZ |
2720 | /* |
2721 | * The expensive part of numa migration is done from task_work context. | |
2722 | * Triggered from task_tick_numa(). | |
2723 | */ | |
9434f9f5 | 2724 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
2725 | { |
2726 | unsigned long migrate, next_scan, now = jiffies; | |
2727 | struct task_struct *p = current; | |
2728 | struct mm_struct *mm = p->mm; | |
51170840 | 2729 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2730 | struct vm_area_struct *vma; |
9f40604c | 2731 | unsigned long start, end; |
598f0ec0 | 2732 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2733 | long pages, virtpages; |
cbee9f88 | 2734 | |
9148a3a1 | 2735 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 2736 | |
b34920d4 | 2737 | work->next = work; |
cbee9f88 PZ |
2738 | /* |
2739 | * Who cares about NUMA placement when they're dying. | |
2740 | * | |
2741 | * NOTE: make sure not to dereference p->mm before this check, | |
2742 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2743 | * without p->mm even though we still had it when we enqueued this | |
2744 | * work. | |
2745 | */ | |
2746 | if (p->flags & PF_EXITING) | |
2747 | return; | |
2748 | ||
930aa174 | 2749 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2750 | mm->numa_next_scan = now + |
2751 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2752 | } |
2753 | ||
cbee9f88 PZ |
2754 | /* |
2755 | * Enforce maximal scan/migration frequency.. | |
2756 | */ | |
2757 | migrate = mm->numa_next_scan; | |
2758 | if (time_before(now, migrate)) | |
2759 | return; | |
2760 | ||
598f0ec0 MG |
2761 | if (p->numa_scan_period == 0) { |
2762 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2763 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2764 | } |
cbee9f88 | 2765 | |
fb003b80 | 2766 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2767 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2768 | return; | |
2769 | ||
19a78d11 PZ |
2770 | /* |
2771 | * Delay this task enough that another task of this mm will likely win | |
2772 | * the next time around. | |
2773 | */ | |
2774 | p->node_stamp += 2 * TICK_NSEC; | |
2775 | ||
9f40604c MG |
2776 | start = mm->numa_scan_offset; |
2777 | pages = sysctl_numa_balancing_scan_size; | |
2778 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2779 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2780 | if (!pages) |
2781 | return; | |
cbee9f88 | 2782 | |
4620f8c1 | 2783 | |
d8ed45c5 | 2784 | if (!mmap_read_trylock(mm)) |
8655d549 | 2785 | return; |
9f40604c | 2786 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2787 | if (!vma) { |
2788 | reset_ptenuma_scan(p); | |
9f40604c | 2789 | start = 0; |
6e5fb223 PZ |
2790 | vma = mm->mmap; |
2791 | } | |
9f40604c | 2792 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2793 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2794 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2795 | continue; |
6b79c57b | 2796 | } |
6e5fb223 | 2797 | |
4591ce4f MG |
2798 | /* |
2799 | * Shared library pages mapped by multiple processes are not | |
2800 | * migrated as it is expected they are cache replicated. Avoid | |
2801 | * hinting faults in read-only file-backed mappings or the vdso | |
2802 | * as migrating the pages will be of marginal benefit. | |
2803 | */ | |
2804 | if (!vma->vm_mm || | |
2805 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2806 | continue; | |
2807 | ||
3c67f474 MG |
2808 | /* |
2809 | * Skip inaccessible VMAs to avoid any confusion between | |
2810 | * PROT_NONE and NUMA hinting ptes | |
2811 | */ | |
3122e80e | 2812 | if (!vma_is_accessible(vma)) |
3c67f474 | 2813 | continue; |
4591ce4f | 2814 | |
9f40604c MG |
2815 | do { |
2816 | start = max(start, vma->vm_start); | |
2817 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2818 | end = min(end, vma->vm_end); | |
4620f8c1 | 2819 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2820 | |
2821 | /* | |
4620f8c1 RR |
2822 | * Try to scan sysctl_numa_balancing_size worth of |
2823 | * hpages that have at least one present PTE that | |
2824 | * is not already pte-numa. If the VMA contains | |
2825 | * areas that are unused or already full of prot_numa | |
2826 | * PTEs, scan up to virtpages, to skip through those | |
2827 | * areas faster. | |
598f0ec0 MG |
2828 | */ |
2829 | if (nr_pte_updates) | |
2830 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2831 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2832 | |
9f40604c | 2833 | start = end; |
4620f8c1 | 2834 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2835 | goto out; |
3cf1962c RR |
2836 | |
2837 | cond_resched(); | |
9f40604c | 2838 | } while (end != vma->vm_end); |
cbee9f88 | 2839 | } |
6e5fb223 | 2840 | |
9f40604c | 2841 | out: |
6e5fb223 | 2842 | /* |
c69307d5 PZ |
2843 | * It is possible to reach the end of the VMA list but the last few |
2844 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2845 | * would find the !migratable VMA on the next scan but not reset the | |
2846 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2847 | */ |
2848 | if (vma) | |
9f40604c | 2849 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2850 | else |
2851 | reset_ptenuma_scan(p); | |
d8ed45c5 | 2852 | mmap_read_unlock(mm); |
51170840 RR |
2853 | |
2854 | /* | |
2855 | * Make sure tasks use at least 32x as much time to run other code | |
2856 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2857 | * Usually update_task_scan_period slows down scanning enough; on an | |
2858 | * overloaded system we need to limit overhead on a per task basis. | |
2859 | */ | |
2860 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2861 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2862 | p->node_stamp += 32 * diff; | |
2863 | } | |
cbee9f88 PZ |
2864 | } |
2865 | ||
d35927a1 VS |
2866 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
2867 | { | |
2868 | int mm_users = 0; | |
2869 | struct mm_struct *mm = p->mm; | |
2870 | ||
2871 | if (mm) { | |
2872 | mm_users = atomic_read(&mm->mm_users); | |
2873 | if (mm_users == 1) { | |
2874 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
2875 | mm->numa_scan_seq = 0; | |
2876 | } | |
2877 | } | |
2878 | p->node_stamp = 0; | |
2879 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
2880 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
b34920d4 | 2881 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
2882 | p->numa_work.next = &p->numa_work; |
2883 | p->numa_faults = NULL; | |
2884 | RCU_INIT_POINTER(p->numa_group, NULL); | |
2885 | p->last_task_numa_placement = 0; | |
2886 | p->last_sum_exec_runtime = 0; | |
2887 | ||
b34920d4 VS |
2888 | init_task_work(&p->numa_work, task_numa_work); |
2889 | ||
d35927a1 VS |
2890 | /* New address space, reset the preferred nid */ |
2891 | if (!(clone_flags & CLONE_VM)) { | |
2892 | p->numa_preferred_nid = NUMA_NO_NODE; | |
2893 | return; | |
2894 | } | |
2895 | ||
2896 | /* | |
2897 | * New thread, keep existing numa_preferred_nid which should be copied | |
2898 | * already by arch_dup_task_struct but stagger when scans start. | |
2899 | */ | |
2900 | if (mm) { | |
2901 | unsigned int delay; | |
2902 | ||
2903 | delay = min_t(unsigned int, task_scan_max(current), | |
2904 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
2905 | delay += 2 * TICK_NSEC; | |
2906 | p->node_stamp = delay; | |
2907 | } | |
2908 | } | |
2909 | ||
cbee9f88 PZ |
2910 | /* |
2911 | * Drive the periodic memory faults.. | |
2912 | */ | |
b1546edc | 2913 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
2914 | { |
2915 | struct callback_head *work = &curr->numa_work; | |
2916 | u64 period, now; | |
2917 | ||
2918 | /* | |
2919 | * We don't care about NUMA placement if we don't have memory. | |
2920 | */ | |
18f855e5 | 2921 | if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) |
cbee9f88 PZ |
2922 | return; |
2923 | ||
2924 | /* | |
2925 | * Using runtime rather than walltime has the dual advantage that | |
2926 | * we (mostly) drive the selection from busy threads and that the | |
2927 | * task needs to have done some actual work before we bother with | |
2928 | * NUMA placement. | |
2929 | */ | |
2930 | now = curr->se.sum_exec_runtime; | |
2931 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2932 | ||
25b3e5a3 | 2933 | if (now > curr->node_stamp + period) { |
4b96a29b | 2934 | if (!curr->node_stamp) |
b5dd77c8 | 2935 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2936 | curr->node_stamp += period; |
cbee9f88 | 2937 | |
b34920d4 | 2938 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
91989c70 | 2939 | task_work_add(curr, work, TWA_RESUME); |
cbee9f88 PZ |
2940 | } |
2941 | } | |
3fed382b | 2942 | |
3f9672ba SD |
2943 | static void update_scan_period(struct task_struct *p, int new_cpu) |
2944 | { | |
2945 | int src_nid = cpu_to_node(task_cpu(p)); | |
2946 | int dst_nid = cpu_to_node(new_cpu); | |
2947 | ||
05cbdf4f MG |
2948 | if (!static_branch_likely(&sched_numa_balancing)) |
2949 | return; | |
2950 | ||
3f9672ba SD |
2951 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
2952 | return; | |
2953 | ||
05cbdf4f MG |
2954 | if (src_nid == dst_nid) |
2955 | return; | |
2956 | ||
2957 | /* | |
2958 | * Allow resets if faults have been trapped before one scan | |
2959 | * has completed. This is most likely due to a new task that | |
2960 | * is pulled cross-node due to wakeups or load balancing. | |
2961 | */ | |
2962 | if (p->numa_scan_seq) { | |
2963 | /* | |
2964 | * Avoid scan adjustments if moving to the preferred | |
2965 | * node or if the task was not previously running on | |
2966 | * the preferred node. | |
2967 | */ | |
2968 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
2969 | (p->numa_preferred_nid != NUMA_NO_NODE && |
2970 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
2971 | return; |
2972 | } | |
2973 | ||
2974 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
2975 | } |
2976 | ||
cbee9f88 PZ |
2977 | #else |
2978 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2979 | { | |
2980 | } | |
0ec8aa00 PZ |
2981 | |
2982 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2983 | { | |
2984 | } | |
2985 | ||
2986 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2987 | { | |
2988 | } | |
3fed382b | 2989 | |
3f9672ba SD |
2990 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
2991 | { | |
2992 | } | |
2993 | ||
cbee9f88 PZ |
2994 | #endif /* CONFIG_NUMA_BALANCING */ |
2995 | ||
30cfdcfc DA |
2996 | static void |
2997 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2998 | { | |
2999 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 3000 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3001 | if (entity_is_task(se)) { |
3002 | struct rq *rq = rq_of(cfs_rq); | |
3003 | ||
3004 | account_numa_enqueue(rq, task_of(se)); | |
3005 | list_add(&se->group_node, &rq->cfs_tasks); | |
3006 | } | |
367456c7 | 3007 | #endif |
30cfdcfc | 3008 | cfs_rq->nr_running++; |
30cfdcfc DA |
3009 | } |
3010 | ||
3011 | static void | |
3012 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3013 | { | |
3014 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 3015 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3016 | if (entity_is_task(se)) { |
3017 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 3018 | list_del_init(&se->group_node); |
0ec8aa00 | 3019 | } |
bfdb198c | 3020 | #endif |
30cfdcfc | 3021 | cfs_rq->nr_running--; |
30cfdcfc DA |
3022 | } |
3023 | ||
8d5b9025 PZ |
3024 | /* |
3025 | * Signed add and clamp on underflow. | |
3026 | * | |
3027 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3028 | * memory. This allows lockless observations without ever seeing the negative | |
3029 | * values. | |
3030 | */ | |
3031 | #define add_positive(_ptr, _val) do { \ | |
3032 | typeof(_ptr) ptr = (_ptr); \ | |
3033 | typeof(_val) val = (_val); \ | |
3034 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3035 | \ | |
3036 | res = var + val; \ | |
3037 | \ | |
3038 | if (val < 0 && res > var) \ | |
3039 | res = 0; \ | |
3040 | \ | |
3041 | WRITE_ONCE(*ptr, res); \ | |
3042 | } while (0) | |
3043 | ||
3044 | /* | |
3045 | * Unsigned subtract and clamp on underflow. | |
3046 | * | |
3047 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3048 | * memory. This allows lockless observations without ever seeing the negative | |
3049 | * values. | |
3050 | */ | |
3051 | #define sub_positive(_ptr, _val) do { \ | |
3052 | typeof(_ptr) ptr = (_ptr); \ | |
3053 | typeof(*ptr) val = (_val); \ | |
3054 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3055 | res = var - val; \ | |
3056 | if (res > var) \ | |
3057 | res = 0; \ | |
3058 | WRITE_ONCE(*ptr, res); \ | |
3059 | } while (0) | |
3060 | ||
b5c0ce7b PB |
3061 | /* |
3062 | * Remove and clamp on negative, from a local variable. | |
3063 | * | |
3064 | * A variant of sub_positive(), which does not use explicit load-store | |
3065 | * and is thus optimized for local variable updates. | |
3066 | */ | |
3067 | #define lsub_positive(_ptr, _val) do { \ | |
3068 | typeof(_ptr) ptr = (_ptr); \ | |
3069 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
3070 | } while (0) | |
3071 | ||
8d5b9025 | 3072 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
3073 | static inline void |
3074 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3075 | { | |
3076 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3077 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
3078 | } | |
3079 | ||
3080 | static inline void | |
3081 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3082 | { | |
3083 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
3084 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); | |
3085 | } | |
3086 | #else | |
3087 | static inline void | |
8d5b9025 PZ |
3088 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } |
3089 | static inline void | |
3090 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
3091 | #endif | |
3092 | ||
9059393e | 3093 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
0dacee1b | 3094 | unsigned long weight) |
9059393e VG |
3095 | { |
3096 | if (se->on_rq) { | |
3097 | /* commit outstanding execution time */ | |
3098 | if (cfs_rq->curr == se) | |
3099 | update_curr(cfs_rq); | |
1724b95b | 3100 | update_load_sub(&cfs_rq->load, se->load.weight); |
9059393e VG |
3101 | } |
3102 | dequeue_load_avg(cfs_rq, se); | |
3103 | ||
3104 | update_load_set(&se->load, weight); | |
3105 | ||
3106 | #ifdef CONFIG_SMP | |
1ea6c46a | 3107 | do { |
87e867b4 | 3108 | u32 divider = get_pelt_divider(&se->avg); |
1ea6c46a PZ |
3109 | |
3110 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
1ea6c46a | 3111 | } while (0); |
9059393e VG |
3112 | #endif |
3113 | ||
3114 | enqueue_load_avg(cfs_rq, se); | |
0dacee1b | 3115 | if (se->on_rq) |
1724b95b | 3116 | update_load_add(&cfs_rq->load, se->load.weight); |
0dacee1b | 3117 | |
9059393e VG |
3118 | } |
3119 | ||
3120 | void reweight_task(struct task_struct *p, int prio) | |
3121 | { | |
3122 | struct sched_entity *se = &p->se; | |
3123 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3124 | struct load_weight *load = &se->load; | |
3125 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
3126 | ||
0dacee1b | 3127 | reweight_entity(cfs_rq, se, weight); |
9059393e VG |
3128 | load->inv_weight = sched_prio_to_wmult[prio]; |
3129 | } | |
3130 | ||
3ff6dcac | 3131 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 3132 | #ifdef CONFIG_SMP |
cef27403 PZ |
3133 | /* |
3134 | * All this does is approximate the hierarchical proportion which includes that | |
3135 | * global sum we all love to hate. | |
3136 | * | |
3137 | * That is, the weight of a group entity, is the proportional share of the | |
3138 | * group weight based on the group runqueue weights. That is: | |
3139 | * | |
3140 | * tg->weight * grq->load.weight | |
3141 | * ge->load.weight = ----------------------------- (1) | |
3142 | * \Sum grq->load.weight | |
3143 | * | |
3144 | * Now, because computing that sum is prohibitively expensive to compute (been | |
3145 | * there, done that) we approximate it with this average stuff. The average | |
3146 | * moves slower and therefore the approximation is cheaper and more stable. | |
3147 | * | |
3148 | * So instead of the above, we substitute: | |
3149 | * | |
3150 | * grq->load.weight -> grq->avg.load_avg (2) | |
3151 | * | |
3152 | * which yields the following: | |
3153 | * | |
3154 | * tg->weight * grq->avg.load_avg | |
3155 | * ge->load.weight = ------------------------------ (3) | |
3156 | * tg->load_avg | |
3157 | * | |
3158 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
3159 | * | |
3160 | * That is shares_avg, and it is right (given the approximation (2)). | |
3161 | * | |
3162 | * The problem with it is that because the average is slow -- it was designed | |
3163 | * to be exactly that of course -- this leads to transients in boundary | |
3164 | * conditions. In specific, the case where the group was idle and we start the | |
3165 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
3166 | * yielding bad latency etc.. | |
3167 | * | |
3168 | * Now, in that special case (1) reduces to: | |
3169 | * | |
3170 | * tg->weight * grq->load.weight | |
17de4ee0 | 3171 | * ge->load.weight = ----------------------------- = tg->weight (4) |
cef27403 PZ |
3172 | * grp->load.weight |
3173 | * | |
3174 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
3175 | * | |
3176 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
3177 | * UP case, like: | |
3178 | * | |
3179 | * ge->load.weight = | |
3180 | * | |
3181 | * tg->weight * grq->load.weight | |
3182 | * --------------------------------------------------- (5) | |
3183 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
3184 | * | |
17de4ee0 PZ |
3185 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
3186 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
3187 | * | |
3188 | * | |
3189 | * tg->weight * grq->load.weight | |
3190 | * ge->load.weight = ----------------------------- (6) | |
3191 | * tg_load_avg' | |
3192 | * | |
3193 | * Where: | |
3194 | * | |
3195 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
3196 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
3197 | * |
3198 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
3199 | * (4) while in the normal case it approaches (3). It consistently | |
3200 | * overestimates the ge->load.weight and therefore: | |
3201 | * | |
3202 | * \Sum ge->load.weight >= tg->weight | |
3203 | * | |
3204 | * hence icky! | |
3205 | */ | |
2c8e4dce | 3206 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 3207 | { |
7c80cfc9 PZ |
3208 | long tg_weight, tg_shares, load, shares; |
3209 | struct task_group *tg = cfs_rq->tg; | |
3210 | ||
3211 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 3212 | |
3d4b60d3 | 3213 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 3214 | |
ea1dc6fc | 3215 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 3216 | |
ea1dc6fc PZ |
3217 | /* Ensure tg_weight >= load */ |
3218 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
3219 | tg_weight += load; | |
3ff6dcac | 3220 | |
7c80cfc9 | 3221 | shares = (tg_shares * load); |
cf5f0acf PZ |
3222 | if (tg_weight) |
3223 | shares /= tg_weight; | |
3ff6dcac | 3224 | |
b8fd8423 DE |
3225 | /* |
3226 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
3227 | * of a group with small tg->shares value. It is a floor value which is | |
3228 | * assigned as a minimum load.weight to the sched_entity representing | |
3229 | * the group on a CPU. | |
3230 | * | |
3231 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
3232 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
3233 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
3234 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
3235 | * instead of 0. | |
3236 | */ | |
7c80cfc9 | 3237 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 3238 | } |
387f77cc | 3239 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 3240 | |
82958366 PT |
3241 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
3242 | ||
1ea6c46a PZ |
3243 | /* |
3244 | * Recomputes the group entity based on the current state of its group | |
3245 | * runqueue. | |
3246 | */ | |
3247 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 3248 | { |
1ea6c46a | 3249 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
0dacee1b | 3250 | long shares; |
2069dd75 | 3251 | |
1ea6c46a | 3252 | if (!gcfs_rq) |
89ee048f VG |
3253 | return; |
3254 | ||
1ea6c46a | 3255 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3256 | return; |
89ee048f | 3257 | |
3ff6dcac | 3258 | #ifndef CONFIG_SMP |
0dacee1b | 3259 | shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
3260 | |
3261 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 3262 | return; |
7c80cfc9 | 3263 | #else |
2c8e4dce | 3264 | shares = calc_group_shares(gcfs_rq); |
3ff6dcac | 3265 | #endif |
2069dd75 | 3266 | |
0dacee1b | 3267 | reweight_entity(cfs_rq_of(se), se, shares); |
2069dd75 | 3268 | } |
89ee048f | 3269 | |
2069dd75 | 3270 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3271 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3272 | { |
3273 | } | |
3274 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3275 | ||
ea14b57e | 3276 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3277 | { |
43964409 LT |
3278 | struct rq *rq = rq_of(cfs_rq); |
3279 | ||
a4f9a0e5 | 3280 | if (&rq->cfs == cfs_rq) { |
a030d738 VK |
3281 | /* |
3282 | * There are a few boundary cases this might miss but it should | |
3283 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3284 | * a real problem. |
a030d738 VK |
3285 | * |
3286 | * It will not get called when we go idle, because the idle | |
3287 | * thread is a different class (!fair), nor will the utilization | |
3288 | * number include things like RT tasks. | |
3289 | * | |
3290 | * As is, the util number is not freq-invariant (we'd have to | |
3291 | * implement arch_scale_freq_capacity() for that). | |
3292 | * | |
3293 | * See cpu_util(). | |
3294 | */ | |
ea14b57e | 3295 | cpufreq_update_util(rq, flags); |
a030d738 VK |
3296 | } |
3297 | } | |
3298 | ||
141965c7 | 3299 | #ifdef CONFIG_SMP |
c566e8e9 | 3300 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7c3edd2c PZ |
3301 | /** |
3302 | * update_tg_load_avg - update the tg's load avg | |
3303 | * @cfs_rq: the cfs_rq whose avg changed | |
7c3edd2c PZ |
3304 | * |
3305 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3306 | * However, because tg->load_avg is a global value there are performance | |
3307 | * considerations. | |
3308 | * | |
3309 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3310 | * differential update where we store the last value we propagated. This in | |
3311 | * turn allows skipping updates if the differential is 'small'. | |
3312 | * | |
815abf5a | 3313 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3314 | */ |
fe749158 | 3315 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) |
bb17f655 | 3316 | { |
9d89c257 | 3317 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3318 | |
aa0b7ae0 WL |
3319 | /* |
3320 | * No need to update load_avg for root_task_group as it is not used. | |
3321 | */ | |
3322 | if (cfs_rq->tg == &root_task_group) | |
3323 | return; | |
3324 | ||
fe749158 | 3325 | if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
9d89c257 YD |
3326 | atomic_long_add(delta, &cfs_rq->tg->load_avg); |
3327 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3328 | } |
8165e145 | 3329 | } |
f5f9739d | 3330 | |
ad936d86 | 3331 | /* |
97fb7a0a | 3332 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3333 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3334 | * including the state of rq->lock, should be made. | |
3335 | */ | |
3336 | void set_task_rq_fair(struct sched_entity *se, | |
3337 | struct cfs_rq *prev, struct cfs_rq *next) | |
3338 | { | |
0ccb977f PZ |
3339 | u64 p_last_update_time; |
3340 | u64 n_last_update_time; | |
3341 | ||
ad936d86 BP |
3342 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3343 | return; | |
3344 | ||
3345 | /* | |
3346 | * We are supposed to update the task to "current" time, then its up to | |
3347 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3348 | * getting what current time is, so simply throw away the out-of-date | |
3349 | * time. This will result in the wakee task is less decayed, but giving | |
3350 | * the wakee more load sounds not bad. | |
3351 | */ | |
0ccb977f PZ |
3352 | if (!(se->avg.last_update_time && prev)) |
3353 | return; | |
ad936d86 BP |
3354 | |
3355 | #ifndef CONFIG_64BIT | |
0ccb977f | 3356 | { |
ad936d86 BP |
3357 | u64 p_last_update_time_copy; |
3358 | u64 n_last_update_time_copy; | |
3359 | ||
3360 | do { | |
3361 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3362 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3363 | ||
3364 | smp_rmb(); | |
3365 | ||
3366 | p_last_update_time = prev->avg.last_update_time; | |
3367 | n_last_update_time = next->avg.last_update_time; | |
3368 | ||
3369 | } while (p_last_update_time != p_last_update_time_copy || | |
3370 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3371 | } |
ad936d86 | 3372 | #else |
0ccb977f PZ |
3373 | p_last_update_time = prev->avg.last_update_time; |
3374 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3375 | #endif |
23127296 | 3376 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 3377 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 3378 | } |
09a43ace | 3379 | |
0e2d2aaa PZ |
3380 | |
3381 | /* | |
3382 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3383 | * propagate its contribution. The key to this propagation is the invariant | |
3384 | * that for each group: | |
3385 | * | |
3386 | * ge->avg == grq->avg (1) | |
3387 | * | |
3388 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3389 | * represent the very same entity, just at different points in the hierarchy. | |
3390 | * | |
9f683953 VG |
3391 | * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial |
3392 | * and simply copies the running/runnable sum over (but still wrong, because | |
3393 | * the group entity and group rq do not have their PELT windows aligned). | |
0e2d2aaa | 3394 | * |
0dacee1b | 3395 | * However, update_tg_cfs_load() is more complex. So we have: |
0e2d2aaa PZ |
3396 | * |
3397 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3398 | * | |
3399 | * And since, like util, the runnable part should be directly transferable, | |
3400 | * the following would _appear_ to be the straight forward approach: | |
3401 | * | |
a4c3c049 | 3402 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3403 | * |
3404 | * And per (1) we have: | |
3405 | * | |
a4c3c049 | 3406 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3407 | * |
3408 | * Which gives: | |
3409 | * | |
3410 | * ge->load.weight * grq->avg.load_avg | |
3411 | * ge->avg.load_avg = ----------------------------------- (4) | |
3412 | * grq->load.weight | |
3413 | * | |
3414 | * Except that is wrong! | |
3415 | * | |
3416 | * Because while for entities historical weight is not important and we | |
3417 | * really only care about our future and therefore can consider a pure | |
3418 | * runnable sum, runqueues can NOT do this. | |
3419 | * | |
3420 | * We specifically want runqueues to have a load_avg that includes | |
3421 | * historical weights. Those represent the blocked load, the load we expect | |
3422 | * to (shortly) return to us. This only works by keeping the weights as | |
3423 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3424 | * | |
a4c3c049 VG |
3425 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3426 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3427 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3428 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3429 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3430 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3431 | * |
a4c3c049 | 3432 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3433 | * |
a4c3c049 | 3434 | * Given the constraint: |
0e2d2aaa | 3435 | * |
a4c3c049 | 3436 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3437 | * |
a4c3c049 VG |
3438 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3439 | * overlap. | |
0e2d2aaa | 3440 | * |
a4c3c049 | 3441 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3442 | * |
a4c3c049 | 3443 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3444 | * |
a4c3c049 | 3445 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3446 | * |
0e2d2aaa PZ |
3447 | */ |
3448 | ||
09a43ace | 3449 | static inline void |
0e2d2aaa | 3450 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3451 | { |
09a43ace | 3452 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; |
87e867b4 | 3453 | u32 divider; |
09a43ace VG |
3454 | |
3455 | /* Nothing to update */ | |
3456 | if (!delta) | |
3457 | return; | |
3458 | ||
87e867b4 VG |
3459 | /* |
3460 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3461 | * See ___update_load_avg() for details. | |
3462 | */ | |
3463 | divider = get_pelt_divider(&cfs_rq->avg); | |
3464 | ||
09a43ace VG |
3465 | /* Set new sched_entity's utilization */ |
3466 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
95d68593 | 3467 | se->avg.util_sum = se->avg.util_avg * divider; |
09a43ace VG |
3468 | |
3469 | /* Update parent cfs_rq utilization */ | |
3470 | add_positive(&cfs_rq->avg.util_avg, delta); | |
95d68593 | 3471 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; |
09a43ace VG |
3472 | } |
3473 | ||
9f683953 VG |
3474 | static inline void |
3475 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) | |
3476 | { | |
3477 | long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; | |
87e867b4 | 3478 | u32 divider; |
9f683953 VG |
3479 | |
3480 | /* Nothing to update */ | |
3481 | if (!delta) | |
3482 | return; | |
3483 | ||
87e867b4 VG |
3484 | /* |
3485 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3486 | * See ___update_load_avg() for details. | |
3487 | */ | |
3488 | divider = get_pelt_divider(&cfs_rq->avg); | |
3489 | ||
9f683953 VG |
3490 | /* Set new sched_entity's runnable */ |
3491 | se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; | |
95d68593 | 3492 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
9f683953 VG |
3493 | |
3494 | /* Update parent cfs_rq runnable */ | |
3495 | add_positive(&cfs_rq->avg.runnable_avg, delta); | |
95d68593 | 3496 | cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; |
9f683953 VG |
3497 | } |
3498 | ||
09a43ace | 3499 | static inline void |
0dacee1b | 3500 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3501 | { |
a4c3c049 | 3502 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
0dacee1b VG |
3503 | unsigned long load_avg; |
3504 | u64 load_sum = 0; | |
a4c3c049 | 3505 | s64 delta_sum; |
95d68593 | 3506 | u32 divider; |
09a43ace | 3507 | |
0e2d2aaa PZ |
3508 | if (!runnable_sum) |
3509 | return; | |
09a43ace | 3510 | |
0e2d2aaa | 3511 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3512 | |
95d68593 VG |
3513 | /* |
3514 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3515 | * See ___update_load_avg() for details. | |
3516 | */ | |
87e867b4 | 3517 | divider = get_pelt_divider(&cfs_rq->avg); |
95d68593 | 3518 | |
a4c3c049 VG |
3519 | if (runnable_sum >= 0) { |
3520 | /* | |
3521 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3522 | * the CPU is saturated running == runnable. | |
3523 | */ | |
3524 | runnable_sum += se->avg.load_sum; | |
95d68593 | 3525 | runnable_sum = min_t(long, runnable_sum, divider); |
a4c3c049 VG |
3526 | } else { |
3527 | /* | |
3528 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3529 | * assuming all tasks are equally runnable. | |
3530 | */ | |
3531 | if (scale_load_down(gcfs_rq->load.weight)) { | |
3532 | load_sum = div_s64(gcfs_rq->avg.load_sum, | |
3533 | scale_load_down(gcfs_rq->load.weight)); | |
3534 | } | |
3535 | ||
3536 | /* But make sure to not inflate se's runnable */ | |
3537 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3538 | } | |
3539 | ||
3540 | /* | |
3541 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
3542 | * Rescale running sum to be in the same range as runnable sum |
3543 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
3544 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 3545 | */ |
23127296 | 3546 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
3547 | runnable_sum = max(runnable_sum, running_sum); |
3548 | ||
0e2d2aaa | 3549 | load_sum = (s64)se_weight(se) * runnable_sum; |
95d68593 | 3550 | load_avg = div_s64(load_sum, divider); |
09a43ace | 3551 | |
a4c3c049 VG |
3552 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
3553 | delta_avg = load_avg - se->avg.load_avg; | |
09a43ace | 3554 | |
a4c3c049 VG |
3555 | se->avg.load_sum = runnable_sum; |
3556 | se->avg.load_avg = load_avg; | |
3557 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
3558 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
09a43ace VG |
3559 | } |
3560 | ||
0e2d2aaa | 3561 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3562 | { |
0e2d2aaa PZ |
3563 | cfs_rq->propagate = 1; |
3564 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3565 | } |
3566 | ||
3567 | /* Update task and its cfs_rq load average */ | |
3568 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3569 | { | |
0e2d2aaa | 3570 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3571 | |
3572 | if (entity_is_task(se)) | |
3573 | return 0; | |
3574 | ||
0e2d2aaa PZ |
3575 | gcfs_rq = group_cfs_rq(se); |
3576 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3577 | return 0; |
3578 | ||
0e2d2aaa PZ |
3579 | gcfs_rq->propagate = 0; |
3580 | ||
09a43ace VG |
3581 | cfs_rq = cfs_rq_of(se); |
3582 | ||
0e2d2aaa | 3583 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3584 | |
0e2d2aaa | 3585 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
9f683953 | 3586 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); |
0dacee1b | 3587 | update_tg_cfs_load(cfs_rq, se, gcfs_rq); |
09a43ace | 3588 | |
ba19f51f | 3589 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 3590 | trace_pelt_se_tp(se); |
ba19f51f | 3591 | |
09a43ace VG |
3592 | return 1; |
3593 | } | |
3594 | ||
bc427898 VG |
3595 | /* |
3596 | * Check if we need to update the load and the utilization of a blocked | |
3597 | * group_entity: | |
3598 | */ | |
3599 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3600 | { | |
3601 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3602 | ||
3603 | /* | |
3604 | * If sched_entity still have not zero load or utilization, we have to | |
3605 | * decay it: | |
3606 | */ | |
3607 | if (se->avg.load_avg || se->avg.util_avg) | |
3608 | return false; | |
3609 | ||
3610 | /* | |
3611 | * If there is a pending propagation, we have to update the load and | |
3612 | * the utilization of the sched_entity: | |
3613 | */ | |
0e2d2aaa | 3614 | if (gcfs_rq->propagate) |
bc427898 VG |
3615 | return false; |
3616 | ||
3617 | /* | |
3618 | * Otherwise, the load and the utilization of the sched_entity is | |
3619 | * already zero and there is no pending propagation, so it will be a | |
3620 | * waste of time to try to decay it: | |
3621 | */ | |
3622 | return true; | |
3623 | } | |
3624 | ||
6e83125c | 3625 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3626 | |
fe749158 | 3627 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} |
09a43ace VG |
3628 | |
3629 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3630 | { | |
3631 | return 0; | |
3632 | } | |
3633 | ||
0e2d2aaa | 3634 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3635 | |
6e83125c | 3636 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3637 | |
3d30544f PZ |
3638 | /** |
3639 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 3640 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 3641 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
3642 | * |
3643 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3644 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3645 | * post_init_entity_util_avg(). | |
3646 | * | |
3647 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3648 | * | |
7c3edd2c PZ |
3649 | * Returns true if the load decayed or we removed load. |
3650 | * | |
3651 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3652 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3653 | */ |
a2c6c91f | 3654 | static inline int |
3a123bbb | 3655 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3656 | { |
9f683953 | 3657 | unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; |
9d89c257 | 3658 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3659 | int decayed = 0; |
2dac754e | 3660 | |
2a2f5d4e PZ |
3661 | if (cfs_rq->removed.nr) { |
3662 | unsigned long r; | |
87e867b4 | 3663 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
2a2f5d4e PZ |
3664 | |
3665 | raw_spin_lock(&cfs_rq->removed.lock); | |
3666 | swap(cfs_rq->removed.util_avg, removed_util); | |
3667 | swap(cfs_rq->removed.load_avg, removed_load); | |
9f683953 | 3668 | swap(cfs_rq->removed.runnable_avg, removed_runnable); |
2a2f5d4e PZ |
3669 | cfs_rq->removed.nr = 0; |
3670 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3671 | ||
2a2f5d4e | 3672 | r = removed_load; |
89741892 | 3673 | sub_positive(&sa->load_avg, r); |
9a2dd585 | 3674 | sub_positive(&sa->load_sum, r * divider); |
2dac754e | 3675 | |
2a2f5d4e | 3676 | r = removed_util; |
89741892 | 3677 | sub_positive(&sa->util_avg, r); |
9a2dd585 | 3678 | sub_positive(&sa->util_sum, r * divider); |
2a2f5d4e | 3679 | |
9f683953 VG |
3680 | r = removed_runnable; |
3681 | sub_positive(&sa->runnable_avg, r); | |
3682 | sub_positive(&sa->runnable_sum, r * divider); | |
3683 | ||
3684 | /* | |
3685 | * removed_runnable is the unweighted version of removed_load so we | |
3686 | * can use it to estimate removed_load_sum. | |
3687 | */ | |
3688 | add_tg_cfs_propagate(cfs_rq, | |
3689 | -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); | |
2a2f5d4e PZ |
3690 | |
3691 | decayed = 1; | |
9d89c257 | 3692 | } |
36ee28e4 | 3693 | |
23127296 | 3694 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
36ee28e4 | 3695 | |
9d89c257 YD |
3696 | #ifndef CONFIG_64BIT |
3697 | smp_wmb(); | |
3698 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3699 | #endif | |
36ee28e4 | 3700 | |
2a2f5d4e | 3701 | return decayed; |
21e96f88 SM |
3702 | } |
3703 | ||
3d30544f PZ |
3704 | /** |
3705 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3706 | * @cfs_rq: cfs_rq to attach to | |
3707 | * @se: sched_entity to attach | |
3708 | * | |
3709 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3710 | * cfs_rq->avg.last_update_time being current. | |
3711 | */ | |
a4f9a0e5 | 3712 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
a05e8c51 | 3713 | { |
95d68593 VG |
3714 | /* |
3715 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3716 | * See ___update_load_avg() for details. | |
3717 | */ | |
87e867b4 | 3718 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
f207934f PZ |
3719 | |
3720 | /* | |
3721 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3722 | * window because without that, really weird and wonderful things can | |
3723 | * happen. | |
3724 | * | |
3725 | * XXX illustrate | |
3726 | */ | |
a05e8c51 | 3727 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3728 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3729 | ||
3730 | /* | |
3731 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3732 | * period_contrib. This isn't strictly correct, but since we're | |
3733 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3734 | * _sum a little. | |
3735 | */ | |
3736 | se->avg.util_sum = se->avg.util_avg * divider; | |
3737 | ||
9f683953 VG |
3738 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
3739 | ||
f207934f PZ |
3740 | se->avg.load_sum = divider; |
3741 | if (se_weight(se)) { | |
3742 | se->avg.load_sum = | |
3743 | div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); | |
3744 | } | |
3745 | ||
8d5b9025 | 3746 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3747 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3748 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
9f683953 VG |
3749 | cfs_rq->avg.runnable_avg += se->avg.runnable_avg; |
3750 | cfs_rq->avg.runnable_sum += se->avg.runnable_sum; | |
0e2d2aaa PZ |
3751 | |
3752 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 3753 | |
a4f9a0e5 | 3754 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3755 | |
3756 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3757 | } |
3758 | ||
3d30544f PZ |
3759 | /** |
3760 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3761 | * @cfs_rq: cfs_rq to detach from | |
3762 | * @se: sched_entity to detach | |
3763 | * | |
3764 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3765 | * cfs_rq->avg.last_update_time being current. | |
3766 | */ | |
a05e8c51 BP |
3767 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3768 | { | |
8d5b9025 | 3769 | dequeue_load_avg(cfs_rq, se); |
89741892 PZ |
3770 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
3771 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
9f683953 VG |
3772 | sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); |
3773 | sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); | |
0e2d2aaa PZ |
3774 | |
3775 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 3776 | |
ea14b57e | 3777 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3778 | |
3779 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3780 | } |
3781 | ||
b382a531 PZ |
3782 | /* |
3783 | * Optional action to be done while updating the load average | |
3784 | */ | |
3785 | #define UPDATE_TG 0x1 | |
3786 | #define SKIP_AGE_LOAD 0x2 | |
3787 | #define DO_ATTACH 0x4 | |
3788 | ||
3789 | /* Update task and its cfs_rq load average */ | |
3790 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
3791 | { | |
23127296 | 3792 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
3793 | int decayed; |
3794 | ||
3795 | /* | |
3796 | * Track task load average for carrying it to new CPU after migrated, and | |
3797 | * track group sched_entity load average for task_h_load calc in migration | |
3798 | */ | |
3799 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 3800 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
3801 | |
3802 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
3803 | decayed |= propagate_entity_load_avg(se); | |
3804 | ||
3805 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
3806 | ||
ea14b57e PZ |
3807 | /* |
3808 | * DO_ATTACH means we're here from enqueue_entity(). | |
3809 | * !last_update_time means we've passed through | |
3810 | * migrate_task_rq_fair() indicating we migrated. | |
3811 | * | |
3812 | * IOW we're enqueueing a task on a new CPU. | |
3813 | */ | |
a4f9a0e5 | 3814 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 3815 | update_tg_load_avg(cfs_rq); |
b382a531 | 3816 | |
bef69dd8 VG |
3817 | } else if (decayed) { |
3818 | cfs_rq_util_change(cfs_rq, 0); | |
3819 | ||
3820 | if (flags & UPDATE_TG) | |
fe749158 | 3821 | update_tg_load_avg(cfs_rq); |
bef69dd8 | 3822 | } |
b382a531 PZ |
3823 | } |
3824 | ||
9d89c257 | 3825 | #ifndef CONFIG_64BIT |
0905f04e YD |
3826 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3827 | { | |
9d89c257 | 3828 | u64 last_update_time_copy; |
0905f04e | 3829 | u64 last_update_time; |
9ee474f5 | 3830 | |
9d89c257 YD |
3831 | do { |
3832 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3833 | smp_rmb(); | |
3834 | last_update_time = cfs_rq->avg.last_update_time; | |
3835 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3836 | |
3837 | return last_update_time; | |
3838 | } | |
9d89c257 | 3839 | #else |
0905f04e YD |
3840 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3841 | { | |
3842 | return cfs_rq->avg.last_update_time; | |
3843 | } | |
9d89c257 YD |
3844 | #endif |
3845 | ||
104cb16d MR |
3846 | /* |
3847 | * Synchronize entity load avg of dequeued entity without locking | |
3848 | * the previous rq. | |
3849 | */ | |
71b47eaf | 3850 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
3851 | { |
3852 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3853 | u64 last_update_time; | |
3854 | ||
3855 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 3856 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
3857 | } |
3858 | ||
0905f04e YD |
3859 | /* |
3860 | * Task first catches up with cfs_rq, and then subtract | |
3861 | * itself from the cfs_rq (task must be off the queue now). | |
3862 | */ | |
71b47eaf | 3863 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
3864 | { |
3865 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 3866 | unsigned long flags; |
0905f04e YD |
3867 | |
3868 | /* | |
7dc603c9 PZ |
3869 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3870 | * post_init_entity_util_avg() which will have added things to the | |
3871 | * cfs_rq, so we can remove unconditionally. | |
0905f04e | 3872 | */ |
0905f04e | 3873 | |
104cb16d | 3874 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
3875 | |
3876 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
3877 | ++cfs_rq->removed.nr; | |
3878 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
3879 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
9f683953 | 3880 | cfs_rq->removed.runnable_avg += se->avg.runnable_avg; |
2a2f5d4e | 3881 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 3882 | } |
642dbc39 | 3883 | |
9f683953 VG |
3884 | static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) |
3885 | { | |
3886 | return cfs_rq->avg.runnable_avg; | |
3887 | } | |
3888 | ||
7ea241af YD |
3889 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) |
3890 | { | |
3891 | return cfs_rq->avg.load_avg; | |
3892 | } | |
3893 | ||
d91cecc1 CY |
3894 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); |
3895 | ||
7f65ea42 PB |
3896 | static inline unsigned long task_util(struct task_struct *p) |
3897 | { | |
3898 | return READ_ONCE(p->se.avg.util_avg); | |
3899 | } | |
3900 | ||
3901 | static inline unsigned long _task_util_est(struct task_struct *p) | |
3902 | { | |
3903 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
3904 | ||
92a801e5 | 3905 | return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED); |
7f65ea42 PB |
3906 | } |
3907 | ||
3908 | static inline unsigned long task_util_est(struct task_struct *p) | |
3909 | { | |
3910 | return max(task_util(p), _task_util_est(p)); | |
3911 | } | |
3912 | ||
a7008c07 VS |
3913 | #ifdef CONFIG_UCLAMP_TASK |
3914 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
3915 | { | |
3916 | return clamp(task_util_est(p), | |
3917 | uclamp_eff_value(p, UCLAMP_MIN), | |
3918 | uclamp_eff_value(p, UCLAMP_MAX)); | |
3919 | } | |
3920 | #else | |
3921 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
3922 | { | |
3923 | return task_util_est(p); | |
3924 | } | |
3925 | #endif | |
3926 | ||
7f65ea42 PB |
3927 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, |
3928 | struct task_struct *p) | |
3929 | { | |
3930 | unsigned int enqueued; | |
3931 | ||
3932 | if (!sched_feat(UTIL_EST)) | |
3933 | return; | |
3934 | ||
3935 | /* Update root cfs_rq's estimated utilization */ | |
3936 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3937 | enqueued += _task_util_est(p); |
7f65ea42 | 3938 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
4581bea8 VD |
3939 | |
3940 | trace_sched_util_est_cfs_tp(cfs_rq); | |
7f65ea42 PB |
3941 | } |
3942 | ||
8c1f560c XY |
3943 | static inline void util_est_dequeue(struct cfs_rq *cfs_rq, |
3944 | struct task_struct *p) | |
3945 | { | |
3946 | unsigned int enqueued; | |
3947 | ||
3948 | if (!sched_feat(UTIL_EST)) | |
3949 | return; | |
3950 | ||
3951 | /* Update root cfs_rq's estimated utilization */ | |
3952 | enqueued = cfs_rq->avg.util_est.enqueued; | |
3953 | enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); | |
3954 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); | |
3955 | ||
3956 | trace_sched_util_est_cfs_tp(cfs_rq); | |
3957 | } | |
3958 | ||
b89997aa VD |
3959 | #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100) |
3960 | ||
7f65ea42 PB |
3961 | /* |
3962 | * Check if a (signed) value is within a specified (unsigned) margin, | |
3963 | * based on the observation that: | |
3964 | * | |
3965 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
3966 | * | |
3b03706f | 3967 | * NOTE: this only works when value + margin < INT_MAX. |
7f65ea42 PB |
3968 | */ |
3969 | static inline bool within_margin(int value, int margin) | |
3970 | { | |
3971 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
3972 | } | |
3973 | ||
8c1f560c XY |
3974 | static inline void util_est_update(struct cfs_rq *cfs_rq, |
3975 | struct task_struct *p, | |
3976 | bool task_sleep) | |
7f65ea42 | 3977 | { |
b89997aa | 3978 | long last_ewma_diff, last_enqueued_diff; |
7f65ea42 PB |
3979 | struct util_est ue; |
3980 | ||
3981 | if (!sched_feat(UTIL_EST)) | |
3982 | return; | |
3983 | ||
7f65ea42 PB |
3984 | /* |
3985 | * Skip update of task's estimated utilization when the task has not | |
3986 | * yet completed an activation, e.g. being migrated. | |
3987 | */ | |
3988 | if (!task_sleep) | |
3989 | return; | |
3990 | ||
d519329f PB |
3991 | /* |
3992 | * If the PELT values haven't changed since enqueue time, | |
3993 | * skip the util_est update. | |
3994 | */ | |
3995 | ue = p->se.avg.util_est; | |
3996 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
3997 | return; | |
3998 | ||
b89997aa VD |
3999 | last_enqueued_diff = ue.enqueued; |
4000 | ||
b8c96361 PB |
4001 | /* |
4002 | * Reset EWMA on utilization increases, the moving average is used only | |
4003 | * to smooth utilization decreases. | |
4004 | */ | |
4005 | ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED); | |
4006 | if (sched_feat(UTIL_EST_FASTUP)) { | |
4007 | if (ue.ewma < ue.enqueued) { | |
4008 | ue.ewma = ue.enqueued; | |
4009 | goto done; | |
4010 | } | |
4011 | } | |
4012 | ||
7f65ea42 | 4013 | /* |
b89997aa | 4014 | * Skip update of task's estimated utilization when its members are |
7f65ea42 PB |
4015 | * already ~1% close to its last activation value. |
4016 | */ | |
7f65ea42 | 4017 | last_ewma_diff = ue.enqueued - ue.ewma; |
b89997aa VD |
4018 | last_enqueued_diff -= ue.enqueued; |
4019 | if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) { | |
4020 | if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN)) | |
4021 | goto done; | |
4022 | ||
7f65ea42 | 4023 | return; |
b89997aa | 4024 | } |
7f65ea42 | 4025 | |
10a35e68 VG |
4026 | /* |
4027 | * To avoid overestimation of actual task utilization, skip updates if | |
4028 | * we cannot grant there is idle time in this CPU. | |
4029 | */ | |
8c1f560c | 4030 | if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq)))) |
10a35e68 VG |
4031 | return; |
4032 | ||
7f65ea42 PB |
4033 | /* |
4034 | * Update Task's estimated utilization | |
4035 | * | |
4036 | * When *p completes an activation we can consolidate another sample | |
4037 | * of the task size. This is done by storing the current PELT value | |
4038 | * as ue.enqueued and by using this value to update the Exponential | |
4039 | * Weighted Moving Average (EWMA): | |
4040 | * | |
4041 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
4042 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
4043 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
4044 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
4045 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
4046 | * | |
4047 | * Where 'w' is the weight of new samples, which is configured to be | |
4048 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
4049 | */ | |
4050 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
4051 | ue.ewma += last_ewma_diff; | |
4052 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
b8c96361 | 4053 | done: |
7f65ea42 | 4054 | WRITE_ONCE(p->se.avg.util_est, ue); |
4581bea8 VD |
4055 | |
4056 | trace_sched_util_est_se_tp(&p->se); | |
7f65ea42 PB |
4057 | } |
4058 | ||
3b1baa64 MR |
4059 | static inline int task_fits_capacity(struct task_struct *p, long capacity) |
4060 | { | |
a7008c07 | 4061 | return fits_capacity(uclamp_task_util(p), capacity); |
3b1baa64 MR |
4062 | } |
4063 | ||
4064 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
4065 | { | |
4066 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) | |
4067 | return; | |
4068 | ||
0ae78eec | 4069 | if (!p || p->nr_cpus_allowed == 1) { |
3b1baa64 MR |
4070 | rq->misfit_task_load = 0; |
4071 | return; | |
4072 | } | |
4073 | ||
4074 | if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { | |
4075 | rq->misfit_task_load = 0; | |
4076 | return; | |
4077 | } | |
4078 | ||
01cfcde9 VG |
4079 | /* |
4080 | * Make sure that misfit_task_load will not be null even if | |
4081 | * task_h_load() returns 0. | |
4082 | */ | |
4083 | rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); | |
3b1baa64 MR |
4084 | } |
4085 | ||
38033c37 PZ |
4086 | #else /* CONFIG_SMP */ |
4087 | ||
d31b1a66 VG |
4088 | #define UPDATE_TG 0x0 |
4089 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 4090 | #define DO_ATTACH 0x0 |
d31b1a66 | 4091 | |
88c0616e | 4092 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 4093 | { |
ea14b57e | 4094 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
4095 | } |
4096 | ||
9d89c257 | 4097 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 4098 | |
a05e8c51 | 4099 | static inline void |
a4f9a0e5 | 4100 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} |
a05e8c51 BP |
4101 | static inline void |
4102 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
4103 | ||
d91cecc1 | 4104 | static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
4105 | { |
4106 | return 0; | |
4107 | } | |
4108 | ||
7f65ea42 PB |
4109 | static inline void |
4110 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
4111 | ||
4112 | static inline void | |
8c1f560c XY |
4113 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {} |
4114 | ||
4115 | static inline void | |
4116 | util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p, | |
4117 | bool task_sleep) {} | |
3b1baa64 | 4118 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 4119 | |
38033c37 | 4120 | #endif /* CONFIG_SMP */ |
9d85f21c | 4121 | |
ddc97297 PZ |
4122 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4123 | { | |
4124 | #ifdef CONFIG_SCHED_DEBUG | |
4125 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
4126 | ||
4127 | if (d < 0) | |
4128 | d = -d; | |
4129 | ||
4130 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 4131 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
4132 | #endif |
4133 | } | |
4134 | ||
aeb73b04 PZ |
4135 | static void |
4136 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
4137 | { | |
1af5f730 | 4138 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 4139 | |
2cb8600e PZ |
4140 | /* |
4141 | * The 'current' period is already promised to the current tasks, | |
4142 | * however the extra weight of the new task will slow them down a | |
4143 | * little, place the new task so that it fits in the slot that | |
4144 | * stays open at the end. | |
4145 | */ | |
94dfb5e7 | 4146 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 4147 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 4148 | |
a2e7a7eb | 4149 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 4150 | if (!initial) { |
a2e7a7eb | 4151 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 4152 | |
a2e7a7eb MG |
4153 | /* |
4154 | * Halve their sleep time's effect, to allow | |
4155 | * for a gentler effect of sleepers: | |
4156 | */ | |
4157 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
4158 | thresh >>= 1; | |
51e0304c | 4159 | |
a2e7a7eb | 4160 | vruntime -= thresh; |
aeb73b04 PZ |
4161 | } |
4162 | ||
b5d9d734 | 4163 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 4164 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
4165 | } |
4166 | ||
d3d9dc33 PT |
4167 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
4168 | ||
cb251765 MG |
4169 | static inline void check_schedstat_required(void) |
4170 | { | |
4171 | #ifdef CONFIG_SCHEDSTATS | |
4172 | if (schedstat_enabled()) | |
4173 | return; | |
4174 | ||
4175 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
4176 | if (trace_sched_stat_wait_enabled() || | |
4177 | trace_sched_stat_sleep_enabled() || | |
4178 | trace_sched_stat_iowait_enabled() || | |
4179 | trace_sched_stat_blocked_enabled() || | |
4180 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 4181 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 | 4182 | "stat_blocked and stat_runtime require the " |
f67abed5 | 4183 | "kernel parameter schedstats=enable or " |
cb251765 MG |
4184 | "kernel.sched_schedstats=1\n"); |
4185 | } | |
4186 | #endif | |
4187 | } | |
4188 | ||
fe61468b | 4189 | static inline bool cfs_bandwidth_used(void); |
b5179ac7 PZ |
4190 | |
4191 | /* | |
4192 | * MIGRATION | |
4193 | * | |
4194 | * dequeue | |
4195 | * update_curr() | |
4196 | * update_min_vruntime() | |
4197 | * vruntime -= min_vruntime | |
4198 | * | |
4199 | * enqueue | |
4200 | * update_curr() | |
4201 | * update_min_vruntime() | |
4202 | * vruntime += min_vruntime | |
4203 | * | |
4204 | * this way the vruntime transition between RQs is done when both | |
4205 | * min_vruntime are up-to-date. | |
4206 | * | |
4207 | * WAKEUP (remote) | |
4208 | * | |
59efa0ba | 4209 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
4210 | * vruntime -= min_vruntime |
4211 | * | |
4212 | * enqueue | |
4213 | * update_curr() | |
4214 | * update_min_vruntime() | |
4215 | * vruntime += min_vruntime | |
4216 | * | |
4217 | * this way we don't have the most up-to-date min_vruntime on the originating | |
4218 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
4219 | */ | |
4220 | ||
bf0f6f24 | 4221 | static void |
88ec22d3 | 4222 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4223 | { |
2f950354 PZ |
4224 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
4225 | bool curr = cfs_rq->curr == se; | |
4226 | ||
88ec22d3 | 4227 | /* |
2f950354 PZ |
4228 | * If we're the current task, we must renormalise before calling |
4229 | * update_curr(). | |
88ec22d3 | 4230 | */ |
2f950354 | 4231 | if (renorm && curr) |
88ec22d3 PZ |
4232 | se->vruntime += cfs_rq->min_vruntime; |
4233 | ||
2f950354 PZ |
4234 | update_curr(cfs_rq); |
4235 | ||
bf0f6f24 | 4236 | /* |
2f950354 PZ |
4237 | * Otherwise, renormalise after, such that we're placed at the current |
4238 | * moment in time, instead of some random moment in the past. Being | |
4239 | * placed in the past could significantly boost this task to the | |
4240 | * fairness detriment of existing tasks. | |
bf0f6f24 | 4241 | */ |
2f950354 PZ |
4242 | if (renorm && !curr) |
4243 | se->vruntime += cfs_rq->min_vruntime; | |
4244 | ||
89ee048f VG |
4245 | /* |
4246 | * When enqueuing a sched_entity, we must: | |
4247 | * - Update loads to have both entity and cfs_rq synced with now. | |
9f683953 | 4248 | * - Add its load to cfs_rq->runnable_avg |
89ee048f VG |
4249 | * - For group_entity, update its weight to reflect the new share of |
4250 | * its group cfs_rq | |
4251 | * - Add its new weight to cfs_rq->load.weight | |
4252 | */ | |
b382a531 | 4253 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
9f683953 | 4254 | se_update_runnable(se); |
1ea6c46a | 4255 | update_cfs_group(se); |
17bc14b7 | 4256 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 4257 | |
1a3d027c | 4258 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 4259 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 4260 | |
cb251765 | 4261 | check_schedstat_required(); |
4fa8d299 JP |
4262 | update_stats_enqueue(cfs_rq, se, flags); |
4263 | check_spread(cfs_rq, se); | |
2f950354 | 4264 | if (!curr) |
83b699ed | 4265 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4266 | se->on_rq = 1; |
3d4b47b4 | 4267 | |
fe61468b VG |
4268 | /* |
4269 | * When bandwidth control is enabled, cfs might have been removed | |
4270 | * because of a parent been throttled but cfs->nr_running > 1. Try to | |
3b03706f | 4271 | * add it unconditionally. |
fe61468b VG |
4272 | */ |
4273 | if (cfs_rq->nr_running == 1 || cfs_bandwidth_used()) | |
3d4b47b4 | 4274 | list_add_leaf_cfs_rq(cfs_rq); |
fe61468b VG |
4275 | |
4276 | if (cfs_rq->nr_running == 1) | |
d3d9dc33 | 4277 | check_enqueue_throttle(cfs_rq); |
bf0f6f24 IM |
4278 | } |
4279 | ||
2c13c919 | 4280 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4281 | { |
2c13c919 RR |
4282 | for_each_sched_entity(se) { |
4283 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4284 | if (cfs_rq->last != se) |
2c13c919 | 4285 | break; |
f1044799 PZ |
4286 | |
4287 | cfs_rq->last = NULL; | |
2c13c919 RR |
4288 | } |
4289 | } | |
2002c695 | 4290 | |
2c13c919 RR |
4291 | static void __clear_buddies_next(struct sched_entity *se) |
4292 | { | |
4293 | for_each_sched_entity(se) { | |
4294 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4295 | if (cfs_rq->next != se) |
2c13c919 | 4296 | break; |
f1044799 PZ |
4297 | |
4298 | cfs_rq->next = NULL; | |
2c13c919 | 4299 | } |
2002c695 PZ |
4300 | } |
4301 | ||
ac53db59 RR |
4302 | static void __clear_buddies_skip(struct sched_entity *se) |
4303 | { | |
4304 | for_each_sched_entity(se) { | |
4305 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4306 | if (cfs_rq->skip != se) |
ac53db59 | 4307 | break; |
f1044799 PZ |
4308 | |
4309 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4310 | } |
4311 | } | |
4312 | ||
a571bbea PZ |
4313 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4314 | { | |
2c13c919 RR |
4315 | if (cfs_rq->last == se) |
4316 | __clear_buddies_last(se); | |
4317 | ||
4318 | if (cfs_rq->next == se) | |
4319 | __clear_buddies_next(se); | |
ac53db59 RR |
4320 | |
4321 | if (cfs_rq->skip == se) | |
4322 | __clear_buddies_skip(se); | |
a571bbea PZ |
4323 | } |
4324 | ||
6c16a6dc | 4325 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4326 | |
bf0f6f24 | 4327 | static void |
371fd7e7 | 4328 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4329 | { |
a2a2d680 DA |
4330 | /* |
4331 | * Update run-time statistics of the 'current'. | |
4332 | */ | |
4333 | update_curr(cfs_rq); | |
89ee048f VG |
4334 | |
4335 | /* | |
4336 | * When dequeuing a sched_entity, we must: | |
4337 | * - Update loads to have both entity and cfs_rq synced with now. | |
9f683953 | 4338 | * - Subtract its load from the cfs_rq->runnable_avg. |
dfcb245e | 4339 | * - Subtract its previous weight from cfs_rq->load.weight. |
89ee048f VG |
4340 | * - For group entity, update its weight to reflect the new share |
4341 | * of its group cfs_rq. | |
4342 | */ | |
88c0616e | 4343 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 4344 | se_update_runnable(se); |
a2a2d680 | 4345 | |
4fa8d299 | 4346 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 4347 | |
2002c695 | 4348 | clear_buddies(cfs_rq, se); |
4793241b | 4349 | |
83b699ed | 4350 | if (se != cfs_rq->curr) |
30cfdcfc | 4351 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4352 | se->on_rq = 0; |
30cfdcfc | 4353 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4354 | |
4355 | /* | |
b60205c7 PZ |
4356 | * Normalize after update_curr(); which will also have moved |
4357 | * min_vruntime if @se is the one holding it back. But before doing | |
4358 | * update_min_vruntime() again, which will discount @se's position and | |
4359 | * can move min_vruntime forward still more. | |
88ec22d3 | 4360 | */ |
371fd7e7 | 4361 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4362 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4363 | |
d8b4986d PT |
4364 | /* return excess runtime on last dequeue */ |
4365 | return_cfs_rq_runtime(cfs_rq); | |
4366 | ||
1ea6c46a | 4367 | update_cfs_group(se); |
b60205c7 PZ |
4368 | |
4369 | /* | |
4370 | * Now advance min_vruntime if @se was the entity holding it back, | |
4371 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4372 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4373 | * further than we started -- ie. we'll be penalized. | |
4374 | */ | |
9845c49c | 4375 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4376 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
4377 | } |
4378 | ||
4379 | /* | |
4380 | * Preempt the current task with a newly woken task if needed: | |
4381 | */ | |
7c92e54f | 4382 | static void |
2e09bf55 | 4383 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4384 | { |
11697830 | 4385 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4386 | struct sched_entity *se; |
4387 | s64 delta; | |
11697830 | 4388 | |
6d0f0ebd | 4389 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4390 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4391 | if (delta_exec > ideal_runtime) { |
8875125e | 4392 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4393 | /* |
4394 | * The current task ran long enough, ensure it doesn't get | |
4395 | * re-elected due to buddy favours. | |
4396 | */ | |
4397 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4398 | return; |
4399 | } | |
4400 | ||
4401 | /* | |
4402 | * Ensure that a task that missed wakeup preemption by a | |
4403 | * narrow margin doesn't have to wait for a full slice. | |
4404 | * This also mitigates buddy induced latencies under load. | |
4405 | */ | |
f685ceac MG |
4406 | if (delta_exec < sysctl_sched_min_granularity) |
4407 | return; | |
4408 | ||
f4cfb33e WX |
4409 | se = __pick_first_entity(cfs_rq); |
4410 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4411 | |
f4cfb33e WX |
4412 | if (delta < 0) |
4413 | return; | |
d7d82944 | 4414 | |
f4cfb33e | 4415 | if (delta > ideal_runtime) |
8875125e | 4416 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4417 | } |
4418 | ||
83b699ed | 4419 | static void |
8494f412 | 4420 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4421 | { |
83b699ed SV |
4422 | /* 'current' is not kept within the tree. */ |
4423 | if (se->on_rq) { | |
4424 | /* | |
4425 | * Any task has to be enqueued before it get to execute on | |
4426 | * a CPU. So account for the time it spent waiting on the | |
4427 | * runqueue. | |
4428 | */ | |
4fa8d299 | 4429 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 4430 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4431 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4432 | } |
4433 | ||
79303e9e | 4434 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4435 | cfs_rq->curr = se; |
4fa8d299 | 4436 | |
eba1ed4b IM |
4437 | /* |
4438 | * Track our maximum slice length, if the CPU's load is at | |
4439 | * least twice that of our own weight (i.e. dont track it | |
4440 | * when there are only lesser-weight tasks around): | |
4441 | */ | |
f2bedc47 DE |
4442 | if (schedstat_enabled() && |
4443 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
4fa8d299 JP |
4444 | schedstat_set(se->statistics.slice_max, |
4445 | max((u64)schedstat_val(se->statistics.slice_max), | |
4446 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 4447 | } |
4fa8d299 | 4448 | |
4a55b450 | 4449 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4450 | } |
4451 | ||
3f3a4904 PZ |
4452 | static int |
4453 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4454 | ||
ac53db59 RR |
4455 | /* |
4456 | * Pick the next process, keeping these things in mind, in this order: | |
4457 | * 1) keep things fair between processes/task groups | |
4458 | * 2) pick the "next" process, since someone really wants that to run | |
4459 | * 3) pick the "last" process, for cache locality | |
4460 | * 4) do not run the "skip" process, if something else is available | |
4461 | */ | |
678d5718 PZ |
4462 | static struct sched_entity * |
4463 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4464 | { |
678d5718 PZ |
4465 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4466 | struct sched_entity *se; | |
4467 | ||
4468 | /* | |
4469 | * If curr is set we have to see if its left of the leftmost entity | |
4470 | * still in the tree, provided there was anything in the tree at all. | |
4471 | */ | |
4472 | if (!left || (curr && entity_before(curr, left))) | |
4473 | left = curr; | |
4474 | ||
4475 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4476 | |
ac53db59 RR |
4477 | /* |
4478 | * Avoid running the skip buddy, if running something else can | |
4479 | * be done without getting too unfair. | |
4480 | */ | |
4481 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
4482 | struct sched_entity *second; |
4483 | ||
4484 | if (se == curr) { | |
4485 | second = __pick_first_entity(cfs_rq); | |
4486 | } else { | |
4487 | second = __pick_next_entity(se); | |
4488 | if (!second || (curr && entity_before(curr, second))) | |
4489 | second = curr; | |
4490 | } | |
4491 | ||
ac53db59 RR |
4492 | if (second && wakeup_preempt_entity(second, left) < 1) |
4493 | se = second; | |
4494 | } | |
aa2ac252 | 4495 | |
9abb8973 PO |
4496 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { |
4497 | /* | |
4498 | * Someone really wants this to run. If it's not unfair, run it. | |
4499 | */ | |
ac53db59 | 4500 | se = cfs_rq->next; |
9abb8973 PO |
4501 | } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { |
4502 | /* | |
4503 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4504 | */ | |
4505 | se = cfs_rq->last; | |
4506 | } | |
ac53db59 | 4507 | |
f685ceac | 4508 | clear_buddies(cfs_rq, se); |
4793241b PZ |
4509 | |
4510 | return se; | |
aa2ac252 PZ |
4511 | } |
4512 | ||
678d5718 | 4513 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4514 | |
ab6cde26 | 4515 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4516 | { |
4517 | /* | |
4518 | * If still on the runqueue then deactivate_task() | |
4519 | * was not called and update_curr() has to be done: | |
4520 | */ | |
4521 | if (prev->on_rq) | |
b7cc0896 | 4522 | update_curr(cfs_rq); |
bf0f6f24 | 4523 | |
d3d9dc33 PT |
4524 | /* throttle cfs_rqs exceeding runtime */ |
4525 | check_cfs_rq_runtime(cfs_rq); | |
4526 | ||
4fa8d299 | 4527 | check_spread(cfs_rq, prev); |
cb251765 | 4528 | |
30cfdcfc | 4529 | if (prev->on_rq) { |
4fa8d299 | 4530 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
4531 | /* Put 'current' back into the tree. */ |
4532 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4533 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4534 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4535 | } |
429d43bc | 4536 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4537 | } |
4538 | ||
8f4d37ec PZ |
4539 | static void |
4540 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4541 | { |
bf0f6f24 | 4542 | /* |
30cfdcfc | 4543 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4544 | */ |
30cfdcfc | 4545 | update_curr(cfs_rq); |
bf0f6f24 | 4546 | |
9d85f21c PT |
4547 | /* |
4548 | * Ensure that runnable average is periodically updated. | |
4549 | */ | |
88c0616e | 4550 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4551 | update_cfs_group(curr); |
9d85f21c | 4552 | |
8f4d37ec PZ |
4553 | #ifdef CONFIG_SCHED_HRTICK |
4554 | /* | |
4555 | * queued ticks are scheduled to match the slice, so don't bother | |
4556 | * validating it and just reschedule. | |
4557 | */ | |
983ed7a6 | 4558 | if (queued) { |
8875125e | 4559 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4560 | return; |
4561 | } | |
8f4d37ec PZ |
4562 | /* |
4563 | * don't let the period tick interfere with the hrtick preemption | |
4564 | */ | |
4565 | if (!sched_feat(DOUBLE_TICK) && | |
4566 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4567 | return; | |
4568 | #endif | |
4569 | ||
2c2efaed | 4570 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4571 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4572 | } |
4573 | ||
ab84d31e PT |
4574 | |
4575 | /************************************************** | |
4576 | * CFS bandwidth control machinery | |
4577 | */ | |
4578 | ||
4579 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 4580 | |
e9666d10 | 4581 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 4582 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4583 | |
4584 | static inline bool cfs_bandwidth_used(void) | |
4585 | { | |
c5905afb | 4586 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4587 | } |
4588 | ||
1ee14e6c | 4589 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4590 | { |
ce48c146 | 4591 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
4592 | } |
4593 | ||
4594 | void cfs_bandwidth_usage_dec(void) | |
4595 | { | |
ce48c146 | 4596 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 4597 | } |
e9666d10 | 4598 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
4599 | static bool cfs_bandwidth_used(void) |
4600 | { | |
4601 | return true; | |
4602 | } | |
4603 | ||
1ee14e6c BS |
4604 | void cfs_bandwidth_usage_inc(void) {} |
4605 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 4606 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 4607 | |
ab84d31e PT |
4608 | /* |
4609 | * default period for cfs group bandwidth. | |
4610 | * default: 0.1s, units: nanoseconds | |
4611 | */ | |
4612 | static inline u64 default_cfs_period(void) | |
4613 | { | |
4614 | return 100000000ULL; | |
4615 | } | |
ec12cb7f PT |
4616 | |
4617 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4618 | { | |
4619 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4620 | } | |
4621 | ||
a9cf55b2 | 4622 | /* |
763a9ec0 QC |
4623 | * Replenish runtime according to assigned quota. We use sched_clock_cpu |
4624 | * directly instead of rq->clock to avoid adding additional synchronization | |
4625 | * around rq->lock. | |
a9cf55b2 PT |
4626 | * |
4627 | * requires cfs_b->lock | |
4628 | */ | |
029632fb | 4629 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 | 4630 | { |
763a9ec0 QC |
4631 | if (cfs_b->quota != RUNTIME_INF) |
4632 | cfs_b->runtime = cfs_b->quota; | |
a9cf55b2 PT |
4633 | } |
4634 | ||
029632fb PZ |
4635 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4636 | { | |
4637 | return &tg->cfs_bandwidth; | |
4638 | } | |
4639 | ||
85dac906 | 4640 | /* returns 0 on failure to allocate runtime */ |
e98fa02c PT |
4641 | static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, |
4642 | struct cfs_rq *cfs_rq, u64 target_runtime) | |
ec12cb7f | 4643 | { |
e98fa02c PT |
4644 | u64 min_amount, amount = 0; |
4645 | ||
4646 | lockdep_assert_held(&cfs_b->lock); | |
ec12cb7f PT |
4647 | |
4648 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
e98fa02c | 4649 | min_amount = target_runtime - cfs_rq->runtime_remaining; |
ec12cb7f | 4650 | |
ec12cb7f PT |
4651 | if (cfs_b->quota == RUNTIME_INF) |
4652 | amount = min_amount; | |
58088ad0 | 4653 | else { |
77a4d1a1 | 4654 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4655 | |
4656 | if (cfs_b->runtime > 0) { | |
4657 | amount = min(cfs_b->runtime, min_amount); | |
4658 | cfs_b->runtime -= amount; | |
4659 | cfs_b->idle = 0; | |
4660 | } | |
ec12cb7f | 4661 | } |
ec12cb7f PT |
4662 | |
4663 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
4664 | |
4665 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4666 | } |
4667 | ||
e98fa02c PT |
4668 | /* returns 0 on failure to allocate runtime */ |
4669 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4670 | { | |
4671 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4672 | int ret; | |
4673 | ||
4674 | raw_spin_lock(&cfs_b->lock); | |
4675 | ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); | |
4676 | raw_spin_unlock(&cfs_b->lock); | |
4677 | ||
4678 | return ret; | |
4679 | } | |
4680 | ||
9dbdb155 | 4681 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4682 | { |
4683 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4684 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4685 | |
4686 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4687 | return; |
4688 | ||
5e2d2cc2 L |
4689 | if (cfs_rq->throttled) |
4690 | return; | |
85dac906 PT |
4691 | /* |
4692 | * if we're unable to extend our runtime we resched so that the active | |
4693 | * hierarchy can be throttled | |
4694 | */ | |
4695 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4696 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4697 | } |
4698 | ||
6c16a6dc | 4699 | static __always_inline |
9dbdb155 | 4700 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4701 | { |
56f570e5 | 4702 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4703 | return; |
4704 | ||
4705 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4706 | } | |
4707 | ||
85dac906 PT |
4708 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4709 | { | |
56f570e5 | 4710 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4711 | } |
4712 | ||
64660c86 PT |
4713 | /* check whether cfs_rq, or any parent, is throttled */ |
4714 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4715 | { | |
56f570e5 | 4716 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4717 | } |
4718 | ||
4719 | /* | |
4720 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4721 | * dest_cpu are members of a throttled hierarchy when performing group | |
4722 | * load-balance operations. | |
4723 | */ | |
4724 | static inline int throttled_lb_pair(struct task_group *tg, | |
4725 | int src_cpu, int dest_cpu) | |
4726 | { | |
4727 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4728 | ||
4729 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4730 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4731 | ||
4732 | return throttled_hierarchy(src_cfs_rq) || | |
4733 | throttled_hierarchy(dest_cfs_rq); | |
4734 | } | |
4735 | ||
64660c86 PT |
4736 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
4737 | { | |
4738 | struct rq *rq = data; | |
4739 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4740 | ||
4741 | cfs_rq->throttle_count--; | |
64660c86 | 4742 | if (!cfs_rq->throttle_count) { |
78becc27 | 4743 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4744 | cfs_rq->throttled_clock_task; |
31bc6aea VG |
4745 | |
4746 | /* Add cfs_rq with already running entity in the list */ | |
4747 | if (cfs_rq->nr_running >= 1) | |
4748 | list_add_leaf_cfs_rq(cfs_rq); | |
64660c86 | 4749 | } |
64660c86 PT |
4750 | |
4751 | return 0; | |
4752 | } | |
4753 | ||
4754 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4755 | { | |
4756 | struct rq *rq = data; | |
4757 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4758 | ||
82958366 | 4759 | /* group is entering throttled state, stop time */ |
31bc6aea | 4760 | if (!cfs_rq->throttle_count) { |
78becc27 | 4761 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
31bc6aea VG |
4762 | list_del_leaf_cfs_rq(cfs_rq); |
4763 | } | |
64660c86 PT |
4764 | cfs_rq->throttle_count++; |
4765 | ||
4766 | return 0; | |
4767 | } | |
4768 | ||
e98fa02c | 4769 | static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4770 | { |
4771 | struct rq *rq = rq_of(cfs_rq); | |
4772 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4773 | struct sched_entity *se; | |
43e9f7f2 | 4774 | long task_delta, idle_task_delta, dequeue = 1; |
e98fa02c PT |
4775 | |
4776 | raw_spin_lock(&cfs_b->lock); | |
4777 | /* This will start the period timer if necessary */ | |
4778 | if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { | |
4779 | /* | |
4780 | * We have raced with bandwidth becoming available, and if we | |
4781 | * actually throttled the timer might not unthrottle us for an | |
4782 | * entire period. We additionally needed to make sure that any | |
4783 | * subsequent check_cfs_rq_runtime calls agree not to throttle | |
4784 | * us, as we may commit to do cfs put_prev+pick_next, so we ask | |
4785 | * for 1ns of runtime rather than just check cfs_b. | |
4786 | */ | |
4787 | dequeue = 0; | |
4788 | } else { | |
4789 | list_add_tail_rcu(&cfs_rq->throttled_list, | |
4790 | &cfs_b->throttled_cfs_rq); | |
4791 | } | |
4792 | raw_spin_unlock(&cfs_b->lock); | |
4793 | ||
4794 | if (!dequeue) | |
4795 | return false; /* Throttle no longer required. */ | |
85dac906 PT |
4796 | |
4797 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4798 | ||
f1b17280 | 4799 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4800 | rcu_read_lock(); |
4801 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4802 | rcu_read_unlock(); | |
85dac906 PT |
4803 | |
4804 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4805 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
4806 | for_each_sched_entity(se) { |
4807 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4808 | /* throttled entity or throttle-on-deactivate */ | |
4809 | if (!se->on_rq) | |
b6d37a76 | 4810 | goto done; |
85dac906 | 4811 | |
b6d37a76 | 4812 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
6212437f | 4813 | |
85dac906 | 4814 | qcfs_rq->h_nr_running -= task_delta; |
43e9f7f2 | 4815 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 | 4816 | |
b6d37a76 PW |
4817 | if (qcfs_rq->load.weight) { |
4818 | /* Avoid re-evaluating load for this entity: */ | |
4819 | se = parent_entity(se); | |
4820 | break; | |
4821 | } | |
4822 | } | |
4823 | ||
4824 | for_each_sched_entity(se) { | |
4825 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4826 | /* throttled entity or throttle-on-deactivate */ | |
4827 | if (!se->on_rq) | |
4828 | goto done; | |
4829 | ||
4830 | update_load_avg(qcfs_rq, se, 0); | |
4831 | se_update_runnable(se); | |
4832 | ||
4833 | qcfs_rq->h_nr_running -= task_delta; | |
4834 | qcfs_rq->idle_h_nr_running -= idle_task_delta; | |
85dac906 PT |
4835 | } |
4836 | ||
b6d37a76 PW |
4837 | /* At this point se is NULL and we are at root level*/ |
4838 | sub_nr_running(rq, task_delta); | |
85dac906 | 4839 | |
b6d37a76 | 4840 | done: |
c06f04c7 | 4841 | /* |
e98fa02c PT |
4842 | * Note: distribution will already see us throttled via the |
4843 | * throttled-list. rq->lock protects completion. | |
c06f04c7 | 4844 | */ |
e98fa02c PT |
4845 | cfs_rq->throttled = 1; |
4846 | cfs_rq->throttled_clock = rq_clock(rq); | |
4847 | return true; | |
85dac906 PT |
4848 | } |
4849 | ||
029632fb | 4850 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4851 | { |
4852 | struct rq *rq = rq_of(cfs_rq); | |
4853 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4854 | struct sched_entity *se; | |
43e9f7f2 | 4855 | long task_delta, idle_task_delta; |
671fd9da | 4856 | |
22b958d8 | 4857 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4858 | |
4859 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4860 | |
4861 | update_rq_clock(rq); | |
4862 | ||
671fd9da | 4863 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4864 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4865 | list_del_rcu(&cfs_rq->throttled_list); |
4866 | raw_spin_unlock(&cfs_b->lock); | |
4867 | ||
64660c86 PT |
4868 | /* update hierarchical throttle state */ |
4869 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4870 | ||
671fd9da PT |
4871 | if (!cfs_rq->load.weight) |
4872 | return; | |
4873 | ||
4874 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4875 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da PT |
4876 | for_each_sched_entity(se) { |
4877 | if (se->on_rq) | |
39f23ce0 VG |
4878 | break; |
4879 | cfs_rq = cfs_rq_of(se); | |
4880 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
4881 | ||
4882 | cfs_rq->h_nr_running += task_delta; | |
4883 | cfs_rq->idle_h_nr_running += idle_task_delta; | |
4884 | ||
4885 | /* end evaluation on encountering a throttled cfs_rq */ | |
4886 | if (cfs_rq_throttled(cfs_rq)) | |
4887 | goto unthrottle_throttle; | |
4888 | } | |
671fd9da | 4889 | |
39f23ce0 | 4890 | for_each_sched_entity(se) { |
671fd9da | 4891 | cfs_rq = cfs_rq_of(se); |
39f23ce0 VG |
4892 | |
4893 | update_load_avg(cfs_rq, se, UPDATE_TG); | |
4894 | se_update_runnable(se); | |
6212437f | 4895 | |
671fd9da | 4896 | cfs_rq->h_nr_running += task_delta; |
43e9f7f2 | 4897 | cfs_rq->idle_h_nr_running += idle_task_delta; |
671fd9da | 4898 | |
39f23ce0 VG |
4899 | |
4900 | /* end evaluation on encountering a throttled cfs_rq */ | |
671fd9da | 4901 | if (cfs_rq_throttled(cfs_rq)) |
39f23ce0 VG |
4902 | goto unthrottle_throttle; |
4903 | ||
4904 | /* | |
4905 | * One parent has been throttled and cfs_rq removed from the | |
4906 | * list. Add it back to not break the leaf list. | |
4907 | */ | |
4908 | if (throttled_hierarchy(cfs_rq)) | |
4909 | list_add_leaf_cfs_rq(cfs_rq); | |
671fd9da PT |
4910 | } |
4911 | ||
39f23ce0 VG |
4912 | /* At this point se is NULL and we are at root level*/ |
4913 | add_nr_running(rq, task_delta); | |
671fd9da | 4914 | |
39f23ce0 | 4915 | unthrottle_throttle: |
fe61468b VG |
4916 | /* |
4917 | * The cfs_rq_throttled() breaks in the above iteration can result in | |
4918 | * incomplete leaf list maintenance, resulting in triggering the | |
4919 | * assertion below. | |
4920 | */ | |
4921 | for_each_sched_entity(se) { | |
4922 | cfs_rq = cfs_rq_of(se); | |
4923 | ||
39f23ce0 VG |
4924 | if (list_add_leaf_cfs_rq(cfs_rq)) |
4925 | break; | |
fe61468b VG |
4926 | } |
4927 | ||
4928 | assert_list_leaf_cfs_rq(rq); | |
4929 | ||
97fb7a0a | 4930 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 4931 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 4932 | resched_curr(rq); |
671fd9da PT |
4933 | } |
4934 | ||
26a8b127 | 4935 | static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) |
671fd9da PT |
4936 | { |
4937 | struct cfs_rq *cfs_rq; | |
26a8b127 | 4938 | u64 runtime, remaining = 1; |
671fd9da PT |
4939 | |
4940 | rcu_read_lock(); | |
4941 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4942 | throttled_list) { | |
4943 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 4944 | struct rq_flags rf; |
671fd9da | 4945 | |
c0ad4aa4 | 4946 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
4947 | if (!cfs_rq_throttled(cfs_rq)) |
4948 | goto next; | |
4949 | ||
5e2d2cc2 L |
4950 | /* By the above check, this should never be true */ |
4951 | SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); | |
4952 | ||
26a8b127 | 4953 | raw_spin_lock(&cfs_b->lock); |
671fd9da | 4954 | runtime = -cfs_rq->runtime_remaining + 1; |
26a8b127 HC |
4955 | if (runtime > cfs_b->runtime) |
4956 | runtime = cfs_b->runtime; | |
4957 | cfs_b->runtime -= runtime; | |
4958 | remaining = cfs_b->runtime; | |
4959 | raw_spin_unlock(&cfs_b->lock); | |
671fd9da PT |
4960 | |
4961 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
4962 | |
4963 | /* we check whether we're throttled above */ | |
4964 | if (cfs_rq->runtime_remaining > 0) | |
4965 | unthrottle_cfs_rq(cfs_rq); | |
4966 | ||
4967 | next: | |
c0ad4aa4 | 4968 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
4969 | |
4970 | if (!remaining) | |
4971 | break; | |
4972 | } | |
4973 | rcu_read_unlock(); | |
671fd9da PT |
4974 | } |
4975 | ||
58088ad0 PT |
4976 | /* |
4977 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4978 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4979 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4980 | * used to track this state. | |
4981 | */ | |
c0ad4aa4 | 4982 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 4983 | { |
51f2176d | 4984 | int throttled; |
58088ad0 | 4985 | |
58088ad0 PT |
4986 | /* no need to continue the timer with no bandwidth constraint */ |
4987 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4988 | goto out_deactivate; |
58088ad0 | 4989 | |
671fd9da | 4990 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4991 | cfs_b->nr_periods += overrun; |
671fd9da | 4992 | |
51f2176d BS |
4993 | /* |
4994 | * idle depends on !throttled (for the case of a large deficit), and if | |
4995 | * we're going inactive then everything else can be deferred | |
4996 | */ | |
4997 | if (cfs_b->idle && !throttled) | |
4998 | goto out_deactivate; | |
a9cf55b2 PT |
4999 | |
5000 | __refill_cfs_bandwidth_runtime(cfs_b); | |
5001 | ||
671fd9da PT |
5002 | if (!throttled) { |
5003 | /* mark as potentially idle for the upcoming period */ | |
5004 | cfs_b->idle = 1; | |
51f2176d | 5005 | return 0; |
671fd9da PT |
5006 | } |
5007 | ||
e8da1b18 NR |
5008 | /* account preceding periods in which throttling occurred */ |
5009 | cfs_b->nr_throttled += overrun; | |
5010 | ||
671fd9da | 5011 | /* |
26a8b127 | 5012 | * This check is repeated as we release cfs_b->lock while we unthrottle. |
671fd9da | 5013 | */ |
ab93a4bc | 5014 | while (throttled && cfs_b->runtime > 0) { |
c0ad4aa4 | 5015 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 5016 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
26a8b127 | 5017 | distribute_cfs_runtime(cfs_b); |
c0ad4aa4 | 5018 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da PT |
5019 | |
5020 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
5021 | } | |
58088ad0 | 5022 | |
671fd9da PT |
5023 | /* |
5024 | * While we are ensured activity in the period following an | |
5025 | * unthrottle, this also covers the case in which the new bandwidth is | |
5026 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
5027 | * timer to remain active while there are any throttled entities.) | |
5028 | */ | |
5029 | cfs_b->idle = 0; | |
58088ad0 | 5030 | |
51f2176d BS |
5031 | return 0; |
5032 | ||
5033 | out_deactivate: | |
51f2176d | 5034 | return 1; |
58088ad0 | 5035 | } |
d3d9dc33 | 5036 | |
d8b4986d PT |
5037 | /* a cfs_rq won't donate quota below this amount */ |
5038 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
5039 | /* minimum remaining period time to redistribute slack quota */ | |
5040 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
5041 | /* how long we wait to gather additional slack before distributing */ | |
5042 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
5043 | ||
db06e78c BS |
5044 | /* |
5045 | * Are we near the end of the current quota period? | |
5046 | * | |
5047 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 5048 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
5049 | * migrate_hrtimers, base is never cleared, so we are fine. |
5050 | */ | |
d8b4986d PT |
5051 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
5052 | { | |
5053 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
5054 | u64 remaining; | |
5055 | ||
5056 | /* if the call-back is running a quota refresh is already occurring */ | |
5057 | if (hrtimer_callback_running(refresh_timer)) | |
5058 | return 1; | |
5059 | ||
5060 | /* is a quota refresh about to occur? */ | |
5061 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
5062 | if (remaining < min_expire) | |
5063 | return 1; | |
5064 | ||
5065 | return 0; | |
5066 | } | |
5067 | ||
5068 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
5069 | { | |
5070 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
5071 | ||
5072 | /* if there's a quota refresh soon don't bother with slack */ | |
5073 | if (runtime_refresh_within(cfs_b, min_left)) | |
5074 | return; | |
5075 | ||
66567fcb | 5076 | /* don't push forwards an existing deferred unthrottle */ |
5077 | if (cfs_b->slack_started) | |
5078 | return; | |
5079 | cfs_b->slack_started = true; | |
5080 | ||
4cfafd30 PZ |
5081 | hrtimer_start(&cfs_b->slack_timer, |
5082 | ns_to_ktime(cfs_bandwidth_slack_period), | |
5083 | HRTIMER_MODE_REL); | |
d8b4986d PT |
5084 | } |
5085 | ||
5086 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
5087 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5088 | { | |
5089 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5090 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
5091 | ||
5092 | if (slack_runtime <= 0) | |
5093 | return; | |
5094 | ||
5095 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 5096 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
5097 | cfs_b->runtime += slack_runtime; |
5098 | ||
5099 | /* we are under rq->lock, defer unthrottling using a timer */ | |
5100 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
5101 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
5102 | start_cfs_slack_bandwidth(cfs_b); | |
5103 | } | |
5104 | raw_spin_unlock(&cfs_b->lock); | |
5105 | ||
5106 | /* even if it's not valid for return we don't want to try again */ | |
5107 | cfs_rq->runtime_remaining -= slack_runtime; | |
5108 | } | |
5109 | ||
5110 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5111 | { | |
56f570e5 PT |
5112 | if (!cfs_bandwidth_used()) |
5113 | return; | |
5114 | ||
fccfdc6f | 5115 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
5116 | return; |
5117 | ||
5118 | __return_cfs_rq_runtime(cfs_rq); | |
5119 | } | |
5120 | ||
5121 | /* | |
5122 | * This is done with a timer (instead of inline with bandwidth return) since | |
5123 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
5124 | */ | |
5125 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
5126 | { | |
5127 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 5128 | unsigned long flags; |
d8b4986d PT |
5129 | |
5130 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 5131 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 5132 | cfs_b->slack_started = false; |
baa9be4f | 5133 | |
db06e78c | 5134 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 5135 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 5136 | return; |
db06e78c | 5137 | } |
d8b4986d | 5138 | |
c06f04c7 | 5139 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 5140 | runtime = cfs_b->runtime; |
c06f04c7 | 5141 | |
c0ad4aa4 | 5142 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
5143 | |
5144 | if (!runtime) | |
5145 | return; | |
5146 | ||
26a8b127 | 5147 | distribute_cfs_runtime(cfs_b); |
d8b4986d PT |
5148 | } |
5149 | ||
d3d9dc33 PT |
5150 | /* |
5151 | * When a group wakes up we want to make sure that its quota is not already | |
5152 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
c034f48e | 5153 | * runtime as update_curr() throttling can not trigger until it's on-rq. |
d3d9dc33 PT |
5154 | */ |
5155 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
5156 | { | |
56f570e5 PT |
5157 | if (!cfs_bandwidth_used()) |
5158 | return; | |
5159 | ||
d3d9dc33 PT |
5160 | /* an active group must be handled by the update_curr()->put() path */ |
5161 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
5162 | return; | |
5163 | ||
5164 | /* ensure the group is not already throttled */ | |
5165 | if (cfs_rq_throttled(cfs_rq)) | |
5166 | return; | |
5167 | ||
5168 | /* update runtime allocation */ | |
5169 | account_cfs_rq_runtime(cfs_rq, 0); | |
5170 | if (cfs_rq->runtime_remaining <= 0) | |
5171 | throttle_cfs_rq(cfs_rq); | |
5172 | } | |
5173 | ||
55e16d30 PZ |
5174 | static void sync_throttle(struct task_group *tg, int cpu) |
5175 | { | |
5176 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
5177 | ||
5178 | if (!cfs_bandwidth_used()) | |
5179 | return; | |
5180 | ||
5181 | if (!tg->parent) | |
5182 | return; | |
5183 | ||
5184 | cfs_rq = tg->cfs_rq[cpu]; | |
5185 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
5186 | ||
5187 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 5188 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
5189 | } |
5190 | ||
d3d9dc33 | 5191 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 5192 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 5193 | { |
56f570e5 | 5194 | if (!cfs_bandwidth_used()) |
678d5718 | 5195 | return false; |
56f570e5 | 5196 | |
d3d9dc33 | 5197 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 5198 | return false; |
d3d9dc33 PT |
5199 | |
5200 | /* | |
5201 | * it's possible for a throttled entity to be forced into a running | |
5202 | * state (e.g. set_curr_task), in this case we're finished. | |
5203 | */ | |
5204 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 5205 | return true; |
d3d9dc33 | 5206 | |
e98fa02c | 5207 | return throttle_cfs_rq(cfs_rq); |
d3d9dc33 | 5208 | } |
029632fb | 5209 | |
029632fb PZ |
5210 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
5211 | { | |
5212 | struct cfs_bandwidth *cfs_b = | |
5213 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 5214 | |
029632fb PZ |
5215 | do_sched_cfs_slack_timer(cfs_b); |
5216 | ||
5217 | return HRTIMER_NORESTART; | |
5218 | } | |
5219 | ||
2e8e1922 PA |
5220 | extern const u64 max_cfs_quota_period; |
5221 | ||
029632fb PZ |
5222 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
5223 | { | |
5224 | struct cfs_bandwidth *cfs_b = | |
5225 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 5226 | unsigned long flags; |
029632fb PZ |
5227 | int overrun; |
5228 | int idle = 0; | |
2e8e1922 | 5229 | int count = 0; |
029632fb | 5230 | |
c0ad4aa4 | 5231 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 5232 | for (;;) { |
77a4d1a1 | 5233 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
5234 | if (!overrun) |
5235 | break; | |
5236 | ||
5a6d6a6c HC |
5237 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
5238 | ||
2e8e1922 PA |
5239 | if (++count > 3) { |
5240 | u64 new, old = ktime_to_ns(cfs_b->period); | |
5241 | ||
4929a4e6 XZ |
5242 | /* |
5243 | * Grow period by a factor of 2 to avoid losing precision. | |
5244 | * Precision loss in the quota/period ratio can cause __cfs_schedulable | |
5245 | * to fail. | |
5246 | */ | |
5247 | new = old * 2; | |
5248 | if (new < max_cfs_quota_period) { | |
5249 | cfs_b->period = ns_to_ktime(new); | |
5250 | cfs_b->quota *= 2; | |
5251 | ||
5252 | pr_warn_ratelimited( | |
5253 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5254 | smp_processor_id(), | |
5255 | div_u64(new, NSEC_PER_USEC), | |
5256 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5257 | } else { | |
5258 | pr_warn_ratelimited( | |
5259 | "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5260 | smp_processor_id(), | |
5261 | div_u64(old, NSEC_PER_USEC), | |
5262 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5263 | } | |
2e8e1922 PA |
5264 | |
5265 | /* reset count so we don't come right back in here */ | |
5266 | count = 0; | |
5267 | } | |
029632fb | 5268 | } |
4cfafd30 PZ |
5269 | if (idle) |
5270 | cfs_b->period_active = 0; | |
c0ad4aa4 | 5271 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
5272 | |
5273 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
5274 | } | |
5275 | ||
5276 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5277 | { | |
5278 | raw_spin_lock_init(&cfs_b->lock); | |
5279 | cfs_b->runtime = 0; | |
5280 | cfs_b->quota = RUNTIME_INF; | |
5281 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
5282 | ||
5283 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 5284 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5285 | cfs_b->period_timer.function = sched_cfs_period_timer; |
5286 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5287 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
66567fcb | 5288 | cfs_b->slack_started = false; |
029632fb PZ |
5289 | } |
5290 | ||
5291 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5292 | { | |
5293 | cfs_rq->runtime_enabled = 0; | |
5294 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
5295 | } | |
5296 | ||
77a4d1a1 | 5297 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 5298 | { |
4cfafd30 | 5299 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 5300 | |
f1d1be8a XP |
5301 | if (cfs_b->period_active) |
5302 | return; | |
5303 | ||
5304 | cfs_b->period_active = 1; | |
763a9ec0 | 5305 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); |
f1d1be8a | 5306 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5307 | } |
5308 | ||
5309 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5310 | { | |
7f1a169b TH |
5311 | /* init_cfs_bandwidth() was not called */ |
5312 | if (!cfs_b->throttled_cfs_rq.next) | |
5313 | return; | |
5314 | ||
029632fb PZ |
5315 | hrtimer_cancel(&cfs_b->period_timer); |
5316 | hrtimer_cancel(&cfs_b->slack_timer); | |
5317 | } | |
5318 | ||
502ce005 | 5319 | /* |
97fb7a0a | 5320 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
5321 | * |
5322 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5323 | * bits doesn't do much. | |
5324 | */ | |
5325 | ||
3b03706f | 5326 | /* cpu online callback */ |
0e59bdae KT |
5327 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5328 | { | |
502ce005 | 5329 | struct task_group *tg; |
0e59bdae | 5330 | |
502ce005 PZ |
5331 | lockdep_assert_held(&rq->lock); |
5332 | ||
5333 | rcu_read_lock(); | |
5334 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5335 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5336 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5337 | |
5338 | raw_spin_lock(&cfs_b->lock); | |
5339 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5340 | raw_spin_unlock(&cfs_b->lock); | |
5341 | } | |
502ce005 | 5342 | rcu_read_unlock(); |
0e59bdae KT |
5343 | } |
5344 | ||
502ce005 | 5345 | /* cpu offline callback */ |
38dc3348 | 5346 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5347 | { |
502ce005 PZ |
5348 | struct task_group *tg; |
5349 | ||
5350 | lockdep_assert_held(&rq->lock); | |
5351 | ||
5352 | rcu_read_lock(); | |
5353 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5354 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5355 | |
029632fb PZ |
5356 | if (!cfs_rq->runtime_enabled) |
5357 | continue; | |
5358 | ||
5359 | /* | |
5360 | * clock_task is not advancing so we just need to make sure | |
5361 | * there's some valid quota amount | |
5362 | */ | |
51f2176d | 5363 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 5364 | /* |
97fb7a0a | 5365 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
5366 | * in take_cpu_down(), so we prevent new cfs throttling here. |
5367 | */ | |
5368 | cfs_rq->runtime_enabled = 0; | |
5369 | ||
029632fb PZ |
5370 | if (cfs_rq_throttled(cfs_rq)) |
5371 | unthrottle_cfs_rq(cfs_rq); | |
5372 | } | |
502ce005 | 5373 | rcu_read_unlock(); |
029632fb PZ |
5374 | } |
5375 | ||
5376 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f6783319 VG |
5377 | |
5378 | static inline bool cfs_bandwidth_used(void) | |
5379 | { | |
5380 | return false; | |
5381 | } | |
5382 | ||
9dbdb155 | 5383 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5384 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5385 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5386 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5387 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5388 | |
5389 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5390 | { | |
5391 | return 0; | |
5392 | } | |
64660c86 PT |
5393 | |
5394 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5395 | { | |
5396 | return 0; | |
5397 | } | |
5398 | ||
5399 | static inline int throttled_lb_pair(struct task_group *tg, | |
5400 | int src_cpu, int dest_cpu) | |
5401 | { | |
5402 | return 0; | |
5403 | } | |
029632fb PZ |
5404 | |
5405 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5406 | ||
5407 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5408 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5409 | #endif |
5410 | ||
029632fb PZ |
5411 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5412 | { | |
5413 | return NULL; | |
5414 | } | |
5415 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5416 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5417 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5418 | |
5419 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5420 | ||
bf0f6f24 IM |
5421 | /************************************************** |
5422 | * CFS operations on tasks: | |
5423 | */ | |
5424 | ||
8f4d37ec PZ |
5425 | #ifdef CONFIG_SCHED_HRTICK |
5426 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5427 | { | |
8f4d37ec PZ |
5428 | struct sched_entity *se = &p->se; |
5429 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5430 | ||
9148a3a1 | 5431 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5432 | |
8bf46a39 | 5433 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5434 | u64 slice = sched_slice(cfs_rq, se); |
5435 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5436 | s64 delta = slice - ran; | |
5437 | ||
5438 | if (delta < 0) { | |
65bcf072 | 5439 | if (task_current(rq, p)) |
8875125e | 5440 | resched_curr(rq); |
8f4d37ec PZ |
5441 | return; |
5442 | } | |
31656519 | 5443 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5444 | } |
5445 | } | |
a4c2f00f PZ |
5446 | |
5447 | /* | |
5448 | * called from enqueue/dequeue and updates the hrtick when the | |
5449 | * current task is from our class and nr_running is low enough | |
5450 | * to matter. | |
5451 | */ | |
5452 | static void hrtick_update(struct rq *rq) | |
5453 | { | |
5454 | struct task_struct *curr = rq->curr; | |
5455 | ||
e0ee463c | 5456 | if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5457 | return; |
5458 | ||
5459 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5460 | hrtick_start_fair(rq, curr); | |
5461 | } | |
55e12e5e | 5462 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5463 | static inline void |
5464 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5465 | { | |
5466 | } | |
a4c2f00f PZ |
5467 | |
5468 | static inline void hrtick_update(struct rq *rq) | |
5469 | { | |
5470 | } | |
8f4d37ec PZ |
5471 | #endif |
5472 | ||
2802bf3c MR |
5473 | #ifdef CONFIG_SMP |
5474 | static inline unsigned long cpu_util(int cpu); | |
2802bf3c MR |
5475 | |
5476 | static inline bool cpu_overutilized(int cpu) | |
5477 | { | |
60e17f5c | 5478 | return !fits_capacity(cpu_util(cpu), capacity_of(cpu)); |
2802bf3c MR |
5479 | } |
5480 | ||
5481 | static inline void update_overutilized_status(struct rq *rq) | |
5482 | { | |
f9f240f9 | 5483 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { |
2802bf3c | 5484 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); |
f9f240f9 QY |
5485 | trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); |
5486 | } | |
2802bf3c MR |
5487 | } |
5488 | #else | |
5489 | static inline void update_overutilized_status(struct rq *rq) { } | |
5490 | #endif | |
5491 | ||
323af6de VK |
5492 | /* Runqueue only has SCHED_IDLE tasks enqueued */ |
5493 | static int sched_idle_rq(struct rq *rq) | |
5494 | { | |
5495 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
5496 | rq->nr_running); | |
5497 | } | |
5498 | ||
afa70d94 | 5499 | #ifdef CONFIG_SMP |
323af6de VK |
5500 | static int sched_idle_cpu(int cpu) |
5501 | { | |
5502 | return sched_idle_rq(cpu_rq(cpu)); | |
5503 | } | |
afa70d94 | 5504 | #endif |
323af6de | 5505 | |
bf0f6f24 IM |
5506 | /* |
5507 | * The enqueue_task method is called before nr_running is | |
5508 | * increased. Here we update the fair scheduling stats and | |
5509 | * then put the task into the rbtree: | |
5510 | */ | |
ea87bb78 | 5511 | static void |
371fd7e7 | 5512 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5513 | { |
5514 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5515 | struct sched_entity *se = &p->se; |
43e9f7f2 | 5516 | int idle_h_nr_running = task_has_idle_policy(p); |
8e1ac429 | 5517 | int task_new = !(flags & ENQUEUE_WAKEUP); |
bf0f6f24 | 5518 | |
2539fc82 PB |
5519 | /* |
5520 | * The code below (indirectly) updates schedutil which looks at | |
5521 | * the cfs_rq utilization to select a frequency. | |
5522 | * Let's add the task's estimated utilization to the cfs_rq's | |
5523 | * estimated utilization, before we update schedutil. | |
5524 | */ | |
5525 | util_est_enqueue(&rq->cfs, p); | |
5526 | ||
8c34ab19 RW |
5527 | /* |
5528 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5529 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5530 | * passed. | |
5531 | */ | |
5532 | if (p->in_iowait) | |
674e7541 | 5533 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5534 | |
bf0f6f24 | 5535 | for_each_sched_entity(se) { |
62fb1851 | 5536 | if (se->on_rq) |
bf0f6f24 IM |
5537 | break; |
5538 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5539 | enqueue_entity(cfs_rq, se, flags); |
85dac906 | 5540 | |
953bfcd1 | 5541 | cfs_rq->h_nr_running++; |
43e9f7f2 | 5542 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 5543 | |
6d4d2246 VG |
5544 | /* end evaluation on encountering a throttled cfs_rq */ |
5545 | if (cfs_rq_throttled(cfs_rq)) | |
5546 | goto enqueue_throttle; | |
5547 | ||
88ec22d3 | 5548 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5549 | } |
8f4d37ec | 5550 | |
2069dd75 | 5551 | for_each_sched_entity(se) { |
0f317143 | 5552 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5553 | |
88c0616e | 5554 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5555 | se_update_runnable(se); |
1ea6c46a | 5556 | update_cfs_group(se); |
6d4d2246 VG |
5557 | |
5558 | cfs_rq->h_nr_running++; | |
5559 | cfs_rq->idle_h_nr_running += idle_h_nr_running; | |
5ab297ba VG |
5560 | |
5561 | /* end evaluation on encountering a throttled cfs_rq */ | |
5562 | if (cfs_rq_throttled(cfs_rq)) | |
5563 | goto enqueue_throttle; | |
b34cb07d PA |
5564 | |
5565 | /* | |
5566 | * One parent has been throttled and cfs_rq removed from the | |
5567 | * list. Add it back to not break the leaf list. | |
5568 | */ | |
5569 | if (throttled_hierarchy(cfs_rq)) | |
5570 | list_add_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
5571 | } |
5572 | ||
7d148be6 VG |
5573 | /* At this point se is NULL and we are at root level*/ |
5574 | add_nr_running(rq, 1); | |
2802bf3c | 5575 | |
7d148be6 VG |
5576 | /* |
5577 | * Since new tasks are assigned an initial util_avg equal to | |
5578 | * half of the spare capacity of their CPU, tiny tasks have the | |
5579 | * ability to cross the overutilized threshold, which will | |
5580 | * result in the load balancer ruining all the task placement | |
5581 | * done by EAS. As a way to mitigate that effect, do not account | |
5582 | * for the first enqueue operation of new tasks during the | |
5583 | * overutilized flag detection. | |
5584 | * | |
5585 | * A better way of solving this problem would be to wait for | |
5586 | * the PELT signals of tasks to converge before taking them | |
5587 | * into account, but that is not straightforward to implement, | |
5588 | * and the following generally works well enough in practice. | |
5589 | */ | |
8e1ac429 | 5590 | if (!task_new) |
7d148be6 | 5591 | update_overutilized_status(rq); |
cd126afe | 5592 | |
7d148be6 | 5593 | enqueue_throttle: |
f6783319 VG |
5594 | if (cfs_bandwidth_used()) { |
5595 | /* | |
5596 | * When bandwidth control is enabled; the cfs_rq_throttled() | |
5597 | * breaks in the above iteration can result in incomplete | |
5598 | * leaf list maintenance, resulting in triggering the assertion | |
5599 | * below. | |
5600 | */ | |
5601 | for_each_sched_entity(se) { | |
5602 | cfs_rq = cfs_rq_of(se); | |
5603 | ||
5604 | if (list_add_leaf_cfs_rq(cfs_rq)) | |
5605 | break; | |
5606 | } | |
5607 | } | |
5608 | ||
5d299eab PZ |
5609 | assert_list_leaf_cfs_rq(rq); |
5610 | ||
a4c2f00f | 5611 | hrtick_update(rq); |
bf0f6f24 IM |
5612 | } |
5613 | ||
2f36825b VP |
5614 | static void set_next_buddy(struct sched_entity *se); |
5615 | ||
bf0f6f24 IM |
5616 | /* |
5617 | * The dequeue_task method is called before nr_running is | |
5618 | * decreased. We remove the task from the rbtree and | |
5619 | * update the fair scheduling stats: | |
5620 | */ | |
371fd7e7 | 5621 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5622 | { |
5623 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5624 | struct sched_entity *se = &p->se; |
2f36825b | 5625 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 5626 | int idle_h_nr_running = task_has_idle_policy(p); |
323af6de | 5627 | bool was_sched_idle = sched_idle_rq(rq); |
bf0f6f24 | 5628 | |
8c1f560c XY |
5629 | util_est_dequeue(&rq->cfs, p); |
5630 | ||
bf0f6f24 IM |
5631 | for_each_sched_entity(se) { |
5632 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5633 | dequeue_entity(cfs_rq, se, flags); |
85dac906 | 5634 | |
953bfcd1 | 5635 | cfs_rq->h_nr_running--; |
43e9f7f2 | 5636 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 5637 | |
6d4d2246 VG |
5638 | /* end evaluation on encountering a throttled cfs_rq */ |
5639 | if (cfs_rq_throttled(cfs_rq)) | |
5640 | goto dequeue_throttle; | |
5641 | ||
bf0f6f24 | 5642 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5643 | if (cfs_rq->load.weight) { |
754bd598 KK |
5644 | /* Avoid re-evaluating load for this entity: */ |
5645 | se = parent_entity(se); | |
2f36825b VP |
5646 | /* |
5647 | * Bias pick_next to pick a task from this cfs_rq, as | |
5648 | * p is sleeping when it is within its sched_slice. | |
5649 | */ | |
754bd598 KK |
5650 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5651 | set_next_buddy(se); | |
bf0f6f24 | 5652 | break; |
2f36825b | 5653 | } |
371fd7e7 | 5654 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5655 | } |
8f4d37ec | 5656 | |
2069dd75 | 5657 | for_each_sched_entity(se) { |
0f317143 | 5658 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5659 | |
88c0616e | 5660 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5661 | se_update_runnable(se); |
1ea6c46a | 5662 | update_cfs_group(se); |
6d4d2246 VG |
5663 | |
5664 | cfs_rq->h_nr_running--; | |
5665 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; | |
5ab297ba VG |
5666 | |
5667 | /* end evaluation on encountering a throttled cfs_rq */ | |
5668 | if (cfs_rq_throttled(cfs_rq)) | |
5669 | goto dequeue_throttle; | |
5670 | ||
2069dd75 PZ |
5671 | } |
5672 | ||
423d02e1 PW |
5673 | /* At this point se is NULL and we are at root level*/ |
5674 | sub_nr_running(rq, 1); | |
cd126afe | 5675 | |
323af6de VK |
5676 | /* balance early to pull high priority tasks */ |
5677 | if (unlikely(!was_sched_idle && sched_idle_rq(rq))) | |
5678 | rq->next_balance = jiffies; | |
5679 | ||
423d02e1 | 5680 | dequeue_throttle: |
8c1f560c | 5681 | util_est_update(&rq->cfs, p, task_sleep); |
a4c2f00f | 5682 | hrtick_update(rq); |
bf0f6f24 IM |
5683 | } |
5684 | ||
e7693a36 | 5685 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5686 | |
5687 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
5688 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5689 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5690 | ||
9fd81dd5 | 5691 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
5692 | |
5693 | static struct { | |
5694 | cpumask_var_t idle_cpus_mask; | |
5695 | atomic_t nr_cpus; | |
f643ea22 | 5696 | int has_blocked; /* Idle CPUS has blocked load */ |
e022e0d3 | 5697 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 5698 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
5699 | } nohz ____cacheline_aligned; |
5700 | ||
9fd81dd5 | 5701 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5702 | |
b0fb1eb4 VG |
5703 | static unsigned long cpu_load(struct rq *rq) |
5704 | { | |
5705 | return cfs_rq_load_avg(&rq->cfs); | |
5706 | } | |
5707 | ||
3318544b VG |
5708 | /* |
5709 | * cpu_load_without - compute CPU load without any contributions from *p | |
5710 | * @cpu: the CPU which load is requested | |
5711 | * @p: the task which load should be discounted | |
5712 | * | |
5713 | * The load of a CPU is defined by the load of tasks currently enqueued on that | |
5714 | * CPU as well as tasks which are currently sleeping after an execution on that | |
5715 | * CPU. | |
5716 | * | |
5717 | * This method returns the load of the specified CPU by discounting the load of | |
5718 | * the specified task, whenever the task is currently contributing to the CPU | |
5719 | * load. | |
5720 | */ | |
5721 | static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) | |
5722 | { | |
5723 | struct cfs_rq *cfs_rq; | |
5724 | unsigned int load; | |
5725 | ||
5726 | /* Task has no contribution or is new */ | |
5727 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5728 | return cpu_load(rq); | |
5729 | ||
5730 | cfs_rq = &rq->cfs; | |
5731 | load = READ_ONCE(cfs_rq->avg.load_avg); | |
5732 | ||
5733 | /* Discount task's util from CPU's util */ | |
5734 | lsub_positive(&load, task_h_load(p)); | |
5735 | ||
5736 | return load; | |
5737 | } | |
5738 | ||
9f683953 VG |
5739 | static unsigned long cpu_runnable(struct rq *rq) |
5740 | { | |
5741 | return cfs_rq_runnable_avg(&rq->cfs); | |
5742 | } | |
5743 | ||
070f5e86 VG |
5744 | static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) |
5745 | { | |
5746 | struct cfs_rq *cfs_rq; | |
5747 | unsigned int runnable; | |
5748 | ||
5749 | /* Task has no contribution or is new */ | |
5750 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5751 | return cpu_runnable(rq); | |
5752 | ||
5753 | cfs_rq = &rq->cfs; | |
5754 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
5755 | ||
5756 | /* Discount task's runnable from CPU's runnable */ | |
5757 | lsub_positive(&runnable, p->se.avg.runnable_avg); | |
5758 | ||
5759 | return runnable; | |
5760 | } | |
5761 | ||
ced549fa | 5762 | static unsigned long capacity_of(int cpu) |
029632fb | 5763 | { |
ced549fa | 5764 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5765 | } |
5766 | ||
c58d25f3 PZ |
5767 | static void record_wakee(struct task_struct *p) |
5768 | { | |
5769 | /* | |
5770 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5771 | * jiffy will not have built up many flips. | |
5772 | */ | |
5773 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5774 | current->wakee_flips >>= 1; | |
5775 | current->wakee_flip_decay_ts = jiffies; | |
5776 | } | |
5777 | ||
5778 | if (current->last_wakee != p) { | |
5779 | current->last_wakee = p; | |
5780 | current->wakee_flips++; | |
5781 | } | |
5782 | } | |
5783 | ||
63b0e9ed MG |
5784 | /* |
5785 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5786 | * |
63b0e9ed | 5787 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5788 | * at a frequency roughly N times higher than one of its wakees. |
5789 | * | |
5790 | * In order to determine whether we should let the load spread vs consolidating | |
5791 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5792 | * partner, and a factor of lls_size higher frequency in the other. | |
5793 | * | |
5794 | * With both conditions met, we can be relatively sure that the relationship is | |
5795 | * non-monogamous, with partner count exceeding socket size. | |
5796 | * | |
5797 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5798 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5799 | * socket size. | |
63b0e9ed | 5800 | */ |
62470419 MW |
5801 | static int wake_wide(struct task_struct *p) |
5802 | { | |
63b0e9ed MG |
5803 | unsigned int master = current->wakee_flips; |
5804 | unsigned int slave = p->wakee_flips; | |
17c891ab | 5805 | int factor = __this_cpu_read(sd_llc_size); |
62470419 | 5806 | |
63b0e9ed MG |
5807 | if (master < slave) |
5808 | swap(master, slave); | |
5809 | if (slave < factor || master < slave * factor) | |
5810 | return 0; | |
5811 | return 1; | |
62470419 MW |
5812 | } |
5813 | ||
90001d67 | 5814 | /* |
d153b153 PZ |
5815 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
5816 | * soonest. For the purpose of speed we only consider the waking and previous | |
5817 | * CPU. | |
90001d67 | 5818 | * |
7332dec0 MG |
5819 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
5820 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
5821 | * |
5822 | * wake_affine_weight() - considers the weight to reflect the average | |
5823 | * scheduling latency of the CPUs. This seems to work | |
5824 | * for the overloaded case. | |
90001d67 | 5825 | */ |
3b76c4a3 | 5826 | static int |
89a55f56 | 5827 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 5828 | { |
7332dec0 MG |
5829 | /* |
5830 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
5831 | * context. Only allow the move if cache is shared. Otherwise an | |
5832 | * interrupt intensive workload could force all tasks onto one | |
5833 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
5834 | * |
5835 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
5836 | * There is no guarantee that the cache hot data from an interrupt | |
5837 | * is more important than cache hot data on the prev_cpu and from | |
5838 | * a cpufreq perspective, it's better to have higher utilisation | |
5839 | * on one CPU. | |
7332dec0 | 5840 | */ |
943d355d RJ |
5841 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
5842 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 5843 | |
d153b153 | 5844 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 5845 | return this_cpu; |
90001d67 | 5846 | |
d8fcb81f JL |
5847 | if (available_idle_cpu(prev_cpu)) |
5848 | return prev_cpu; | |
5849 | ||
3b76c4a3 | 5850 | return nr_cpumask_bits; |
90001d67 PZ |
5851 | } |
5852 | ||
3b76c4a3 | 5853 | static int |
f2cdd9cc PZ |
5854 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
5855 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5856 | { |
90001d67 PZ |
5857 | s64 this_eff_load, prev_eff_load; |
5858 | unsigned long task_load; | |
5859 | ||
11f10e54 | 5860 | this_eff_load = cpu_load(cpu_rq(this_cpu)); |
90001d67 | 5861 | |
90001d67 PZ |
5862 | if (sync) { |
5863 | unsigned long current_load = task_h_load(current); | |
5864 | ||
f2cdd9cc | 5865 | if (current_load > this_eff_load) |
3b76c4a3 | 5866 | return this_cpu; |
90001d67 | 5867 | |
f2cdd9cc | 5868 | this_eff_load -= current_load; |
90001d67 PZ |
5869 | } |
5870 | ||
90001d67 PZ |
5871 | task_load = task_h_load(p); |
5872 | ||
f2cdd9cc PZ |
5873 | this_eff_load += task_load; |
5874 | if (sched_feat(WA_BIAS)) | |
5875 | this_eff_load *= 100; | |
5876 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 5877 | |
11f10e54 | 5878 | prev_eff_load = cpu_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
5879 | prev_eff_load -= task_load; |
5880 | if (sched_feat(WA_BIAS)) | |
5881 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
5882 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 5883 | |
082f764a MG |
5884 | /* |
5885 | * If sync, adjust the weight of prev_eff_load such that if | |
5886 | * prev_eff == this_eff that select_idle_sibling() will consider | |
5887 | * stacking the wakee on top of the waker if no other CPU is | |
5888 | * idle. | |
5889 | */ | |
5890 | if (sync) | |
5891 | prev_eff_load += 1; | |
5892 | ||
5893 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
5894 | } |
5895 | ||
772bd008 | 5896 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 5897 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 5898 | { |
3b76c4a3 | 5899 | int target = nr_cpumask_bits; |
098fb9db | 5900 | |
89a55f56 | 5901 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 5902 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 5903 | |
3b76c4a3 MG |
5904 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
5905 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 5906 | |
ae92882e | 5907 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
3b76c4a3 MG |
5908 | if (target == nr_cpumask_bits) |
5909 | return prev_cpu; | |
098fb9db | 5910 | |
3b76c4a3 MG |
5911 | schedstat_inc(sd->ttwu_move_affine); |
5912 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
5913 | return target; | |
098fb9db IM |
5914 | } |
5915 | ||
aaee1203 | 5916 | static struct sched_group * |
45da2773 | 5917 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); |
aaee1203 PZ |
5918 | |
5919 | /* | |
97fb7a0a | 5920 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
5921 | */ |
5922 | static int | |
18bd1b4b | 5923 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
5924 | { |
5925 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5926 | unsigned int min_exit_latency = UINT_MAX; |
5927 | u64 latest_idle_timestamp = 0; | |
5928 | int least_loaded_cpu = this_cpu; | |
17346452 | 5929 | int shallowest_idle_cpu = -1; |
aaee1203 PZ |
5930 | int i; |
5931 | ||
eaecf41f MR |
5932 | /* Check if we have any choice: */ |
5933 | if (group->group_weight == 1) | |
ae4df9d6 | 5934 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 5935 | |
aaee1203 | 5936 | /* Traverse only the allowed CPUs */ |
3bd37062 | 5937 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
17346452 VK |
5938 | if (sched_idle_cpu(i)) |
5939 | return i; | |
5940 | ||
943d355d | 5941 | if (available_idle_cpu(i)) { |
83a0a96a NP |
5942 | struct rq *rq = cpu_rq(i); |
5943 | struct cpuidle_state *idle = idle_get_state(rq); | |
5944 | if (idle && idle->exit_latency < min_exit_latency) { | |
5945 | /* | |
5946 | * We give priority to a CPU whose idle state | |
5947 | * has the smallest exit latency irrespective | |
5948 | * of any idle timestamp. | |
5949 | */ | |
5950 | min_exit_latency = idle->exit_latency; | |
5951 | latest_idle_timestamp = rq->idle_stamp; | |
5952 | shallowest_idle_cpu = i; | |
5953 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5954 | rq->idle_stamp > latest_idle_timestamp) { | |
5955 | /* | |
5956 | * If equal or no active idle state, then | |
5957 | * the most recently idled CPU might have | |
5958 | * a warmer cache. | |
5959 | */ | |
5960 | latest_idle_timestamp = rq->idle_stamp; | |
5961 | shallowest_idle_cpu = i; | |
5962 | } | |
17346452 | 5963 | } else if (shallowest_idle_cpu == -1) { |
11f10e54 | 5964 | load = cpu_load(cpu_rq(i)); |
18cec7e0 | 5965 | if (load < min_load) { |
83a0a96a NP |
5966 | min_load = load; |
5967 | least_loaded_cpu = i; | |
5968 | } | |
e7693a36 GH |
5969 | } |
5970 | } | |
5971 | ||
17346452 | 5972 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 5973 | } |
e7693a36 | 5974 | |
18bd1b4b BJ |
5975 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
5976 | int cpu, int prev_cpu, int sd_flag) | |
5977 | { | |
93f50f90 | 5978 | int new_cpu = cpu; |
18bd1b4b | 5979 | |
3bd37062 | 5980 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
5981 | return prev_cpu; |
5982 | ||
c976a862 | 5983 | /* |
57abff06 | 5984 | * We need task's util for cpu_util_without, sync it up to |
c469933e | 5985 | * prev_cpu's last_update_time. |
c976a862 VK |
5986 | */ |
5987 | if (!(sd_flag & SD_BALANCE_FORK)) | |
5988 | sync_entity_load_avg(&p->se); | |
5989 | ||
18bd1b4b BJ |
5990 | while (sd) { |
5991 | struct sched_group *group; | |
5992 | struct sched_domain *tmp; | |
5993 | int weight; | |
5994 | ||
5995 | if (!(sd->flags & sd_flag)) { | |
5996 | sd = sd->child; | |
5997 | continue; | |
5998 | } | |
5999 | ||
45da2773 | 6000 | group = find_idlest_group(sd, p, cpu); |
18bd1b4b BJ |
6001 | if (!group) { |
6002 | sd = sd->child; | |
6003 | continue; | |
6004 | } | |
6005 | ||
6006 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 6007 | if (new_cpu == cpu) { |
97fb7a0a | 6008 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
6009 | sd = sd->child; |
6010 | continue; | |
6011 | } | |
6012 | ||
97fb7a0a | 6013 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
6014 | cpu = new_cpu; |
6015 | weight = sd->span_weight; | |
6016 | sd = NULL; | |
6017 | for_each_domain(cpu, tmp) { | |
6018 | if (weight <= tmp->span_weight) | |
6019 | break; | |
6020 | if (tmp->flags & sd_flag) | |
6021 | sd = tmp; | |
6022 | } | |
18bd1b4b BJ |
6023 | } |
6024 | ||
6025 | return new_cpu; | |
6026 | } | |
6027 | ||
9fe1f127 MG |
6028 | static inline int __select_idle_cpu(int cpu) |
6029 | { | |
6030 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) | |
6031 | return cpu; | |
6032 | ||
6033 | return -1; | |
6034 | } | |
6035 | ||
10e2f1ac | 6036 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 6037 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 6038 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
6039 | |
6040 | static inline void set_idle_cores(int cpu, int val) | |
6041 | { | |
6042 | struct sched_domain_shared *sds; | |
6043 | ||
6044 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
6045 | if (sds) | |
6046 | WRITE_ONCE(sds->has_idle_cores, val); | |
6047 | } | |
6048 | ||
6049 | static inline bool test_idle_cores(int cpu, bool def) | |
6050 | { | |
6051 | struct sched_domain_shared *sds; | |
6052 | ||
c722f35b RR |
6053 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
6054 | if (sds) | |
6055 | return READ_ONCE(sds->has_idle_cores); | |
10e2f1ac PZ |
6056 | |
6057 | return def; | |
6058 | } | |
6059 | ||
6060 | /* | |
6061 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
6062 | * information in sd_llc_shared->has_idle_cores. | |
6063 | * | |
6064 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
6065 | * state should be fairly cheap. | |
6066 | */ | |
1b568f0a | 6067 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
6068 | { |
6069 | int core = cpu_of(rq); | |
6070 | int cpu; | |
6071 | ||
6072 | rcu_read_lock(); | |
6073 | if (test_idle_cores(core, true)) | |
6074 | goto unlock; | |
6075 | ||
6076 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6077 | if (cpu == core) | |
6078 | continue; | |
6079 | ||
943d355d | 6080 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6081 | goto unlock; |
6082 | } | |
6083 | ||
6084 | set_idle_cores(core, 1); | |
6085 | unlock: | |
6086 | rcu_read_unlock(); | |
6087 | } | |
6088 | ||
6089 | /* | |
6090 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
6091 | * there are no idle cores left in the system; tracked through | |
6092 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
6093 | */ | |
9fe1f127 | 6094 | static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) |
10e2f1ac | 6095 | { |
9fe1f127 MG |
6096 | bool idle = true; |
6097 | int cpu; | |
10e2f1ac | 6098 | |
1b568f0a | 6099 | if (!static_branch_likely(&sched_smt_present)) |
9fe1f127 | 6100 | return __select_idle_cpu(core); |
10e2f1ac | 6101 | |
9fe1f127 MG |
6102 | for_each_cpu(cpu, cpu_smt_mask(core)) { |
6103 | if (!available_idle_cpu(cpu)) { | |
6104 | idle = false; | |
6105 | if (*idle_cpu == -1) { | |
6106 | if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) { | |
6107 | *idle_cpu = cpu; | |
6108 | break; | |
6109 | } | |
6110 | continue; | |
bec2860a | 6111 | } |
9fe1f127 | 6112 | break; |
10e2f1ac | 6113 | } |
9fe1f127 MG |
6114 | if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr)) |
6115 | *idle_cpu = cpu; | |
10e2f1ac PZ |
6116 | } |
6117 | ||
9fe1f127 MG |
6118 | if (idle) |
6119 | return core; | |
10e2f1ac | 6120 | |
9fe1f127 | 6121 | cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); |
10e2f1ac PZ |
6122 | return -1; |
6123 | } | |
6124 | ||
c722f35b RR |
6125 | /* |
6126 | * Scan the local SMT mask for idle CPUs. | |
6127 | */ | |
6128 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6129 | { | |
6130 | int cpu; | |
6131 | ||
6132 | for_each_cpu(cpu, cpu_smt_mask(target)) { | |
6133 | if (!cpumask_test_cpu(cpu, p->cpus_ptr) || | |
6134 | !cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
6135 | continue; | |
6136 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) | |
6137 | return cpu; | |
6138 | } | |
6139 | ||
6140 | return -1; | |
6141 | } | |
6142 | ||
10e2f1ac PZ |
6143 | #else /* CONFIG_SCHED_SMT */ |
6144 | ||
9fe1f127 | 6145 | static inline void set_idle_cores(int cpu, int val) |
10e2f1ac | 6146 | { |
9fe1f127 MG |
6147 | } |
6148 | ||
6149 | static inline bool test_idle_cores(int cpu, bool def) | |
6150 | { | |
6151 | return def; | |
6152 | } | |
6153 | ||
6154 | static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) | |
6155 | { | |
6156 | return __select_idle_cpu(core); | |
10e2f1ac PZ |
6157 | } |
6158 | ||
c722f35b RR |
6159 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) |
6160 | { | |
6161 | return -1; | |
6162 | } | |
6163 | ||
10e2f1ac PZ |
6164 | #endif /* CONFIG_SCHED_SMT */ |
6165 | ||
6166 | /* | |
6167 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6168 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6169 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6170 | */ |
c722f35b | 6171 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target) |
10e2f1ac | 6172 | { |
60588bfa | 6173 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); |
9fe1f127 | 6174 | int i, cpu, idle_cpu = -1, nr = INT_MAX; |
9fe1f127 | 6175 | int this = smp_processor_id(); |
9cfb38a7 | 6176 | struct sched_domain *this_sd; |
d76343c6 | 6177 | u64 time; |
10e2f1ac | 6178 | |
9cfb38a7 WL |
6179 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6180 | if (!this_sd) | |
6181 | return -1; | |
6182 | ||
bae4ec13 MG |
6183 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6184 | ||
c722f35b | 6185 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
e6e0dc2d | 6186 | u64 avg_cost, avg_idle, span_avg; |
1ad3aaf3 | 6187 | |
e6e0dc2d MG |
6188 | /* |
6189 | * Due to large variance we need a large fuzz factor; | |
6190 | * hackbench in particularly is sensitive here. | |
6191 | */ | |
6192 | avg_idle = this_rq()->avg_idle / 512; | |
6193 | avg_cost = this_sd->avg_scan_cost + 1; | |
10e2f1ac | 6194 | |
e6e0dc2d | 6195 | span_avg = sd->span_weight * avg_idle; |
1ad3aaf3 PZ |
6196 | if (span_avg > 4*avg_cost) |
6197 | nr = div_u64(span_avg, avg_cost); | |
6198 | else | |
6199 | nr = 4; | |
10e2f1ac | 6200 | |
bae4ec13 MG |
6201 | time = cpu_clock(this); |
6202 | } | |
60588bfa CJ |
6203 | |
6204 | for_each_cpu_wrap(cpu, cpus, target) { | |
c722f35b | 6205 | if (has_idle_core) { |
9fe1f127 MG |
6206 | i = select_idle_core(p, cpu, cpus, &idle_cpu); |
6207 | if ((unsigned int)i < nr_cpumask_bits) | |
6208 | return i; | |
6209 | ||
6210 | } else { | |
6211 | if (!--nr) | |
6212 | return -1; | |
6213 | idle_cpu = __select_idle_cpu(cpu); | |
6214 | if ((unsigned int)idle_cpu < nr_cpumask_bits) | |
6215 | break; | |
6216 | } | |
10e2f1ac PZ |
6217 | } |
6218 | ||
c722f35b | 6219 | if (has_idle_core) |
02dbb724 | 6220 | set_idle_cores(target, false); |
9fe1f127 | 6221 | |
c722f35b | 6222 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
bae4ec13 MG |
6223 | time = cpu_clock(this) - time; |
6224 | update_avg(&this_sd->avg_scan_cost, time); | |
6225 | } | |
10e2f1ac | 6226 | |
9fe1f127 | 6227 | return idle_cpu; |
10e2f1ac PZ |
6228 | } |
6229 | ||
b7a33161 MR |
6230 | /* |
6231 | * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which | |
6232 | * the task fits. If no CPU is big enough, but there are idle ones, try to | |
6233 | * maximize capacity. | |
6234 | */ | |
6235 | static int | |
6236 | select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) | |
6237 | { | |
b4c9c9f1 | 6238 | unsigned long task_util, best_cap = 0; |
b7a33161 MR |
6239 | int cpu, best_cpu = -1; |
6240 | struct cpumask *cpus; | |
6241 | ||
b7a33161 MR |
6242 | cpus = this_cpu_cpumask_var_ptr(select_idle_mask); |
6243 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); | |
6244 | ||
b4c9c9f1 VG |
6245 | task_util = uclamp_task_util(p); |
6246 | ||
b7a33161 MR |
6247 | for_each_cpu_wrap(cpu, cpus, target) { |
6248 | unsigned long cpu_cap = capacity_of(cpu); | |
6249 | ||
6250 | if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) | |
6251 | continue; | |
b4c9c9f1 | 6252 | if (fits_capacity(task_util, cpu_cap)) |
b7a33161 MR |
6253 | return cpu; |
6254 | ||
6255 | if (cpu_cap > best_cap) { | |
6256 | best_cap = cpu_cap; | |
6257 | best_cpu = cpu; | |
6258 | } | |
6259 | } | |
6260 | ||
6261 | return best_cpu; | |
6262 | } | |
6263 | ||
b4c9c9f1 VG |
6264 | static inline bool asym_fits_capacity(int task_util, int cpu) |
6265 | { | |
6266 | if (static_branch_unlikely(&sched_asym_cpucapacity)) | |
6267 | return fits_capacity(task_util, capacity_of(cpu)); | |
6268 | ||
6269 | return true; | |
6270 | } | |
6271 | ||
10e2f1ac PZ |
6272 | /* |
6273 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6274 | */ |
772bd008 | 6275 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6276 | { |
c722f35b | 6277 | bool has_idle_core = false; |
99bd5e2f | 6278 | struct sched_domain *sd; |
b4c9c9f1 | 6279 | unsigned long task_util; |
32e839dd | 6280 | int i, recent_used_cpu; |
a50bde51 | 6281 | |
b7a33161 | 6282 | /* |
b4c9c9f1 VG |
6283 | * On asymmetric system, update task utilization because we will check |
6284 | * that the task fits with cpu's capacity. | |
b7a33161 MR |
6285 | */ |
6286 | if (static_branch_unlikely(&sched_asym_cpucapacity)) { | |
b4c9c9f1 VG |
6287 | sync_entity_load_avg(&p->se); |
6288 | task_util = uclamp_task_util(p); | |
b7a33161 MR |
6289 | } |
6290 | ||
b4c9c9f1 VG |
6291 | if ((available_idle_cpu(target) || sched_idle_cpu(target)) && |
6292 | asym_fits_capacity(task_util, target)) | |
e0a79f52 | 6293 | return target; |
99bd5e2f SS |
6294 | |
6295 | /* | |
97fb7a0a | 6296 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6297 | */ |
3c29e651 | 6298 | if (prev != target && cpus_share_cache(prev, target) && |
b4c9c9f1 VG |
6299 | (available_idle_cpu(prev) || sched_idle_cpu(prev)) && |
6300 | asym_fits_capacity(task_util, prev)) | |
772bd008 | 6301 | return prev; |
a50bde51 | 6302 | |
52262ee5 MG |
6303 | /* |
6304 | * Allow a per-cpu kthread to stack with the wakee if the | |
6305 | * kworker thread and the tasks previous CPUs are the same. | |
6306 | * The assumption is that the wakee queued work for the | |
6307 | * per-cpu kthread that is now complete and the wakeup is | |
6308 | * essentially a sync wakeup. An obvious example of this | |
6309 | * pattern is IO completions. | |
6310 | */ | |
6311 | if (is_per_cpu_kthread(current) && | |
6312 | prev == smp_processor_id() && | |
6313 | this_rq()->nr_running <= 1) { | |
6314 | return prev; | |
6315 | } | |
6316 | ||
97fb7a0a | 6317 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd MG |
6318 | recent_used_cpu = p->recent_used_cpu; |
6319 | if (recent_used_cpu != prev && | |
6320 | recent_used_cpu != target && | |
6321 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 6322 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
b4c9c9f1 VG |
6323 | cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) && |
6324 | asym_fits_capacity(task_util, recent_used_cpu)) { | |
32e839dd MG |
6325 | /* |
6326 | * Replace recent_used_cpu with prev as it is a potential | |
97fb7a0a | 6327 | * candidate for the next wake: |
32e839dd MG |
6328 | */ |
6329 | p->recent_used_cpu = prev; | |
6330 | return recent_used_cpu; | |
6331 | } | |
6332 | ||
b4c9c9f1 VG |
6333 | /* |
6334 | * For asymmetric CPU capacity systems, our domain of interest is | |
6335 | * sd_asym_cpucapacity rather than sd_llc. | |
6336 | */ | |
6337 | if (static_branch_unlikely(&sched_asym_cpucapacity)) { | |
6338 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); | |
6339 | /* | |
6340 | * On an asymmetric CPU capacity system where an exclusive | |
6341 | * cpuset defines a symmetric island (i.e. one unique | |
6342 | * capacity_orig value through the cpuset), the key will be set | |
6343 | * but the CPUs within that cpuset will not have a domain with | |
6344 | * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric | |
6345 | * capacity path. | |
6346 | */ | |
6347 | if (sd) { | |
6348 | i = select_idle_capacity(p, sd, target); | |
6349 | return ((unsigned)i < nr_cpumask_bits) ? i : target; | |
6350 | } | |
6351 | } | |
6352 | ||
518cd623 | 6353 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6354 | if (!sd) |
6355 | return target; | |
772bd008 | 6356 | |
c722f35b RR |
6357 | if (sched_smt_active()) { |
6358 | has_idle_core = test_idle_cores(target, false); | |
6359 | ||
6360 | if (!has_idle_core && cpus_share_cache(prev, target)) { | |
6361 | i = select_idle_smt(p, sd, prev); | |
6362 | if ((unsigned int)i < nr_cpumask_bits) | |
6363 | return i; | |
6364 | } | |
6365 | } | |
6366 | ||
6367 | i = select_idle_cpu(p, sd, has_idle_core, target); | |
10e2f1ac PZ |
6368 | if ((unsigned)i < nr_cpumask_bits) |
6369 | return i; | |
6370 | ||
a50bde51 PZ |
6371 | return target; |
6372 | } | |
231678b7 | 6373 | |
f9be3e59 | 6374 | /** |
59a74b15 | 6375 | * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks. |
f9be3e59 PB |
6376 | * @cpu: the CPU to get the utilization of |
6377 | * | |
6378 | * The unit of the return value must be the one of capacity so we can compare | |
6379 | * the utilization with the capacity of the CPU that is available for CFS task | |
6380 | * (ie cpu_capacity). | |
231678b7 DE |
6381 | * |
6382 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
6383 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
6384 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
6385 | * capacity_orig is the cpu_capacity available at the highest frequency | |
6386 | * (arch_scale_freq_capacity()). | |
6387 | * The utilization of a CPU converges towards a sum equal to or less than the | |
6388 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
6389 | * the running time on this CPU scaled by capacity_curr. | |
6390 | * | |
f9be3e59 PB |
6391 | * The estimated utilization of a CPU is defined to be the maximum between its |
6392 | * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks | |
6393 | * currently RUNNABLE on that CPU. | |
6394 | * This allows to properly represent the expected utilization of a CPU which | |
6395 | * has just got a big task running since a long sleep period. At the same time | |
6396 | * however it preserves the benefits of the "blocked utilization" in | |
6397 | * describing the potential for other tasks waking up on the same CPU. | |
6398 | * | |
231678b7 DE |
6399 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even |
6400 | * higher than capacity_orig because of unfortunate rounding in | |
6401 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
6402 | * the average stabilizes with the new running time. We need to check that the | |
6403 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
6404 | * necessary. Without utilization capping, a group could be seen as overloaded | |
6405 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
6406 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
6407 | * capacity_orig) as it useful for predicting the capacity required after task | |
6408 | * migrations (scheduler-driven DVFS). | |
f9be3e59 PB |
6409 | * |
6410 | * Return: the (estimated) utilization for the specified CPU | |
8bb5b00c | 6411 | */ |
f9be3e59 | 6412 | static inline unsigned long cpu_util(int cpu) |
8bb5b00c | 6413 | { |
f9be3e59 PB |
6414 | struct cfs_rq *cfs_rq; |
6415 | unsigned int util; | |
6416 | ||
6417 | cfs_rq = &cpu_rq(cpu)->cfs; | |
6418 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6419 | ||
6420 | if (sched_feat(UTIL_EST)) | |
6421 | util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); | |
8bb5b00c | 6422 | |
f9be3e59 | 6423 | return min_t(unsigned long, util, capacity_orig_of(cpu)); |
8bb5b00c | 6424 | } |
a50bde51 | 6425 | |
104cb16d | 6426 | /* |
c469933e PB |
6427 | * cpu_util_without: compute cpu utilization without any contributions from *p |
6428 | * @cpu: the CPU which utilization is requested | |
6429 | * @p: the task which utilization should be discounted | |
6430 | * | |
6431 | * The utilization of a CPU is defined by the utilization of tasks currently | |
6432 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
6433 | * execution on that CPU. | |
6434 | * | |
6435 | * This method returns the utilization of the specified CPU by discounting the | |
6436 | * utilization of the specified task, whenever the task is currently | |
6437 | * contributing to the CPU utilization. | |
104cb16d | 6438 | */ |
c469933e | 6439 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) |
104cb16d | 6440 | { |
f9be3e59 PB |
6441 | struct cfs_rq *cfs_rq; |
6442 | unsigned int util; | |
104cb16d MR |
6443 | |
6444 | /* Task has no contribution or is new */ | |
f9be3e59 | 6445 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) |
104cb16d MR |
6446 | return cpu_util(cpu); |
6447 | ||
f9be3e59 PB |
6448 | cfs_rq = &cpu_rq(cpu)->cfs; |
6449 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6450 | ||
c469933e | 6451 | /* Discount task's util from CPU's util */ |
b5c0ce7b | 6452 | lsub_positive(&util, task_util(p)); |
104cb16d | 6453 | |
f9be3e59 PB |
6454 | /* |
6455 | * Covered cases: | |
6456 | * | |
6457 | * a) if *p is the only task sleeping on this CPU, then: | |
6458 | * cpu_util (== task_util) > util_est (== 0) | |
6459 | * and thus we return: | |
c469933e | 6460 | * cpu_util_without = (cpu_util - task_util) = 0 |
f9be3e59 PB |
6461 | * |
6462 | * b) if other tasks are SLEEPING on this CPU, which is now exiting | |
6463 | * IDLE, then: | |
6464 | * cpu_util >= task_util | |
6465 | * cpu_util > util_est (== 0) | |
6466 | * and thus we discount *p's blocked utilization to return: | |
c469933e | 6467 | * cpu_util_without = (cpu_util - task_util) >= 0 |
f9be3e59 PB |
6468 | * |
6469 | * c) if other tasks are RUNNABLE on that CPU and | |
6470 | * util_est > cpu_util | |
6471 | * then we use util_est since it returns a more restrictive | |
6472 | * estimation of the spare capacity on that CPU, by just | |
6473 | * considering the expected utilization of tasks already | |
6474 | * runnable on that CPU. | |
6475 | * | |
6476 | * Cases a) and b) are covered by the above code, while case c) is | |
6477 | * covered by the following code when estimated utilization is | |
6478 | * enabled. | |
6479 | */ | |
c469933e PB |
6480 | if (sched_feat(UTIL_EST)) { |
6481 | unsigned int estimated = | |
6482 | READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6483 | ||
6484 | /* | |
6485 | * Despite the following checks we still have a small window | |
6486 | * for a possible race, when an execl's select_task_rq_fair() | |
6487 | * races with LB's detach_task(): | |
6488 | * | |
6489 | * detach_task() | |
6490 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
6491 | * ---------------------------------- A | |
6492 | * deactivate_task() \ | |
6493 | * dequeue_task() + RaceTime | |
6494 | * util_est_dequeue() / | |
6495 | * ---------------------------------- B | |
6496 | * | |
6497 | * The additional check on "current == p" it's required to | |
6498 | * properly fix the execl regression and it helps in further | |
6499 | * reducing the chances for the above race. | |
6500 | */ | |
b5c0ce7b PB |
6501 | if (unlikely(task_on_rq_queued(p) || current == p)) |
6502 | lsub_positive(&estimated, _task_util_est(p)); | |
6503 | ||
c469933e PB |
6504 | util = max(util, estimated); |
6505 | } | |
f9be3e59 PB |
6506 | |
6507 | /* | |
6508 | * Utilization (estimated) can exceed the CPU capacity, thus let's | |
6509 | * clamp to the maximum CPU capacity to ensure consistency with | |
6510 | * the cpu_util call. | |
6511 | */ | |
6512 | return min_t(unsigned long, util, capacity_orig_of(cpu)); | |
104cb16d MR |
6513 | } |
6514 | ||
390031e4 QP |
6515 | /* |
6516 | * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) | |
6517 | * to @dst_cpu. | |
6518 | */ | |
6519 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
6520 | { | |
6521 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
6522 | unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); | |
6523 | ||
6524 | /* | |
6525 | * If @p migrates from @cpu to another, remove its contribution. Or, | |
6526 | * if @p migrates from another CPU to @cpu, add its contribution. In | |
6527 | * the other cases, @cpu is not impacted by the migration, so the | |
6528 | * util_avg should already be correct. | |
6529 | */ | |
6530 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
736cc6b3 | 6531 | lsub_positive(&util, task_util(p)); |
390031e4 QP |
6532 | else if (task_cpu(p) != cpu && dst_cpu == cpu) |
6533 | util += task_util(p); | |
6534 | ||
6535 | if (sched_feat(UTIL_EST)) { | |
6536 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6537 | ||
6538 | /* | |
6539 | * During wake-up, the task isn't enqueued yet and doesn't | |
6540 | * appear in the cfs_rq->avg.util_est.enqueued of any rq, | |
6541 | * so just add it (if needed) to "simulate" what will be | |
6542 | * cpu_util() after the task has been enqueued. | |
6543 | */ | |
6544 | if (dst_cpu == cpu) | |
6545 | util_est += _task_util_est(p); | |
6546 | ||
6547 | util = max(util, util_est); | |
6548 | } | |
6549 | ||
6550 | return min(util, capacity_orig_of(cpu)); | |
6551 | } | |
6552 | ||
6553 | /* | |
eb92692b | 6554 | * compute_energy(): Estimates the energy that @pd would consume if @p was |
390031e4 | 6555 | * migrated to @dst_cpu. compute_energy() predicts what will be the utilization |
eb92692b | 6556 | * landscape of @pd's CPUs after the task migration, and uses the Energy Model |
390031e4 QP |
6557 | * to compute what would be the energy if we decided to actually migrate that |
6558 | * task. | |
6559 | */ | |
6560 | static long | |
6561 | compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) | |
6562 | { | |
eb92692b QP |
6563 | struct cpumask *pd_mask = perf_domain_span(pd); |
6564 | unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); | |
6565 | unsigned long max_util = 0, sum_util = 0; | |
390031e4 QP |
6566 | int cpu; |
6567 | ||
eb92692b QP |
6568 | /* |
6569 | * The capacity state of CPUs of the current rd can be driven by CPUs | |
6570 | * of another rd if they belong to the same pd. So, account for the | |
6571 | * utilization of these CPUs too by masking pd with cpu_online_mask | |
6572 | * instead of the rd span. | |
6573 | * | |
6574 | * If an entire pd is outside of the current rd, it will not appear in | |
6575 | * its pd list and will not be accounted by compute_energy(). | |
6576 | */ | |
6577 | for_each_cpu_and(cpu, pd_mask, cpu_online_mask) { | |
0372e1cf VD |
6578 | unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu); |
6579 | unsigned long cpu_util, util_running = util_freq; | |
6580 | struct task_struct *tsk = NULL; | |
6581 | ||
6582 | /* | |
6583 | * When @p is placed on @cpu: | |
6584 | * | |
6585 | * util_running = max(cpu_util, cpu_util_est) + | |
6586 | * max(task_util, _task_util_est) | |
6587 | * | |
6588 | * while cpu_util_next is: max(cpu_util + task_util, | |
6589 | * cpu_util_est + _task_util_est) | |
6590 | */ | |
6591 | if (cpu == dst_cpu) { | |
6592 | tsk = p; | |
6593 | util_running = | |
6594 | cpu_util_next(cpu, p, -1) + task_util_est(p); | |
6595 | } | |
af24bde8 PB |
6596 | |
6597 | /* | |
eb92692b QP |
6598 | * Busy time computation: utilization clamping is not |
6599 | * required since the ratio (sum_util / cpu_capacity) | |
6600 | * is already enough to scale the EM reported power | |
6601 | * consumption at the (eventually clamped) cpu_capacity. | |
af24bde8 | 6602 | */ |
0372e1cf | 6603 | sum_util += effective_cpu_util(cpu, util_running, cpu_cap, |
eb92692b | 6604 | ENERGY_UTIL, NULL); |
af24bde8 | 6605 | |
390031e4 | 6606 | /* |
eb92692b QP |
6607 | * Performance domain frequency: utilization clamping |
6608 | * must be considered since it affects the selection | |
6609 | * of the performance domain frequency. | |
6610 | * NOTE: in case RT tasks are running, by default the | |
6611 | * FREQUENCY_UTIL's utilization can be max OPP. | |
390031e4 | 6612 | */ |
0372e1cf | 6613 | cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap, |
eb92692b QP |
6614 | FREQUENCY_UTIL, tsk); |
6615 | max_util = max(max_util, cpu_util); | |
390031e4 QP |
6616 | } |
6617 | ||
f0b56947 | 6618 | return em_cpu_energy(pd->em_pd, max_util, sum_util); |
390031e4 QP |
6619 | } |
6620 | ||
732cd75b QP |
6621 | /* |
6622 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
6623 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
6624 | * spare capacity in each performance domain and uses it as a potential | |
6625 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
6626 | * out which of the CPU candidates is the most energy-efficient. | |
6627 | * | |
6628 | * The rationale for this heuristic is as follows. In a performance domain, | |
6629 | * all the most energy efficient CPU candidates (according to the Energy | |
6630 | * Model) are those for which we'll request a low frequency. When there are | |
6631 | * several CPUs for which the frequency request will be the same, we don't | |
6632 | * have enough data to break the tie between them, because the Energy Model | |
6633 | * only includes active power costs. With this model, if we assume that | |
6634 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
6635 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
6636 | * the best candidates of the performance domain. | |
6637 | * | |
6638 | * In practice, it could be preferable from an energy standpoint to pack | |
6639 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
6640 | * but that could also hurt our chances to go cluster idle, and we have no | |
6641 | * ways to tell with the current Energy Model if this is actually a good | |
6642 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
6643 | * cluster-packing, and spreading inside a cluster. That should at least be | |
6644 | * a good thing for latency, and this is consistent with the idea that most | |
6645 | * of the energy savings of EAS come from the asymmetry of the system, and | |
6646 | * not so much from breaking the tie between identical CPUs. That's also the | |
6647 | * reason why EAS is enabled in the topology code only for systems where | |
6648 | * SD_ASYM_CPUCAPACITY is set. | |
6649 | * | |
6650 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
6651 | * they don't have any useful utilization data yet and it's not possible to | |
6652 | * forecast their impact on energy consumption. Consequently, they will be | |
6653 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
6654 | * to be energy-inefficient in some use-cases. The alternative would be to | |
6655 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
6656 | * their util_avg from the parent task, but those heuristics could hurt | |
6657 | * other use-cases too. So, until someone finds a better way to solve this, | |
6658 | * let's keep things simple by re-using the existing slow path. | |
6659 | */ | |
732cd75b QP |
6660 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) |
6661 | { | |
eb92692b | 6662 | unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; |
732cd75b | 6663 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; |
eb92692b | 6664 | unsigned long cpu_cap, util, base_energy = 0; |
732cd75b | 6665 | int cpu, best_energy_cpu = prev_cpu; |
732cd75b | 6666 | struct sched_domain *sd; |
eb92692b | 6667 | struct perf_domain *pd; |
732cd75b QP |
6668 | |
6669 | rcu_read_lock(); | |
6670 | pd = rcu_dereference(rd->pd); | |
6671 | if (!pd || READ_ONCE(rd->overutilized)) | |
6672 | goto fail; | |
732cd75b QP |
6673 | |
6674 | /* | |
6675 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
6676 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
6677 | */ | |
6678 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
6679 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
6680 | sd = sd->parent; | |
6681 | if (!sd) | |
6682 | goto fail; | |
6683 | ||
6684 | sync_entity_load_avg(&p->se); | |
6685 | if (!task_util_est(p)) | |
6686 | goto unlock; | |
6687 | ||
6688 | for (; pd; pd = pd->next) { | |
eb92692b QP |
6689 | unsigned long cur_delta, spare_cap, max_spare_cap = 0; |
6690 | unsigned long base_energy_pd; | |
732cd75b QP |
6691 | int max_spare_cap_cpu = -1; |
6692 | ||
eb92692b QP |
6693 | /* Compute the 'base' energy of the pd, without @p */ |
6694 | base_energy_pd = compute_energy(p, -1, pd); | |
6695 | base_energy += base_energy_pd; | |
6696 | ||
732cd75b | 6697 | for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { |
3bd37062 | 6698 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
6699 | continue; |
6700 | ||
732cd75b QP |
6701 | util = cpu_util_next(cpu, p, cpu); |
6702 | cpu_cap = capacity_of(cpu); | |
da0777d3 LL |
6703 | spare_cap = cpu_cap; |
6704 | lsub_positive(&spare_cap, util); | |
1d42509e VS |
6705 | |
6706 | /* | |
6707 | * Skip CPUs that cannot satisfy the capacity request. | |
6708 | * IOW, placing the task there would make the CPU | |
6709 | * overutilized. Take uclamp into account to see how | |
6710 | * much capacity we can get out of the CPU; this is | |
a5418be9 | 6711 | * aligned with sched_cpu_util(). |
1d42509e VS |
6712 | */ |
6713 | util = uclamp_rq_util_with(cpu_rq(cpu), util, p); | |
60e17f5c | 6714 | if (!fits_capacity(util, cpu_cap)) |
732cd75b QP |
6715 | continue; |
6716 | ||
6717 | /* Always use prev_cpu as a candidate. */ | |
6718 | if (cpu == prev_cpu) { | |
eb92692b QP |
6719 | prev_delta = compute_energy(p, prev_cpu, pd); |
6720 | prev_delta -= base_energy_pd; | |
6721 | best_delta = min(best_delta, prev_delta); | |
732cd75b QP |
6722 | } |
6723 | ||
6724 | /* | |
6725 | * Find the CPU with the maximum spare capacity in | |
6726 | * the performance domain | |
6727 | */ | |
732cd75b QP |
6728 | if (spare_cap > max_spare_cap) { |
6729 | max_spare_cap = spare_cap; | |
6730 | max_spare_cap_cpu = cpu; | |
6731 | } | |
6732 | } | |
6733 | ||
6734 | /* Evaluate the energy impact of using this CPU. */ | |
4892f51a | 6735 | if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) { |
eb92692b QP |
6736 | cur_delta = compute_energy(p, max_spare_cap_cpu, pd); |
6737 | cur_delta -= base_energy_pd; | |
6738 | if (cur_delta < best_delta) { | |
6739 | best_delta = cur_delta; | |
732cd75b QP |
6740 | best_energy_cpu = max_spare_cap_cpu; |
6741 | } | |
6742 | } | |
6743 | } | |
6744 | unlock: | |
6745 | rcu_read_unlock(); | |
6746 | ||
6747 | /* | |
6748 | * Pick the best CPU if prev_cpu cannot be used, or if it saves at | |
6749 | * least 6% of the energy used by prev_cpu. | |
6750 | */ | |
eb92692b | 6751 | if (prev_delta == ULONG_MAX) |
732cd75b QP |
6752 | return best_energy_cpu; |
6753 | ||
eb92692b | 6754 | if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) |
732cd75b QP |
6755 | return best_energy_cpu; |
6756 | ||
6757 | return prev_cpu; | |
6758 | ||
6759 | fail: | |
6760 | rcu_read_unlock(); | |
6761 | ||
6762 | return -1; | |
6763 | } | |
6764 | ||
aaee1203 | 6765 | /* |
de91b9cb | 6766 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
3aef1551 | 6767 | * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE, |
de91b9cb | 6768 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. |
aaee1203 | 6769 | * |
97fb7a0a IM |
6770 | * Balances load by selecting the idlest CPU in the idlest group, or under |
6771 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 6772 | * |
97fb7a0a | 6773 | * Returns the target CPU number. |
aaee1203 PZ |
6774 | * |
6775 | * preempt must be disabled. | |
6776 | */ | |
0017d735 | 6777 | static int |
3aef1551 | 6778 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags) |
aaee1203 | 6779 | { |
3aef1551 | 6780 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
f1d88b44 | 6781 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 6782 | int cpu = smp_processor_id(); |
63b0e9ed | 6783 | int new_cpu = prev_cpu; |
99bd5e2f | 6784 | int want_affine = 0; |
3aef1551 VS |
6785 | /* SD_flags and WF_flags share the first nibble */ |
6786 | int sd_flag = wake_flags & 0xF; | |
c88d5910 | 6787 | |
dc824eb8 | 6788 | if (wake_flags & WF_TTWU) { |
c58d25f3 | 6789 | record_wakee(p); |
732cd75b | 6790 | |
f8a696f2 | 6791 | if (sched_energy_enabled()) { |
732cd75b QP |
6792 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
6793 | if (new_cpu >= 0) | |
6794 | return new_cpu; | |
6795 | new_cpu = prev_cpu; | |
6796 | } | |
6797 | ||
00061968 | 6798 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 6799 | } |
aaee1203 | 6800 | |
dce840a0 | 6801 | rcu_read_lock(); |
aaee1203 | 6802 | for_each_domain(cpu, tmp) { |
fe3bcfe1 | 6803 | /* |
97fb7a0a | 6804 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 6805 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 6806 | */ |
99bd5e2f SS |
6807 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6808 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
6809 | if (cpu != prev_cpu) |
6810 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
6811 | ||
6812 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 6813 | break; |
f03542a7 | 6814 | } |
29cd8bae | 6815 | |
f03542a7 | 6816 | if (tmp->flags & sd_flag) |
29cd8bae | 6817 | sd = tmp; |
63b0e9ed MG |
6818 | else if (!want_affine) |
6819 | break; | |
29cd8bae PZ |
6820 | } |
6821 | ||
f1d88b44 VK |
6822 | if (unlikely(sd)) { |
6823 | /* Slow path */ | |
18bd1b4b | 6824 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
dc824eb8 | 6825 | } else if (wake_flags & WF_TTWU) { /* XXX always ? */ |
f1d88b44 | 6826 | /* Fast path */ |
f1d88b44 VK |
6827 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
6828 | ||
6829 | if (want_affine) | |
6830 | current->recent_used_cpu = cpu; | |
e7693a36 | 6831 | } |
dce840a0 | 6832 | rcu_read_unlock(); |
e7693a36 | 6833 | |
c88d5910 | 6834 | return new_cpu; |
e7693a36 | 6835 | } |
0a74bef8 | 6836 | |
144d8487 PZ |
6837 | static void detach_entity_cfs_rq(struct sched_entity *se); |
6838 | ||
0a74bef8 | 6839 | /* |
97fb7a0a | 6840 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 6841 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 6842 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6843 | */ |
3f9672ba | 6844 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 6845 | { |
59efa0ba PZ |
6846 | /* |
6847 | * As blocked tasks retain absolute vruntime the migration needs to | |
6848 | * deal with this by subtracting the old and adding the new | |
6849 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6850 | * the task on the new runqueue. | |
6851 | */ | |
6852 | if (p->state == TASK_WAKING) { | |
6853 | struct sched_entity *se = &p->se; | |
6854 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6855 | u64 min_vruntime; | |
6856 | ||
6857 | #ifndef CONFIG_64BIT | |
6858 | u64 min_vruntime_copy; | |
6859 | ||
6860 | do { | |
6861 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
6862 | smp_rmb(); | |
6863 | min_vruntime = cfs_rq->min_vruntime; | |
6864 | } while (min_vruntime != min_vruntime_copy); | |
6865 | #else | |
6866 | min_vruntime = cfs_rq->min_vruntime; | |
6867 | #endif | |
6868 | ||
6869 | se->vruntime -= min_vruntime; | |
6870 | } | |
6871 | ||
144d8487 PZ |
6872 | if (p->on_rq == TASK_ON_RQ_MIGRATING) { |
6873 | /* | |
6874 | * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' | |
6875 | * rq->lock and can modify state directly. | |
6876 | */ | |
6877 | lockdep_assert_held(&task_rq(p)->lock); | |
6878 | detach_entity_cfs_rq(&p->se); | |
6879 | ||
6880 | } else { | |
6881 | /* | |
6882 | * We are supposed to update the task to "current" time, then | |
6883 | * its up to date and ready to go to new CPU/cfs_rq. But we | |
6884 | * have difficulty in getting what current time is, so simply | |
6885 | * throw away the out-of-date time. This will result in the | |
6886 | * wakee task is less decayed, but giving the wakee more load | |
6887 | * sounds not bad. | |
6888 | */ | |
6889 | remove_entity_load_avg(&p->se); | |
6890 | } | |
9d89c257 YD |
6891 | |
6892 | /* Tell new CPU we are migrated */ | |
6893 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
6894 | |
6895 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 6896 | p->se.exec_start = 0; |
3f9672ba SD |
6897 | |
6898 | update_scan_period(p, new_cpu); | |
0a74bef8 | 6899 | } |
12695578 YD |
6900 | |
6901 | static void task_dead_fair(struct task_struct *p) | |
6902 | { | |
6903 | remove_entity_load_avg(&p->se); | |
6904 | } | |
6e2df058 PZ |
6905 | |
6906 | static int | |
6907 | balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
6908 | { | |
6909 | if (rq->nr_running) | |
6910 | return 1; | |
6911 | ||
6912 | return newidle_balance(rq, rf) != 0; | |
6913 | } | |
e7693a36 GH |
6914 | #endif /* CONFIG_SMP */ |
6915 | ||
a555e9d8 | 6916 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
6917 | { |
6918 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
6919 | ||
6920 | /* | |
e52fb7c0 PZ |
6921 | * Since its curr running now, convert the gran from real-time |
6922 | * to virtual-time in his units. | |
13814d42 MG |
6923 | * |
6924 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
6925 | * they get preempted easier. That is, if 'se' < 'curr' then | |
6926 | * the resulting gran will be larger, therefore penalizing the | |
6927 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
6928 | * be smaller, again penalizing the lighter task. | |
6929 | * | |
6930 | * This is especially important for buddies when the leftmost | |
6931 | * task is higher priority than the buddy. | |
0bbd3336 | 6932 | */ |
f4ad9bd2 | 6933 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
6934 | } |
6935 | ||
464b7527 PZ |
6936 | /* |
6937 | * Should 'se' preempt 'curr'. | |
6938 | * | |
6939 | * |s1 | |
6940 | * |s2 | |
6941 | * |s3 | |
6942 | * g | |
6943 | * |<--->|c | |
6944 | * | |
6945 | * w(c, s1) = -1 | |
6946 | * w(c, s2) = 0 | |
6947 | * w(c, s3) = 1 | |
6948 | * | |
6949 | */ | |
6950 | static int | |
6951 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
6952 | { | |
6953 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
6954 | ||
6955 | if (vdiff <= 0) | |
6956 | return -1; | |
6957 | ||
a555e9d8 | 6958 | gran = wakeup_gran(se); |
464b7527 PZ |
6959 | if (vdiff > gran) |
6960 | return 1; | |
6961 | ||
6962 | return 0; | |
6963 | } | |
6964 | ||
02479099 PZ |
6965 | static void set_last_buddy(struct sched_entity *se) |
6966 | { | |
1da1843f | 6967 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6968 | return; |
6969 | ||
c5ae366e DA |
6970 | for_each_sched_entity(se) { |
6971 | if (SCHED_WARN_ON(!se->on_rq)) | |
6972 | return; | |
69c80f3e | 6973 | cfs_rq_of(se)->last = se; |
c5ae366e | 6974 | } |
02479099 PZ |
6975 | } |
6976 | ||
6977 | static void set_next_buddy(struct sched_entity *se) | |
6978 | { | |
1da1843f | 6979 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6980 | return; |
6981 | ||
c5ae366e DA |
6982 | for_each_sched_entity(se) { |
6983 | if (SCHED_WARN_ON(!se->on_rq)) | |
6984 | return; | |
69c80f3e | 6985 | cfs_rq_of(se)->next = se; |
c5ae366e | 6986 | } |
02479099 PZ |
6987 | } |
6988 | ||
ac53db59 RR |
6989 | static void set_skip_buddy(struct sched_entity *se) |
6990 | { | |
69c80f3e VP |
6991 | for_each_sched_entity(se) |
6992 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
6993 | } |
6994 | ||
bf0f6f24 IM |
6995 | /* |
6996 | * Preempt the current task with a newly woken task if needed: | |
6997 | */ | |
5a9b86f6 | 6998 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
6999 | { |
7000 | struct task_struct *curr = rq->curr; | |
8651a86c | 7001 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 7002 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 7003 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 7004 | int next_buddy_marked = 0; |
bf0f6f24 | 7005 | |
4ae7d5ce IM |
7006 | if (unlikely(se == pse)) |
7007 | return; | |
7008 | ||
5238cdd3 | 7009 | /* |
163122b7 | 7010 | * This is possible from callers such as attach_tasks(), in which we |
3b03706f | 7011 | * unconditionally check_preempt_curr() after an enqueue (which may have |
5238cdd3 PT |
7012 | * lead to a throttle). This both saves work and prevents false |
7013 | * next-buddy nomination below. | |
7014 | */ | |
7015 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
7016 | return; | |
7017 | ||
2f36825b | 7018 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 7019 | set_next_buddy(pse); |
2f36825b VP |
7020 | next_buddy_marked = 1; |
7021 | } | |
57fdc26d | 7022 | |
aec0a514 BR |
7023 | /* |
7024 | * We can come here with TIF_NEED_RESCHED already set from new task | |
7025 | * wake up path. | |
5238cdd3 PT |
7026 | * |
7027 | * Note: this also catches the edge-case of curr being in a throttled | |
7028 | * group (e.g. via set_curr_task), since update_curr() (in the | |
7029 | * enqueue of curr) will have resulted in resched being set. This | |
7030 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
7031 | * below. | |
aec0a514 BR |
7032 | */ |
7033 | if (test_tsk_need_resched(curr)) | |
7034 | return; | |
7035 | ||
a2f5c9ab | 7036 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
7037 | if (unlikely(task_has_idle_policy(curr)) && |
7038 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
7039 | goto preempt; |
7040 | ||
91c234b4 | 7041 | /* |
a2f5c9ab DH |
7042 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
7043 | * is driven by the tick): | |
91c234b4 | 7044 | */ |
8ed92e51 | 7045 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 7046 | return; |
bf0f6f24 | 7047 | |
464b7527 | 7048 | find_matching_se(&se, &pse); |
9bbd7374 | 7049 | update_curr(cfs_rq_of(se)); |
002f128b | 7050 | BUG_ON(!pse); |
2f36825b VP |
7051 | if (wakeup_preempt_entity(se, pse) == 1) { |
7052 | /* | |
7053 | * Bias pick_next to pick the sched entity that is | |
7054 | * triggering this preemption. | |
7055 | */ | |
7056 | if (!next_buddy_marked) | |
7057 | set_next_buddy(pse); | |
3a7e73a2 | 7058 | goto preempt; |
2f36825b | 7059 | } |
464b7527 | 7060 | |
3a7e73a2 | 7061 | return; |
a65ac745 | 7062 | |
3a7e73a2 | 7063 | preempt: |
8875125e | 7064 | resched_curr(rq); |
3a7e73a2 PZ |
7065 | /* |
7066 | * Only set the backward buddy when the current task is still | |
7067 | * on the rq. This can happen when a wakeup gets interleaved | |
7068 | * with schedule on the ->pre_schedule() or idle_balance() | |
7069 | * point, either of which can * drop the rq lock. | |
7070 | * | |
7071 | * Also, during early boot the idle thread is in the fair class, | |
7072 | * for obvious reasons its a bad idea to schedule back to it. | |
7073 | */ | |
7074 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
7075 | return; | |
7076 | ||
7077 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
7078 | set_last_buddy(se); | |
bf0f6f24 IM |
7079 | } |
7080 | ||
5d7d6056 | 7081 | struct task_struct * |
d8ac8971 | 7082 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
7083 | { |
7084 | struct cfs_rq *cfs_rq = &rq->cfs; | |
7085 | struct sched_entity *se; | |
678d5718 | 7086 | struct task_struct *p; |
37e117c0 | 7087 | int new_tasks; |
678d5718 | 7088 | |
6e83125c | 7089 | again: |
6e2df058 | 7090 | if (!sched_fair_runnable(rq)) |
38033c37 | 7091 | goto idle; |
678d5718 | 7092 | |
9674f5ca | 7093 | #ifdef CONFIG_FAIR_GROUP_SCHED |
67692435 | 7094 | if (!prev || prev->sched_class != &fair_sched_class) |
678d5718 PZ |
7095 | goto simple; |
7096 | ||
7097 | /* | |
7098 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
7099 | * likely that a next task is from the same cgroup as the current. | |
7100 | * | |
7101 | * Therefore attempt to avoid putting and setting the entire cgroup | |
7102 | * hierarchy, only change the part that actually changes. | |
7103 | */ | |
7104 | ||
7105 | do { | |
7106 | struct sched_entity *curr = cfs_rq->curr; | |
7107 | ||
7108 | /* | |
7109 | * Since we got here without doing put_prev_entity() we also | |
7110 | * have to consider cfs_rq->curr. If it is still a runnable | |
7111 | * entity, update_curr() will update its vruntime, otherwise | |
7112 | * forget we've ever seen it. | |
7113 | */ | |
54d27365 BS |
7114 | if (curr) { |
7115 | if (curr->on_rq) | |
7116 | update_curr(cfs_rq); | |
7117 | else | |
7118 | curr = NULL; | |
678d5718 | 7119 | |
54d27365 BS |
7120 | /* |
7121 | * This call to check_cfs_rq_runtime() will do the | |
7122 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 7123 | * Therefore the nr_running test will indeed |
54d27365 BS |
7124 | * be correct. |
7125 | */ | |
9674f5ca VK |
7126 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
7127 | cfs_rq = &rq->cfs; | |
7128 | ||
7129 | if (!cfs_rq->nr_running) | |
7130 | goto idle; | |
7131 | ||
54d27365 | 7132 | goto simple; |
9674f5ca | 7133 | } |
54d27365 | 7134 | } |
678d5718 PZ |
7135 | |
7136 | se = pick_next_entity(cfs_rq, curr); | |
7137 | cfs_rq = group_cfs_rq(se); | |
7138 | } while (cfs_rq); | |
7139 | ||
7140 | p = task_of(se); | |
7141 | ||
7142 | /* | |
7143 | * Since we haven't yet done put_prev_entity and if the selected task | |
7144 | * is a different task than we started out with, try and touch the | |
7145 | * least amount of cfs_rqs. | |
7146 | */ | |
7147 | if (prev != p) { | |
7148 | struct sched_entity *pse = &prev->se; | |
7149 | ||
7150 | while (!(cfs_rq = is_same_group(se, pse))) { | |
7151 | int se_depth = se->depth; | |
7152 | int pse_depth = pse->depth; | |
7153 | ||
7154 | if (se_depth <= pse_depth) { | |
7155 | put_prev_entity(cfs_rq_of(pse), pse); | |
7156 | pse = parent_entity(pse); | |
7157 | } | |
7158 | if (se_depth >= pse_depth) { | |
7159 | set_next_entity(cfs_rq_of(se), se); | |
7160 | se = parent_entity(se); | |
7161 | } | |
7162 | } | |
7163 | ||
7164 | put_prev_entity(cfs_rq, pse); | |
7165 | set_next_entity(cfs_rq, se); | |
7166 | } | |
7167 | ||
93824900 | 7168 | goto done; |
678d5718 | 7169 | simple: |
678d5718 | 7170 | #endif |
67692435 PZ |
7171 | if (prev) |
7172 | put_prev_task(rq, prev); | |
606dba2e | 7173 | |
bf0f6f24 | 7174 | do { |
678d5718 | 7175 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 7176 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
7177 | cfs_rq = group_cfs_rq(se); |
7178 | } while (cfs_rq); | |
7179 | ||
8f4d37ec | 7180 | p = task_of(se); |
678d5718 | 7181 | |
13a453c2 | 7182 | done: __maybe_unused; |
93824900 UR |
7183 | #ifdef CONFIG_SMP |
7184 | /* | |
7185 | * Move the next running task to the front of | |
7186 | * the list, so our cfs_tasks list becomes MRU | |
7187 | * one. | |
7188 | */ | |
7189 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
7190 | #endif | |
7191 | ||
e0ee463c | 7192 | if (hrtick_enabled_fair(rq)) |
b39e66ea | 7193 | hrtick_start_fair(rq, p); |
8f4d37ec | 7194 | |
3b1baa64 MR |
7195 | update_misfit_status(p, rq); |
7196 | ||
8f4d37ec | 7197 | return p; |
38033c37 PZ |
7198 | |
7199 | idle: | |
67692435 PZ |
7200 | if (!rf) |
7201 | return NULL; | |
7202 | ||
5ba553ef | 7203 | new_tasks = newidle_balance(rq, rf); |
46f69fa3 | 7204 | |
37e117c0 | 7205 | /* |
5ba553ef | 7206 | * Because newidle_balance() releases (and re-acquires) rq->lock, it is |
37e117c0 PZ |
7207 | * possible for any higher priority task to appear. In that case we |
7208 | * must re-start the pick_next_entity() loop. | |
7209 | */ | |
e4aa358b | 7210 | if (new_tasks < 0) |
37e117c0 PZ |
7211 | return RETRY_TASK; |
7212 | ||
e4aa358b | 7213 | if (new_tasks > 0) |
38033c37 | 7214 | goto again; |
38033c37 | 7215 | |
23127296 VG |
7216 | /* |
7217 | * rq is about to be idle, check if we need to update the | |
7218 | * lost_idle_time of clock_pelt | |
7219 | */ | |
7220 | update_idle_rq_clock_pelt(rq); | |
7221 | ||
38033c37 | 7222 | return NULL; |
bf0f6f24 IM |
7223 | } |
7224 | ||
98c2f700 PZ |
7225 | static struct task_struct *__pick_next_task_fair(struct rq *rq) |
7226 | { | |
7227 | return pick_next_task_fair(rq, NULL, NULL); | |
7228 | } | |
7229 | ||
bf0f6f24 IM |
7230 | /* |
7231 | * Account for a descheduled task: | |
7232 | */ | |
6e2df058 | 7233 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
7234 | { |
7235 | struct sched_entity *se = &prev->se; | |
7236 | struct cfs_rq *cfs_rq; | |
7237 | ||
7238 | for_each_sched_entity(se) { | |
7239 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 7240 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
7241 | } |
7242 | } | |
7243 | ||
ac53db59 RR |
7244 | /* |
7245 | * sched_yield() is very simple | |
7246 | * | |
7247 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
7248 | */ | |
7249 | static void yield_task_fair(struct rq *rq) | |
7250 | { | |
7251 | struct task_struct *curr = rq->curr; | |
7252 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
7253 | struct sched_entity *se = &curr->se; | |
7254 | ||
7255 | /* | |
7256 | * Are we the only task in the tree? | |
7257 | */ | |
7258 | if (unlikely(rq->nr_running == 1)) | |
7259 | return; | |
7260 | ||
7261 | clear_buddies(cfs_rq, se); | |
7262 | ||
7263 | if (curr->policy != SCHED_BATCH) { | |
7264 | update_rq_clock(rq); | |
7265 | /* | |
7266 | * Update run-time statistics of the 'current'. | |
7267 | */ | |
7268 | update_curr(cfs_rq); | |
916671c0 MG |
7269 | /* |
7270 | * Tell update_rq_clock() that we've just updated, | |
7271 | * so we don't do microscopic update in schedule() | |
7272 | * and double the fastpath cost. | |
7273 | */ | |
adcc8da8 | 7274 | rq_clock_skip_update(rq); |
ac53db59 RR |
7275 | } |
7276 | ||
7277 | set_skip_buddy(se); | |
7278 | } | |
7279 | ||
0900acf2 | 7280 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) |
d95f4122 MG |
7281 | { |
7282 | struct sched_entity *se = &p->se; | |
7283 | ||
5238cdd3 PT |
7284 | /* throttled hierarchies are not runnable */ |
7285 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
7286 | return false; |
7287 | ||
7288 | /* Tell the scheduler that we'd really like pse to run next. */ | |
7289 | set_next_buddy(se); | |
7290 | ||
d95f4122 MG |
7291 | yield_task_fair(rq); |
7292 | ||
7293 | return true; | |
7294 | } | |
7295 | ||
681f3e68 | 7296 | #ifdef CONFIG_SMP |
bf0f6f24 | 7297 | /************************************************** |
e9c84cb8 PZ |
7298 | * Fair scheduling class load-balancing methods. |
7299 | * | |
7300 | * BASICS | |
7301 | * | |
7302 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 7303 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
7304 | * time to each task. This is expressed in the following equation: |
7305 | * | |
7306 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
7307 | * | |
97fb7a0a | 7308 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
7309 | * W_i,0 is defined as: |
7310 | * | |
7311 | * W_i,0 = \Sum_j w_i,j (2) | |
7312 | * | |
97fb7a0a | 7313 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 7314 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
7315 | * |
7316 | * The weight average is an exponential decay average of the instantaneous | |
7317 | * weight: | |
7318 | * | |
7319 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
7320 | * | |
97fb7a0a | 7321 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
7322 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
7323 | * can also include other factors [XXX]. | |
7324 | * | |
7325 | * To achieve this balance we define a measure of imbalance which follows | |
7326 | * directly from (1): | |
7327 | * | |
ced549fa | 7328 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
7329 | * |
7330 | * We them move tasks around to minimize the imbalance. In the continuous | |
7331 | * function space it is obvious this converges, in the discrete case we get | |
7332 | * a few fun cases generally called infeasible weight scenarios. | |
7333 | * | |
7334 | * [XXX expand on: | |
7335 | * - infeasible weights; | |
7336 | * - local vs global optima in the discrete case. ] | |
7337 | * | |
7338 | * | |
7339 | * SCHED DOMAINS | |
7340 | * | |
7341 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 7342 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 7343 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 7344 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 7345 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 7346 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
7347 | * the groups. |
7348 | * | |
7349 | * This yields: | |
7350 | * | |
7351 | * log_2 n 1 n | |
7352 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
7353 | * i = 0 2^i 2^i | |
7354 | * `- size of each group | |
97fb7a0a | 7355 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
7356 | * | `- freq |
7357 | * `- sum over all levels | |
7358 | * | |
7359 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
7360 | * this makes (5) the runtime complexity of the balancer. | |
7361 | * | |
7362 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 7363 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
7364 | * |
7365 | * The adjacency matrix of the resulting graph is given by: | |
7366 | * | |
97a7142f | 7367 | * log_2 n |
e9c84cb8 PZ |
7368 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
7369 | * k = 0 | |
7370 | * | |
7371 | * And you'll find that: | |
7372 | * | |
7373 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
7374 | * | |
97fb7a0a | 7375 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
7376 | * The task movement gives a factor of O(m), giving a convergence complexity |
7377 | * of: | |
7378 | * | |
7379 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
7380 | * | |
7381 | * | |
7382 | * WORK CONSERVING | |
7383 | * | |
7384 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 7385 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
7386 | * tree itself instead of relying on other CPUs to bring it work. |
7387 | * | |
7388 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
7389 | * time. | |
7390 | * | |
7391 | * [XXX more?] | |
7392 | * | |
7393 | * | |
7394 | * CGROUPS | |
7395 | * | |
7396 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
7397 | * | |
7398 | * s_k,i | |
7399 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
7400 | * S_k | |
7401 | * | |
7402 | * Where | |
7403 | * | |
7404 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
7405 | * | |
97fb7a0a | 7406 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
7407 | * |
7408 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
7409 | * property. | |
7410 | * | |
7411 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
7412 | * rewrite all of this once again.] | |
97a7142f | 7413 | */ |
bf0f6f24 | 7414 | |
ed387b78 HS |
7415 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
7416 | ||
0ec8aa00 PZ |
7417 | enum fbq_type { regular, remote, all }; |
7418 | ||
0b0695f2 | 7419 | /* |
a9723389 VG |
7420 | * 'group_type' describes the group of CPUs at the moment of load balancing. |
7421 | * | |
0b0695f2 | 7422 | * The enum is ordered by pulling priority, with the group with lowest priority |
a9723389 VG |
7423 | * first so the group_type can simply be compared when selecting the busiest |
7424 | * group. See update_sd_pick_busiest(). | |
0b0695f2 | 7425 | */ |
3b1baa64 | 7426 | enum group_type { |
a9723389 | 7427 | /* The group has spare capacity that can be used to run more tasks. */ |
0b0695f2 | 7428 | group_has_spare = 0, |
a9723389 VG |
7429 | /* |
7430 | * The group is fully used and the tasks don't compete for more CPU | |
7431 | * cycles. Nevertheless, some tasks might wait before running. | |
7432 | */ | |
0b0695f2 | 7433 | group_fully_busy, |
a9723389 VG |
7434 | /* |
7435 | * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity | |
7436 | * and must be migrated to a more powerful CPU. | |
7437 | */ | |
3b1baa64 | 7438 | group_misfit_task, |
a9723389 VG |
7439 | /* |
7440 | * SD_ASYM_PACKING only: One local CPU with higher capacity is available, | |
7441 | * and the task should be migrated to it instead of running on the | |
7442 | * current CPU. | |
7443 | */ | |
0b0695f2 | 7444 | group_asym_packing, |
a9723389 VG |
7445 | /* |
7446 | * The tasks' affinity constraints previously prevented the scheduler | |
7447 | * from balancing the load across the system. | |
7448 | */ | |
3b1baa64 | 7449 | group_imbalanced, |
a9723389 VG |
7450 | /* |
7451 | * The CPU is overloaded and can't provide expected CPU cycles to all | |
7452 | * tasks. | |
7453 | */ | |
0b0695f2 VG |
7454 | group_overloaded |
7455 | }; | |
7456 | ||
7457 | enum migration_type { | |
7458 | migrate_load = 0, | |
7459 | migrate_util, | |
7460 | migrate_task, | |
7461 | migrate_misfit | |
3b1baa64 MR |
7462 | }; |
7463 | ||
ddcdf6e7 | 7464 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 7465 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
7466 | #define LBF_DST_PINNED 0x04 |
7467 | #define LBF_SOME_PINNED 0x08 | |
23fb06d9 | 7468 | #define LBF_ACTIVE_LB 0x10 |
ddcdf6e7 PZ |
7469 | |
7470 | struct lb_env { | |
7471 | struct sched_domain *sd; | |
7472 | ||
ddcdf6e7 | 7473 | struct rq *src_rq; |
85c1e7da | 7474 | int src_cpu; |
ddcdf6e7 PZ |
7475 | |
7476 | int dst_cpu; | |
7477 | struct rq *dst_rq; | |
7478 | ||
88b8dac0 SV |
7479 | struct cpumask *dst_grpmask; |
7480 | int new_dst_cpu; | |
ddcdf6e7 | 7481 | enum cpu_idle_type idle; |
bd939f45 | 7482 | long imbalance; |
b9403130 MW |
7483 | /* The set of CPUs under consideration for load-balancing */ |
7484 | struct cpumask *cpus; | |
7485 | ||
ddcdf6e7 | 7486 | unsigned int flags; |
367456c7 PZ |
7487 | |
7488 | unsigned int loop; | |
7489 | unsigned int loop_break; | |
7490 | unsigned int loop_max; | |
0ec8aa00 PZ |
7491 | |
7492 | enum fbq_type fbq_type; | |
0b0695f2 | 7493 | enum migration_type migration_type; |
163122b7 | 7494 | struct list_head tasks; |
ddcdf6e7 PZ |
7495 | }; |
7496 | ||
029632fb PZ |
7497 | /* |
7498 | * Is this task likely cache-hot: | |
7499 | */ | |
5d5e2b1b | 7500 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
7501 | { |
7502 | s64 delta; | |
7503 | ||
e5673f28 KT |
7504 | lockdep_assert_held(&env->src_rq->lock); |
7505 | ||
029632fb PZ |
7506 | if (p->sched_class != &fair_sched_class) |
7507 | return 0; | |
7508 | ||
1da1843f | 7509 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
7510 | return 0; |
7511 | ||
ec73240b JD |
7512 | /* SMT siblings share cache */ |
7513 | if (env->sd->flags & SD_SHARE_CPUCAPACITY) | |
7514 | return 0; | |
7515 | ||
029632fb PZ |
7516 | /* |
7517 | * Buddy candidates are cache hot: | |
7518 | */ | |
5d5e2b1b | 7519 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
7520 | (&p->se == cfs_rq_of(&p->se)->next || |
7521 | &p->se == cfs_rq_of(&p->se)->last)) | |
7522 | return 1; | |
7523 | ||
7524 | if (sysctl_sched_migration_cost == -1) | |
7525 | return 1; | |
7526 | if (sysctl_sched_migration_cost == 0) | |
7527 | return 0; | |
7528 | ||
5d5e2b1b | 7529 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
7530 | |
7531 | return delta < (s64)sysctl_sched_migration_cost; | |
7532 | } | |
7533 | ||
3a7053b3 | 7534 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 7535 | /* |
2a1ed24c SD |
7536 | * Returns 1, if task migration degrades locality |
7537 | * Returns 0, if task migration improves locality i.e migration preferred. | |
7538 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 7539 | */ |
2a1ed24c | 7540 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 7541 | { |
b1ad065e | 7542 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
7543 | unsigned long src_weight, dst_weight; |
7544 | int src_nid, dst_nid, dist; | |
3a7053b3 | 7545 | |
2a595721 | 7546 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
7547 | return -1; |
7548 | ||
c3b9bc5b | 7549 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 7550 | return -1; |
7a0f3083 MG |
7551 | |
7552 | src_nid = cpu_to_node(env->src_cpu); | |
7553 | dst_nid = cpu_to_node(env->dst_cpu); | |
7554 | ||
83e1d2cd | 7555 | if (src_nid == dst_nid) |
2a1ed24c | 7556 | return -1; |
7a0f3083 | 7557 | |
2a1ed24c SD |
7558 | /* Migrating away from the preferred node is always bad. */ |
7559 | if (src_nid == p->numa_preferred_nid) { | |
7560 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
7561 | return 1; | |
7562 | else | |
7563 | return -1; | |
7564 | } | |
b1ad065e | 7565 | |
c1ceac62 RR |
7566 | /* Encourage migration to the preferred node. */ |
7567 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 7568 | return 0; |
b1ad065e | 7569 | |
739294fb | 7570 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 7571 | if (env->idle == CPU_IDLE) |
739294fb RR |
7572 | return -1; |
7573 | ||
f35678b6 | 7574 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 7575 | if (numa_group) { |
f35678b6 SD |
7576 | src_weight = group_weight(p, src_nid, dist); |
7577 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 7578 | } else { |
f35678b6 SD |
7579 | src_weight = task_weight(p, src_nid, dist); |
7580 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
7581 | } |
7582 | ||
f35678b6 | 7583 | return dst_weight < src_weight; |
7a0f3083 MG |
7584 | } |
7585 | ||
3a7053b3 | 7586 | #else |
2a1ed24c | 7587 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
7588 | struct lb_env *env) |
7589 | { | |
2a1ed24c | 7590 | return -1; |
7a0f3083 | 7591 | } |
3a7053b3 MG |
7592 | #endif |
7593 | ||
1e3c88bd PZ |
7594 | /* |
7595 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
7596 | */ | |
7597 | static | |
8e45cb54 | 7598 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 7599 | { |
2a1ed24c | 7600 | int tsk_cache_hot; |
e5673f28 KT |
7601 | |
7602 | lockdep_assert_held(&env->src_rq->lock); | |
7603 | ||
1e3c88bd PZ |
7604 | /* |
7605 | * We do not migrate tasks that are: | |
d3198084 | 7606 | * 1) throttled_lb_pair, or |
3bd37062 | 7607 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
7608 | * 3) running (obviously), or |
7609 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7610 | */ |
d3198084 JK |
7611 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7612 | return 0; | |
7613 | ||
9bcb959d | 7614 | /* Disregard pcpu kthreads; they are where they need to be. */ |
3a7956e2 | 7615 | if (kthread_is_per_cpu(p)) |
9bcb959d LC |
7616 | return 0; |
7617 | ||
3bd37062 | 7618 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 7619 | int cpu; |
88b8dac0 | 7620 | |
ae92882e | 7621 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 7622 | |
6263322c PZ |
7623 | env->flags |= LBF_SOME_PINNED; |
7624 | ||
88b8dac0 | 7625 | /* |
97fb7a0a | 7626 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
7627 | * our sched_group. We may want to revisit it if we couldn't |
7628 | * meet load balance goals by pulling other tasks on src_cpu. | |
7629 | * | |
23fb06d9 VS |
7630 | * Avoid computing new_dst_cpu |
7631 | * - for NEWLY_IDLE | |
7632 | * - if we have already computed one in current iteration | |
7633 | * - if it's an active balance | |
88b8dac0 | 7634 | */ |
23fb06d9 VS |
7635 | if (env->idle == CPU_NEWLY_IDLE || |
7636 | env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) | |
88b8dac0 SV |
7637 | return 0; |
7638 | ||
97fb7a0a | 7639 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 7640 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 7641 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 7642 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7643 | env->new_dst_cpu = cpu; |
7644 | break; | |
7645 | } | |
88b8dac0 | 7646 | } |
e02e60c1 | 7647 | |
1e3c88bd PZ |
7648 | return 0; |
7649 | } | |
88b8dac0 | 7650 | |
3b03706f | 7651 | /* Record that we found at least one task that could run on dst_cpu */ |
8e45cb54 | 7652 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7653 | |
ddcdf6e7 | 7654 | if (task_running(env->src_rq, p)) { |
ae92882e | 7655 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
7656 | return 0; |
7657 | } | |
7658 | ||
7659 | /* | |
7660 | * Aggressive migration if: | |
23fb06d9 VS |
7661 | * 1) active balance |
7662 | * 2) destination numa is preferred | |
7663 | * 3) task is cache cold, or | |
7664 | * 4) too many balance attempts have failed. | |
1e3c88bd | 7665 | */ |
23fb06d9 VS |
7666 | if (env->flags & LBF_ACTIVE_LB) |
7667 | return 1; | |
7668 | ||
2a1ed24c SD |
7669 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7670 | if (tsk_cache_hot == -1) | |
7671 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7672 | |
2a1ed24c | 7673 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7674 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7675 | if (tsk_cache_hot == 1) { |
ae92882e JP |
7676 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
7677 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 7678 | } |
1e3c88bd PZ |
7679 | return 1; |
7680 | } | |
7681 | ||
ae92882e | 7682 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 7683 | return 0; |
1e3c88bd PZ |
7684 | } |
7685 | ||
897c395f | 7686 | /* |
163122b7 KT |
7687 | * detach_task() -- detach the task for the migration specified in env |
7688 | */ | |
7689 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7690 | { | |
7691 | lockdep_assert_held(&env->src_rq->lock); | |
7692 | ||
5704ac0a | 7693 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7694 | set_task_cpu(p, env->dst_cpu); |
7695 | } | |
7696 | ||
897c395f | 7697 | /* |
e5673f28 | 7698 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7699 | * part of active balancing operations within "domain". |
897c395f | 7700 | * |
e5673f28 | 7701 | * Returns a task if successful and NULL otherwise. |
897c395f | 7702 | */ |
e5673f28 | 7703 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7704 | { |
93824900 | 7705 | struct task_struct *p; |
897c395f | 7706 | |
e5673f28 KT |
7707 | lockdep_assert_held(&env->src_rq->lock); |
7708 | ||
93824900 UR |
7709 | list_for_each_entry_reverse(p, |
7710 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7711 | if (!can_migrate_task(p, env)) |
7712 | continue; | |
897c395f | 7713 | |
163122b7 | 7714 | detach_task(p, env); |
e5673f28 | 7715 | |
367456c7 | 7716 | /* |
e5673f28 | 7717 | * Right now, this is only the second place where |
163122b7 | 7718 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 7719 | * so we can safely collect stats here rather than |
163122b7 | 7720 | * inside detach_tasks(). |
367456c7 | 7721 | */ |
ae92882e | 7722 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 7723 | return p; |
897c395f | 7724 | } |
e5673f28 | 7725 | return NULL; |
897c395f PZ |
7726 | } |
7727 | ||
eb95308e PZ |
7728 | static const unsigned int sched_nr_migrate_break = 32; |
7729 | ||
5d6523eb | 7730 | /* |
0b0695f2 | 7731 | * detach_tasks() -- tries to detach up to imbalance load/util/tasks from |
163122b7 | 7732 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 7733 | * |
163122b7 | 7734 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 7735 | */ |
163122b7 | 7736 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 7737 | { |
5d6523eb | 7738 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
0b0695f2 | 7739 | unsigned long util, load; |
5d6523eb | 7740 | struct task_struct *p; |
163122b7 KT |
7741 | int detached = 0; |
7742 | ||
7743 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 7744 | |
acb4decc AL |
7745 | /* |
7746 | * Source run queue has been emptied by another CPU, clear | |
7747 | * LBF_ALL_PINNED flag as we will not test any task. | |
7748 | */ | |
7749 | if (env->src_rq->nr_running <= 1) { | |
7750 | env->flags &= ~LBF_ALL_PINNED; | |
7751 | return 0; | |
7752 | } | |
7753 | ||
bd939f45 | 7754 | if (env->imbalance <= 0) |
5d6523eb | 7755 | return 0; |
1e3c88bd | 7756 | |
5d6523eb | 7757 | while (!list_empty(tasks)) { |
985d3a4c YD |
7758 | /* |
7759 | * We don't want to steal all, otherwise we may be treated likewise, | |
7760 | * which could at worst lead to a livelock crash. | |
7761 | */ | |
7762 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
7763 | break; | |
7764 | ||
93824900 | 7765 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 7766 | |
367456c7 PZ |
7767 | env->loop++; |
7768 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 7769 | if (env->loop > env->loop_max) |
367456c7 | 7770 | break; |
5d6523eb PZ |
7771 | |
7772 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 7773 | if (env->loop > env->loop_break) { |
eb95308e | 7774 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 7775 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 7776 | break; |
a195f004 | 7777 | } |
1e3c88bd | 7778 | |
d3198084 | 7779 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
7780 | goto next; |
7781 | ||
0b0695f2 VG |
7782 | switch (env->migration_type) { |
7783 | case migrate_load: | |
01cfcde9 VG |
7784 | /* |
7785 | * Depending of the number of CPUs and tasks and the | |
7786 | * cgroup hierarchy, task_h_load() can return a null | |
7787 | * value. Make sure that env->imbalance decreases | |
7788 | * otherwise detach_tasks() will stop only after | |
7789 | * detaching up to loop_max tasks. | |
7790 | */ | |
7791 | load = max_t(unsigned long, task_h_load(p), 1); | |
5d6523eb | 7792 | |
0b0695f2 VG |
7793 | if (sched_feat(LB_MIN) && |
7794 | load < 16 && !env->sd->nr_balance_failed) | |
7795 | goto next; | |
367456c7 | 7796 | |
6cf82d55 VG |
7797 | /* |
7798 | * Make sure that we don't migrate too much load. | |
7799 | * Nevertheless, let relax the constraint if | |
7800 | * scheduler fails to find a good waiting task to | |
7801 | * migrate. | |
7802 | */ | |
39a2a6eb | 7803 | if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) |
0b0695f2 VG |
7804 | goto next; |
7805 | ||
7806 | env->imbalance -= load; | |
7807 | break; | |
7808 | ||
7809 | case migrate_util: | |
7810 | util = task_util_est(p); | |
7811 | ||
7812 | if (util > env->imbalance) | |
7813 | goto next; | |
7814 | ||
7815 | env->imbalance -= util; | |
7816 | break; | |
7817 | ||
7818 | case migrate_task: | |
7819 | env->imbalance--; | |
7820 | break; | |
7821 | ||
7822 | case migrate_misfit: | |
c63be7be VG |
7823 | /* This is not a misfit task */ |
7824 | if (task_fits_capacity(p, capacity_of(env->src_cpu))) | |
0b0695f2 VG |
7825 | goto next; |
7826 | ||
7827 | env->imbalance = 0; | |
7828 | break; | |
7829 | } | |
1e3c88bd | 7830 | |
163122b7 KT |
7831 | detach_task(p, env); |
7832 | list_add(&p->se.group_node, &env->tasks); | |
7833 | ||
7834 | detached++; | |
1e3c88bd | 7835 | |
c1a280b6 | 7836 | #ifdef CONFIG_PREEMPTION |
ee00e66f PZ |
7837 | /* |
7838 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 7839 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
7840 | * the critical section. |
7841 | */ | |
5d6523eb | 7842 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 7843 | break; |
1e3c88bd PZ |
7844 | #endif |
7845 | ||
ee00e66f PZ |
7846 | /* |
7847 | * We only want to steal up to the prescribed amount of | |
0b0695f2 | 7848 | * load/util/tasks. |
ee00e66f | 7849 | */ |
bd939f45 | 7850 | if (env->imbalance <= 0) |
ee00e66f | 7851 | break; |
367456c7 PZ |
7852 | |
7853 | continue; | |
7854 | next: | |
93824900 | 7855 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 7856 | } |
5d6523eb | 7857 | |
1e3c88bd | 7858 | /* |
163122b7 KT |
7859 | * Right now, this is one of only two places we collect this stat |
7860 | * so we can safely collect detach_one_task() stats here rather | |
7861 | * than inside detach_one_task(). | |
1e3c88bd | 7862 | */ |
ae92882e | 7863 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 7864 | |
163122b7 KT |
7865 | return detached; |
7866 | } | |
7867 | ||
7868 | /* | |
7869 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
7870 | */ | |
7871 | static void attach_task(struct rq *rq, struct task_struct *p) | |
7872 | { | |
7873 | lockdep_assert_held(&rq->lock); | |
7874 | ||
7875 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 7876 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
163122b7 KT |
7877 | check_preempt_curr(rq, p, 0); |
7878 | } | |
7879 | ||
7880 | /* | |
7881 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
7882 | * its new rq. | |
7883 | */ | |
7884 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
7885 | { | |
8a8c69c3 PZ |
7886 | struct rq_flags rf; |
7887 | ||
7888 | rq_lock(rq, &rf); | |
5704ac0a | 7889 | update_rq_clock(rq); |
163122b7 | 7890 | attach_task(rq, p); |
8a8c69c3 | 7891 | rq_unlock(rq, &rf); |
163122b7 KT |
7892 | } |
7893 | ||
7894 | /* | |
7895 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
7896 | * new rq. | |
7897 | */ | |
7898 | static void attach_tasks(struct lb_env *env) | |
7899 | { | |
7900 | struct list_head *tasks = &env->tasks; | |
7901 | struct task_struct *p; | |
8a8c69c3 | 7902 | struct rq_flags rf; |
163122b7 | 7903 | |
8a8c69c3 | 7904 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 7905 | update_rq_clock(env->dst_rq); |
163122b7 KT |
7906 | |
7907 | while (!list_empty(tasks)) { | |
7908 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
7909 | list_del_init(&p->se.group_node); | |
1e3c88bd | 7910 | |
163122b7 KT |
7911 | attach_task(env->dst_rq, p); |
7912 | } | |
7913 | ||
8a8c69c3 | 7914 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
7915 | } |
7916 | ||
b0c79224 | 7917 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
7918 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
7919 | { | |
7920 | if (cfs_rq->avg.load_avg) | |
7921 | return true; | |
7922 | ||
7923 | if (cfs_rq->avg.util_avg) | |
7924 | return true; | |
7925 | ||
7926 | return false; | |
7927 | } | |
7928 | ||
91c27493 | 7929 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
7930 | { |
7931 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
7932 | return true; | |
7933 | ||
3727e0e1 VG |
7934 | if (READ_ONCE(rq->avg_dl.util_avg)) |
7935 | return true; | |
7936 | ||
b4eccf5f TG |
7937 | if (thermal_load_avg(rq)) |
7938 | return true; | |
7939 | ||
11d4afd4 | 7940 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
7941 | if (READ_ONCE(rq->avg_irq.util_avg)) |
7942 | return true; | |
7943 | #endif | |
7944 | ||
371bf427 VG |
7945 | return false; |
7946 | } | |
7947 | ||
39b6a429 | 7948 | static inline void update_blocked_load_tick(struct rq *rq) |
b0c79224 | 7949 | { |
39b6a429 VG |
7950 | WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); |
7951 | } | |
b0c79224 | 7952 | |
39b6a429 VG |
7953 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
7954 | { | |
b0c79224 VS |
7955 | if (!has_blocked) |
7956 | rq->has_blocked_load = 0; | |
7957 | } | |
7958 | #else | |
7959 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
7960 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
39b6a429 | 7961 | static inline void update_blocked_load_tick(struct rq *rq) {} |
b0c79224 VS |
7962 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} |
7963 | #endif | |
7964 | ||
bef69dd8 VG |
7965 | static bool __update_blocked_others(struct rq *rq, bool *done) |
7966 | { | |
7967 | const struct sched_class *curr_class; | |
7968 | u64 now = rq_clock_pelt(rq); | |
b4eccf5f | 7969 | unsigned long thermal_pressure; |
bef69dd8 VG |
7970 | bool decayed; |
7971 | ||
7972 | /* | |
7973 | * update_load_avg() can call cpufreq_update_util(). Make sure that RT, | |
7974 | * DL and IRQ signals have been updated before updating CFS. | |
7975 | */ | |
7976 | curr_class = rq->curr->sched_class; | |
7977 | ||
b4eccf5f TG |
7978 | thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); |
7979 | ||
bef69dd8 VG |
7980 | decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | |
7981 | update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | | |
05289b90 | 7982 | update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | |
bef69dd8 VG |
7983 | update_irq_load_avg(rq, 0); |
7984 | ||
7985 | if (others_have_blocked(rq)) | |
7986 | *done = false; | |
7987 | ||
7988 | return decayed; | |
7989 | } | |
7990 | ||
1936c53c VG |
7991 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7992 | ||
039ae8bc VG |
7993 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
7994 | { | |
7995 | if (cfs_rq->load.weight) | |
7996 | return false; | |
7997 | ||
7998 | if (cfs_rq->avg.load_sum) | |
7999 | return false; | |
8000 | ||
8001 | if (cfs_rq->avg.util_sum) | |
8002 | return false; | |
8003 | ||
9f683953 VG |
8004 | if (cfs_rq->avg.runnable_sum) |
8005 | return false; | |
8006 | ||
039ae8bc VG |
8007 | return true; |
8008 | } | |
8009 | ||
bef69dd8 | 8010 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8011 | { |
039ae8bc | 8012 | struct cfs_rq *cfs_rq, *pos; |
bef69dd8 VG |
8013 | bool decayed = false; |
8014 | int cpu = cpu_of(rq); | |
b90f7c9d | 8015 | |
9763b67f PZ |
8016 | /* |
8017 | * Iterates the task_group tree in a bottom up fashion, see | |
8018 | * list_add_leaf_cfs_rq() for details. | |
8019 | */ | |
039ae8bc | 8020 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
8021 | struct sched_entity *se; |
8022 | ||
bef69dd8 | 8023 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { |
fe749158 | 8024 | update_tg_load_avg(cfs_rq); |
4e516076 | 8025 | |
bef69dd8 VG |
8026 | if (cfs_rq == &rq->cfs) |
8027 | decayed = true; | |
8028 | } | |
8029 | ||
bc427898 VG |
8030 | /* Propagate pending load changes to the parent, if any: */ |
8031 | se = cfs_rq->tg->se[cpu]; | |
8032 | if (se && !skip_blocked_update(se)) | |
88c0616e | 8033 | update_load_avg(cfs_rq_of(se), se, 0); |
a9e7f654 | 8034 | |
039ae8bc VG |
8035 | /* |
8036 | * There can be a lot of idle CPU cgroups. Don't let fully | |
8037 | * decayed cfs_rqs linger on the list. | |
8038 | */ | |
8039 | if (cfs_rq_is_decayed(cfs_rq)) | |
8040 | list_del_leaf_cfs_rq(cfs_rq); | |
8041 | ||
1936c53c VG |
8042 | /* Don't need periodic decay once load/util_avg are null */ |
8043 | if (cfs_rq_has_blocked(cfs_rq)) | |
bef69dd8 | 8044 | *done = false; |
9d89c257 | 8045 | } |
12b04875 | 8046 | |
bef69dd8 | 8047 | return decayed; |
9e3081ca PZ |
8048 | } |
8049 | ||
9763b67f | 8050 | /* |
68520796 | 8051 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
8052 | * This needs to be done in a top-down fashion because the load of a child |
8053 | * group is a fraction of its parents load. | |
8054 | */ | |
68520796 | 8055 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 8056 | { |
68520796 VD |
8057 | struct rq *rq = rq_of(cfs_rq); |
8058 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 8059 | unsigned long now = jiffies; |
68520796 | 8060 | unsigned long load; |
a35b6466 | 8061 | |
68520796 | 8062 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
8063 | return; |
8064 | ||
0e9f0245 | 8065 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
8066 | for_each_sched_entity(se) { |
8067 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 8068 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
8069 | if (cfs_rq->last_h_load_update == now) |
8070 | break; | |
8071 | } | |
a35b6466 | 8072 | |
68520796 | 8073 | if (!se) { |
7ea241af | 8074 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
8075 | cfs_rq->last_h_load_update = now; |
8076 | } | |
8077 | ||
0e9f0245 | 8078 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 8079 | load = cfs_rq->h_load; |
7ea241af YD |
8080 | load = div64_ul(load * se->avg.load_avg, |
8081 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
8082 | cfs_rq = group_cfs_rq(se); |
8083 | cfs_rq->h_load = load; | |
8084 | cfs_rq->last_h_load_update = now; | |
8085 | } | |
9763b67f PZ |
8086 | } |
8087 | ||
367456c7 | 8088 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 8089 | { |
367456c7 | 8090 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 8091 | |
68520796 | 8092 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 8093 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 8094 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
8095 | } |
8096 | #else | |
bef69dd8 | 8097 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8098 | { |
6c1d47c0 | 8099 | struct cfs_rq *cfs_rq = &rq->cfs; |
bef69dd8 | 8100 | bool decayed; |
b90f7c9d | 8101 | |
bef69dd8 VG |
8102 | decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
8103 | if (cfs_rq_has_blocked(cfs_rq)) | |
8104 | *done = false; | |
b90f7c9d | 8105 | |
bef69dd8 | 8106 | return decayed; |
9e3081ca PZ |
8107 | } |
8108 | ||
367456c7 | 8109 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 8110 | { |
9d89c257 | 8111 | return p->se.avg.load_avg; |
1e3c88bd | 8112 | } |
230059de | 8113 | #endif |
1e3c88bd | 8114 | |
bef69dd8 VG |
8115 | static void update_blocked_averages(int cpu) |
8116 | { | |
8117 | bool decayed = false, done = true; | |
8118 | struct rq *rq = cpu_rq(cpu); | |
8119 | struct rq_flags rf; | |
8120 | ||
8121 | rq_lock_irqsave(rq, &rf); | |
39b6a429 | 8122 | update_blocked_load_tick(rq); |
bef69dd8 VG |
8123 | update_rq_clock(rq); |
8124 | ||
8125 | decayed |= __update_blocked_others(rq, &done); | |
8126 | decayed |= __update_blocked_fair(rq, &done); | |
8127 | ||
8128 | update_blocked_load_status(rq, !done); | |
8129 | if (decayed) | |
8130 | cpufreq_update_util(rq, 0); | |
8131 | rq_unlock_irqrestore(rq, &rf); | |
8132 | } | |
8133 | ||
1e3c88bd | 8134 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 8135 | |
1e3c88bd PZ |
8136 | /* |
8137 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
8138 | */ | |
8139 | struct sg_lb_stats { | |
8140 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
8141 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
63b2ca30 | 8142 | unsigned long group_capacity; |
070f5e86 VG |
8143 | unsigned long group_util; /* Total utilization over the CPUs of the group */ |
8144 | unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ | |
5e23e474 | 8145 | unsigned int sum_nr_running; /* Nr of tasks running in the group */ |
a3498347 | 8146 | unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ |
147c5fc2 PZ |
8147 | unsigned int idle_cpus; |
8148 | unsigned int group_weight; | |
caeb178c | 8149 | enum group_type group_type; |
490ba971 | 8150 | unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ |
3b1baa64 | 8151 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
8152 | #ifdef CONFIG_NUMA_BALANCING |
8153 | unsigned int nr_numa_running; | |
8154 | unsigned int nr_preferred_running; | |
8155 | #endif | |
1e3c88bd PZ |
8156 | }; |
8157 | ||
56cf515b JK |
8158 | /* |
8159 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
8160 | * during load balancing. | |
8161 | */ | |
8162 | struct sd_lb_stats { | |
8163 | struct sched_group *busiest; /* Busiest group in this sd */ | |
8164 | struct sched_group *local; /* Local group in this sd */ | |
8165 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 8166 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b | 8167 | unsigned long avg_load; /* Average load across all groups in sd */ |
0b0695f2 | 8168 | unsigned int prefer_sibling; /* tasks should go to sibling first */ |
56cf515b | 8169 | |
56cf515b | 8170 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 8171 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
8172 | }; |
8173 | ||
147c5fc2 PZ |
8174 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
8175 | { | |
8176 | /* | |
8177 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
8178 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
0b0695f2 VG |
8179 | * We must however set busiest_stat::group_type and |
8180 | * busiest_stat::idle_cpus to the worst busiest group because | |
8181 | * update_sd_pick_busiest() reads these before assignment. | |
147c5fc2 PZ |
8182 | */ |
8183 | *sds = (struct sd_lb_stats){ | |
8184 | .busiest = NULL, | |
8185 | .local = NULL, | |
8186 | .total_load = 0UL, | |
63b2ca30 | 8187 | .total_capacity = 0UL, |
147c5fc2 | 8188 | .busiest_stat = { |
0b0695f2 VG |
8189 | .idle_cpus = UINT_MAX, |
8190 | .group_type = group_has_spare, | |
147c5fc2 PZ |
8191 | }, |
8192 | }; | |
8193 | } | |
8194 | ||
1ca2034e | 8195 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
8196 | { |
8197 | struct rq *rq = cpu_rq(cpu); | |
8ec59c0f | 8198 | unsigned long max = arch_scale_cpu_capacity(cpu); |
523e979d | 8199 | unsigned long used, free; |
523e979d | 8200 | unsigned long irq; |
b654f7de | 8201 | |
2e62c474 | 8202 | irq = cpu_util_irq(rq); |
cadefd3d | 8203 | |
523e979d VG |
8204 | if (unlikely(irq >= max)) |
8205 | return 1; | |
aa483808 | 8206 | |
467b7d01 TG |
8207 | /* |
8208 | * avg_rt.util_avg and avg_dl.util_avg track binary signals | |
8209 | * (running and not running) with weights 0 and 1024 respectively. | |
8210 | * avg_thermal.load_avg tracks thermal pressure and the weighted | |
8211 | * average uses the actual delta max capacity(load). | |
8212 | */ | |
523e979d VG |
8213 | used = READ_ONCE(rq->avg_rt.util_avg); |
8214 | used += READ_ONCE(rq->avg_dl.util_avg); | |
467b7d01 | 8215 | used += thermal_load_avg(rq); |
1e3c88bd | 8216 | |
523e979d VG |
8217 | if (unlikely(used >= max)) |
8218 | return 1; | |
1e3c88bd | 8219 | |
523e979d | 8220 | free = max - used; |
2e62c474 VG |
8221 | |
8222 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
8223 | } |
8224 | ||
ced549fa | 8225 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 8226 | { |
1ca2034e | 8227 | unsigned long capacity = scale_rt_capacity(cpu); |
1e3c88bd PZ |
8228 | struct sched_group *sdg = sd->groups; |
8229 | ||
8ec59c0f | 8230 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); |
1e3c88bd | 8231 | |
ced549fa NP |
8232 | if (!capacity) |
8233 | capacity = 1; | |
1e3c88bd | 8234 | |
ced549fa | 8235 | cpu_rq(cpu)->cpu_capacity = capacity; |
51cf18c9 VD |
8236 | trace_sched_cpu_capacity_tp(cpu_rq(cpu)); |
8237 | ||
ced549fa | 8238 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8239 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 8240 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
8241 | } |
8242 | ||
63b2ca30 | 8243 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
8244 | { |
8245 | struct sched_domain *child = sd->child; | |
8246 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 8247 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
8248 | unsigned long interval; |
8249 | ||
8250 | interval = msecs_to_jiffies(sd->balance_interval); | |
8251 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 8252 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
8253 | |
8254 | if (!child) { | |
ced549fa | 8255 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
8256 | return; |
8257 | } | |
8258 | ||
dc7ff76e | 8259 | capacity = 0; |
bf475ce0 | 8260 | min_capacity = ULONG_MAX; |
e3d6d0cb | 8261 | max_capacity = 0; |
1e3c88bd | 8262 | |
74a5ce20 PZ |
8263 | if (child->flags & SD_OVERLAP) { |
8264 | /* | |
8265 | * SD_OVERLAP domains cannot assume that child groups | |
8266 | * span the current group. | |
8267 | */ | |
8268 | ||
ae4df9d6 | 8269 | for_each_cpu(cpu, sched_group_span(sdg)) { |
4c58f57f | 8270 | unsigned long cpu_cap = capacity_of(cpu); |
863bffc8 | 8271 | |
4c58f57f PL |
8272 | capacity += cpu_cap; |
8273 | min_capacity = min(cpu_cap, min_capacity); | |
8274 | max_capacity = max(cpu_cap, max_capacity); | |
863bffc8 | 8275 | } |
74a5ce20 PZ |
8276 | } else { |
8277 | /* | |
8278 | * !SD_OVERLAP domains can assume that child groups | |
8279 | * span the current group. | |
97a7142f | 8280 | */ |
74a5ce20 PZ |
8281 | |
8282 | group = child->groups; | |
8283 | do { | |
bf475ce0 MR |
8284 | struct sched_group_capacity *sgc = group->sgc; |
8285 | ||
8286 | capacity += sgc->capacity; | |
8287 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 8288 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
8289 | group = group->next; |
8290 | } while (group != child->groups); | |
8291 | } | |
1e3c88bd | 8292 | |
63b2ca30 | 8293 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8294 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 8295 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
8296 | } |
8297 | ||
9d5efe05 | 8298 | /* |
ea67821b VG |
8299 | * Check whether the capacity of the rq has been noticeably reduced by side |
8300 | * activity. The imbalance_pct is used for the threshold. | |
8301 | * Return true is the capacity is reduced | |
9d5efe05 SV |
8302 | */ |
8303 | static inline int | |
ea67821b | 8304 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 8305 | { |
ea67821b VG |
8306 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
8307 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
8308 | } |
8309 | ||
a0fe2cf0 VS |
8310 | /* |
8311 | * Check whether a rq has a misfit task and if it looks like we can actually | |
8312 | * help that task: we can migrate the task to a CPU of higher capacity, or | |
8313 | * the task's current CPU is heavily pressured. | |
8314 | */ | |
8315 | static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) | |
8316 | { | |
8317 | return rq->misfit_task_load && | |
8318 | (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || | |
8319 | check_cpu_capacity(rq, sd)); | |
8320 | } | |
8321 | ||
30ce5dab PZ |
8322 | /* |
8323 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 8324 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 8325 | * |
97fb7a0a IM |
8326 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
8327 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
8328 | * Something like: |
8329 | * | |
2b4d5b25 IM |
8330 | * { 0 1 2 3 } { 4 5 6 7 } |
8331 | * * * * * | |
30ce5dab PZ |
8332 | * |
8333 | * If we were to balance group-wise we'd place two tasks in the first group and | |
8334 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 8335 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
8336 | * |
8337 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
8338 | * by noticing the lower domain failed to reach balance and had difficulty |
8339 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
8340 | * |
8341 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 8342 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 8343 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
8344 | * to create an effective group imbalance. |
8345 | * | |
8346 | * This is a somewhat tricky proposition since the next run might not find the | |
8347 | * group imbalance and decide the groups need to be balanced again. A most | |
8348 | * subtle and fragile situation. | |
8349 | */ | |
8350 | ||
6263322c | 8351 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 8352 | { |
63b2ca30 | 8353 | return group->sgc->imbalance; |
30ce5dab PZ |
8354 | } |
8355 | ||
b37d9316 | 8356 | /* |
ea67821b VG |
8357 | * group_has_capacity returns true if the group has spare capacity that could |
8358 | * be used by some tasks. | |
8359 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
8360 | * smaller than the number of CPUs or if the utilization is lower than the |
8361 | * available capacity for CFS tasks. | |
ea67821b VG |
8362 | * For the latter, we use a threshold to stabilize the state, to take into |
8363 | * account the variance of the tasks' load and to return true if the available | |
8364 | * capacity in meaningful for the load balancer. | |
8365 | * As an example, an available capacity of 1% can appear but it doesn't make | |
8366 | * any benefit for the load balance. | |
b37d9316 | 8367 | */ |
ea67821b | 8368 | static inline bool |
57abff06 | 8369 | group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
b37d9316 | 8370 | { |
5e23e474 | 8371 | if (sgs->sum_nr_running < sgs->group_weight) |
ea67821b | 8372 | return true; |
c61037e9 | 8373 | |
070f5e86 VG |
8374 | if ((sgs->group_capacity * imbalance_pct) < |
8375 | (sgs->group_runnable * 100)) | |
8376 | return false; | |
8377 | ||
ea67821b | 8378 | if ((sgs->group_capacity * 100) > |
57abff06 | 8379 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8380 | return true; |
b37d9316 | 8381 | |
ea67821b VG |
8382 | return false; |
8383 | } | |
8384 | ||
8385 | /* | |
8386 | * group_is_overloaded returns true if the group has more tasks than it can | |
8387 | * handle. | |
8388 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
8389 | * with the exact right number of tasks, has no more spare capacity but is not | |
8390 | * overloaded so both group_has_capacity and group_is_overloaded return | |
8391 | * false. | |
8392 | */ | |
8393 | static inline bool | |
57abff06 | 8394 | group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
ea67821b | 8395 | { |
5e23e474 | 8396 | if (sgs->sum_nr_running <= sgs->group_weight) |
ea67821b | 8397 | return false; |
b37d9316 | 8398 | |
ea67821b | 8399 | if ((sgs->group_capacity * 100) < |
57abff06 | 8400 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8401 | return true; |
b37d9316 | 8402 | |
070f5e86 VG |
8403 | if ((sgs->group_capacity * imbalance_pct) < |
8404 | (sgs->group_runnable * 100)) | |
8405 | return true; | |
8406 | ||
ea67821b | 8407 | return false; |
b37d9316 PZ |
8408 | } |
8409 | ||
79a89f92 | 8410 | static inline enum |
57abff06 | 8411 | group_type group_classify(unsigned int imbalance_pct, |
0b0695f2 | 8412 | struct sched_group *group, |
79a89f92 | 8413 | struct sg_lb_stats *sgs) |
caeb178c | 8414 | { |
57abff06 | 8415 | if (group_is_overloaded(imbalance_pct, sgs)) |
caeb178c RR |
8416 | return group_overloaded; |
8417 | ||
8418 | if (sg_imbalanced(group)) | |
8419 | return group_imbalanced; | |
8420 | ||
0b0695f2 VG |
8421 | if (sgs->group_asym_packing) |
8422 | return group_asym_packing; | |
8423 | ||
3b1baa64 MR |
8424 | if (sgs->group_misfit_task_load) |
8425 | return group_misfit_task; | |
8426 | ||
57abff06 | 8427 | if (!group_has_capacity(imbalance_pct, sgs)) |
0b0695f2 VG |
8428 | return group_fully_busy; |
8429 | ||
8430 | return group_has_spare; | |
caeb178c RR |
8431 | } |
8432 | ||
1e3c88bd PZ |
8433 | /** |
8434 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 8435 | * @env: The load balancing environment. |
1e3c88bd | 8436 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 8437 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 8438 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 8439 | */ |
bd939f45 | 8440 | static inline void update_sg_lb_stats(struct lb_env *env, |
630246a0 QP |
8441 | struct sched_group *group, |
8442 | struct sg_lb_stats *sgs, | |
8443 | int *sg_status) | |
1e3c88bd | 8444 | { |
0b0695f2 | 8445 | int i, nr_running, local_group; |
1e3c88bd | 8446 | |
b72ff13c PZ |
8447 | memset(sgs, 0, sizeof(*sgs)); |
8448 | ||
0b0695f2 VG |
8449 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group)); |
8450 | ||
ae4df9d6 | 8451 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd PZ |
8452 | struct rq *rq = cpu_rq(i); |
8453 | ||
b0fb1eb4 | 8454 | sgs->group_load += cpu_load(rq); |
9e91d61d | 8455 | sgs->group_util += cpu_util(i); |
070f5e86 | 8456 | sgs->group_runnable += cpu_runnable(rq); |
a3498347 | 8457 | sgs->sum_h_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 8458 | |
a426f99c | 8459 | nr_running = rq->nr_running; |
5e23e474 VG |
8460 | sgs->sum_nr_running += nr_running; |
8461 | ||
a426f99c | 8462 | if (nr_running > 1) |
630246a0 | 8463 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 8464 | |
2802bf3c MR |
8465 | if (cpu_overutilized(i)) |
8466 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 8467 | |
0ec8aa00 PZ |
8468 | #ifdef CONFIG_NUMA_BALANCING |
8469 | sgs->nr_numa_running += rq->nr_numa_running; | |
8470 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
8471 | #endif | |
a426f99c WL |
8472 | /* |
8473 | * No need to call idle_cpu() if nr_running is not 0 | |
8474 | */ | |
0b0695f2 | 8475 | if (!nr_running && idle_cpu(i)) { |
aae6d3dd | 8476 | sgs->idle_cpus++; |
0b0695f2 VG |
8477 | /* Idle cpu can't have misfit task */ |
8478 | continue; | |
8479 | } | |
8480 | ||
8481 | if (local_group) | |
8482 | continue; | |
3b1baa64 | 8483 | |
0b0695f2 | 8484 | /* Check for a misfit task on the cpu */ |
3b1baa64 | 8485 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && |
757ffdd7 | 8486 | sgs->group_misfit_task_load < rq->misfit_task_load) { |
3b1baa64 | 8487 | sgs->group_misfit_task_load = rq->misfit_task_load; |
630246a0 | 8488 | *sg_status |= SG_OVERLOAD; |
757ffdd7 | 8489 | } |
1e3c88bd PZ |
8490 | } |
8491 | ||
0b0695f2 VG |
8492 | /* Check if dst CPU is idle and preferred to this group */ |
8493 | if (env->sd->flags & SD_ASYM_PACKING && | |
8494 | env->idle != CPU_NOT_IDLE && | |
8495 | sgs->sum_h_nr_running && | |
8496 | sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) { | |
8497 | sgs->group_asym_packing = 1; | |
8498 | } | |
8499 | ||
63b2ca30 | 8500 | sgs->group_capacity = group->sgc->capacity; |
1e3c88bd | 8501 | |
aae6d3dd | 8502 | sgs->group_weight = group->group_weight; |
b37d9316 | 8503 | |
57abff06 | 8504 | sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); |
0b0695f2 VG |
8505 | |
8506 | /* Computing avg_load makes sense only when group is overloaded */ | |
8507 | if (sgs->group_type == group_overloaded) | |
8508 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / | |
8509 | sgs->group_capacity; | |
1e3c88bd PZ |
8510 | } |
8511 | ||
532cb4c4 MN |
8512 | /** |
8513 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 8514 | * @env: The load balancing environment. |
532cb4c4 MN |
8515 | * @sds: sched_domain statistics |
8516 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 8517 | * @sgs: sched_group statistics |
532cb4c4 MN |
8518 | * |
8519 | * Determine if @sg is a busier group than the previously selected | |
8520 | * busiest group. | |
e69f6186 YB |
8521 | * |
8522 | * Return: %true if @sg is a busier group than the previously selected | |
8523 | * busiest group. %false otherwise. | |
532cb4c4 | 8524 | */ |
bd939f45 | 8525 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
8526 | struct sd_lb_stats *sds, |
8527 | struct sched_group *sg, | |
bd939f45 | 8528 | struct sg_lb_stats *sgs) |
532cb4c4 | 8529 | { |
caeb178c | 8530 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 8531 | |
0b0695f2 VG |
8532 | /* Make sure that there is at least one task to pull */ |
8533 | if (!sgs->sum_h_nr_running) | |
8534 | return false; | |
8535 | ||
cad68e55 MR |
8536 | /* |
8537 | * Don't try to pull misfit tasks we can't help. | |
8538 | * We can use max_capacity here as reduction in capacity on some | |
8539 | * CPUs in the group should either be possible to resolve | |
8540 | * internally or be covered by avg_load imbalance (eventually). | |
8541 | */ | |
8542 | if (sgs->group_type == group_misfit_task && | |
4aed8aa4 | 8543 | (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || |
0b0695f2 | 8544 | sds->local_stat.group_type != group_has_spare)) |
cad68e55 MR |
8545 | return false; |
8546 | ||
caeb178c | 8547 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
8548 | return true; |
8549 | ||
caeb178c RR |
8550 | if (sgs->group_type < busiest->group_type) |
8551 | return false; | |
8552 | ||
9e0994c0 | 8553 | /* |
0b0695f2 VG |
8554 | * The candidate and the current busiest group are the same type of |
8555 | * group. Let check which one is the busiest according to the type. | |
9e0994c0 | 8556 | */ |
9e0994c0 | 8557 | |
0b0695f2 VG |
8558 | switch (sgs->group_type) { |
8559 | case group_overloaded: | |
8560 | /* Select the overloaded group with highest avg_load. */ | |
8561 | if (sgs->avg_load <= busiest->avg_load) | |
8562 | return false; | |
8563 | break; | |
8564 | ||
8565 | case group_imbalanced: | |
8566 | /* | |
8567 | * Select the 1st imbalanced group as we don't have any way to | |
8568 | * choose one more than another. | |
8569 | */ | |
9e0994c0 MR |
8570 | return false; |
8571 | ||
0b0695f2 VG |
8572 | case group_asym_packing: |
8573 | /* Prefer to move from lowest priority CPU's work */ | |
8574 | if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) | |
8575 | return false; | |
8576 | break; | |
532cb4c4 | 8577 | |
0b0695f2 VG |
8578 | case group_misfit_task: |
8579 | /* | |
8580 | * If we have more than one misfit sg go with the biggest | |
8581 | * misfit. | |
8582 | */ | |
8583 | if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
8584 | return false; | |
8585 | break; | |
532cb4c4 | 8586 | |
0b0695f2 VG |
8587 | case group_fully_busy: |
8588 | /* | |
8589 | * Select the fully busy group with highest avg_load. In | |
8590 | * theory, there is no need to pull task from such kind of | |
8591 | * group because tasks have all compute capacity that they need | |
8592 | * but we can still improve the overall throughput by reducing | |
8593 | * contention when accessing shared HW resources. | |
8594 | * | |
8595 | * XXX for now avg_load is not computed and always 0 so we | |
8596 | * select the 1st one. | |
8597 | */ | |
8598 | if (sgs->avg_load <= busiest->avg_load) | |
8599 | return false; | |
8600 | break; | |
8601 | ||
8602 | case group_has_spare: | |
8603 | /* | |
5f68eb19 VG |
8604 | * Select not overloaded group with lowest number of idle cpus |
8605 | * and highest number of running tasks. We could also compare | |
8606 | * the spare capacity which is more stable but it can end up | |
8607 | * that the group has less spare capacity but finally more idle | |
0b0695f2 VG |
8608 | * CPUs which means less opportunity to pull tasks. |
8609 | */ | |
5f68eb19 | 8610 | if (sgs->idle_cpus > busiest->idle_cpus) |
0b0695f2 | 8611 | return false; |
5f68eb19 VG |
8612 | else if ((sgs->idle_cpus == busiest->idle_cpus) && |
8613 | (sgs->sum_nr_running <= busiest->sum_nr_running)) | |
8614 | return false; | |
8615 | ||
0b0695f2 | 8616 | break; |
532cb4c4 MN |
8617 | } |
8618 | ||
0b0695f2 VG |
8619 | /* |
8620 | * Candidate sg has no more than one task per CPU and has higher | |
8621 | * per-CPU capacity. Migrating tasks to less capable CPUs may harm | |
8622 | * throughput. Maximize throughput, power/energy consequences are not | |
8623 | * considered. | |
8624 | */ | |
8625 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && | |
8626 | (sgs->group_type <= group_fully_busy) && | |
4aed8aa4 | 8627 | (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) |
0b0695f2 VG |
8628 | return false; |
8629 | ||
8630 | return true; | |
532cb4c4 MN |
8631 | } |
8632 | ||
0ec8aa00 PZ |
8633 | #ifdef CONFIG_NUMA_BALANCING |
8634 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8635 | { | |
a3498347 | 8636 | if (sgs->sum_h_nr_running > sgs->nr_numa_running) |
0ec8aa00 | 8637 | return regular; |
a3498347 | 8638 | if (sgs->sum_h_nr_running > sgs->nr_preferred_running) |
0ec8aa00 PZ |
8639 | return remote; |
8640 | return all; | |
8641 | } | |
8642 | ||
8643 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8644 | { | |
8645 | if (rq->nr_running > rq->nr_numa_running) | |
8646 | return regular; | |
8647 | if (rq->nr_running > rq->nr_preferred_running) | |
8648 | return remote; | |
8649 | return all; | |
8650 | } | |
8651 | #else | |
8652 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8653 | { | |
8654 | return all; | |
8655 | } | |
8656 | ||
8657 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8658 | { | |
8659 | return regular; | |
8660 | } | |
8661 | #endif /* CONFIG_NUMA_BALANCING */ | |
8662 | ||
57abff06 VG |
8663 | |
8664 | struct sg_lb_stats; | |
8665 | ||
3318544b VG |
8666 | /* |
8667 | * task_running_on_cpu - return 1 if @p is running on @cpu. | |
8668 | */ | |
8669 | ||
8670 | static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) | |
8671 | { | |
8672 | /* Task has no contribution or is new */ | |
8673 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
8674 | return 0; | |
8675 | ||
8676 | if (task_on_rq_queued(p)) | |
8677 | return 1; | |
8678 | ||
8679 | return 0; | |
8680 | } | |
8681 | ||
8682 | /** | |
8683 | * idle_cpu_without - would a given CPU be idle without p ? | |
8684 | * @cpu: the processor on which idleness is tested. | |
8685 | * @p: task which should be ignored. | |
8686 | * | |
8687 | * Return: 1 if the CPU would be idle. 0 otherwise. | |
8688 | */ | |
8689 | static int idle_cpu_without(int cpu, struct task_struct *p) | |
8690 | { | |
8691 | struct rq *rq = cpu_rq(cpu); | |
8692 | ||
8693 | if (rq->curr != rq->idle && rq->curr != p) | |
8694 | return 0; | |
8695 | ||
8696 | /* | |
8697 | * rq->nr_running can't be used but an updated version without the | |
8698 | * impact of p on cpu must be used instead. The updated nr_running | |
8699 | * be computed and tested before calling idle_cpu_without(). | |
8700 | */ | |
8701 | ||
8702 | #ifdef CONFIG_SMP | |
126c2092 | 8703 | if (rq->ttwu_pending) |
3318544b VG |
8704 | return 0; |
8705 | #endif | |
8706 | ||
8707 | return 1; | |
8708 | } | |
8709 | ||
57abff06 VG |
8710 | /* |
8711 | * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. | |
3318544b | 8712 | * @sd: The sched_domain level to look for idlest group. |
57abff06 VG |
8713 | * @group: sched_group whose statistics are to be updated. |
8714 | * @sgs: variable to hold the statistics for this group. | |
3318544b | 8715 | * @p: The task for which we look for the idlest group/CPU. |
57abff06 VG |
8716 | */ |
8717 | static inline void update_sg_wakeup_stats(struct sched_domain *sd, | |
8718 | struct sched_group *group, | |
8719 | struct sg_lb_stats *sgs, | |
8720 | struct task_struct *p) | |
8721 | { | |
8722 | int i, nr_running; | |
8723 | ||
8724 | memset(sgs, 0, sizeof(*sgs)); | |
8725 | ||
8726 | for_each_cpu(i, sched_group_span(group)) { | |
8727 | struct rq *rq = cpu_rq(i); | |
3318544b | 8728 | unsigned int local; |
57abff06 | 8729 | |
3318544b | 8730 | sgs->group_load += cpu_load_without(rq, p); |
57abff06 | 8731 | sgs->group_util += cpu_util_without(i, p); |
070f5e86 | 8732 | sgs->group_runnable += cpu_runnable_without(rq, p); |
3318544b VG |
8733 | local = task_running_on_cpu(i, p); |
8734 | sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; | |
57abff06 | 8735 | |
3318544b | 8736 | nr_running = rq->nr_running - local; |
57abff06 VG |
8737 | sgs->sum_nr_running += nr_running; |
8738 | ||
8739 | /* | |
3318544b | 8740 | * No need to call idle_cpu_without() if nr_running is not 0 |
57abff06 | 8741 | */ |
3318544b | 8742 | if (!nr_running && idle_cpu_without(i, p)) |
57abff06 VG |
8743 | sgs->idle_cpus++; |
8744 | ||
57abff06 VG |
8745 | } |
8746 | ||
8747 | /* Check if task fits in the group */ | |
8748 | if (sd->flags & SD_ASYM_CPUCAPACITY && | |
8749 | !task_fits_capacity(p, group->sgc->max_capacity)) { | |
8750 | sgs->group_misfit_task_load = 1; | |
8751 | } | |
8752 | ||
8753 | sgs->group_capacity = group->sgc->capacity; | |
8754 | ||
289de359 VG |
8755 | sgs->group_weight = group->group_weight; |
8756 | ||
57abff06 VG |
8757 | sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); |
8758 | ||
8759 | /* | |
8760 | * Computing avg_load makes sense only when group is fully busy or | |
8761 | * overloaded | |
8762 | */ | |
6c8116c9 TZ |
8763 | if (sgs->group_type == group_fully_busy || |
8764 | sgs->group_type == group_overloaded) | |
57abff06 VG |
8765 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / |
8766 | sgs->group_capacity; | |
8767 | } | |
8768 | ||
8769 | static bool update_pick_idlest(struct sched_group *idlest, | |
8770 | struct sg_lb_stats *idlest_sgs, | |
8771 | struct sched_group *group, | |
8772 | struct sg_lb_stats *sgs) | |
8773 | { | |
8774 | if (sgs->group_type < idlest_sgs->group_type) | |
8775 | return true; | |
8776 | ||
8777 | if (sgs->group_type > idlest_sgs->group_type) | |
8778 | return false; | |
8779 | ||
8780 | /* | |
8781 | * The candidate and the current idlest group are the same type of | |
8782 | * group. Let check which one is the idlest according to the type. | |
8783 | */ | |
8784 | ||
8785 | switch (sgs->group_type) { | |
8786 | case group_overloaded: | |
8787 | case group_fully_busy: | |
8788 | /* Select the group with lowest avg_load. */ | |
8789 | if (idlest_sgs->avg_load <= sgs->avg_load) | |
8790 | return false; | |
8791 | break; | |
8792 | ||
8793 | case group_imbalanced: | |
8794 | case group_asym_packing: | |
8795 | /* Those types are not used in the slow wakeup path */ | |
8796 | return false; | |
8797 | ||
8798 | case group_misfit_task: | |
8799 | /* Select group with the highest max capacity */ | |
8800 | if (idlest->sgc->max_capacity >= group->sgc->max_capacity) | |
8801 | return false; | |
8802 | break; | |
8803 | ||
8804 | case group_has_spare: | |
8805 | /* Select group with most idle CPUs */ | |
3edecfef | 8806 | if (idlest_sgs->idle_cpus > sgs->idle_cpus) |
57abff06 | 8807 | return false; |
3edecfef PP |
8808 | |
8809 | /* Select group with lowest group_util */ | |
8810 | if (idlest_sgs->idle_cpus == sgs->idle_cpus && | |
8811 | idlest_sgs->group_util <= sgs->group_util) | |
8812 | return false; | |
8813 | ||
57abff06 VG |
8814 | break; |
8815 | } | |
8816 | ||
8817 | return true; | |
8818 | } | |
8819 | ||
23e6082a MG |
8820 | /* |
8821 | * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain. | |
8822 | * This is an approximation as the number of running tasks may not be | |
8823 | * related to the number of busy CPUs due to sched_setaffinity. | |
8824 | */ | |
8825 | static inline bool allow_numa_imbalance(int dst_running, int dst_weight) | |
8826 | { | |
8827 | return (dst_running < (dst_weight >> 2)); | |
8828 | } | |
8829 | ||
57abff06 VG |
8830 | /* |
8831 | * find_idlest_group() finds and returns the least busy CPU group within the | |
8832 | * domain. | |
8833 | * | |
8834 | * Assumes p is allowed on at least one CPU in sd. | |
8835 | */ | |
8836 | static struct sched_group * | |
45da2773 | 8837 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
57abff06 VG |
8838 | { |
8839 | struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; | |
8840 | struct sg_lb_stats local_sgs, tmp_sgs; | |
8841 | struct sg_lb_stats *sgs; | |
8842 | unsigned long imbalance; | |
8843 | struct sg_lb_stats idlest_sgs = { | |
8844 | .avg_load = UINT_MAX, | |
8845 | .group_type = group_overloaded, | |
8846 | }; | |
8847 | ||
57abff06 VG |
8848 | do { |
8849 | int local_group; | |
8850 | ||
8851 | /* Skip over this group if it has no CPUs allowed */ | |
8852 | if (!cpumask_intersects(sched_group_span(group), | |
8853 | p->cpus_ptr)) | |
8854 | continue; | |
8855 | ||
8856 | local_group = cpumask_test_cpu(this_cpu, | |
8857 | sched_group_span(group)); | |
8858 | ||
8859 | if (local_group) { | |
8860 | sgs = &local_sgs; | |
8861 | local = group; | |
8862 | } else { | |
8863 | sgs = &tmp_sgs; | |
8864 | } | |
8865 | ||
8866 | update_sg_wakeup_stats(sd, group, sgs, p); | |
8867 | ||
8868 | if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { | |
8869 | idlest = group; | |
8870 | idlest_sgs = *sgs; | |
8871 | } | |
8872 | ||
8873 | } while (group = group->next, group != sd->groups); | |
8874 | ||
8875 | ||
8876 | /* There is no idlest group to push tasks to */ | |
8877 | if (!idlest) | |
8878 | return NULL; | |
8879 | ||
7ed735c3 VG |
8880 | /* The local group has been skipped because of CPU affinity */ |
8881 | if (!local) | |
8882 | return idlest; | |
8883 | ||
57abff06 VG |
8884 | /* |
8885 | * If the local group is idler than the selected idlest group | |
8886 | * don't try and push the task. | |
8887 | */ | |
8888 | if (local_sgs.group_type < idlest_sgs.group_type) | |
8889 | return NULL; | |
8890 | ||
8891 | /* | |
8892 | * If the local group is busier than the selected idlest group | |
8893 | * try and push the task. | |
8894 | */ | |
8895 | if (local_sgs.group_type > idlest_sgs.group_type) | |
8896 | return idlest; | |
8897 | ||
8898 | switch (local_sgs.group_type) { | |
8899 | case group_overloaded: | |
8900 | case group_fully_busy: | |
5c339005 MG |
8901 | |
8902 | /* Calculate allowed imbalance based on load */ | |
8903 | imbalance = scale_load_down(NICE_0_LOAD) * | |
8904 | (sd->imbalance_pct-100) / 100; | |
8905 | ||
57abff06 VG |
8906 | /* |
8907 | * When comparing groups across NUMA domains, it's possible for | |
8908 | * the local domain to be very lightly loaded relative to the | |
8909 | * remote domains but "imbalance" skews the comparison making | |
8910 | * remote CPUs look much more favourable. When considering | |
8911 | * cross-domain, add imbalance to the load on the remote node | |
8912 | * and consider staying local. | |
8913 | */ | |
8914 | ||
8915 | if ((sd->flags & SD_NUMA) && | |
8916 | ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) | |
8917 | return NULL; | |
8918 | ||
8919 | /* | |
8920 | * If the local group is less loaded than the selected | |
8921 | * idlest group don't try and push any tasks. | |
8922 | */ | |
8923 | if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) | |
8924 | return NULL; | |
8925 | ||
8926 | if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) | |
8927 | return NULL; | |
8928 | break; | |
8929 | ||
8930 | case group_imbalanced: | |
8931 | case group_asym_packing: | |
8932 | /* Those type are not used in the slow wakeup path */ | |
8933 | return NULL; | |
8934 | ||
8935 | case group_misfit_task: | |
8936 | /* Select group with the highest max capacity */ | |
8937 | if (local->sgc->max_capacity >= idlest->sgc->max_capacity) | |
8938 | return NULL; | |
8939 | break; | |
8940 | ||
8941 | case group_has_spare: | |
8942 | if (sd->flags & SD_NUMA) { | |
8943 | #ifdef CONFIG_NUMA_BALANCING | |
8944 | int idlest_cpu; | |
8945 | /* | |
8946 | * If there is spare capacity at NUMA, try to select | |
8947 | * the preferred node | |
8948 | */ | |
8949 | if (cpu_to_node(this_cpu) == p->numa_preferred_nid) | |
8950 | return NULL; | |
8951 | ||
8952 | idlest_cpu = cpumask_first(sched_group_span(idlest)); | |
8953 | if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) | |
8954 | return idlest; | |
8955 | #endif | |
8956 | /* | |
8957 | * Otherwise, keep the task on this node to stay close | |
8958 | * its wakeup source and improve locality. If there is | |
8959 | * a real need of migration, periodic load balance will | |
8960 | * take care of it. | |
8961 | */ | |
23e6082a | 8962 | if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight)) |
57abff06 VG |
8963 | return NULL; |
8964 | } | |
8965 | ||
8966 | /* | |
8967 | * Select group with highest number of idle CPUs. We could also | |
8968 | * compare the utilization which is more stable but it can end | |
8969 | * up that the group has less spare capacity but finally more | |
8970 | * idle CPUs which means more opportunity to run task. | |
8971 | */ | |
8972 | if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) | |
8973 | return NULL; | |
8974 | break; | |
8975 | } | |
8976 | ||
8977 | return idlest; | |
8978 | } | |
8979 | ||
1e3c88bd | 8980 | /** |
461819ac | 8981 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 8982 | * @env: The load balancing environment. |
1e3c88bd PZ |
8983 | * @sds: variable to hold the statistics for this sched_domain. |
8984 | */ | |
0b0695f2 | 8985 | |
0ec8aa00 | 8986 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8987 | { |
bd939f45 PZ |
8988 | struct sched_domain *child = env->sd->child; |
8989 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 8990 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 8991 | struct sg_lb_stats tmp_sgs; |
630246a0 | 8992 | int sg_status = 0; |
1e3c88bd | 8993 | |
1e3c88bd | 8994 | do { |
56cf515b | 8995 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
8996 | int local_group; |
8997 | ||
ae4df9d6 | 8998 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
8999 | if (local_group) { |
9000 | sds->local = sg; | |
05b40e05 | 9001 | sgs = local; |
b72ff13c PZ |
9002 | |
9003 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
9004 | time_after_eq(jiffies, sg->sgc->next_update)) |
9005 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 9006 | } |
1e3c88bd | 9007 | |
630246a0 | 9008 | update_sg_lb_stats(env, sg, sgs, &sg_status); |
1e3c88bd | 9009 | |
b72ff13c PZ |
9010 | if (local_group) |
9011 | goto next_group; | |
9012 | ||
1e3c88bd | 9013 | |
b72ff13c | 9014 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 9015 | sds->busiest = sg; |
56cf515b | 9016 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
9017 | } |
9018 | ||
b72ff13c PZ |
9019 | next_group: |
9020 | /* Now, start updating sd_lb_stats */ | |
9021 | sds->total_load += sgs->group_load; | |
63b2ca30 | 9022 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 9023 | |
532cb4c4 | 9024 | sg = sg->next; |
bd939f45 | 9025 | } while (sg != env->sd->groups); |
0ec8aa00 | 9026 | |
0b0695f2 VG |
9027 | /* Tag domain that child domain prefers tasks go to siblings first */ |
9028 | sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; | |
9029 | ||
f643ea22 | 9030 | |
0ec8aa00 PZ |
9031 | if (env->sd->flags & SD_NUMA) |
9032 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
9033 | |
9034 | if (!env->sd->parent) { | |
2802bf3c MR |
9035 | struct root_domain *rd = env->dst_rq->rd; |
9036 | ||
4486edd1 | 9037 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
9038 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
9039 | ||
9040 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
9041 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
f9f240f9 | 9042 | trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); |
2802bf3c | 9043 | } else if (sg_status & SG_OVERUTILIZED) { |
f9f240f9 QY |
9044 | struct root_domain *rd = env->dst_rq->rd; |
9045 | ||
9046 | WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); | |
9047 | trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); | |
4486edd1 | 9048 | } |
532cb4c4 MN |
9049 | } |
9050 | ||
abeae76a MG |
9051 | #define NUMA_IMBALANCE_MIN 2 |
9052 | ||
7d2b5dd0 MG |
9053 | static inline long adjust_numa_imbalance(int imbalance, |
9054 | int dst_running, int dst_weight) | |
fb86f5b2 | 9055 | { |
23e6082a MG |
9056 | if (!allow_numa_imbalance(dst_running, dst_weight)) |
9057 | return imbalance; | |
9058 | ||
fb86f5b2 MG |
9059 | /* |
9060 | * Allow a small imbalance based on a simple pair of communicating | |
7d2b5dd0 | 9061 | * tasks that remain local when the destination is lightly loaded. |
fb86f5b2 | 9062 | */ |
23e6082a | 9063 | if (imbalance <= NUMA_IMBALANCE_MIN) |
fb86f5b2 MG |
9064 | return 0; |
9065 | ||
9066 | return imbalance; | |
9067 | } | |
9068 | ||
1e3c88bd PZ |
9069 | /** |
9070 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
9071 | * groups of a given sched_domain during load balance. | |
bd939f45 | 9072 | * @env: load balance environment |
1e3c88bd | 9073 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 9074 | */ |
bd939f45 | 9075 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9076 | { |
56cf515b JK |
9077 | struct sg_lb_stats *local, *busiest; |
9078 | ||
9079 | local = &sds->local_stat; | |
56cf515b | 9080 | busiest = &sds->busiest_stat; |
dd5feea1 | 9081 | |
0b0695f2 VG |
9082 | if (busiest->group_type == group_misfit_task) { |
9083 | /* Set imbalance to allow misfit tasks to be balanced. */ | |
9084 | env->migration_type = migrate_misfit; | |
c63be7be | 9085 | env->imbalance = 1; |
0b0695f2 VG |
9086 | return; |
9087 | } | |
9088 | ||
9089 | if (busiest->group_type == group_asym_packing) { | |
9090 | /* | |
9091 | * In case of asym capacity, we will try to migrate all load to | |
9092 | * the preferred CPU. | |
9093 | */ | |
9094 | env->migration_type = migrate_task; | |
9095 | env->imbalance = busiest->sum_h_nr_running; | |
9096 | return; | |
9097 | } | |
9098 | ||
9099 | if (busiest->group_type == group_imbalanced) { | |
9100 | /* | |
9101 | * In the group_imb case we cannot rely on group-wide averages | |
9102 | * to ensure CPU-load equilibrium, try to move any task to fix | |
9103 | * the imbalance. The next load balance will take care of | |
9104 | * balancing back the system. | |
9105 | */ | |
9106 | env->migration_type = migrate_task; | |
9107 | env->imbalance = 1; | |
490ba971 VG |
9108 | return; |
9109 | } | |
9110 | ||
1e3c88bd | 9111 | /* |
0b0695f2 | 9112 | * Try to use spare capacity of local group without overloading it or |
a9723389 | 9113 | * emptying busiest. |
1e3c88bd | 9114 | */ |
0b0695f2 | 9115 | if (local->group_type == group_has_spare) { |
16b0a7a1 VG |
9116 | if ((busiest->group_type > group_fully_busy) && |
9117 | !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { | |
0b0695f2 VG |
9118 | /* |
9119 | * If busiest is overloaded, try to fill spare | |
9120 | * capacity. This might end up creating spare capacity | |
9121 | * in busiest or busiest still being overloaded but | |
9122 | * there is no simple way to directly compute the | |
9123 | * amount of load to migrate in order to balance the | |
9124 | * system. | |
9125 | */ | |
9126 | env->migration_type = migrate_util; | |
9127 | env->imbalance = max(local->group_capacity, local->group_util) - | |
9128 | local->group_util; | |
9129 | ||
9130 | /* | |
9131 | * In some cases, the group's utilization is max or even | |
9132 | * higher than capacity because of migrations but the | |
9133 | * local CPU is (newly) idle. There is at least one | |
9134 | * waiting task in this overloaded busiest group. Let's | |
9135 | * try to pull it. | |
9136 | */ | |
9137 | if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { | |
9138 | env->migration_type = migrate_task; | |
9139 | env->imbalance = 1; | |
9140 | } | |
9141 | ||
9142 | return; | |
9143 | } | |
9144 | ||
9145 | if (busiest->group_weight == 1 || sds->prefer_sibling) { | |
5e23e474 | 9146 | unsigned int nr_diff = busiest->sum_nr_running; |
0b0695f2 VG |
9147 | /* |
9148 | * When prefer sibling, evenly spread running tasks on | |
9149 | * groups. | |
9150 | */ | |
9151 | env->migration_type = migrate_task; | |
5e23e474 | 9152 | lsub_positive(&nr_diff, local->sum_nr_running); |
0b0695f2 | 9153 | env->imbalance = nr_diff >> 1; |
b396f523 | 9154 | } else { |
0b0695f2 | 9155 | |
b396f523 MG |
9156 | /* |
9157 | * If there is no overload, we just want to even the number of | |
9158 | * idle cpus. | |
9159 | */ | |
9160 | env->migration_type = migrate_task; | |
9161 | env->imbalance = max_t(long, 0, (local->idle_cpus - | |
0b0695f2 | 9162 | busiest->idle_cpus) >> 1); |
b396f523 MG |
9163 | } |
9164 | ||
9165 | /* Consider allowing a small imbalance between NUMA groups */ | |
7d2b5dd0 | 9166 | if (env->sd->flags & SD_NUMA) { |
fb86f5b2 | 9167 | env->imbalance = adjust_numa_imbalance(env->imbalance, |
7d2b5dd0 MG |
9168 | busiest->sum_nr_running, busiest->group_weight); |
9169 | } | |
b396f523 | 9170 | |
fcf0553d | 9171 | return; |
1e3c88bd PZ |
9172 | } |
9173 | ||
9a5d9ba6 | 9174 | /* |
0b0695f2 VG |
9175 | * Local is fully busy but has to take more load to relieve the |
9176 | * busiest group | |
9a5d9ba6 | 9177 | */ |
0b0695f2 VG |
9178 | if (local->group_type < group_overloaded) { |
9179 | /* | |
9180 | * Local will become overloaded so the avg_load metrics are | |
9181 | * finally needed. | |
9182 | */ | |
9183 | ||
9184 | local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / | |
9185 | local->group_capacity; | |
9186 | ||
9187 | sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / | |
9188 | sds->total_capacity; | |
111688ca AL |
9189 | /* |
9190 | * If the local group is more loaded than the selected | |
9191 | * busiest group don't try to pull any tasks. | |
9192 | */ | |
9193 | if (local->avg_load >= busiest->avg_load) { | |
9194 | env->imbalance = 0; | |
9195 | return; | |
9196 | } | |
dd5feea1 SS |
9197 | } |
9198 | ||
9199 | /* | |
0b0695f2 VG |
9200 | * Both group are or will become overloaded and we're trying to get all |
9201 | * the CPUs to the average_load, so we don't want to push ourselves | |
9202 | * above the average load, nor do we wish to reduce the max loaded CPU | |
9203 | * below the average load. At the same time, we also don't want to | |
9204 | * reduce the group load below the group capacity. Thus we look for | |
9205 | * the minimum possible imbalance. | |
dd5feea1 | 9206 | */ |
0b0695f2 | 9207 | env->migration_type = migrate_load; |
56cf515b | 9208 | env->imbalance = min( |
0b0695f2 | 9209 | (busiest->avg_load - sds->avg_load) * busiest->group_capacity, |
63b2ca30 | 9210 | (sds->avg_load - local->avg_load) * local->group_capacity |
ca8ce3d0 | 9211 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 9212 | } |
fab47622 | 9213 | |
1e3c88bd PZ |
9214 | /******* find_busiest_group() helpers end here *********************/ |
9215 | ||
0b0695f2 VG |
9216 | /* |
9217 | * Decision matrix according to the local and busiest group type: | |
9218 | * | |
9219 | * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded | |
9220 | * has_spare nr_idle balanced N/A N/A balanced balanced | |
9221 | * fully_busy nr_idle nr_idle N/A N/A balanced balanced | |
9222 | * misfit_task force N/A N/A N/A force force | |
9223 | * asym_packing force force N/A N/A force force | |
9224 | * imbalanced force force N/A N/A force force | |
9225 | * overloaded force force N/A N/A force avg_load | |
9226 | * | |
9227 | * N/A : Not Applicable because already filtered while updating | |
9228 | * statistics. | |
9229 | * balanced : The system is balanced for these 2 groups. | |
9230 | * force : Calculate the imbalance as load migration is probably needed. | |
9231 | * avg_load : Only if imbalance is significant enough. | |
9232 | * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite | |
9233 | * different in groups. | |
9234 | */ | |
9235 | ||
1e3c88bd PZ |
9236 | /** |
9237 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 9238 | * if there is an imbalance. |
1e3c88bd | 9239 | * |
a3df0679 | 9240 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
9241 | * to restore balance. |
9242 | * | |
cd96891d | 9243 | * @env: The load balancing environment. |
1e3c88bd | 9244 | * |
e69f6186 | 9245 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 9246 | */ |
56cf515b | 9247 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 9248 | { |
56cf515b | 9249 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
9250 | struct sd_lb_stats sds; |
9251 | ||
147c5fc2 | 9252 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
9253 | |
9254 | /* | |
b0fb1eb4 | 9255 | * Compute the various statistics relevant for load balancing at |
1e3c88bd PZ |
9256 | * this level. |
9257 | */ | |
23f0d209 | 9258 | update_sd_lb_stats(env, &sds); |
2802bf3c | 9259 | |
f8a696f2 | 9260 | if (sched_energy_enabled()) { |
2802bf3c MR |
9261 | struct root_domain *rd = env->dst_rq->rd; |
9262 | ||
9263 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
9264 | goto out_balanced; | |
9265 | } | |
9266 | ||
56cf515b JK |
9267 | local = &sds.local_stat; |
9268 | busiest = &sds.busiest_stat; | |
1e3c88bd | 9269 | |
cc57aa8f | 9270 | /* There is no busy sibling group to pull tasks from */ |
0b0695f2 | 9271 | if (!sds.busiest) |
1e3c88bd PZ |
9272 | goto out_balanced; |
9273 | ||
0b0695f2 VG |
9274 | /* Misfit tasks should be dealt with regardless of the avg load */ |
9275 | if (busiest->group_type == group_misfit_task) | |
9276 | goto force_balance; | |
9277 | ||
9278 | /* ASYM feature bypasses nice load balance check */ | |
9279 | if (busiest->group_type == group_asym_packing) | |
9280 | goto force_balance; | |
b0432d8f | 9281 | |
866ab43e PZ |
9282 | /* |
9283 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 9284 | * work because they assume all things are equal, which typically |
3bd37062 | 9285 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 9286 | */ |
caeb178c | 9287 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
9288 | goto force_balance; |
9289 | ||
cc57aa8f | 9290 | /* |
9c58c79a | 9291 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
9292 | * don't try and pull any tasks. |
9293 | */ | |
0b0695f2 | 9294 | if (local->group_type > busiest->group_type) |
1e3c88bd PZ |
9295 | goto out_balanced; |
9296 | ||
cc57aa8f | 9297 | /* |
0b0695f2 VG |
9298 | * When groups are overloaded, use the avg_load to ensure fairness |
9299 | * between tasks. | |
cc57aa8f | 9300 | */ |
0b0695f2 VG |
9301 | if (local->group_type == group_overloaded) { |
9302 | /* | |
9303 | * If the local group is more loaded than the selected | |
9304 | * busiest group don't try to pull any tasks. | |
9305 | */ | |
9306 | if (local->avg_load >= busiest->avg_load) | |
9307 | goto out_balanced; | |
9308 | ||
9309 | /* XXX broken for overlapping NUMA groups */ | |
9310 | sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / | |
9311 | sds.total_capacity; | |
1e3c88bd | 9312 | |
aae6d3dd | 9313 | /* |
0b0695f2 VG |
9314 | * Don't pull any tasks if this group is already above the |
9315 | * domain average load. | |
aae6d3dd | 9316 | */ |
0b0695f2 | 9317 | if (local->avg_load >= sds.avg_load) |
aae6d3dd | 9318 | goto out_balanced; |
0b0695f2 | 9319 | |
c186fafe | 9320 | /* |
0b0695f2 VG |
9321 | * If the busiest group is more loaded, use imbalance_pct to be |
9322 | * conservative. | |
c186fafe | 9323 | */ |
56cf515b JK |
9324 | if (100 * busiest->avg_load <= |
9325 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 9326 | goto out_balanced; |
aae6d3dd | 9327 | } |
1e3c88bd | 9328 | |
0b0695f2 VG |
9329 | /* Try to move all excess tasks to child's sibling domain */ |
9330 | if (sds.prefer_sibling && local->group_type == group_has_spare && | |
5e23e474 | 9331 | busiest->sum_nr_running > local->sum_nr_running + 1) |
0b0695f2 VG |
9332 | goto force_balance; |
9333 | ||
2ab4092f VG |
9334 | if (busiest->group_type != group_overloaded) { |
9335 | if (env->idle == CPU_NOT_IDLE) | |
9336 | /* | |
9337 | * If the busiest group is not overloaded (and as a | |
9338 | * result the local one too) but this CPU is already | |
9339 | * busy, let another idle CPU try to pull task. | |
9340 | */ | |
9341 | goto out_balanced; | |
9342 | ||
9343 | if (busiest->group_weight > 1 && | |
9344 | local->idle_cpus <= (busiest->idle_cpus + 1)) | |
9345 | /* | |
9346 | * If the busiest group is not overloaded | |
9347 | * and there is no imbalance between this and busiest | |
9348 | * group wrt idle CPUs, it is balanced. The imbalance | |
9349 | * becomes significant if the diff is greater than 1 | |
9350 | * otherwise we might end up to just move the imbalance | |
9351 | * on another group. Of course this applies only if | |
9352 | * there is more than 1 CPU per group. | |
9353 | */ | |
9354 | goto out_balanced; | |
9355 | ||
9356 | if (busiest->sum_h_nr_running == 1) | |
9357 | /* | |
9358 | * busiest doesn't have any tasks waiting to run | |
9359 | */ | |
9360 | goto out_balanced; | |
9361 | } | |
0b0695f2 | 9362 | |
fab47622 | 9363 | force_balance: |
1e3c88bd | 9364 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 9365 | calculate_imbalance(env, &sds); |
bb3485c8 | 9366 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
9367 | |
9368 | out_balanced: | |
bd939f45 | 9369 | env->imbalance = 0; |
1e3c88bd PZ |
9370 | return NULL; |
9371 | } | |
9372 | ||
9373 | /* | |
97fb7a0a | 9374 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 9375 | */ |
bd939f45 | 9376 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 9377 | struct sched_group *group) |
1e3c88bd PZ |
9378 | { |
9379 | struct rq *busiest = NULL, *rq; | |
0b0695f2 VG |
9380 | unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; |
9381 | unsigned int busiest_nr = 0; | |
1e3c88bd PZ |
9382 | int i; |
9383 | ||
ae4df9d6 | 9384 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
0b0695f2 VG |
9385 | unsigned long capacity, load, util; |
9386 | unsigned int nr_running; | |
0ec8aa00 PZ |
9387 | enum fbq_type rt; |
9388 | ||
9389 | rq = cpu_rq(i); | |
9390 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 9391 | |
0ec8aa00 PZ |
9392 | /* |
9393 | * We classify groups/runqueues into three groups: | |
9394 | * - regular: there are !numa tasks | |
9395 | * - remote: there are numa tasks that run on the 'wrong' node | |
9396 | * - all: there is no distinction | |
9397 | * | |
9398 | * In order to avoid migrating ideally placed numa tasks, | |
9399 | * ignore those when there's better options. | |
9400 | * | |
9401 | * If we ignore the actual busiest queue to migrate another | |
9402 | * task, the next balance pass can still reduce the busiest | |
9403 | * queue by moving tasks around inside the node. | |
9404 | * | |
9405 | * If we cannot move enough load due to this classification | |
9406 | * the next pass will adjust the group classification and | |
9407 | * allow migration of more tasks. | |
9408 | * | |
9409 | * Both cases only affect the total convergence complexity. | |
9410 | */ | |
9411 | if (rt > env->fbq_type) | |
9412 | continue; | |
9413 | ||
0b0695f2 | 9414 | nr_running = rq->cfs.h_nr_running; |
fc488ffd VG |
9415 | if (!nr_running) |
9416 | continue; | |
9417 | ||
9418 | capacity = capacity_of(i); | |
9d5efe05 | 9419 | |
4ad3831a CR |
9420 | /* |
9421 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
9422 | * eventually lead to active_balancing high->low capacity. | |
9423 | * Higher per-CPU capacity is considered better than balancing | |
9424 | * average load. | |
9425 | */ | |
9426 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
4aed8aa4 | 9427 | !capacity_greater(capacity_of(env->dst_cpu), capacity) && |
0b0695f2 | 9428 | nr_running == 1) |
4ad3831a CR |
9429 | continue; |
9430 | ||
0b0695f2 VG |
9431 | switch (env->migration_type) { |
9432 | case migrate_load: | |
9433 | /* | |
b0fb1eb4 VG |
9434 | * When comparing with load imbalance, use cpu_load() |
9435 | * which is not scaled with the CPU capacity. | |
0b0695f2 | 9436 | */ |
b0fb1eb4 | 9437 | load = cpu_load(rq); |
1e3c88bd | 9438 | |
0b0695f2 VG |
9439 | if (nr_running == 1 && load > env->imbalance && |
9440 | !check_cpu_capacity(rq, env->sd)) | |
9441 | break; | |
ea67821b | 9442 | |
0b0695f2 VG |
9443 | /* |
9444 | * For the load comparisons with the other CPUs, | |
b0fb1eb4 VG |
9445 | * consider the cpu_load() scaled with the CPU |
9446 | * capacity, so that the load can be moved away | |
9447 | * from the CPU that is potentially running at a | |
9448 | * lower capacity. | |
0b0695f2 VG |
9449 | * |
9450 | * Thus we're looking for max(load_i / capacity_i), | |
9451 | * crosswise multiplication to rid ourselves of the | |
9452 | * division works out to: | |
9453 | * load_i * capacity_j > load_j * capacity_i; | |
9454 | * where j is our previous maximum. | |
9455 | */ | |
9456 | if (load * busiest_capacity > busiest_load * capacity) { | |
9457 | busiest_load = load; | |
9458 | busiest_capacity = capacity; | |
9459 | busiest = rq; | |
9460 | } | |
9461 | break; | |
9462 | ||
9463 | case migrate_util: | |
9464 | util = cpu_util(cpu_of(rq)); | |
9465 | ||
c32b4308 VG |
9466 | /* |
9467 | * Don't try to pull utilization from a CPU with one | |
9468 | * running task. Whatever its utilization, we will fail | |
9469 | * detach the task. | |
9470 | */ | |
9471 | if (nr_running <= 1) | |
9472 | continue; | |
9473 | ||
0b0695f2 VG |
9474 | if (busiest_util < util) { |
9475 | busiest_util = util; | |
9476 | busiest = rq; | |
9477 | } | |
9478 | break; | |
9479 | ||
9480 | case migrate_task: | |
9481 | if (busiest_nr < nr_running) { | |
9482 | busiest_nr = nr_running; | |
9483 | busiest = rq; | |
9484 | } | |
9485 | break; | |
9486 | ||
9487 | case migrate_misfit: | |
9488 | /* | |
9489 | * For ASYM_CPUCAPACITY domains with misfit tasks we | |
9490 | * simply seek the "biggest" misfit task. | |
9491 | */ | |
9492 | if (rq->misfit_task_load > busiest_load) { | |
9493 | busiest_load = rq->misfit_task_load; | |
9494 | busiest = rq; | |
9495 | } | |
9496 | ||
9497 | break; | |
1e3c88bd | 9498 | |
1e3c88bd PZ |
9499 | } |
9500 | } | |
9501 | ||
9502 | return busiest; | |
9503 | } | |
9504 | ||
9505 | /* | |
9506 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
9507 | * so long as it is large enough. | |
9508 | */ | |
9509 | #define MAX_PINNED_INTERVAL 512 | |
9510 | ||
46a745d9 VG |
9511 | static inline bool |
9512 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 9513 | { |
46a745d9 VG |
9514 | /* |
9515 | * ASYM_PACKING needs to force migrate tasks from busy but | |
9516 | * lower priority CPUs in order to pack all tasks in the | |
9517 | * highest priority CPUs. | |
9518 | */ | |
9519 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
9520 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
9521 | } | |
bd939f45 | 9522 | |
46a745d9 | 9523 | static inline bool |
e9b9734b VG |
9524 | imbalanced_active_balance(struct lb_env *env) |
9525 | { | |
9526 | struct sched_domain *sd = env->sd; | |
9527 | ||
9528 | /* | |
9529 | * The imbalanced case includes the case of pinned tasks preventing a fair | |
9530 | * distribution of the load on the system but also the even distribution of the | |
9531 | * threads on a system with spare capacity | |
9532 | */ | |
9533 | if ((env->migration_type == migrate_task) && | |
9534 | (sd->nr_balance_failed > sd->cache_nice_tries+2)) | |
9535 | return 1; | |
9536 | ||
9537 | return 0; | |
9538 | } | |
9539 | ||
9540 | static int need_active_balance(struct lb_env *env) | |
46a745d9 VG |
9541 | { |
9542 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 9543 | |
46a745d9 VG |
9544 | if (asym_active_balance(env)) |
9545 | return 1; | |
1af3ed3d | 9546 | |
e9b9734b VG |
9547 | if (imbalanced_active_balance(env)) |
9548 | return 1; | |
9549 | ||
1aaf90a4 VG |
9550 | /* |
9551 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
9552 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
9553 | * because of other sched_class or IRQs if more capacity stays | |
9554 | * available on dst_cpu. | |
9555 | */ | |
9556 | if ((env->idle != CPU_NOT_IDLE) && | |
9557 | (env->src_rq->cfs.h_nr_running == 1)) { | |
9558 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
9559 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
9560 | return 1; | |
9561 | } | |
9562 | ||
0b0695f2 | 9563 | if (env->migration_type == migrate_misfit) |
cad68e55 MR |
9564 | return 1; |
9565 | ||
46a745d9 VG |
9566 | return 0; |
9567 | } | |
9568 | ||
969c7921 TH |
9569 | static int active_load_balance_cpu_stop(void *data); |
9570 | ||
23f0d209 JK |
9571 | static int should_we_balance(struct lb_env *env) |
9572 | { | |
9573 | struct sched_group *sg = env->sd->groups; | |
64297f2b | 9574 | int cpu; |
23f0d209 | 9575 | |
024c9d2f PZ |
9576 | /* |
9577 | * Ensure the balancing environment is consistent; can happen | |
9578 | * when the softirq triggers 'during' hotplug. | |
9579 | */ | |
9580 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
9581 | return 0; | |
9582 | ||
23f0d209 | 9583 | /* |
97fb7a0a | 9584 | * In the newly idle case, we will allow all the CPUs |
23f0d209 JK |
9585 | * to do the newly idle load balance. |
9586 | */ | |
9587 | if (env->idle == CPU_NEWLY_IDLE) | |
9588 | return 1; | |
9589 | ||
97fb7a0a | 9590 | /* Try to find first idle CPU */ |
e5c14b1f | 9591 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 9592 | if (!idle_cpu(cpu)) |
23f0d209 JK |
9593 | continue; |
9594 | ||
64297f2b PW |
9595 | /* Are we the first idle CPU? */ |
9596 | return cpu == env->dst_cpu; | |
23f0d209 JK |
9597 | } |
9598 | ||
64297f2b PW |
9599 | /* Are we the first CPU of this group ? */ |
9600 | return group_balance_cpu(sg) == env->dst_cpu; | |
23f0d209 JK |
9601 | } |
9602 | ||
1e3c88bd PZ |
9603 | /* |
9604 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
9605 | * tasks if there is an imbalance. | |
9606 | */ | |
9607 | static int load_balance(int this_cpu, struct rq *this_rq, | |
9608 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 9609 | int *continue_balancing) |
1e3c88bd | 9610 | { |
88b8dac0 | 9611 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 9612 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 9613 | struct sched_group *group; |
1e3c88bd | 9614 | struct rq *busiest; |
8a8c69c3 | 9615 | struct rq_flags rf; |
4ba29684 | 9616 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 9617 | |
8e45cb54 PZ |
9618 | struct lb_env env = { |
9619 | .sd = sd, | |
ddcdf6e7 PZ |
9620 | .dst_cpu = this_cpu, |
9621 | .dst_rq = this_rq, | |
ae4df9d6 | 9622 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 9623 | .idle = idle, |
eb95308e | 9624 | .loop_break = sched_nr_migrate_break, |
b9403130 | 9625 | .cpus = cpus, |
0ec8aa00 | 9626 | .fbq_type = all, |
163122b7 | 9627 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
9628 | }; |
9629 | ||
65a4433a | 9630 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 9631 | |
ae92882e | 9632 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
9633 | |
9634 | redo: | |
23f0d209 JK |
9635 | if (!should_we_balance(&env)) { |
9636 | *continue_balancing = 0; | |
1e3c88bd | 9637 | goto out_balanced; |
23f0d209 | 9638 | } |
1e3c88bd | 9639 | |
23f0d209 | 9640 | group = find_busiest_group(&env); |
1e3c88bd | 9641 | if (!group) { |
ae92882e | 9642 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
9643 | goto out_balanced; |
9644 | } | |
9645 | ||
b9403130 | 9646 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 9647 | if (!busiest) { |
ae92882e | 9648 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
9649 | goto out_balanced; |
9650 | } | |
9651 | ||
78feefc5 | 9652 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 9653 | |
ae92882e | 9654 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 9655 | |
1aaf90a4 VG |
9656 | env.src_cpu = busiest->cpu; |
9657 | env.src_rq = busiest; | |
9658 | ||
1e3c88bd | 9659 | ld_moved = 0; |
8a41dfcd VG |
9660 | /* Clear this flag as soon as we find a pullable task */ |
9661 | env.flags |= LBF_ALL_PINNED; | |
1e3c88bd PZ |
9662 | if (busiest->nr_running > 1) { |
9663 | /* | |
9664 | * Attempt to move tasks. If find_busiest_group has found | |
9665 | * an imbalance but busiest->nr_running <= 1, the group is | |
9666 | * still unbalanced. ld_moved simply stays zero, so it is | |
9667 | * correctly treated as an imbalance. | |
9668 | */ | |
c82513e5 | 9669 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 9670 | |
5d6523eb | 9671 | more_balance: |
8a8c69c3 | 9672 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 9673 | update_rq_clock(busiest); |
88b8dac0 SV |
9674 | |
9675 | /* | |
9676 | * cur_ld_moved - load moved in current iteration | |
9677 | * ld_moved - cumulative load moved across iterations | |
9678 | */ | |
163122b7 | 9679 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
9680 | |
9681 | /* | |
163122b7 KT |
9682 | * We've detached some tasks from busiest_rq. Every |
9683 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
9684 | * unlock busiest->lock, and we are able to be sure | |
9685 | * that nobody can manipulate the tasks in parallel. | |
9686 | * See task_rq_lock() family for the details. | |
1e3c88bd | 9687 | */ |
163122b7 | 9688 | |
8a8c69c3 | 9689 | rq_unlock(busiest, &rf); |
163122b7 KT |
9690 | |
9691 | if (cur_ld_moved) { | |
9692 | attach_tasks(&env); | |
9693 | ld_moved += cur_ld_moved; | |
9694 | } | |
9695 | ||
8a8c69c3 | 9696 | local_irq_restore(rf.flags); |
88b8dac0 | 9697 | |
f1cd0858 JK |
9698 | if (env.flags & LBF_NEED_BREAK) { |
9699 | env.flags &= ~LBF_NEED_BREAK; | |
9700 | goto more_balance; | |
9701 | } | |
9702 | ||
88b8dac0 SV |
9703 | /* |
9704 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
9705 | * us and move them to an alternate dst_cpu in our sched_group | |
9706 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 9707 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
9708 | * sched_group. |
9709 | * | |
9710 | * This changes load balance semantics a bit on who can move | |
9711 | * load to a given_cpu. In addition to the given_cpu itself | |
9712 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
9713 | * nohz-idle), we now have balance_cpu in a position to move | |
9714 | * load to given_cpu. In rare situations, this may cause | |
9715 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
9716 | * _independently_ and at _same_ time to move some load to | |
3b03706f | 9717 | * given_cpu) causing excess load to be moved to given_cpu. |
88b8dac0 SV |
9718 | * This however should not happen so much in practice and |
9719 | * moreover subsequent load balance cycles should correct the | |
9720 | * excess load moved. | |
9721 | */ | |
6263322c | 9722 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 9723 | |
97fb7a0a | 9724 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 9725 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 9726 | |
78feefc5 | 9727 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 9728 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 9729 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
9730 | env.loop = 0; |
9731 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 9732 | |
88b8dac0 SV |
9733 | /* |
9734 | * Go back to "more_balance" rather than "redo" since we | |
9735 | * need to continue with same src_cpu. | |
9736 | */ | |
9737 | goto more_balance; | |
9738 | } | |
1e3c88bd | 9739 | |
6263322c PZ |
9740 | /* |
9741 | * We failed to reach balance because of affinity. | |
9742 | */ | |
9743 | if (sd_parent) { | |
63b2ca30 | 9744 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 9745 | |
afdeee05 | 9746 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 9747 | *group_imbalance = 1; |
6263322c PZ |
9748 | } |
9749 | ||
1e3c88bd | 9750 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 9751 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 9752 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
9753 | /* |
9754 | * Attempting to continue load balancing at the current | |
9755 | * sched_domain level only makes sense if there are | |
9756 | * active CPUs remaining as possible busiest CPUs to | |
9757 | * pull load from which are not contained within the | |
9758 | * destination group that is receiving any migrated | |
9759 | * load. | |
9760 | */ | |
9761 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
9762 | env.loop = 0; |
9763 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 9764 | goto redo; |
bbf18b19 | 9765 | } |
afdeee05 | 9766 | goto out_all_pinned; |
1e3c88bd PZ |
9767 | } |
9768 | } | |
9769 | ||
9770 | if (!ld_moved) { | |
ae92882e | 9771 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
9772 | /* |
9773 | * Increment the failure counter only on periodic balance. | |
9774 | * We do not want newidle balance, which can be very | |
9775 | * frequent, pollute the failure counter causing | |
9776 | * excessive cache_hot migrations and active balances. | |
9777 | */ | |
9778 | if (idle != CPU_NEWLY_IDLE) | |
9779 | sd->nr_balance_failed++; | |
1e3c88bd | 9780 | |
bd939f45 | 9781 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
9782 | unsigned long flags; |
9783 | ||
1e3c88bd PZ |
9784 | raw_spin_lock_irqsave(&busiest->lock, flags); |
9785 | ||
97fb7a0a IM |
9786 | /* |
9787 | * Don't kick the active_load_balance_cpu_stop, | |
9788 | * if the curr task on busiest CPU can't be | |
9789 | * moved to this_cpu: | |
1e3c88bd | 9790 | */ |
3bd37062 | 9791 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
1e3c88bd PZ |
9792 | raw_spin_unlock_irqrestore(&busiest->lock, |
9793 | flags); | |
1e3c88bd PZ |
9794 | goto out_one_pinned; |
9795 | } | |
9796 | ||
8a41dfcd VG |
9797 | /* Record that we found at least one task that could run on this_cpu */ |
9798 | env.flags &= ~LBF_ALL_PINNED; | |
9799 | ||
969c7921 TH |
9800 | /* |
9801 | * ->active_balance synchronizes accesses to | |
9802 | * ->active_balance_work. Once set, it's cleared | |
9803 | * only after active load balance is finished. | |
9804 | */ | |
1e3c88bd PZ |
9805 | if (!busiest->active_balance) { |
9806 | busiest->active_balance = 1; | |
9807 | busiest->push_cpu = this_cpu; | |
9808 | active_balance = 1; | |
9809 | } | |
9810 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 9811 | |
bd939f45 | 9812 | if (active_balance) { |
969c7921 TH |
9813 | stop_one_cpu_nowait(cpu_of(busiest), |
9814 | active_load_balance_cpu_stop, busiest, | |
9815 | &busiest->active_balance_work); | |
bd939f45 | 9816 | } |
1e3c88bd | 9817 | } |
e9b9734b | 9818 | } else { |
1e3c88bd | 9819 | sd->nr_balance_failed = 0; |
e9b9734b | 9820 | } |
1e3c88bd | 9821 | |
e9b9734b | 9822 | if (likely(!active_balance) || need_active_balance(&env)) { |
1e3c88bd PZ |
9823 | /* We were unbalanced, so reset the balancing interval */ |
9824 | sd->balance_interval = sd->min_interval; | |
1e3c88bd PZ |
9825 | } |
9826 | ||
1e3c88bd PZ |
9827 | goto out; |
9828 | ||
9829 | out_balanced: | |
afdeee05 VG |
9830 | /* |
9831 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
9832 | * constraints. Clear the imbalance flag only if other tasks got |
9833 | * a chance to move and fix the imbalance. | |
afdeee05 | 9834 | */ |
f6cad8df | 9835 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
9836 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
9837 | ||
9838 | if (*group_imbalance) | |
9839 | *group_imbalance = 0; | |
9840 | } | |
9841 | ||
9842 | out_all_pinned: | |
9843 | /* | |
9844 | * We reach balance because all tasks are pinned at this level so | |
9845 | * we can't migrate them. Let the imbalance flag set so parent level | |
9846 | * can try to migrate them. | |
9847 | */ | |
ae92882e | 9848 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
9849 | |
9850 | sd->nr_balance_failed = 0; | |
9851 | ||
9852 | out_one_pinned: | |
3f130a37 VS |
9853 | ld_moved = 0; |
9854 | ||
9855 | /* | |
5ba553ef PZ |
9856 | * newidle_balance() disregards balance intervals, so we could |
9857 | * repeatedly reach this code, which would lead to balance_interval | |
3b03706f | 9858 | * skyrocketing in a short amount of time. Skip the balance_interval |
5ba553ef | 9859 | * increase logic to avoid that. |
3f130a37 VS |
9860 | */ |
9861 | if (env.idle == CPU_NEWLY_IDLE) | |
9862 | goto out; | |
9863 | ||
1e3c88bd | 9864 | /* tune up the balancing interval */ |
47b7aee1 VS |
9865 | if ((env.flags & LBF_ALL_PINNED && |
9866 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
9867 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 9868 | sd->balance_interval *= 2; |
1e3c88bd | 9869 | out: |
1e3c88bd PZ |
9870 | return ld_moved; |
9871 | } | |
9872 | ||
52a08ef1 JL |
9873 | static inline unsigned long |
9874 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
9875 | { | |
9876 | unsigned long interval = sd->balance_interval; | |
9877 | ||
9878 | if (cpu_busy) | |
9879 | interval *= sd->busy_factor; | |
9880 | ||
9881 | /* scale ms to jiffies */ | |
9882 | interval = msecs_to_jiffies(interval); | |
e4d32e4d VG |
9883 | |
9884 | /* | |
9885 | * Reduce likelihood of busy balancing at higher domains racing with | |
9886 | * balancing at lower domains by preventing their balancing periods | |
9887 | * from being multiples of each other. | |
9888 | */ | |
9889 | if (cpu_busy) | |
9890 | interval -= 1; | |
9891 | ||
52a08ef1 JL |
9892 | interval = clamp(interval, 1UL, max_load_balance_interval); |
9893 | ||
9894 | return interval; | |
9895 | } | |
9896 | ||
9897 | static inline void | |
31851a98 | 9898 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
9899 | { |
9900 | unsigned long interval, next; | |
9901 | ||
31851a98 LY |
9902 | /* used by idle balance, so cpu_busy = 0 */ |
9903 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
9904 | next = sd->last_balance + interval; |
9905 | ||
9906 | if (time_after(*next_balance, next)) | |
9907 | *next_balance = next; | |
9908 | } | |
9909 | ||
1e3c88bd | 9910 | /* |
97fb7a0a | 9911 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
9912 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
9913 | * least 1 task to be running on each physical CPU where possible, and | |
9914 | * avoids physical / logical imbalances. | |
1e3c88bd | 9915 | */ |
969c7921 | 9916 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 9917 | { |
969c7921 TH |
9918 | struct rq *busiest_rq = data; |
9919 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 9920 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 9921 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 9922 | struct sched_domain *sd; |
e5673f28 | 9923 | struct task_struct *p = NULL; |
8a8c69c3 | 9924 | struct rq_flags rf; |
969c7921 | 9925 | |
8a8c69c3 | 9926 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
9927 | /* |
9928 | * Between queueing the stop-work and running it is a hole in which | |
9929 | * CPUs can become inactive. We should not move tasks from or to | |
9930 | * inactive CPUs. | |
9931 | */ | |
9932 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
9933 | goto out_unlock; | |
969c7921 | 9934 | |
97fb7a0a | 9935 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
9936 | if (unlikely(busiest_cpu != smp_processor_id() || |
9937 | !busiest_rq->active_balance)) | |
9938 | goto out_unlock; | |
1e3c88bd PZ |
9939 | |
9940 | /* Is there any task to move? */ | |
9941 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 9942 | goto out_unlock; |
1e3c88bd PZ |
9943 | |
9944 | /* | |
9945 | * This condition is "impossible", if it occurs | |
9946 | * we need to fix it. Originally reported by | |
97fb7a0a | 9947 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd PZ |
9948 | */ |
9949 | BUG_ON(busiest_rq == target_rq); | |
9950 | ||
1e3c88bd | 9951 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 9952 | rcu_read_lock(); |
1e3c88bd | 9953 | for_each_domain(target_cpu, sd) { |
e669ac8a VS |
9954 | if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
9955 | break; | |
1e3c88bd PZ |
9956 | } |
9957 | ||
9958 | if (likely(sd)) { | |
8e45cb54 PZ |
9959 | struct lb_env env = { |
9960 | .sd = sd, | |
ddcdf6e7 PZ |
9961 | .dst_cpu = target_cpu, |
9962 | .dst_rq = target_rq, | |
9963 | .src_cpu = busiest_rq->cpu, | |
9964 | .src_rq = busiest_rq, | |
8e45cb54 | 9965 | .idle = CPU_IDLE, |
23fb06d9 | 9966 | .flags = LBF_ACTIVE_LB, |
8e45cb54 PZ |
9967 | }; |
9968 | ||
ae92882e | 9969 | schedstat_inc(sd->alb_count); |
3bed5e21 | 9970 | update_rq_clock(busiest_rq); |
1e3c88bd | 9971 | |
e5673f28 | 9972 | p = detach_one_task(&env); |
d02c0711 | 9973 | if (p) { |
ae92882e | 9974 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
9975 | /* Active balancing done, reset the failure counter. */ |
9976 | sd->nr_balance_failed = 0; | |
9977 | } else { | |
ae92882e | 9978 | schedstat_inc(sd->alb_failed); |
d02c0711 | 9979 | } |
1e3c88bd | 9980 | } |
dce840a0 | 9981 | rcu_read_unlock(); |
969c7921 TH |
9982 | out_unlock: |
9983 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 9984 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
9985 | |
9986 | if (p) | |
9987 | attach_one_task(target_rq, p); | |
9988 | ||
9989 | local_irq_enable(); | |
9990 | ||
969c7921 | 9991 | return 0; |
1e3c88bd PZ |
9992 | } |
9993 | ||
af3fe03c PZ |
9994 | static DEFINE_SPINLOCK(balancing); |
9995 | ||
9996 | /* | |
9997 | * Scale the max load_balance interval with the number of CPUs in the system. | |
9998 | * This trades load-balance latency on larger machines for less cross talk. | |
9999 | */ | |
10000 | void update_max_interval(void) | |
10001 | { | |
10002 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
10003 | } | |
10004 | ||
10005 | /* | |
10006 | * It checks each scheduling domain to see if it is due to be balanced, | |
10007 | * and initiates a balancing operation if so. | |
10008 | * | |
10009 | * Balancing parameters are set up in init_sched_domains. | |
10010 | */ | |
10011 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
10012 | { | |
10013 | int continue_balancing = 1; | |
10014 | int cpu = rq->cpu; | |
323af6de | 10015 | int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10016 | unsigned long interval; |
10017 | struct sched_domain *sd; | |
10018 | /* Earliest time when we have to do rebalance again */ | |
10019 | unsigned long next_balance = jiffies + 60*HZ; | |
10020 | int update_next_balance = 0; | |
10021 | int need_serialize, need_decay = 0; | |
10022 | u64 max_cost = 0; | |
10023 | ||
10024 | rcu_read_lock(); | |
10025 | for_each_domain(cpu, sd) { | |
10026 | /* | |
10027 | * Decay the newidle max times here because this is a regular | |
10028 | * visit to all the domains. Decay ~1% per second. | |
10029 | */ | |
10030 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
10031 | sd->max_newidle_lb_cost = | |
10032 | (sd->max_newidle_lb_cost * 253) / 256; | |
10033 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
10034 | need_decay = 1; | |
10035 | } | |
10036 | max_cost += sd->max_newidle_lb_cost; | |
10037 | ||
af3fe03c PZ |
10038 | /* |
10039 | * Stop the load balance at this level. There is another | |
10040 | * CPU in our sched group which is doing load balancing more | |
10041 | * actively. | |
10042 | */ | |
10043 | if (!continue_balancing) { | |
10044 | if (need_decay) | |
10045 | continue; | |
10046 | break; | |
10047 | } | |
10048 | ||
323af6de | 10049 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
10050 | |
10051 | need_serialize = sd->flags & SD_SERIALIZE; | |
10052 | if (need_serialize) { | |
10053 | if (!spin_trylock(&balancing)) | |
10054 | goto out; | |
10055 | } | |
10056 | ||
10057 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
10058 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
10059 | /* | |
10060 | * The LBF_DST_PINNED logic could have changed | |
10061 | * env->dst_cpu, so we can't know our idle | |
10062 | * state even if we migrated tasks. Update it. | |
10063 | */ | |
10064 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
323af6de | 10065 | busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10066 | } |
10067 | sd->last_balance = jiffies; | |
323af6de | 10068 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
10069 | } |
10070 | if (need_serialize) | |
10071 | spin_unlock(&balancing); | |
10072 | out: | |
10073 | if (time_after(next_balance, sd->last_balance + interval)) { | |
10074 | next_balance = sd->last_balance + interval; | |
10075 | update_next_balance = 1; | |
10076 | } | |
10077 | } | |
10078 | if (need_decay) { | |
10079 | /* | |
10080 | * Ensure the rq-wide value also decays but keep it at a | |
10081 | * reasonable floor to avoid funnies with rq->avg_idle. | |
10082 | */ | |
10083 | rq->max_idle_balance_cost = | |
10084 | max((u64)sysctl_sched_migration_cost, max_cost); | |
10085 | } | |
10086 | rcu_read_unlock(); | |
10087 | ||
10088 | /* | |
10089 | * next_balance will be updated only when there is a need. | |
10090 | * When the cpu is attached to null domain for ex, it will not be | |
10091 | * updated. | |
10092 | */ | |
7a82e5f5 | 10093 | if (likely(update_next_balance)) |
af3fe03c PZ |
10094 | rq->next_balance = next_balance; |
10095 | ||
af3fe03c PZ |
10096 | } |
10097 | ||
d987fc7f MG |
10098 | static inline int on_null_domain(struct rq *rq) |
10099 | { | |
10100 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
10101 | } | |
10102 | ||
3451d024 | 10103 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
10104 | /* |
10105 | * idle load balancing details | |
83cd4fe2 VP |
10106 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
10107 | * needed, they will kick the idle load balancer, which then does idle | |
10108 | * load balancing for all the idle CPUs. | |
9b019acb NP |
10109 | * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set |
10110 | * anywhere yet. | |
83cd4fe2 | 10111 | */ |
1e3c88bd | 10112 | |
3dd0337d | 10113 | static inline int find_new_ilb(void) |
1e3c88bd | 10114 | { |
9b019acb | 10115 | int ilb; |
1e3c88bd | 10116 | |
9b019acb NP |
10117 | for_each_cpu_and(ilb, nohz.idle_cpus_mask, |
10118 | housekeeping_cpumask(HK_FLAG_MISC)) { | |
45da7a2b PZ |
10119 | |
10120 | if (ilb == smp_processor_id()) | |
10121 | continue; | |
10122 | ||
9b019acb NP |
10123 | if (idle_cpu(ilb)) |
10124 | return ilb; | |
10125 | } | |
786d6dc7 SS |
10126 | |
10127 | return nr_cpu_ids; | |
1e3c88bd | 10128 | } |
1e3c88bd | 10129 | |
83cd4fe2 | 10130 | /* |
9b019acb NP |
10131 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick any |
10132 | * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one). | |
83cd4fe2 | 10133 | */ |
a4064fb6 | 10134 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
10135 | { |
10136 | int ilb_cpu; | |
10137 | ||
3ea2f097 VG |
10138 | /* |
10139 | * Increase nohz.next_balance only when if full ilb is triggered but | |
10140 | * not if we only update stats. | |
10141 | */ | |
10142 | if (flags & NOHZ_BALANCE_KICK) | |
10143 | nohz.next_balance = jiffies+1; | |
83cd4fe2 | 10144 | |
3dd0337d | 10145 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 10146 | |
0b005cf5 SS |
10147 | if (ilb_cpu >= nr_cpu_ids) |
10148 | return; | |
83cd4fe2 | 10149 | |
19a1f5ec PZ |
10150 | /* |
10151 | * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets | |
10152 | * the first flag owns it; cleared by nohz_csd_func(). | |
10153 | */ | |
a4064fb6 | 10154 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 10155 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 10156 | return; |
4550487a | 10157 | |
1c792db7 | 10158 | /* |
90b5363a | 10159 | * This way we generate an IPI on the target CPU which |
1c792db7 SS |
10160 | * is idle. And the softirq performing nohz idle load balance |
10161 | * will be run before returning from the IPI. | |
10162 | */ | |
90b5363a | 10163 | smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); |
4550487a PZ |
10164 | } |
10165 | ||
10166 | /* | |
9f132742 VS |
10167 | * Current decision point for kicking the idle load balancer in the presence |
10168 | * of idle CPUs in the system. | |
4550487a PZ |
10169 | */ |
10170 | static void nohz_balancer_kick(struct rq *rq) | |
10171 | { | |
10172 | unsigned long now = jiffies; | |
10173 | struct sched_domain_shared *sds; | |
10174 | struct sched_domain *sd; | |
10175 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 10176 | unsigned int flags = 0; |
4550487a PZ |
10177 | |
10178 | if (unlikely(rq->idle_balance)) | |
10179 | return; | |
10180 | ||
10181 | /* | |
10182 | * We may be recently in ticked or tickless idle mode. At the first | |
10183 | * busy tick after returning from idle, we will update the busy stats. | |
10184 | */ | |
00357f5e | 10185 | nohz_balance_exit_idle(rq); |
4550487a PZ |
10186 | |
10187 | /* | |
10188 | * None are in tickless mode and hence no need for NOHZ idle load | |
10189 | * balancing. | |
10190 | */ | |
10191 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
10192 | return; | |
10193 | ||
f643ea22 VG |
10194 | if (READ_ONCE(nohz.has_blocked) && |
10195 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
10196 | flags = NOHZ_STATS_KICK; |
10197 | ||
4550487a | 10198 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 10199 | goto out; |
4550487a | 10200 | |
a0fe2cf0 | 10201 | if (rq->nr_running >= 2) { |
a4064fb6 | 10202 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10203 | goto out; |
10204 | } | |
10205 | ||
10206 | rcu_read_lock(); | |
4550487a PZ |
10207 | |
10208 | sd = rcu_dereference(rq->sd); | |
10209 | if (sd) { | |
e25a7a94 VS |
10210 | /* |
10211 | * If there's a CFS task and the current CPU has reduced | |
10212 | * capacity; kick the ILB to see if there's a better CPU to run | |
10213 | * on. | |
10214 | */ | |
10215 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
a4064fb6 | 10216 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10217 | goto unlock; |
10218 | } | |
10219 | } | |
10220 | ||
011b27bb | 10221 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 10222 | if (sd) { |
b9a7b883 VS |
10223 | /* |
10224 | * When ASYM_PACKING; see if there's a more preferred CPU | |
10225 | * currently idle; in which case, kick the ILB to move tasks | |
10226 | * around. | |
10227 | */ | |
7edab78d | 10228 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
4550487a | 10229 | if (sched_asym_prefer(i, cpu)) { |
a4064fb6 | 10230 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10231 | goto unlock; |
10232 | } | |
10233 | } | |
10234 | } | |
b9a7b883 | 10235 | |
a0fe2cf0 VS |
10236 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
10237 | if (sd) { | |
10238 | /* | |
10239 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
10240 | * to run the misfit task on. | |
10241 | */ | |
10242 | if (check_misfit_status(rq, sd)) { | |
10243 | flags = NOHZ_KICK_MASK; | |
10244 | goto unlock; | |
10245 | } | |
b9a7b883 VS |
10246 | |
10247 | /* | |
10248 | * For asymmetric systems, we do not want to nicely balance | |
10249 | * cache use, instead we want to embrace asymmetry and only | |
10250 | * ensure tasks have enough CPU capacity. | |
10251 | * | |
10252 | * Skip the LLC logic because it's not relevant in that case. | |
10253 | */ | |
10254 | goto unlock; | |
a0fe2cf0 VS |
10255 | } |
10256 | ||
b9a7b883 VS |
10257 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
10258 | if (sds) { | |
e25a7a94 | 10259 | /* |
b9a7b883 VS |
10260 | * If there is an imbalance between LLC domains (IOW we could |
10261 | * increase the overall cache use), we need some less-loaded LLC | |
10262 | * domain to pull some load. Likewise, we may need to spread | |
10263 | * load within the current LLC domain (e.g. packed SMT cores but | |
10264 | * other CPUs are idle). We can't really know from here how busy | |
10265 | * the others are - so just get a nohz balance going if it looks | |
10266 | * like this LLC domain has tasks we could move. | |
e25a7a94 | 10267 | */ |
b9a7b883 VS |
10268 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
10269 | if (nr_busy > 1) { | |
10270 | flags = NOHZ_KICK_MASK; | |
10271 | goto unlock; | |
4550487a PZ |
10272 | } |
10273 | } | |
10274 | unlock: | |
10275 | rcu_read_unlock(); | |
10276 | out: | |
a4064fb6 PZ |
10277 | if (flags) |
10278 | kick_ilb(flags); | |
83cd4fe2 VP |
10279 | } |
10280 | ||
00357f5e | 10281 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 10282 | { |
00357f5e | 10283 | struct sched_domain *sd; |
a22e47a4 | 10284 | |
00357f5e PZ |
10285 | rcu_read_lock(); |
10286 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 10287 | |
00357f5e PZ |
10288 | if (!sd || !sd->nohz_idle) |
10289 | goto unlock; | |
10290 | sd->nohz_idle = 0; | |
10291 | ||
10292 | atomic_inc(&sd->shared->nr_busy_cpus); | |
10293 | unlock: | |
10294 | rcu_read_unlock(); | |
71325960 SS |
10295 | } |
10296 | ||
00357f5e PZ |
10297 | void nohz_balance_exit_idle(struct rq *rq) |
10298 | { | |
10299 | SCHED_WARN_ON(rq != this_rq()); | |
10300 | ||
10301 | if (likely(!rq->nohz_tick_stopped)) | |
10302 | return; | |
10303 | ||
10304 | rq->nohz_tick_stopped = 0; | |
10305 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
10306 | atomic_dec(&nohz.nr_cpus); | |
10307 | ||
10308 | set_cpu_sd_state_busy(rq->cpu); | |
10309 | } | |
10310 | ||
10311 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
10312 | { |
10313 | struct sched_domain *sd; | |
69e1e811 | 10314 | |
69e1e811 | 10315 | rcu_read_lock(); |
0e369d75 | 10316 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
10317 | |
10318 | if (!sd || sd->nohz_idle) | |
10319 | goto unlock; | |
10320 | sd->nohz_idle = 1; | |
10321 | ||
0e369d75 | 10322 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 10323 | unlock: |
69e1e811 SS |
10324 | rcu_read_unlock(); |
10325 | } | |
10326 | ||
1e3c88bd | 10327 | /* |
97fb7a0a | 10328 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 10329 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 10330 | */ |
c1cc017c | 10331 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 10332 | { |
00357f5e PZ |
10333 | struct rq *rq = cpu_rq(cpu); |
10334 | ||
10335 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
10336 | ||
97fb7a0a | 10337 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
10338 | if (!cpu_active(cpu)) |
10339 | return; | |
10340 | ||
387bc8b5 | 10341 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
de201559 | 10342 | if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) |
387bc8b5 FW |
10343 | return; |
10344 | ||
f643ea22 VG |
10345 | /* |
10346 | * Can be set safely without rq->lock held | |
10347 | * If a clear happens, it will have evaluated last additions because | |
10348 | * rq->lock is held during the check and the clear | |
10349 | */ | |
10350 | rq->has_blocked_load = 1; | |
10351 | ||
10352 | /* | |
10353 | * The tick is still stopped but load could have been added in the | |
10354 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
10355 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
10356 | * of nohz.has_blocked can only happen after checking the new load | |
10357 | */ | |
00357f5e | 10358 | if (rq->nohz_tick_stopped) |
f643ea22 | 10359 | goto out; |
1e3c88bd | 10360 | |
97fb7a0a | 10361 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 10362 | if (on_null_domain(rq)) |
d987fc7f MG |
10363 | return; |
10364 | ||
00357f5e PZ |
10365 | rq->nohz_tick_stopped = 1; |
10366 | ||
c1cc017c AS |
10367 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
10368 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 10369 | |
f643ea22 VG |
10370 | /* |
10371 | * Ensures that if nohz_idle_balance() fails to observe our | |
10372 | * @idle_cpus_mask store, it must observe the @has_blocked | |
10373 | * store. | |
10374 | */ | |
10375 | smp_mb__after_atomic(); | |
10376 | ||
00357f5e | 10377 | set_cpu_sd_state_idle(cpu); |
f643ea22 VG |
10378 | |
10379 | out: | |
10380 | /* | |
10381 | * Each time a cpu enter idle, we assume that it has blocked load and | |
10382 | * enable the periodic update of the load of idle cpus | |
10383 | */ | |
10384 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 10385 | } |
1e3c88bd | 10386 | |
3f5ad914 Y |
10387 | static bool update_nohz_stats(struct rq *rq) |
10388 | { | |
10389 | unsigned int cpu = rq->cpu; | |
10390 | ||
10391 | if (!rq->has_blocked_load) | |
10392 | return false; | |
10393 | ||
10394 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) | |
10395 | return false; | |
10396 | ||
10397 | if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) | |
10398 | return true; | |
10399 | ||
10400 | update_blocked_averages(cpu); | |
10401 | ||
10402 | return rq->has_blocked_load; | |
10403 | } | |
10404 | ||
1e3c88bd | 10405 | /* |
31e77c93 VG |
10406 | * Internal function that runs load balance for all idle cpus. The load balance |
10407 | * can be a simple update of blocked load or a complete load balance with | |
10408 | * tasks movement depending of flags. | |
1e3c88bd | 10409 | */ |
ab2dde5e | 10410 | static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags, |
31e77c93 | 10411 | enum cpu_idle_type idle) |
83cd4fe2 | 10412 | { |
c5afb6a8 | 10413 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
10414 | unsigned long now = jiffies; |
10415 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 10416 | bool has_blocked_load = false; |
c5afb6a8 | 10417 | int update_next_balance = 0; |
b7031a02 | 10418 | int this_cpu = this_rq->cpu; |
b7031a02 PZ |
10419 | int balance_cpu; |
10420 | struct rq *rq; | |
83cd4fe2 | 10421 | |
b7031a02 | 10422 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 10423 | |
f643ea22 VG |
10424 | /* |
10425 | * We assume there will be no idle load after this update and clear | |
10426 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
10427 | * set the has_blocked flag and trig another update of idle load. | |
10428 | * Because a cpu that becomes idle, is added to idle_cpus_mask before | |
10429 | * setting the flag, we are sure to not clear the state and not | |
10430 | * check the load of an idle cpu. | |
10431 | */ | |
10432 | WRITE_ONCE(nohz.has_blocked, 0); | |
10433 | ||
10434 | /* | |
10435 | * Ensures that if we miss the CPU, we must see the has_blocked | |
10436 | * store from nohz_balance_enter_idle(). | |
10437 | */ | |
10438 | smp_mb(); | |
10439 | ||
7a82e5f5 VG |
10440 | /* |
10441 | * Start with the next CPU after this_cpu so we will end with this_cpu and let a | |
10442 | * chance for other idle cpu to pull load. | |
10443 | */ | |
10444 | for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) { | |
10445 | if (!idle_cpu(balance_cpu)) | |
83cd4fe2 VP |
10446 | continue; |
10447 | ||
10448 | /* | |
97fb7a0a IM |
10449 | * If this CPU gets work to do, stop the load balancing |
10450 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
10451 | * balancing owner will pick it up. |
10452 | */ | |
f643ea22 VG |
10453 | if (need_resched()) { |
10454 | has_blocked_load = true; | |
10455 | goto abort; | |
10456 | } | |
83cd4fe2 | 10457 | |
5ed4f1d9 VG |
10458 | rq = cpu_rq(balance_cpu); |
10459 | ||
64f84f27 | 10460 | has_blocked_load |= update_nohz_stats(rq); |
f643ea22 | 10461 | |
ed61bbc6 TC |
10462 | /* |
10463 | * If time for next balance is due, | |
10464 | * do the balance. | |
10465 | */ | |
10466 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
10467 | struct rq_flags rf; |
10468 | ||
31e77c93 | 10469 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 10470 | update_rq_clock(rq); |
31e77c93 | 10471 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 10472 | |
b7031a02 PZ |
10473 | if (flags & NOHZ_BALANCE_KICK) |
10474 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 10475 | } |
83cd4fe2 | 10476 | |
c5afb6a8 VG |
10477 | if (time_after(next_balance, rq->next_balance)) { |
10478 | next_balance = rq->next_balance; | |
10479 | update_next_balance = 1; | |
10480 | } | |
83cd4fe2 | 10481 | } |
c5afb6a8 | 10482 | |
3ea2f097 VG |
10483 | /* |
10484 | * next_balance will be updated only when there is a need. | |
10485 | * When the CPU is attached to null domain for ex, it will not be | |
10486 | * updated. | |
10487 | */ | |
10488 | if (likely(update_next_balance)) | |
10489 | nohz.next_balance = next_balance; | |
10490 | ||
f643ea22 VG |
10491 | WRITE_ONCE(nohz.next_blocked, |
10492 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
10493 | ||
10494 | abort: | |
10495 | /* There is still blocked load, enable periodic update */ | |
10496 | if (has_blocked_load) | |
10497 | WRITE_ONCE(nohz.has_blocked, 1); | |
31e77c93 VG |
10498 | } |
10499 | ||
10500 | /* | |
10501 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
10502 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
10503 | */ | |
10504 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
10505 | { | |
19a1f5ec | 10506 | unsigned int flags = this_rq->nohz_idle_balance; |
31e77c93 | 10507 | |
19a1f5ec | 10508 | if (!flags) |
31e77c93 VG |
10509 | return false; |
10510 | ||
19a1f5ec | 10511 | this_rq->nohz_idle_balance = 0; |
31e77c93 | 10512 | |
19a1f5ec | 10513 | if (idle != CPU_IDLE) |
31e77c93 VG |
10514 | return false; |
10515 | ||
10516 | _nohz_idle_balance(this_rq, flags, idle); | |
10517 | ||
b7031a02 | 10518 | return true; |
83cd4fe2 | 10519 | } |
31e77c93 | 10520 | |
c6f88654 VG |
10521 | /* |
10522 | * Check if we need to run the ILB for updating blocked load before entering | |
10523 | * idle state. | |
10524 | */ | |
10525 | void nohz_run_idle_balance(int cpu) | |
10526 | { | |
10527 | unsigned int flags; | |
10528 | ||
10529 | flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu)); | |
10530 | ||
10531 | /* | |
10532 | * Update the blocked load only if no SCHED_SOFTIRQ is about to happen | |
10533 | * (ie NOHZ_STATS_KICK set) and will do the same. | |
10534 | */ | |
10535 | if ((flags == NOHZ_NEWILB_KICK) && !need_resched()) | |
10536 | _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE); | |
10537 | } | |
10538 | ||
31e77c93 VG |
10539 | static void nohz_newidle_balance(struct rq *this_rq) |
10540 | { | |
10541 | int this_cpu = this_rq->cpu; | |
10542 | ||
10543 | /* | |
10544 | * This CPU doesn't want to be disturbed by scheduler | |
10545 | * housekeeping | |
10546 | */ | |
10547 | if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) | |
10548 | return; | |
10549 | ||
10550 | /* Will wake up very soon. No time for doing anything else*/ | |
10551 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
10552 | return; | |
10553 | ||
10554 | /* Don't need to update blocked load of idle CPUs*/ | |
10555 | if (!READ_ONCE(nohz.has_blocked) || | |
10556 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
10557 | return; | |
10558 | ||
31e77c93 | 10559 | /* |
c6f88654 VG |
10560 | * Set the need to trigger ILB in order to update blocked load |
10561 | * before entering idle state. | |
31e77c93 | 10562 | */ |
c6f88654 | 10563 | atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu)); |
31e77c93 VG |
10564 | } |
10565 | ||
dd707247 PZ |
10566 | #else /* !CONFIG_NO_HZ_COMMON */ |
10567 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
10568 | ||
31e77c93 | 10569 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
10570 | { |
10571 | return false; | |
10572 | } | |
31e77c93 VG |
10573 | |
10574 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 10575 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 10576 | |
47ea5412 | 10577 | /* |
5b78f2dc | 10578 | * newidle_balance is called by schedule() if this_cpu is about to become |
47ea5412 | 10579 | * idle. Attempts to pull tasks from other CPUs. |
7277a34c PZ |
10580 | * |
10581 | * Returns: | |
10582 | * < 0 - we released the lock and there are !fair tasks present | |
10583 | * 0 - failed, no new tasks | |
10584 | * > 0 - success, new (fair) tasks present | |
47ea5412 | 10585 | */ |
d91cecc1 | 10586 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) |
47ea5412 PZ |
10587 | { |
10588 | unsigned long next_balance = jiffies + HZ; | |
10589 | int this_cpu = this_rq->cpu; | |
10590 | struct sched_domain *sd; | |
10591 | int pulled_task = 0; | |
10592 | u64 curr_cost = 0; | |
10593 | ||
5ba553ef | 10594 | update_misfit_status(NULL, this_rq); |
e5e678e4 RR |
10595 | |
10596 | /* | |
10597 | * There is a task waiting to run. No need to search for one. | |
10598 | * Return 0; the task will be enqueued when switching to idle. | |
10599 | */ | |
10600 | if (this_rq->ttwu_pending) | |
10601 | return 0; | |
10602 | ||
47ea5412 PZ |
10603 | /* |
10604 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
10605 | * measure the duration of idle_balance() as idle time. | |
10606 | */ | |
10607 | this_rq->idle_stamp = rq_clock(this_rq); | |
10608 | ||
10609 | /* | |
10610 | * Do not pull tasks towards !active CPUs... | |
10611 | */ | |
10612 | if (!cpu_active(this_cpu)) | |
10613 | return 0; | |
10614 | ||
10615 | /* | |
10616 | * This is OK, because current is on_cpu, which avoids it being picked | |
10617 | * for load-balance and preemption/IRQs are still disabled avoiding | |
10618 | * further scheduler activity on it and we're being very careful to | |
10619 | * re-start the picking loop. | |
10620 | */ | |
10621 | rq_unpin_lock(this_rq, rf); | |
10622 | ||
10623 | if (this_rq->avg_idle < sysctl_sched_migration_cost || | |
e90c8fe1 | 10624 | !READ_ONCE(this_rq->rd->overload)) { |
31e77c93 | 10625 | |
47ea5412 PZ |
10626 | rcu_read_lock(); |
10627 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
10628 | if (sd) | |
10629 | update_next_balance(sd, &next_balance); | |
10630 | rcu_read_unlock(); | |
10631 | ||
10632 | goto out; | |
10633 | } | |
10634 | ||
10635 | raw_spin_unlock(&this_rq->lock); | |
10636 | ||
10637 | update_blocked_averages(this_cpu); | |
10638 | rcu_read_lock(); | |
10639 | for_each_domain(this_cpu, sd) { | |
10640 | int continue_balancing = 1; | |
10641 | u64 t0, domain_cost; | |
10642 | ||
47ea5412 PZ |
10643 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
10644 | update_next_balance(sd, &next_balance); | |
10645 | break; | |
10646 | } | |
10647 | ||
10648 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
10649 | t0 = sched_clock_cpu(this_cpu); | |
10650 | ||
10651 | pulled_task = load_balance(this_cpu, this_rq, | |
10652 | sd, CPU_NEWLY_IDLE, | |
10653 | &continue_balancing); | |
10654 | ||
10655 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
10656 | if (domain_cost > sd->max_newidle_lb_cost) | |
10657 | sd->max_newidle_lb_cost = domain_cost; | |
10658 | ||
10659 | curr_cost += domain_cost; | |
10660 | } | |
10661 | ||
10662 | update_next_balance(sd, &next_balance); | |
10663 | ||
10664 | /* | |
10665 | * Stop searching for tasks to pull if there are | |
10666 | * now runnable tasks on this rq. | |
10667 | */ | |
e5e678e4 RR |
10668 | if (pulled_task || this_rq->nr_running > 0 || |
10669 | this_rq->ttwu_pending) | |
47ea5412 PZ |
10670 | break; |
10671 | } | |
10672 | rcu_read_unlock(); | |
10673 | ||
10674 | raw_spin_lock(&this_rq->lock); | |
10675 | ||
10676 | if (curr_cost > this_rq->max_idle_balance_cost) | |
10677 | this_rq->max_idle_balance_cost = curr_cost; | |
10678 | ||
10679 | /* | |
10680 | * While browsing the domains, we released the rq lock, a task could | |
10681 | * have been enqueued in the meantime. Since we're not going idle, | |
10682 | * pretend we pulled a task. | |
10683 | */ | |
10684 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
10685 | pulled_task = 1; | |
10686 | ||
47ea5412 PZ |
10687 | /* Is there a task of a high priority class? */ |
10688 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
10689 | pulled_task = -1; | |
10690 | ||
6553fc18 VG |
10691 | out: |
10692 | /* Move the next balance forward */ | |
10693 | if (time_after(this_rq->next_balance, next_balance)) | |
10694 | this_rq->next_balance = next_balance; | |
10695 | ||
47ea5412 PZ |
10696 | if (pulled_task) |
10697 | this_rq->idle_stamp = 0; | |
0826530d VG |
10698 | else |
10699 | nohz_newidle_balance(this_rq); | |
47ea5412 PZ |
10700 | |
10701 | rq_repin_lock(this_rq, rf); | |
10702 | ||
10703 | return pulled_task; | |
10704 | } | |
10705 | ||
83cd4fe2 VP |
10706 | /* |
10707 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
10708 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
10709 | */ | |
0766f788 | 10710 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 10711 | { |
208cb16b | 10712 | struct rq *this_rq = this_rq(); |
6eb57e0d | 10713 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
10714 | CPU_IDLE : CPU_NOT_IDLE; |
10715 | ||
1e3c88bd | 10716 | /* |
97fb7a0a IM |
10717 | * If this CPU has a pending nohz_balance_kick, then do the |
10718 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 10719 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 10720 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
10721 | * load balance only within the local sched_domain hierarchy |
10722 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 10723 | */ |
b7031a02 PZ |
10724 | if (nohz_idle_balance(this_rq, idle)) |
10725 | return; | |
10726 | ||
10727 | /* normal load balance */ | |
10728 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 10729 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
10730 | } |
10731 | ||
1e3c88bd PZ |
10732 | /* |
10733 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 10734 | */ |
7caff66f | 10735 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 10736 | { |
e0b257c3 AMB |
10737 | /* |
10738 | * Don't need to rebalance while attached to NULL domain or | |
10739 | * runqueue CPU is not active | |
10740 | */ | |
10741 | if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq)))) | |
c726099e DL |
10742 | return; |
10743 | ||
10744 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 10745 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
10746 | |
10747 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
10748 | } |
10749 | ||
0bcdcf28 CE |
10750 | static void rq_online_fair(struct rq *rq) |
10751 | { | |
10752 | update_sysctl(); | |
0e59bdae KT |
10753 | |
10754 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
10755 | } |
10756 | ||
10757 | static void rq_offline_fair(struct rq *rq) | |
10758 | { | |
10759 | update_sysctl(); | |
a4c96ae3 PB |
10760 | |
10761 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
10762 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
10763 | } |
10764 | ||
55e12e5e | 10765 | #endif /* CONFIG_SMP */ |
e1d1484f | 10766 | |
bf0f6f24 | 10767 | /* |
d84b3131 FW |
10768 | * scheduler tick hitting a task of our scheduling class. |
10769 | * | |
10770 | * NOTE: This function can be called remotely by the tick offload that | |
10771 | * goes along full dynticks. Therefore no local assumption can be made | |
10772 | * and everything must be accessed through the @rq and @curr passed in | |
10773 | * parameters. | |
bf0f6f24 | 10774 | */ |
8f4d37ec | 10775 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
10776 | { |
10777 | struct cfs_rq *cfs_rq; | |
10778 | struct sched_entity *se = &curr->se; | |
10779 | ||
10780 | for_each_sched_entity(se) { | |
10781 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 10782 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 10783 | } |
18bf2805 | 10784 | |
b52da86e | 10785 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 10786 | task_tick_numa(rq, curr); |
3b1baa64 MR |
10787 | |
10788 | update_misfit_status(curr, rq); | |
2802bf3c | 10789 | update_overutilized_status(task_rq(curr)); |
bf0f6f24 IM |
10790 | } |
10791 | ||
10792 | /* | |
cd29fe6f PZ |
10793 | * called on fork with the child task as argument from the parent's context |
10794 | * - child not yet on the tasklist | |
10795 | * - preemption disabled | |
bf0f6f24 | 10796 | */ |
cd29fe6f | 10797 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 10798 | { |
4fc420c9 DN |
10799 | struct cfs_rq *cfs_rq; |
10800 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 10801 | struct rq *rq = this_rq(); |
8a8c69c3 | 10802 | struct rq_flags rf; |
bf0f6f24 | 10803 | |
8a8c69c3 | 10804 | rq_lock(rq, &rf); |
861d034e PZ |
10805 | update_rq_clock(rq); |
10806 | ||
4fc420c9 DN |
10807 | cfs_rq = task_cfs_rq(current); |
10808 | curr = cfs_rq->curr; | |
e210bffd PZ |
10809 | if (curr) { |
10810 | update_curr(cfs_rq); | |
b5d9d734 | 10811 | se->vruntime = curr->vruntime; |
e210bffd | 10812 | } |
aeb73b04 | 10813 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 10814 | |
cd29fe6f | 10815 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 10816 | /* |
edcb60a3 IM |
10817 | * Upon rescheduling, sched_class::put_prev_task() will place |
10818 | * 'current' within the tree based on its new key value. | |
10819 | */ | |
4d78e7b6 | 10820 | swap(curr->vruntime, se->vruntime); |
8875125e | 10821 | resched_curr(rq); |
4d78e7b6 | 10822 | } |
bf0f6f24 | 10823 | |
88ec22d3 | 10824 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 10825 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
10826 | } |
10827 | ||
cb469845 SR |
10828 | /* |
10829 | * Priority of the task has changed. Check to see if we preempt | |
10830 | * the current task. | |
10831 | */ | |
da7a735e PZ |
10832 | static void |
10833 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 10834 | { |
da0c1e65 | 10835 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
10836 | return; |
10837 | ||
7c2e8bbd FW |
10838 | if (rq->cfs.nr_running == 1) |
10839 | return; | |
10840 | ||
cb469845 SR |
10841 | /* |
10842 | * Reschedule if we are currently running on this runqueue and | |
10843 | * our priority decreased, or if we are not currently running on | |
10844 | * this runqueue and our priority is higher than the current's | |
10845 | */ | |
65bcf072 | 10846 | if (task_current(rq, p)) { |
cb469845 | 10847 | if (p->prio > oldprio) |
8875125e | 10848 | resched_curr(rq); |
cb469845 | 10849 | } else |
15afe09b | 10850 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
10851 | } |
10852 | ||
daa59407 | 10853 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
10854 | { |
10855 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
10856 | |
10857 | /* | |
daa59407 BP |
10858 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
10859 | * the dequeue_entity(.flags=0) will already have normalized the | |
10860 | * vruntime. | |
10861 | */ | |
10862 | if (p->on_rq) | |
10863 | return true; | |
10864 | ||
10865 | /* | |
10866 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
10867 | * But there are some cases where it has already been normalized: | |
da7a735e | 10868 | * |
daa59407 BP |
10869 | * - A forked child which is waiting for being woken up by |
10870 | * wake_up_new_task(). | |
10871 | * - A task which has been woken up by try_to_wake_up() and | |
10872 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 10873 | */ |
d0cdb3ce SM |
10874 | if (!se->sum_exec_runtime || |
10875 | (p->state == TASK_WAKING && p->sched_remote_wakeup)) | |
daa59407 BP |
10876 | return true; |
10877 | ||
10878 | return false; | |
10879 | } | |
10880 | ||
09a43ace VG |
10881 | #ifdef CONFIG_FAIR_GROUP_SCHED |
10882 | /* | |
10883 | * Propagate the changes of the sched_entity across the tg tree to make it | |
10884 | * visible to the root | |
10885 | */ | |
10886 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
10887 | { | |
10888 | struct cfs_rq *cfs_rq; | |
10889 | ||
0258bdfa OU |
10890 | list_add_leaf_cfs_rq(cfs_rq_of(se)); |
10891 | ||
09a43ace VG |
10892 | /* Start to propagate at parent */ |
10893 | se = se->parent; | |
10894 | ||
10895 | for_each_sched_entity(se) { | |
10896 | cfs_rq = cfs_rq_of(se); | |
10897 | ||
0258bdfa OU |
10898 | if (!cfs_rq_throttled(cfs_rq)){ |
10899 | update_load_avg(cfs_rq, se, UPDATE_TG); | |
10900 | list_add_leaf_cfs_rq(cfs_rq); | |
10901 | continue; | |
10902 | } | |
09a43ace | 10903 | |
0258bdfa OU |
10904 | if (list_add_leaf_cfs_rq(cfs_rq)) |
10905 | break; | |
09a43ace VG |
10906 | } |
10907 | } | |
10908 | #else | |
10909 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
10910 | #endif | |
10911 | ||
df217913 | 10912 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 10913 | { |
daa59407 BP |
10914 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
10915 | ||
9d89c257 | 10916 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 10917 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 10918 | detach_entity_load_avg(cfs_rq, se); |
fe749158 | 10919 | update_tg_load_avg(cfs_rq); |
09a43ace | 10920 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
10921 | } |
10922 | ||
df217913 | 10923 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 10924 | { |
daa59407 | 10925 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
10926 | |
10927 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
10928 | /* |
10929 | * Since the real-depth could have been changed (only FAIR | |
10930 | * class maintain depth value), reset depth properly. | |
10931 | */ | |
10932 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10933 | #endif | |
7855a35a | 10934 | |
df217913 | 10935 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 10936 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
a4f9a0e5 | 10937 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 10938 | update_tg_load_avg(cfs_rq); |
09a43ace | 10939 | propagate_entity_cfs_rq(se); |
df217913 VG |
10940 | } |
10941 | ||
10942 | static void detach_task_cfs_rq(struct task_struct *p) | |
10943 | { | |
10944 | struct sched_entity *se = &p->se; | |
10945 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10946 | ||
10947 | if (!vruntime_normalized(p)) { | |
10948 | /* | |
10949 | * Fix up our vruntime so that the current sleep doesn't | |
10950 | * cause 'unlimited' sleep bonus. | |
10951 | */ | |
10952 | place_entity(cfs_rq, se, 0); | |
10953 | se->vruntime -= cfs_rq->min_vruntime; | |
10954 | } | |
10955 | ||
10956 | detach_entity_cfs_rq(se); | |
10957 | } | |
10958 | ||
10959 | static void attach_task_cfs_rq(struct task_struct *p) | |
10960 | { | |
10961 | struct sched_entity *se = &p->se; | |
10962 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10963 | ||
10964 | attach_entity_cfs_rq(se); | |
daa59407 BP |
10965 | |
10966 | if (!vruntime_normalized(p)) | |
10967 | se->vruntime += cfs_rq->min_vruntime; | |
10968 | } | |
6efdb105 | 10969 | |
daa59407 BP |
10970 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
10971 | { | |
10972 | detach_task_cfs_rq(p); | |
10973 | } | |
10974 | ||
10975 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
10976 | { | |
10977 | attach_task_cfs_rq(p); | |
7855a35a | 10978 | |
daa59407 | 10979 | if (task_on_rq_queued(p)) { |
7855a35a | 10980 | /* |
daa59407 BP |
10981 | * We were most likely switched from sched_rt, so |
10982 | * kick off the schedule if running, otherwise just see | |
10983 | * if we can still preempt the current task. | |
7855a35a | 10984 | */ |
65bcf072 | 10985 | if (task_current(rq, p)) |
daa59407 BP |
10986 | resched_curr(rq); |
10987 | else | |
10988 | check_preempt_curr(rq, p, 0); | |
7855a35a | 10989 | } |
cb469845 SR |
10990 | } |
10991 | ||
83b699ed SV |
10992 | /* Account for a task changing its policy or group. |
10993 | * | |
10994 | * This routine is mostly called to set cfs_rq->curr field when a task | |
10995 | * migrates between groups/classes. | |
10996 | */ | |
a0e813f2 | 10997 | static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) |
83b699ed | 10998 | { |
03b7fad1 PZ |
10999 | struct sched_entity *se = &p->se; |
11000 | ||
11001 | #ifdef CONFIG_SMP | |
11002 | if (task_on_rq_queued(p)) { | |
11003 | /* | |
11004 | * Move the next running task to the front of the list, so our | |
11005 | * cfs_tasks list becomes MRU one. | |
11006 | */ | |
11007 | list_move(&se->group_node, &rq->cfs_tasks); | |
11008 | } | |
11009 | #endif | |
83b699ed | 11010 | |
ec12cb7f PT |
11011 | for_each_sched_entity(se) { |
11012 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11013 | ||
11014 | set_next_entity(cfs_rq, se); | |
11015 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
11016 | account_cfs_rq_runtime(cfs_rq, 0); | |
11017 | } | |
83b699ed SV |
11018 | } |
11019 | ||
029632fb PZ |
11020 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
11021 | { | |
bfb06889 | 11022 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
029632fb PZ |
11023 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
11024 | #ifndef CONFIG_64BIT | |
11025 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
11026 | #endif | |
141965c7 | 11027 | #ifdef CONFIG_SMP |
2a2f5d4e | 11028 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 11029 | #endif |
029632fb PZ |
11030 | } |
11031 | ||
810b3817 | 11032 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
11033 | static void task_set_group_fair(struct task_struct *p) |
11034 | { | |
11035 | struct sched_entity *se = &p->se; | |
11036 | ||
11037 | set_task_rq(p, task_cpu(p)); | |
11038 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
11039 | } | |
11040 | ||
bc54da21 | 11041 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 11042 | { |
daa59407 | 11043 | detach_task_cfs_rq(p); |
b2b5ce02 | 11044 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
11045 | |
11046 | #ifdef CONFIG_SMP | |
11047 | /* Tell se's cfs_rq has been changed -- migrated */ | |
11048 | p->se.avg.last_update_time = 0; | |
11049 | #endif | |
daa59407 | 11050 | attach_task_cfs_rq(p); |
810b3817 | 11051 | } |
029632fb | 11052 | |
ea86cb4b VG |
11053 | static void task_change_group_fair(struct task_struct *p, int type) |
11054 | { | |
11055 | switch (type) { | |
11056 | case TASK_SET_GROUP: | |
11057 | task_set_group_fair(p); | |
11058 | break; | |
11059 | ||
11060 | case TASK_MOVE_GROUP: | |
11061 | task_move_group_fair(p); | |
11062 | break; | |
11063 | } | |
11064 | } | |
11065 | ||
029632fb PZ |
11066 | void free_fair_sched_group(struct task_group *tg) |
11067 | { | |
11068 | int i; | |
11069 | ||
11070 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
11071 | ||
11072 | for_each_possible_cpu(i) { | |
11073 | if (tg->cfs_rq) | |
11074 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 11075 | if (tg->se) |
029632fb PZ |
11076 | kfree(tg->se[i]); |
11077 | } | |
11078 | ||
11079 | kfree(tg->cfs_rq); | |
11080 | kfree(tg->se); | |
11081 | } | |
11082 | ||
11083 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11084 | { | |
029632fb | 11085 | struct sched_entity *se; |
b7fa30c9 | 11086 | struct cfs_rq *cfs_rq; |
029632fb PZ |
11087 | int i; |
11088 | ||
6396bb22 | 11089 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
11090 | if (!tg->cfs_rq) |
11091 | goto err; | |
6396bb22 | 11092 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
11093 | if (!tg->se) |
11094 | goto err; | |
11095 | ||
11096 | tg->shares = NICE_0_LOAD; | |
11097 | ||
11098 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
11099 | ||
11100 | for_each_possible_cpu(i) { | |
11101 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
11102 | GFP_KERNEL, cpu_to_node(i)); | |
11103 | if (!cfs_rq) | |
11104 | goto err; | |
11105 | ||
11106 | se = kzalloc_node(sizeof(struct sched_entity), | |
11107 | GFP_KERNEL, cpu_to_node(i)); | |
11108 | if (!se) | |
11109 | goto err_free_rq; | |
11110 | ||
11111 | init_cfs_rq(cfs_rq); | |
11112 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 11113 | init_entity_runnable_average(se); |
029632fb PZ |
11114 | } |
11115 | ||
11116 | return 1; | |
11117 | ||
11118 | err_free_rq: | |
11119 | kfree(cfs_rq); | |
11120 | err: | |
11121 | return 0; | |
11122 | } | |
11123 | ||
8663e24d PZ |
11124 | void online_fair_sched_group(struct task_group *tg) |
11125 | { | |
11126 | struct sched_entity *se; | |
a46d14ec | 11127 | struct rq_flags rf; |
8663e24d PZ |
11128 | struct rq *rq; |
11129 | int i; | |
11130 | ||
11131 | for_each_possible_cpu(i) { | |
11132 | rq = cpu_rq(i); | |
11133 | se = tg->se[i]; | |
a46d14ec | 11134 | rq_lock_irq(rq, &rf); |
4126bad6 | 11135 | update_rq_clock(rq); |
d0326691 | 11136 | attach_entity_cfs_rq(se); |
55e16d30 | 11137 | sync_throttle(tg, i); |
a46d14ec | 11138 | rq_unlock_irq(rq, &rf); |
8663e24d PZ |
11139 | } |
11140 | } | |
11141 | ||
6fe1f348 | 11142 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 11143 | { |
029632fb | 11144 | unsigned long flags; |
6fe1f348 PZ |
11145 | struct rq *rq; |
11146 | int cpu; | |
029632fb | 11147 | |
6fe1f348 PZ |
11148 | for_each_possible_cpu(cpu) { |
11149 | if (tg->se[cpu]) | |
11150 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 11151 | |
6fe1f348 PZ |
11152 | /* |
11153 | * Only empty task groups can be destroyed; so we can speculatively | |
11154 | * check on_list without danger of it being re-added. | |
11155 | */ | |
11156 | if (!tg->cfs_rq[cpu]->on_list) | |
11157 | continue; | |
11158 | ||
11159 | rq = cpu_rq(cpu); | |
11160 | ||
11161 | raw_spin_lock_irqsave(&rq->lock, flags); | |
11162 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
11163 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
11164 | } | |
029632fb PZ |
11165 | } |
11166 | ||
11167 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
11168 | struct sched_entity *se, int cpu, | |
11169 | struct sched_entity *parent) | |
11170 | { | |
11171 | struct rq *rq = cpu_rq(cpu); | |
11172 | ||
11173 | cfs_rq->tg = tg; | |
11174 | cfs_rq->rq = rq; | |
029632fb PZ |
11175 | init_cfs_rq_runtime(cfs_rq); |
11176 | ||
11177 | tg->cfs_rq[cpu] = cfs_rq; | |
11178 | tg->se[cpu] = se; | |
11179 | ||
11180 | /* se could be NULL for root_task_group */ | |
11181 | if (!se) | |
11182 | return; | |
11183 | ||
fed14d45 | 11184 | if (!parent) { |
029632fb | 11185 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
11186 | se->depth = 0; |
11187 | } else { | |
029632fb | 11188 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
11189 | se->depth = parent->depth + 1; |
11190 | } | |
029632fb PZ |
11191 | |
11192 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
11193 | /* guarantee group entities always have weight */ |
11194 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
11195 | se->parent = parent; |
11196 | } | |
11197 | ||
11198 | static DEFINE_MUTEX(shares_mutex); | |
11199 | ||
11200 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
11201 | { | |
11202 | int i; | |
029632fb PZ |
11203 | |
11204 | /* | |
11205 | * We can't change the weight of the root cgroup. | |
11206 | */ | |
11207 | if (!tg->se[0]) | |
11208 | return -EINVAL; | |
11209 | ||
11210 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
11211 | ||
11212 | mutex_lock(&shares_mutex); | |
11213 | if (tg->shares == shares) | |
11214 | goto done; | |
11215 | ||
11216 | tg->shares = shares; | |
11217 | for_each_possible_cpu(i) { | |
11218 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
11219 | struct sched_entity *se = tg->se[i]; |
11220 | struct rq_flags rf; | |
029632fb | 11221 | |
029632fb | 11222 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 11223 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 11224 | update_rq_clock(rq); |
89ee048f | 11225 | for_each_sched_entity(se) { |
88c0616e | 11226 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 11227 | update_cfs_group(se); |
89ee048f | 11228 | } |
8a8c69c3 | 11229 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
11230 | } |
11231 | ||
11232 | done: | |
11233 | mutex_unlock(&shares_mutex); | |
11234 | return 0; | |
11235 | } | |
11236 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
11237 | ||
11238 | void free_fair_sched_group(struct task_group *tg) { } | |
11239 | ||
11240 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11241 | { | |
11242 | return 1; | |
11243 | } | |
11244 | ||
8663e24d PZ |
11245 | void online_fair_sched_group(struct task_group *tg) { } |
11246 | ||
6fe1f348 | 11247 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
11248 | |
11249 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
11250 | ||
810b3817 | 11251 | |
6d686f45 | 11252 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
11253 | { |
11254 | struct sched_entity *se = &task->se; | |
0d721cea PW |
11255 | unsigned int rr_interval = 0; |
11256 | ||
11257 | /* | |
11258 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
11259 | * idle runqueue: | |
11260 | */ | |
0d721cea | 11261 | if (rq->cfs.load.weight) |
a59f4e07 | 11262 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
11263 | |
11264 | return rr_interval; | |
11265 | } | |
11266 | ||
bf0f6f24 IM |
11267 | /* |
11268 | * All the scheduling class methods: | |
11269 | */ | |
43c31ac0 PZ |
11270 | DEFINE_SCHED_CLASS(fair) = { |
11271 | ||
bf0f6f24 IM |
11272 | .enqueue_task = enqueue_task_fair, |
11273 | .dequeue_task = dequeue_task_fair, | |
11274 | .yield_task = yield_task_fair, | |
d95f4122 | 11275 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 11276 | |
2e09bf55 | 11277 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 | 11278 | |
98c2f700 | 11279 | .pick_next_task = __pick_next_task_fair, |
bf0f6f24 | 11280 | .put_prev_task = put_prev_task_fair, |
03b7fad1 | 11281 | .set_next_task = set_next_task_fair, |
bf0f6f24 | 11282 | |
681f3e68 | 11283 | #ifdef CONFIG_SMP |
6e2df058 | 11284 | .balance = balance_fair, |
4ce72a2c | 11285 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 11286 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 11287 | |
0bcdcf28 CE |
11288 | .rq_online = rq_online_fair, |
11289 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 11290 | |
12695578 | 11291 | .task_dead = task_dead_fair, |
c5b28038 | 11292 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 11293 | #endif |
bf0f6f24 | 11294 | |
bf0f6f24 | 11295 | .task_tick = task_tick_fair, |
cd29fe6f | 11296 | .task_fork = task_fork_fair, |
cb469845 SR |
11297 | |
11298 | .prio_changed = prio_changed_fair, | |
da7a735e | 11299 | .switched_from = switched_from_fair, |
cb469845 | 11300 | .switched_to = switched_to_fair, |
810b3817 | 11301 | |
0d721cea PW |
11302 | .get_rr_interval = get_rr_interval_fair, |
11303 | ||
6e998916 SG |
11304 | .update_curr = update_curr_fair, |
11305 | ||
810b3817 | 11306 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 11307 | .task_change_group = task_change_group_fair, |
810b3817 | 11308 | #endif |
982d9cdc PB |
11309 | |
11310 | #ifdef CONFIG_UCLAMP_TASK | |
11311 | .uclamp_enabled = 1, | |
11312 | #endif | |
bf0f6f24 IM |
11313 | }; |
11314 | ||
11315 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 11316 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 11317 | { |
039ae8bc | 11318 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 11319 | |
5973e5b9 | 11320 | rcu_read_lock(); |
039ae8bc | 11321 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 11322 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 11323 | rcu_read_unlock(); |
bf0f6f24 | 11324 | } |
397f2378 SD |
11325 | |
11326 | #ifdef CONFIG_NUMA_BALANCING | |
11327 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
11328 | { | |
11329 | int node; | |
11330 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 11331 | struct numa_group *ng; |
397f2378 | 11332 | |
cb361d8c JH |
11333 | rcu_read_lock(); |
11334 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
11335 | for_each_online_node(node) { |
11336 | if (p->numa_faults) { | |
11337 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
11338 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
11339 | } | |
cb361d8c JH |
11340 | if (ng) { |
11341 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
11342 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
11343 | } |
11344 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
11345 | } | |
cb361d8c | 11346 | rcu_read_unlock(); |
397f2378 SD |
11347 | } |
11348 | #endif /* CONFIG_NUMA_BALANCING */ | |
11349 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
11350 | |
11351 | __init void init_sched_fair_class(void) | |
11352 | { | |
11353 | #ifdef CONFIG_SMP | |
11354 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
11355 | ||
3451d024 | 11356 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 11357 | nohz.next_balance = jiffies; |
f643ea22 | 11358 | nohz.next_blocked = jiffies; |
029632fb | 11359 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
11360 | #endif |
11361 | #endif /* SMP */ | |
11362 | ||
11363 | } | |
3c93a0c0 QY |
11364 | |
11365 | /* | |
11366 | * Helper functions to facilitate extracting info from tracepoints. | |
11367 | */ | |
11368 | ||
11369 | const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq) | |
11370 | { | |
11371 | #ifdef CONFIG_SMP | |
11372 | return cfs_rq ? &cfs_rq->avg : NULL; | |
11373 | #else | |
11374 | return NULL; | |
11375 | #endif | |
11376 | } | |
11377 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg); | |
11378 | ||
11379 | char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len) | |
11380 | { | |
11381 | if (!cfs_rq) { | |
11382 | if (str) | |
11383 | strlcpy(str, "(null)", len); | |
11384 | else | |
11385 | return NULL; | |
11386 | } | |
11387 | ||
11388 | cfs_rq_tg_path(cfs_rq, str, len); | |
11389 | return str; | |
11390 | } | |
11391 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path); | |
11392 | ||
11393 | int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq) | |
11394 | { | |
11395 | return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; | |
11396 | } | |
11397 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu); | |
11398 | ||
11399 | const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq) | |
11400 | { | |
11401 | #ifdef CONFIG_SMP | |
11402 | return rq ? &rq->avg_rt : NULL; | |
11403 | #else | |
11404 | return NULL; | |
11405 | #endif | |
11406 | } | |
11407 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt); | |
11408 | ||
11409 | const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq) | |
11410 | { | |
11411 | #ifdef CONFIG_SMP | |
11412 | return rq ? &rq->avg_dl : NULL; | |
11413 | #else | |
11414 | return NULL; | |
11415 | #endif | |
11416 | } | |
11417 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl); | |
11418 | ||
11419 | const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq) | |
11420 | { | |
11421 | #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ) | |
11422 | return rq ? &rq->avg_irq : NULL; | |
11423 | #else | |
11424 | return NULL; | |
11425 | #endif | |
11426 | } | |
11427 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq); | |
11428 | ||
11429 | int sched_trace_rq_cpu(struct rq *rq) | |
11430 | { | |
11431 | return rq ? cpu_of(rq) : -1; | |
11432 | } | |
11433 | EXPORT_SYMBOL_GPL(sched_trace_rq_cpu); | |
11434 | ||
51cf18c9 VD |
11435 | int sched_trace_rq_cpu_capacity(struct rq *rq) |
11436 | { | |
11437 | return rq ? | |
11438 | #ifdef CONFIG_SMP | |
11439 | rq->cpu_capacity | |
11440 | #else | |
11441 | SCHED_CAPACITY_SCALE | |
11442 | #endif | |
11443 | : -1; | |
11444 | } | |
11445 | EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity); | |
11446 | ||
3c93a0c0 QY |
11447 | const struct cpumask *sched_trace_rd_span(struct root_domain *rd) |
11448 | { | |
11449 | #ifdef CONFIG_SMP | |
11450 | return rd ? rd->span : NULL; | |
11451 | #else | |
11452 | return NULL; | |
11453 | #endif | |
11454 | } | |
11455 | EXPORT_SYMBOL_GPL(sched_trace_rd_span); | |
9d246053 PA |
11456 | |
11457 | int sched_trace_rq_nr_running(struct rq *rq) | |
11458 | { | |
11459 | return rq ? rq->nr_running : -1; | |
11460 | } | |
11461 | EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running); |