]>
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 | 271 | |
b758149c PZ |
272 | /* Walk up scheduling entities hierarchy */ |
273 | #define for_each_sched_entity(se) \ | |
274 | for (; se; se = se->parent) | |
275 | ||
3c93a0c0 QY |
276 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
277 | { | |
278 | if (!path) | |
279 | return; | |
280 | ||
281 | if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) | |
282 | autogroup_path(cfs_rq->tg, path, len); | |
283 | else if (cfs_rq && cfs_rq->tg->css.cgroup) | |
284 | cgroup_path(cfs_rq->tg->css.cgroup, path, len); | |
285 | else | |
286 | strlcpy(path, "(null)", len); | |
287 | } | |
288 | ||
f6783319 | 289 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 290 | { |
5d299eab PZ |
291 | struct rq *rq = rq_of(cfs_rq); |
292 | int cpu = cpu_of(rq); | |
293 | ||
294 | if (cfs_rq->on_list) | |
f6783319 | 295 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
296 | |
297 | cfs_rq->on_list = 1; | |
298 | ||
299 | /* | |
300 | * Ensure we either appear before our parent (if already | |
301 | * enqueued) or force our parent to appear after us when it is | |
302 | * enqueued. The fact that we always enqueue bottom-up | |
303 | * reduces this to two cases and a special case for the root | |
304 | * cfs_rq. Furthermore, it also means that we will always reset | |
305 | * tmp_alone_branch either when the branch is connected | |
306 | * to a tree or when we reach the top of the tree | |
307 | */ | |
308 | if (cfs_rq->tg->parent && | |
309 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 310 | /* |
5d299eab PZ |
311 | * If parent is already on the list, we add the child |
312 | * just before. Thanks to circular linked property of | |
313 | * the list, this means to put the child at the tail | |
314 | * of the list that starts by parent. | |
67e86250 | 315 | */ |
5d299eab PZ |
316 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
317 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
318 | /* | |
319 | * The branch is now connected to its tree so we can | |
320 | * reset tmp_alone_branch to the beginning of the | |
321 | * list. | |
322 | */ | |
323 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 324 | return true; |
5d299eab | 325 | } |
3d4b47b4 | 326 | |
5d299eab PZ |
327 | if (!cfs_rq->tg->parent) { |
328 | /* | |
329 | * cfs rq without parent should be put | |
330 | * at the tail of the list. | |
331 | */ | |
332 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
333 | &rq->leaf_cfs_rq_list); | |
334 | /* | |
335 | * We have reach the top of a tree so we can reset | |
336 | * tmp_alone_branch to the beginning of the list. | |
337 | */ | |
338 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 339 | return true; |
3d4b47b4 | 340 | } |
5d299eab PZ |
341 | |
342 | /* | |
343 | * The parent has not already been added so we want to | |
344 | * make sure that it will be put after us. | |
345 | * tmp_alone_branch points to the begin of the branch | |
346 | * where we will add parent. | |
347 | */ | |
348 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
349 | /* | |
350 | * update tmp_alone_branch to points to the new begin | |
351 | * of the branch | |
352 | */ | |
353 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 354 | return false; |
3d4b47b4 PZ |
355 | } |
356 | ||
357 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
358 | { | |
359 | if (cfs_rq->on_list) { | |
31bc6aea VG |
360 | struct rq *rq = rq_of(cfs_rq); |
361 | ||
362 | /* | |
363 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
364 | * it can happen the tmp_alone_branch points the a leaf that | |
365 | * we finally want to del. In this case, tmp_alone_branch moves | |
366 | * to the prev element but it will point to rq->leaf_cfs_rq_list | |
367 | * at the end of the enqueue. | |
368 | */ | |
369 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
370 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
371 | ||
3d4b47b4 PZ |
372 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
373 | cfs_rq->on_list = 0; | |
374 | } | |
375 | } | |
376 | ||
5d299eab PZ |
377 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
378 | { | |
379 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
380 | } | |
381 | ||
039ae8bc VG |
382 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
383 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ | |
384 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
385 | leaf_cfs_rq_list) | |
b758149c PZ |
386 | |
387 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 388 | static inline struct cfs_rq * |
b758149c PZ |
389 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
390 | { | |
391 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 392 | return se->cfs_rq; |
b758149c | 393 | |
fed14d45 | 394 | return NULL; |
b758149c PZ |
395 | } |
396 | ||
397 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
398 | { | |
399 | return se->parent; | |
400 | } | |
401 | ||
464b7527 PZ |
402 | static void |
403 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
404 | { | |
405 | int se_depth, pse_depth; | |
406 | ||
407 | /* | |
408 | * preemption test can be made between sibling entities who are in the | |
409 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
410 | * both tasks until we find their ancestors who are siblings of common | |
411 | * parent. | |
412 | */ | |
413 | ||
414 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
415 | se_depth = (*se)->depth; |
416 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
417 | |
418 | while (se_depth > pse_depth) { | |
419 | se_depth--; | |
420 | *se = parent_entity(*se); | |
421 | } | |
422 | ||
423 | while (pse_depth > se_depth) { | |
424 | pse_depth--; | |
425 | *pse = parent_entity(*pse); | |
426 | } | |
427 | ||
428 | while (!is_same_group(*se, *pse)) { | |
429 | *se = parent_entity(*se); | |
430 | *pse = parent_entity(*pse); | |
431 | } | |
432 | } | |
433 | ||
30400039 JD |
434 | static int tg_is_idle(struct task_group *tg) |
435 | { | |
436 | return tg->idle > 0; | |
437 | } | |
438 | ||
439 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
440 | { | |
441 | return cfs_rq->idle > 0; | |
442 | } | |
443 | ||
444 | static int se_is_idle(struct sched_entity *se) | |
445 | { | |
446 | if (entity_is_task(se)) | |
447 | return task_has_idle_policy(task_of(se)); | |
448 | return cfs_rq_is_idle(group_cfs_rq(se)); | |
449 | } | |
450 | ||
8f48894f PZ |
451 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
452 | ||
b758149c PZ |
453 | #define for_each_sched_entity(se) \ |
454 | for (; se; se = NULL) | |
bf0f6f24 | 455 | |
3c93a0c0 QY |
456 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
457 | { | |
458 | if (path) | |
459 | strlcpy(path, "(null)", len); | |
460 | } | |
461 | ||
f6783319 | 462 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 463 | { |
f6783319 | 464 | return true; |
3d4b47b4 PZ |
465 | } |
466 | ||
467 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
468 | { | |
469 | } | |
470 | ||
5d299eab PZ |
471 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
472 | { | |
473 | } | |
474 | ||
039ae8bc VG |
475 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
476 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 477 | |
b758149c PZ |
478 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
479 | { | |
480 | return NULL; | |
481 | } | |
482 | ||
464b7527 PZ |
483 | static inline void |
484 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
485 | { | |
486 | } | |
487 | ||
366e7ad6 | 488 | static inline int tg_is_idle(struct task_group *tg) |
30400039 JD |
489 | { |
490 | return 0; | |
491 | } | |
492 | ||
493 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
494 | { | |
495 | return 0; | |
496 | } | |
497 | ||
498 | static int se_is_idle(struct sched_entity *se) | |
499 | { | |
500 | return 0; | |
501 | } | |
502 | ||
b758149c PZ |
503 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
504 | ||
6c16a6dc | 505 | static __always_inline |
9dbdb155 | 506 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
507 | |
508 | /************************************************************** | |
509 | * Scheduling class tree data structure manipulation methods: | |
510 | */ | |
511 | ||
1bf08230 | 512 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 513 | { |
1bf08230 | 514 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 515 | if (delta > 0) |
1bf08230 | 516 | max_vruntime = vruntime; |
02e0431a | 517 | |
1bf08230 | 518 | return max_vruntime; |
02e0431a PZ |
519 | } |
520 | ||
0702e3eb | 521 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
522 | { |
523 | s64 delta = (s64)(vruntime - min_vruntime); | |
524 | if (delta < 0) | |
525 | min_vruntime = vruntime; | |
526 | ||
527 | return min_vruntime; | |
528 | } | |
529 | ||
bf9be9a1 | 530 | static inline bool entity_before(struct sched_entity *a, |
54fdc581 FC |
531 | struct sched_entity *b) |
532 | { | |
533 | return (s64)(a->vruntime - b->vruntime) < 0; | |
534 | } | |
535 | ||
bf9be9a1 PZ |
536 | #define __node_2_se(node) \ |
537 | rb_entry((node), struct sched_entity, run_node) | |
538 | ||
1af5f730 PZ |
539 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
540 | { | |
b60205c7 | 541 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 542 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 543 | |
1af5f730 PZ |
544 | u64 vruntime = cfs_rq->min_vruntime; |
545 | ||
b60205c7 PZ |
546 | if (curr) { |
547 | if (curr->on_rq) | |
548 | vruntime = curr->vruntime; | |
549 | else | |
550 | curr = NULL; | |
551 | } | |
1af5f730 | 552 | |
bfb06889 | 553 | if (leftmost) { /* non-empty tree */ |
bf9be9a1 | 554 | struct sched_entity *se = __node_2_se(leftmost); |
1af5f730 | 555 | |
b60205c7 | 556 | if (!curr) |
1af5f730 PZ |
557 | vruntime = se->vruntime; |
558 | else | |
559 | vruntime = min_vruntime(vruntime, se->vruntime); | |
560 | } | |
561 | ||
1bf08230 | 562 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 563 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
564 | #ifndef CONFIG_64BIT |
565 | smp_wmb(); | |
566 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
567 | #endif | |
1af5f730 PZ |
568 | } |
569 | ||
bf9be9a1 PZ |
570 | static inline bool __entity_less(struct rb_node *a, const struct rb_node *b) |
571 | { | |
572 | return entity_before(__node_2_se(a), __node_2_se(b)); | |
573 | } | |
574 | ||
bf0f6f24 IM |
575 | /* |
576 | * Enqueue an entity into the rb-tree: | |
577 | */ | |
0702e3eb | 578 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 579 | { |
bf9be9a1 | 580 | rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less); |
bf0f6f24 IM |
581 | } |
582 | ||
0702e3eb | 583 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 584 | { |
bfb06889 | 585 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
586 | } |
587 | ||
029632fb | 588 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 589 | { |
bfb06889 | 590 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
591 | |
592 | if (!left) | |
593 | return NULL; | |
594 | ||
bf9be9a1 | 595 | return __node_2_se(left); |
bf0f6f24 IM |
596 | } |
597 | ||
ac53db59 RR |
598 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
599 | { | |
600 | struct rb_node *next = rb_next(&se->run_node); | |
601 | ||
602 | if (!next) | |
603 | return NULL; | |
604 | ||
bf9be9a1 | 605 | return __node_2_se(next); |
ac53db59 RR |
606 | } |
607 | ||
608 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 609 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 610 | { |
bfb06889 | 611 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 612 | |
70eee74b BS |
613 | if (!last) |
614 | return NULL; | |
7eee3e67 | 615 | |
bf9be9a1 | 616 | return __node_2_se(last); |
aeb73b04 PZ |
617 | } |
618 | ||
bf0f6f24 IM |
619 | /************************************************************** |
620 | * Scheduling class statistics methods: | |
621 | */ | |
622 | ||
8a99b683 | 623 | int sched_update_scaling(void) |
b2be5e96 | 624 | { |
58ac93e4 | 625 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 | 626 | |
b2be5e96 PZ |
627 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
628 | sysctl_sched_min_granularity); | |
629 | ||
acb4a848 CE |
630 | #define WRT_SYSCTL(name) \ |
631 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
632 | WRT_SYSCTL(sched_min_granularity); | |
633 | WRT_SYSCTL(sched_latency); | |
634 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
635 | #undef WRT_SYSCTL |
636 | ||
b2be5e96 PZ |
637 | return 0; |
638 | } | |
639 | #endif | |
647e7cac | 640 | |
a7be37ac | 641 | /* |
f9c0b095 | 642 | * delta /= w |
a7be37ac | 643 | */ |
9dbdb155 | 644 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 645 | { |
f9c0b095 | 646 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 647 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
648 | |
649 | return delta; | |
650 | } | |
651 | ||
647e7cac IM |
652 | /* |
653 | * The idea is to set a period in which each task runs once. | |
654 | * | |
532b1858 | 655 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
656 | * this period because otherwise the slices get too small. |
657 | * | |
658 | * p = (nr <= nl) ? l : l*nr/nl | |
659 | */ | |
4d78e7b6 PZ |
660 | static u64 __sched_period(unsigned long nr_running) |
661 | { | |
8e2b0bf3 BF |
662 | if (unlikely(nr_running > sched_nr_latency)) |
663 | return nr_running * sysctl_sched_min_granularity; | |
664 | else | |
665 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
666 | } |
667 | ||
647e7cac IM |
668 | /* |
669 | * We calculate the wall-time slice from the period by taking a part | |
670 | * proportional to the weight. | |
671 | * | |
f9c0b095 | 672 | * s = p*P[w/rw] |
647e7cac | 673 | */ |
6d0f0ebd | 674 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 675 | { |
0c2de3f0 PZ |
676 | unsigned int nr_running = cfs_rq->nr_running; |
677 | u64 slice; | |
678 | ||
679 | if (sched_feat(ALT_PERIOD)) | |
680 | nr_running = rq_of(cfs_rq)->cfs.h_nr_running; | |
681 | ||
682 | slice = __sched_period(nr_running + !se->on_rq); | |
f9c0b095 | 683 | |
0a582440 | 684 | for_each_sched_entity(se) { |
6272d68c | 685 | struct load_weight *load; |
3104bf03 | 686 | struct load_weight lw; |
6272d68c LM |
687 | |
688 | cfs_rq = cfs_rq_of(se); | |
689 | load = &cfs_rq->load; | |
f9c0b095 | 690 | |
0a582440 | 691 | if (unlikely(!se->on_rq)) { |
3104bf03 | 692 | lw = cfs_rq->load; |
0a582440 MG |
693 | |
694 | update_load_add(&lw, se->load.weight); | |
695 | load = &lw; | |
696 | } | |
9dbdb155 | 697 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 | 698 | } |
0c2de3f0 PZ |
699 | |
700 | if (sched_feat(BASE_SLICE)) | |
701 | slice = max(slice, (u64)sysctl_sched_min_granularity); | |
702 | ||
0a582440 | 703 | return slice; |
bf0f6f24 IM |
704 | } |
705 | ||
647e7cac | 706 | /* |
660cc00f | 707 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 708 | * |
f9c0b095 | 709 | * vs = s/w |
647e7cac | 710 | */ |
f9c0b095 | 711 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 712 | { |
f9c0b095 | 713 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
714 | } |
715 | ||
c0796298 | 716 | #include "pelt.h" |
23127296 | 717 | #ifdef CONFIG_SMP |
283e2ed3 | 718 | |
772bd008 | 719 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 720 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 721 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 722 | |
540247fb YD |
723 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
724 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 725 | { |
540247fb | 726 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 727 | |
f207934f PZ |
728 | memset(sa, 0, sizeof(*sa)); |
729 | ||
b5a9b340 | 730 | /* |
dfcb245e | 731 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 732 | * they get a chance to stabilize to their real load level. |
dfcb245e | 733 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
734 | * nothing has been attached to the task group yet. |
735 | */ | |
736 | if (entity_is_task(se)) | |
0dacee1b | 737 | sa->load_avg = scale_load_down(se->load.weight); |
f207934f | 738 | |
9d89c257 | 739 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 740 | } |
7ea241af | 741 | |
df217913 | 742 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 743 | |
2b8c41da YD |
744 | /* |
745 | * With new tasks being created, their initial util_avgs are extrapolated | |
746 | * based on the cfs_rq's current util_avg: | |
747 | * | |
748 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
749 | * | |
750 | * However, in many cases, the above util_avg does not give a desired | |
751 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
752 | * as when the series is a harmonic series. | |
753 | * | |
754 | * To solve this problem, we also cap the util_avg of successive tasks to | |
755 | * only 1/2 of the left utilization budget: | |
756 | * | |
8fe5c5a9 | 757 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 758 | * |
8fe5c5a9 | 759 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 760 | * |
8fe5c5a9 QP |
761 | * For example, for a CPU with 1024 of capacity, a simplest series from |
762 | * the beginning would be like: | |
2b8c41da YD |
763 | * |
764 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
765 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
766 | * | |
767 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
768 | * if util_avg > util_avg_cap. | |
769 | */ | |
d0fe0b9c | 770 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 771 | { |
d0fe0b9c | 772 | struct sched_entity *se = &p->se; |
2b8c41da YD |
773 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
774 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 775 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 776 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
777 | |
778 | if (cap > 0) { | |
779 | if (cfs_rq->avg.util_avg != 0) { | |
780 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
781 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
782 | ||
783 | if (sa->util_avg > cap) | |
784 | sa->util_avg = cap; | |
785 | } else { | |
786 | sa->util_avg = cap; | |
787 | } | |
2b8c41da | 788 | } |
7dc603c9 | 789 | |
e21cf434 | 790 | sa->runnable_avg = sa->util_avg; |
9f683953 | 791 | |
d0fe0b9c DE |
792 | if (p->sched_class != &fair_sched_class) { |
793 | /* | |
794 | * For !fair tasks do: | |
795 | * | |
796 | update_cfs_rq_load_avg(now, cfs_rq); | |
a4f9a0e5 | 797 | attach_entity_load_avg(cfs_rq, se); |
d0fe0b9c DE |
798 | switched_from_fair(rq, p); |
799 | * | |
800 | * such that the next switched_to_fair() has the | |
801 | * expected state. | |
802 | */ | |
803 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
804 | return; | |
7dc603c9 PZ |
805 | } |
806 | ||
df217913 | 807 | attach_entity_cfs_rq(se); |
2b8c41da YD |
808 | } |
809 | ||
7dc603c9 | 810 | #else /* !CONFIG_SMP */ |
540247fb | 811 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
812 | { |
813 | } | |
d0fe0b9c | 814 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
815 | { |
816 | } | |
fe749158 | 817 | static void update_tg_load_avg(struct cfs_rq *cfs_rq) |
3d30544f PZ |
818 | { |
819 | } | |
7dc603c9 | 820 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 821 | |
bf0f6f24 | 822 | /* |
9dbdb155 | 823 | * Update the current task's runtime statistics. |
bf0f6f24 | 824 | */ |
b7cc0896 | 825 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 826 | { |
429d43bc | 827 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 828 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 829 | u64 delta_exec; |
bf0f6f24 IM |
830 | |
831 | if (unlikely(!curr)) | |
832 | return; | |
833 | ||
9dbdb155 PZ |
834 | delta_exec = now - curr->exec_start; |
835 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 836 | return; |
bf0f6f24 | 837 | |
8ebc91d9 | 838 | curr->exec_start = now; |
d842de87 | 839 | |
9dbdb155 PZ |
840 | schedstat_set(curr->statistics.exec_max, |
841 | max(delta_exec, curr->statistics.exec_max)); | |
842 | ||
843 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 844 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
845 | |
846 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
847 | update_min_vruntime(cfs_rq); | |
848 | ||
d842de87 SV |
849 | if (entity_is_task(curr)) { |
850 | struct task_struct *curtask = task_of(curr); | |
851 | ||
f977bb49 | 852 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 853 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 854 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 855 | } |
ec12cb7f PT |
856 | |
857 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
858 | } |
859 | ||
6e998916 SG |
860 | static void update_curr_fair(struct rq *rq) |
861 | { | |
862 | update_curr(cfs_rq_of(&rq->curr->se)); | |
863 | } | |
864 | ||
bf0f6f24 | 865 | static inline void |
5870db5b | 866 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 867 | { |
4fa8d299 JP |
868 | u64 wait_start, prev_wait_start; |
869 | ||
870 | if (!schedstat_enabled()) | |
871 | return; | |
872 | ||
873 | wait_start = rq_clock(rq_of(cfs_rq)); | |
874 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
875 | |
876 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
877 | likely(wait_start > prev_wait_start)) |
878 | wait_start -= prev_wait_start; | |
3ea94de1 | 879 | |
2ed41a55 | 880 | __schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
881 | } |
882 | ||
4fa8d299 | 883 | static inline void |
3ea94de1 JP |
884 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
885 | { | |
886 | struct task_struct *p; | |
cb251765 MG |
887 | u64 delta; |
888 | ||
4fa8d299 JP |
889 | if (!schedstat_enabled()) |
890 | return; | |
891 | ||
b9c88f75 | 892 | /* |
893 | * When the sched_schedstat changes from 0 to 1, some sched se | |
894 | * maybe already in the runqueue, the se->statistics.wait_start | |
895 | * will be 0.So it will let the delta wrong. We need to avoid this | |
896 | * scenario. | |
897 | */ | |
898 | if (unlikely(!schedstat_val(se->statistics.wait_start))) | |
899 | return; | |
900 | ||
4fa8d299 | 901 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); |
3ea94de1 JP |
902 | |
903 | if (entity_is_task(se)) { | |
904 | p = task_of(se); | |
905 | if (task_on_rq_migrating(p)) { | |
906 | /* | |
907 | * Preserve migrating task's wait time so wait_start | |
908 | * time stamp can be adjusted to accumulate wait time | |
909 | * prior to migration. | |
910 | */ | |
2ed41a55 | 911 | __schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
912 | return; |
913 | } | |
914 | trace_sched_stat_wait(p, delta); | |
915 | } | |
916 | ||
2ed41a55 | 917 | __schedstat_set(se->statistics.wait_max, |
4fa8d299 | 918 | max(schedstat_val(se->statistics.wait_max), delta)); |
2ed41a55 PZ |
919 | __schedstat_inc(se->statistics.wait_count); |
920 | __schedstat_add(se->statistics.wait_sum, delta); | |
921 | __schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 922 | } |
3ea94de1 | 923 | |
4fa8d299 | 924 | static inline void |
1a3d027c JP |
925 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
926 | { | |
927 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
928 | u64 sleep_start, block_start; |
929 | ||
930 | if (!schedstat_enabled()) | |
931 | return; | |
932 | ||
933 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
934 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
935 | |
936 | if (entity_is_task(se)) | |
937 | tsk = task_of(se); | |
938 | ||
4fa8d299 JP |
939 | if (sleep_start) { |
940 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
941 | |
942 | if ((s64)delta < 0) | |
943 | delta = 0; | |
944 | ||
4fa8d299 | 945 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
2ed41a55 | 946 | __schedstat_set(se->statistics.sleep_max, delta); |
1a3d027c | 947 | |
2ed41a55 PZ |
948 | __schedstat_set(se->statistics.sleep_start, 0); |
949 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
950 | |
951 | if (tsk) { | |
952 | account_scheduler_latency(tsk, delta >> 10, 1); | |
953 | trace_sched_stat_sleep(tsk, delta); | |
954 | } | |
955 | } | |
4fa8d299 JP |
956 | if (block_start) { |
957 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
958 | |
959 | if ((s64)delta < 0) | |
960 | delta = 0; | |
961 | ||
4fa8d299 | 962 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
2ed41a55 | 963 | __schedstat_set(se->statistics.block_max, delta); |
1a3d027c | 964 | |
2ed41a55 PZ |
965 | __schedstat_set(se->statistics.block_start, 0); |
966 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
967 | |
968 | if (tsk) { | |
969 | if (tsk->in_iowait) { | |
2ed41a55 PZ |
970 | __schedstat_add(se->statistics.iowait_sum, delta); |
971 | __schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
972 | trace_sched_stat_iowait(tsk, delta); |
973 | } | |
974 | ||
975 | trace_sched_stat_blocked(tsk, delta); | |
976 | ||
977 | /* | |
978 | * Blocking time is in units of nanosecs, so shift by | |
979 | * 20 to get a milliseconds-range estimation of the | |
980 | * amount of time that the task spent sleeping: | |
981 | */ | |
982 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
983 | profile_hits(SLEEP_PROFILING, | |
984 | (void *)get_wchan(tsk), | |
985 | delta >> 20); | |
986 | } | |
987 | account_scheduler_latency(tsk, delta >> 10, 0); | |
988 | } | |
989 | } | |
3ea94de1 | 990 | } |
3ea94de1 | 991 | |
bf0f6f24 IM |
992 | /* |
993 | * Task is being enqueued - update stats: | |
994 | */ | |
cb251765 | 995 | static inline void |
1a3d027c | 996 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 997 | { |
4fa8d299 JP |
998 | if (!schedstat_enabled()) |
999 | return; | |
1000 | ||
bf0f6f24 IM |
1001 | /* |
1002 | * Are we enqueueing a waiting task? (for current tasks | |
1003 | * a dequeue/enqueue event is a NOP) | |
1004 | */ | |
429d43bc | 1005 | if (se != cfs_rq->curr) |
5870db5b | 1006 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
1007 | |
1008 | if (flags & ENQUEUE_WAKEUP) | |
1009 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
1010 | } |
1011 | ||
bf0f6f24 | 1012 | static inline void |
cb251765 | 1013 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1014 | { |
4fa8d299 JP |
1015 | |
1016 | if (!schedstat_enabled()) | |
1017 | return; | |
1018 | ||
bf0f6f24 IM |
1019 | /* |
1020 | * Mark the end of the wait period if dequeueing a | |
1021 | * waiting task: | |
1022 | */ | |
429d43bc | 1023 | if (se != cfs_rq->curr) |
9ef0a961 | 1024 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 1025 | |
4fa8d299 JP |
1026 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1027 | struct task_struct *tsk = task_of(se); | |
2f064a59 | 1028 | unsigned int state; |
cb251765 | 1029 | |
2f064a59 PZ |
1030 | /* XXX racy against TTWU */ |
1031 | state = READ_ONCE(tsk->__state); | |
1032 | if (state & TASK_INTERRUPTIBLE) | |
2ed41a55 | 1033 | __schedstat_set(se->statistics.sleep_start, |
4fa8d299 | 1034 | rq_clock(rq_of(cfs_rq))); |
2f064a59 | 1035 | if (state & TASK_UNINTERRUPTIBLE) |
2ed41a55 | 1036 | __schedstat_set(se->statistics.block_start, |
4fa8d299 | 1037 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1038 | } |
cb251765 MG |
1039 | } |
1040 | ||
bf0f6f24 IM |
1041 | /* |
1042 | * We are picking a new current task - update its stats: | |
1043 | */ | |
1044 | static inline void | |
79303e9e | 1045 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1046 | { |
1047 | /* | |
1048 | * We are starting a new run period: | |
1049 | */ | |
78becc27 | 1050 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1051 | } |
1052 | ||
bf0f6f24 IM |
1053 | /************************************************** |
1054 | * Scheduling class queueing methods: | |
1055 | */ | |
1056 | ||
cbee9f88 PZ |
1057 | #ifdef CONFIG_NUMA_BALANCING |
1058 | /* | |
598f0ec0 MG |
1059 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1060 | * calculated based on the tasks virtual memory size and | |
1061 | * numa_balancing_scan_size. | |
cbee9f88 | 1062 | */ |
598f0ec0 MG |
1063 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1064 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1065 | |
1066 | /* Portion of address space to scan in MB */ | |
1067 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1068 | |
4b96a29b PZ |
1069 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1070 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1071 | ||
b5dd77c8 | 1072 | struct numa_group { |
c45a7795 | 1073 | refcount_t refcount; |
b5dd77c8 RR |
1074 | |
1075 | spinlock_t lock; /* nr_tasks, tasks */ | |
1076 | int nr_tasks; | |
1077 | pid_t gid; | |
1078 | int active_nodes; | |
1079 | ||
1080 | struct rcu_head rcu; | |
1081 | unsigned long total_faults; | |
1082 | unsigned long max_faults_cpu; | |
1083 | /* | |
1084 | * Faults_cpu is used to decide whether memory should move | |
1085 | * towards the CPU. As a consequence, these stats are weighted | |
1086 | * more by CPU use than by memory faults. | |
1087 | */ | |
1088 | unsigned long *faults_cpu; | |
04f5c362 | 1089 | unsigned long faults[]; |
b5dd77c8 RR |
1090 | }; |
1091 | ||
cb361d8c JH |
1092 | /* |
1093 | * For functions that can be called in multiple contexts that permit reading | |
1094 | * ->numa_group (see struct task_struct for locking rules). | |
1095 | */ | |
1096 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1097 | { | |
1098 | return rcu_dereference_check(p->numa_group, p == current || | |
9ef7e7e3 | 1099 | (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu))); |
cb361d8c JH |
1100 | } |
1101 | ||
1102 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1103 | { | |
1104 | return rcu_dereference_protected(p->numa_group, p == current); | |
1105 | } | |
1106 | ||
b5dd77c8 RR |
1107 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1108 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1109 | ||
598f0ec0 MG |
1110 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1111 | { | |
1112 | unsigned long rss = 0; | |
1113 | unsigned long nr_scan_pages; | |
1114 | ||
1115 | /* | |
1116 | * Calculations based on RSS as non-present and empty pages are skipped | |
1117 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1118 | * on resident pages | |
1119 | */ | |
1120 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1121 | rss = get_mm_rss(p->mm); | |
1122 | if (!rss) | |
1123 | rss = nr_scan_pages; | |
1124 | ||
1125 | rss = round_up(rss, nr_scan_pages); | |
1126 | return rss / nr_scan_pages; | |
1127 | } | |
1128 | ||
3b03706f | 1129 | /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ |
598f0ec0 MG |
1130 | #define MAX_SCAN_WINDOW 2560 |
1131 | ||
1132 | static unsigned int task_scan_min(struct task_struct *p) | |
1133 | { | |
316c1608 | 1134 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1135 | unsigned int scan, floor; |
1136 | unsigned int windows = 1; | |
1137 | ||
64192658 KT |
1138 | if (scan_size < MAX_SCAN_WINDOW) |
1139 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1140 | floor = 1000 / windows; |
1141 | ||
1142 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1143 | return max_t(unsigned int, floor, scan); | |
1144 | } | |
1145 | ||
b5dd77c8 RR |
1146 | static unsigned int task_scan_start(struct task_struct *p) |
1147 | { | |
1148 | unsigned long smin = task_scan_min(p); | |
1149 | unsigned long period = smin; | |
cb361d8c | 1150 | struct numa_group *ng; |
b5dd77c8 RR |
1151 | |
1152 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1153 | rcu_read_lock(); |
1154 | ng = rcu_dereference(p->numa_group); | |
1155 | if (ng) { | |
b5dd77c8 RR |
1156 | unsigned long shared = group_faults_shared(ng); |
1157 | unsigned long private = group_faults_priv(ng); | |
1158 | ||
c45a7795 | 1159 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1160 | period *= shared + 1; |
1161 | period /= private + shared + 1; | |
1162 | } | |
cb361d8c | 1163 | rcu_read_unlock(); |
b5dd77c8 RR |
1164 | |
1165 | return max(smin, period); | |
1166 | } | |
1167 | ||
598f0ec0 MG |
1168 | static unsigned int task_scan_max(struct task_struct *p) |
1169 | { | |
b5dd77c8 RR |
1170 | unsigned long smin = task_scan_min(p); |
1171 | unsigned long smax; | |
cb361d8c | 1172 | struct numa_group *ng; |
598f0ec0 MG |
1173 | |
1174 | /* Watch for min being lower than max due to floor calculations */ | |
1175 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1176 | |
1177 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1178 | ng = deref_curr_numa_group(p); |
1179 | if (ng) { | |
b5dd77c8 RR |
1180 | unsigned long shared = group_faults_shared(ng); |
1181 | unsigned long private = group_faults_priv(ng); | |
1182 | unsigned long period = smax; | |
1183 | ||
c45a7795 | 1184 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1185 | period *= shared + 1; |
1186 | period /= private + shared + 1; | |
1187 | ||
1188 | smax = max(smax, period); | |
1189 | } | |
1190 | ||
598f0ec0 MG |
1191 | return max(smin, smax); |
1192 | } | |
1193 | ||
0ec8aa00 PZ |
1194 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1195 | { | |
98fa15f3 | 1196 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1197 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1198 | } | |
1199 | ||
1200 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1201 | { | |
98fa15f3 | 1202 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1203 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1204 | } | |
1205 | ||
be1e4e76 RR |
1206 | /* Shared or private faults. */ |
1207 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1208 | ||
1209 | /* Memory and CPU locality */ | |
1210 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1211 | ||
1212 | /* Averaged statistics, and temporary buffers. */ | |
1213 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1214 | ||
e29cf08b MG |
1215 | pid_t task_numa_group_id(struct task_struct *p) |
1216 | { | |
cb361d8c JH |
1217 | struct numa_group *ng; |
1218 | pid_t gid = 0; | |
1219 | ||
1220 | rcu_read_lock(); | |
1221 | ng = rcu_dereference(p->numa_group); | |
1222 | if (ng) | |
1223 | gid = ng->gid; | |
1224 | rcu_read_unlock(); | |
1225 | ||
1226 | return gid; | |
e29cf08b MG |
1227 | } |
1228 | ||
44dba3d5 | 1229 | /* |
97fb7a0a | 1230 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1231 | * occupy the first half of the array. The second half of the |
1232 | * array is for current counters, which are averaged into the | |
1233 | * first set by task_numa_placement. | |
1234 | */ | |
1235 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1236 | { |
44dba3d5 | 1237 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1238 | } |
1239 | ||
1240 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1241 | { | |
44dba3d5 | 1242 | if (!p->numa_faults) |
ac8e895b MG |
1243 | return 0; |
1244 | ||
44dba3d5 IM |
1245 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1246 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1247 | } |
1248 | ||
83e1d2cd MG |
1249 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1250 | { | |
cb361d8c JH |
1251 | struct numa_group *ng = deref_task_numa_group(p); |
1252 | ||
1253 | if (!ng) | |
83e1d2cd MG |
1254 | return 0; |
1255 | ||
cb361d8c JH |
1256 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1257 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1258 | } |
1259 | ||
20e07dea RR |
1260 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1261 | { | |
44dba3d5 IM |
1262 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1263 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1264 | } |
1265 | ||
b5dd77c8 RR |
1266 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1267 | { | |
1268 | unsigned long faults = 0; | |
1269 | int node; | |
1270 | ||
1271 | for_each_online_node(node) { | |
1272 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1273 | } | |
1274 | ||
1275 | return faults; | |
1276 | } | |
1277 | ||
1278 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1279 | { | |
1280 | unsigned long faults = 0; | |
1281 | int node; | |
1282 | ||
1283 | for_each_online_node(node) { | |
1284 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1285 | } | |
1286 | ||
1287 | return faults; | |
1288 | } | |
1289 | ||
4142c3eb RR |
1290 | /* |
1291 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1292 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1293 | * between these nodes are slowed down, to allow things to settle down. | |
1294 | */ | |
1295 | #define ACTIVE_NODE_FRACTION 3 | |
1296 | ||
1297 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1298 | { | |
1299 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1300 | } | |
1301 | ||
6c6b1193 RR |
1302 | /* Handle placement on systems where not all nodes are directly connected. */ |
1303 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1304 | int maxdist, bool task) | |
1305 | { | |
1306 | unsigned long score = 0; | |
1307 | int node; | |
1308 | ||
1309 | /* | |
1310 | * All nodes are directly connected, and the same distance | |
1311 | * from each other. No need for fancy placement algorithms. | |
1312 | */ | |
1313 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1314 | return 0; | |
1315 | ||
1316 | /* | |
1317 | * This code is called for each node, introducing N^2 complexity, | |
1318 | * which should be ok given the number of nodes rarely exceeds 8. | |
1319 | */ | |
1320 | for_each_online_node(node) { | |
1321 | unsigned long faults; | |
1322 | int dist = node_distance(nid, node); | |
1323 | ||
1324 | /* | |
1325 | * The furthest away nodes in the system are not interesting | |
1326 | * for placement; nid was already counted. | |
1327 | */ | |
1328 | if (dist == sched_max_numa_distance || node == nid) | |
1329 | continue; | |
1330 | ||
1331 | /* | |
1332 | * On systems with a backplane NUMA topology, compare groups | |
1333 | * of nodes, and move tasks towards the group with the most | |
1334 | * memory accesses. When comparing two nodes at distance | |
1335 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1336 | * of each group. Skip other nodes. | |
1337 | */ | |
1338 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
0ee7e74d | 1339 | dist >= maxdist) |
6c6b1193 RR |
1340 | continue; |
1341 | ||
1342 | /* Add up the faults from nearby nodes. */ | |
1343 | if (task) | |
1344 | faults = task_faults(p, node); | |
1345 | else | |
1346 | faults = group_faults(p, node); | |
1347 | ||
1348 | /* | |
1349 | * On systems with a glueless mesh NUMA topology, there are | |
1350 | * no fixed "groups of nodes". Instead, nodes that are not | |
1351 | * directly connected bounce traffic through intermediate | |
1352 | * nodes; a numa_group can occupy any set of nodes. | |
1353 | * The further away a node is, the less the faults count. | |
1354 | * This seems to result in good task placement. | |
1355 | */ | |
1356 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1357 | faults *= (sched_max_numa_distance - dist); | |
1358 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1359 | } | |
1360 | ||
1361 | score += faults; | |
1362 | } | |
1363 | ||
1364 | return score; | |
1365 | } | |
1366 | ||
83e1d2cd MG |
1367 | /* |
1368 | * These return the fraction of accesses done by a particular task, or | |
1369 | * task group, on a particular numa node. The group weight is given a | |
1370 | * larger multiplier, in order to group tasks together that are almost | |
1371 | * evenly spread out between numa nodes. | |
1372 | */ | |
7bd95320 RR |
1373 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1374 | int dist) | |
83e1d2cd | 1375 | { |
7bd95320 | 1376 | unsigned long faults, total_faults; |
83e1d2cd | 1377 | |
44dba3d5 | 1378 | if (!p->numa_faults) |
83e1d2cd MG |
1379 | return 0; |
1380 | ||
1381 | total_faults = p->total_numa_faults; | |
1382 | ||
1383 | if (!total_faults) | |
1384 | return 0; | |
1385 | ||
7bd95320 | 1386 | faults = task_faults(p, nid); |
6c6b1193 RR |
1387 | faults += score_nearby_nodes(p, nid, dist, true); |
1388 | ||
7bd95320 | 1389 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1390 | } |
1391 | ||
7bd95320 RR |
1392 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1393 | int dist) | |
83e1d2cd | 1394 | { |
cb361d8c | 1395 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1396 | unsigned long faults, total_faults; |
1397 | ||
cb361d8c | 1398 | if (!ng) |
7bd95320 RR |
1399 | return 0; |
1400 | ||
cb361d8c | 1401 | total_faults = ng->total_faults; |
7bd95320 RR |
1402 | |
1403 | if (!total_faults) | |
83e1d2cd MG |
1404 | return 0; |
1405 | ||
7bd95320 | 1406 | faults = group_faults(p, nid); |
6c6b1193 RR |
1407 | faults += score_nearby_nodes(p, nid, dist, false); |
1408 | ||
7bd95320 | 1409 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1410 | } |
1411 | ||
10f39042 RR |
1412 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1413 | int src_nid, int dst_cpu) | |
1414 | { | |
cb361d8c | 1415 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1416 | int dst_nid = cpu_to_node(dst_cpu); |
1417 | int last_cpupid, this_cpupid; | |
1418 | ||
1419 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
37355bdc MG |
1420 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1421 | ||
1422 | /* | |
1423 | * Allow first faults or private faults to migrate immediately early in | |
1424 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1425 | * two full passes of the "multi-stage node selection" test that is | |
1426 | * executed below. | |
1427 | */ | |
98fa15f3 | 1428 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1429 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1430 | return true; | |
10f39042 RR |
1431 | |
1432 | /* | |
1433 | * Multi-stage node selection is used in conjunction with a periodic | |
1434 | * migration fault to build a temporal task<->page relation. By using | |
1435 | * a two-stage filter we remove short/unlikely relations. | |
1436 | * | |
1437 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1438 | * a task's usage of a particular page (n_p) per total usage of this | |
1439 | * page (n_t) (in a given time-span) to a probability. | |
1440 | * | |
1441 | * Our periodic faults will sample this probability and getting the | |
1442 | * same result twice in a row, given these samples are fully | |
1443 | * independent, is then given by P(n)^2, provided our sample period | |
1444 | * is sufficiently short compared to the usage pattern. | |
1445 | * | |
1446 | * This quadric squishes small probabilities, making it less likely we | |
1447 | * act on an unlikely task<->page relation. | |
1448 | */ | |
10f39042 RR |
1449 | if (!cpupid_pid_unset(last_cpupid) && |
1450 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1451 | return false; | |
1452 | ||
1453 | /* Always allow migrate on private faults */ | |
1454 | if (cpupid_match_pid(p, last_cpupid)) | |
1455 | return true; | |
1456 | ||
1457 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1458 | if (!ng) | |
1459 | return true; | |
1460 | ||
1461 | /* | |
4142c3eb RR |
1462 | * Destination node is much more heavily used than the source |
1463 | * node? Allow migration. | |
10f39042 | 1464 | */ |
4142c3eb RR |
1465 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1466 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1467 | return true; |
1468 | ||
1469 | /* | |
4142c3eb RR |
1470 | * Distribute memory according to CPU & memory use on each node, |
1471 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1472 | * | |
1473 | * faults_cpu(dst) 3 faults_cpu(src) | |
1474 | * --------------- * - > --------------- | |
1475 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1476 | */ |
4142c3eb RR |
1477 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1478 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1479 | } |
1480 | ||
6499b1b2 VG |
1481 | /* |
1482 | * 'numa_type' describes the node at the moment of load balancing. | |
1483 | */ | |
1484 | enum numa_type { | |
1485 | /* The node has spare capacity that can be used to run more tasks. */ | |
1486 | node_has_spare = 0, | |
1487 | /* | |
1488 | * The node is fully used and the tasks don't compete for more CPU | |
1489 | * cycles. Nevertheless, some tasks might wait before running. | |
1490 | */ | |
1491 | node_fully_busy, | |
1492 | /* | |
1493 | * The node is overloaded and can't provide expected CPU cycles to all | |
1494 | * tasks. | |
1495 | */ | |
1496 | node_overloaded | |
1497 | }; | |
58d081b5 | 1498 | |
fb13c7ee | 1499 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1500 | struct numa_stats { |
1501 | unsigned long load; | |
8e0e0eda | 1502 | unsigned long runnable; |
6499b1b2 | 1503 | unsigned long util; |
fb13c7ee | 1504 | /* Total compute capacity of CPUs on a node */ |
5ef20ca1 | 1505 | unsigned long compute_capacity; |
6499b1b2 VG |
1506 | unsigned int nr_running; |
1507 | unsigned int weight; | |
1508 | enum numa_type node_type; | |
ff7db0bf | 1509 | int idle_cpu; |
58d081b5 | 1510 | }; |
e6628d5b | 1511 | |
ff7db0bf MG |
1512 | static inline bool is_core_idle(int cpu) |
1513 | { | |
1514 | #ifdef CONFIG_SCHED_SMT | |
1515 | int sibling; | |
1516 | ||
1517 | for_each_cpu(sibling, cpu_smt_mask(cpu)) { | |
1518 | if (cpu == sibling) | |
1519 | continue; | |
1520 | ||
1c6829cf | 1521 | if (!idle_cpu(sibling)) |
ff7db0bf MG |
1522 | return false; |
1523 | } | |
1524 | #endif | |
1525 | ||
1526 | return true; | |
1527 | } | |
1528 | ||
58d081b5 MG |
1529 | struct task_numa_env { |
1530 | struct task_struct *p; | |
e6628d5b | 1531 | |
58d081b5 MG |
1532 | int src_cpu, src_nid; |
1533 | int dst_cpu, dst_nid; | |
e6628d5b | 1534 | |
58d081b5 | 1535 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1536 | |
40ea2b42 | 1537 | int imbalance_pct; |
7bd95320 | 1538 | int dist; |
fb13c7ee MG |
1539 | |
1540 | struct task_struct *best_task; | |
1541 | long best_imp; | |
58d081b5 MG |
1542 | int best_cpu; |
1543 | }; | |
1544 | ||
6499b1b2 | 1545 | static unsigned long cpu_load(struct rq *rq); |
8e0e0eda | 1546 | static unsigned long cpu_runnable(struct rq *rq); |
6499b1b2 | 1547 | static unsigned long cpu_util(int cpu); |
7d2b5dd0 MG |
1548 | static inline long adjust_numa_imbalance(int imbalance, |
1549 | int dst_running, int dst_weight); | |
6499b1b2 VG |
1550 | |
1551 | static inline enum | |
1552 | numa_type numa_classify(unsigned int imbalance_pct, | |
1553 | struct numa_stats *ns) | |
1554 | { | |
1555 | if ((ns->nr_running > ns->weight) && | |
8e0e0eda VG |
1556 | (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || |
1557 | ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) | |
6499b1b2 VG |
1558 | return node_overloaded; |
1559 | ||
1560 | if ((ns->nr_running < ns->weight) || | |
8e0e0eda VG |
1561 | (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && |
1562 | ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) | |
6499b1b2 VG |
1563 | return node_has_spare; |
1564 | ||
1565 | return node_fully_busy; | |
1566 | } | |
1567 | ||
76c389ab VS |
1568 | #ifdef CONFIG_SCHED_SMT |
1569 | /* Forward declarations of select_idle_sibling helpers */ | |
1570 | static inline bool test_idle_cores(int cpu, bool def); | |
ff7db0bf MG |
1571 | static inline int numa_idle_core(int idle_core, int cpu) |
1572 | { | |
ff7db0bf MG |
1573 | if (!static_branch_likely(&sched_smt_present) || |
1574 | idle_core >= 0 || !test_idle_cores(cpu, false)) | |
1575 | return idle_core; | |
1576 | ||
1577 | /* | |
1578 | * Prefer cores instead of packing HT siblings | |
1579 | * and triggering future load balancing. | |
1580 | */ | |
1581 | if (is_core_idle(cpu)) | |
1582 | idle_core = cpu; | |
ff7db0bf MG |
1583 | |
1584 | return idle_core; | |
1585 | } | |
76c389ab VS |
1586 | #else |
1587 | static inline int numa_idle_core(int idle_core, int cpu) | |
1588 | { | |
1589 | return idle_core; | |
1590 | } | |
1591 | #endif | |
ff7db0bf | 1592 | |
6499b1b2 | 1593 | /* |
ff7db0bf MG |
1594 | * Gather all necessary information to make NUMA balancing placement |
1595 | * decisions that are compatible with standard load balancer. This | |
1596 | * borrows code and logic from update_sg_lb_stats but sharing a | |
1597 | * common implementation is impractical. | |
6499b1b2 VG |
1598 | */ |
1599 | static void update_numa_stats(struct task_numa_env *env, | |
ff7db0bf MG |
1600 | struct numa_stats *ns, int nid, |
1601 | bool find_idle) | |
6499b1b2 | 1602 | { |
ff7db0bf | 1603 | int cpu, idle_core = -1; |
6499b1b2 VG |
1604 | |
1605 | memset(ns, 0, sizeof(*ns)); | |
ff7db0bf MG |
1606 | ns->idle_cpu = -1; |
1607 | ||
0621df31 | 1608 | rcu_read_lock(); |
6499b1b2 VG |
1609 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
1610 | struct rq *rq = cpu_rq(cpu); | |
1611 | ||
1612 | ns->load += cpu_load(rq); | |
8e0e0eda | 1613 | ns->runnable += cpu_runnable(rq); |
6499b1b2 VG |
1614 | ns->util += cpu_util(cpu); |
1615 | ns->nr_running += rq->cfs.h_nr_running; | |
1616 | ns->compute_capacity += capacity_of(cpu); | |
ff7db0bf MG |
1617 | |
1618 | if (find_idle && !rq->nr_running && idle_cpu(cpu)) { | |
1619 | if (READ_ONCE(rq->numa_migrate_on) || | |
1620 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) | |
1621 | continue; | |
1622 | ||
1623 | if (ns->idle_cpu == -1) | |
1624 | ns->idle_cpu = cpu; | |
1625 | ||
1626 | idle_core = numa_idle_core(idle_core, cpu); | |
1627 | } | |
6499b1b2 | 1628 | } |
0621df31 | 1629 | rcu_read_unlock(); |
6499b1b2 VG |
1630 | |
1631 | ns->weight = cpumask_weight(cpumask_of_node(nid)); | |
1632 | ||
1633 | ns->node_type = numa_classify(env->imbalance_pct, ns); | |
ff7db0bf MG |
1634 | |
1635 | if (idle_core >= 0) | |
1636 | ns->idle_cpu = idle_core; | |
6499b1b2 VG |
1637 | } |
1638 | ||
fb13c7ee MG |
1639 | static void task_numa_assign(struct task_numa_env *env, |
1640 | struct task_struct *p, long imp) | |
1641 | { | |
a4739eca SD |
1642 | struct rq *rq = cpu_rq(env->dst_cpu); |
1643 | ||
5fb52dd9 MG |
1644 | /* Check if run-queue part of active NUMA balance. */ |
1645 | if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { | |
1646 | int cpu; | |
1647 | int start = env->dst_cpu; | |
1648 | ||
1649 | /* Find alternative idle CPU. */ | |
1650 | for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { | |
1651 | if (cpu == env->best_cpu || !idle_cpu(cpu) || | |
1652 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { | |
1653 | continue; | |
1654 | } | |
1655 | ||
1656 | env->dst_cpu = cpu; | |
1657 | rq = cpu_rq(env->dst_cpu); | |
1658 | if (!xchg(&rq->numa_migrate_on, 1)) | |
1659 | goto assign; | |
1660 | } | |
1661 | ||
1662 | /* Failed to find an alternative idle CPU */ | |
a4739eca | 1663 | return; |
5fb52dd9 | 1664 | } |
a4739eca | 1665 | |
5fb52dd9 | 1666 | assign: |
a4739eca SD |
1667 | /* |
1668 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1669 | * found a better CPU to move/swap. | |
1670 | */ | |
5fb52dd9 | 1671 | if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { |
a4739eca SD |
1672 | rq = cpu_rq(env->best_cpu); |
1673 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1674 | } | |
1675 | ||
fb13c7ee MG |
1676 | if (env->best_task) |
1677 | put_task_struct(env->best_task); | |
bac78573 ON |
1678 | if (p) |
1679 | get_task_struct(p); | |
fb13c7ee MG |
1680 | |
1681 | env->best_task = p; | |
1682 | env->best_imp = imp; | |
1683 | env->best_cpu = env->dst_cpu; | |
1684 | } | |
1685 | ||
28a21745 | 1686 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1687 | struct task_numa_env *env) |
1688 | { | |
e4991b24 RR |
1689 | long imb, old_imb; |
1690 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1691 | long src_capacity, dst_capacity; |
1692 | ||
1693 | /* | |
1694 | * The load is corrected for the CPU capacity available on each node. | |
1695 | * | |
1696 | * src_load dst_load | |
1697 | * ------------ vs --------- | |
1698 | * src_capacity dst_capacity | |
1699 | */ | |
1700 | src_capacity = env->src_stats.compute_capacity; | |
1701 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1702 | |
5f95ba7a | 1703 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1704 | |
28a21745 | 1705 | orig_src_load = env->src_stats.load; |
e4991b24 | 1706 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1707 | |
5f95ba7a | 1708 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1709 | |
1710 | /* Would this change make things worse? */ | |
1711 | return (imb > old_imb); | |
e63da036 RR |
1712 | } |
1713 | ||
6fd98e77 SD |
1714 | /* |
1715 | * Maximum NUMA importance can be 1998 (2*999); | |
1716 | * SMALLIMP @ 30 would be close to 1998/64. | |
1717 | * Used to deter task migration. | |
1718 | */ | |
1719 | #define SMALLIMP 30 | |
1720 | ||
fb13c7ee MG |
1721 | /* |
1722 | * This checks if the overall compute and NUMA accesses of the system would | |
1723 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1724 | * into account that it might be best if task running on the dst_cpu should | |
1725 | * be exchanged with the source task | |
1726 | */ | |
a0f03b61 | 1727 | static bool task_numa_compare(struct task_numa_env *env, |
305c1fac | 1728 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1729 | { |
cb361d8c | 1730 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 1731 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 1732 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 1733 | struct task_struct *cur; |
28a21745 | 1734 | long src_load, dst_load; |
7bd95320 | 1735 | int dist = env->dist; |
cb361d8c JH |
1736 | long moveimp = imp; |
1737 | long load; | |
a0f03b61 | 1738 | bool stopsearch = false; |
fb13c7ee | 1739 | |
a4739eca | 1740 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
a0f03b61 | 1741 | return false; |
a4739eca | 1742 | |
fb13c7ee | 1743 | rcu_read_lock(); |
154abafc | 1744 | cur = rcu_dereference(dst_rq->curr); |
bac78573 | 1745 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) |
fb13c7ee MG |
1746 | cur = NULL; |
1747 | ||
7af68335 PZ |
1748 | /* |
1749 | * Because we have preemption enabled we can get migrated around and | |
1750 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1751 | */ | |
a0f03b61 MG |
1752 | if (cur == env->p) { |
1753 | stopsearch = true; | |
7af68335 | 1754 | goto unlock; |
a0f03b61 | 1755 | } |
7af68335 | 1756 | |
305c1fac | 1757 | if (!cur) { |
6fd98e77 | 1758 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1759 | goto assign; |
1760 | else | |
1761 | goto unlock; | |
1762 | } | |
1763 | ||
88cca72c MG |
1764 | /* Skip this swap candidate if cannot move to the source cpu. */ |
1765 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) | |
1766 | goto unlock; | |
1767 | ||
1768 | /* | |
1769 | * Skip this swap candidate if it is not moving to its preferred | |
1770 | * node and the best task is. | |
1771 | */ | |
1772 | if (env->best_task && | |
1773 | env->best_task->numa_preferred_nid == env->src_nid && | |
1774 | cur->numa_preferred_nid != env->src_nid) { | |
1775 | goto unlock; | |
1776 | } | |
1777 | ||
fb13c7ee MG |
1778 | /* |
1779 | * "imp" is the fault differential for the source task between the | |
1780 | * source and destination node. Calculate the total differential for | |
1781 | * the source task and potential destination task. The more negative | |
305c1fac | 1782 | * the value is, the more remote accesses that would be expected to |
fb13c7ee | 1783 | * be incurred if the tasks were swapped. |
88cca72c | 1784 | * |
305c1fac SD |
1785 | * If dst and source tasks are in the same NUMA group, or not |
1786 | * in any group then look only at task weights. | |
1787 | */ | |
cb361d8c JH |
1788 | cur_ng = rcu_dereference(cur->numa_group); |
1789 | if (cur_ng == p_ng) { | |
305c1fac SD |
1790 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1791 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1792 | /* |
305c1fac SD |
1793 | * Add some hysteresis to prevent swapping the |
1794 | * tasks within a group over tiny differences. | |
887c290e | 1795 | */ |
cb361d8c | 1796 | if (cur_ng) |
305c1fac SD |
1797 | imp -= imp / 16; |
1798 | } else { | |
1799 | /* | |
1800 | * Compare the group weights. If a task is all by itself | |
1801 | * (not part of a group), use the task weight instead. | |
1802 | */ | |
cb361d8c | 1803 | if (cur_ng && p_ng) |
305c1fac SD |
1804 | imp += group_weight(cur, env->src_nid, dist) - |
1805 | group_weight(cur, env->dst_nid, dist); | |
1806 | else | |
1807 | imp += task_weight(cur, env->src_nid, dist) - | |
1808 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
1809 | } |
1810 | ||
88cca72c MG |
1811 | /* Discourage picking a task already on its preferred node */ |
1812 | if (cur->numa_preferred_nid == env->dst_nid) | |
1813 | imp -= imp / 16; | |
1814 | ||
1815 | /* | |
1816 | * Encourage picking a task that moves to its preferred node. | |
1817 | * This potentially makes imp larger than it's maximum of | |
1818 | * 1998 (see SMALLIMP and task_weight for why) but in this | |
1819 | * case, it does not matter. | |
1820 | */ | |
1821 | if (cur->numa_preferred_nid == env->src_nid) | |
1822 | imp += imp / 8; | |
1823 | ||
305c1fac | 1824 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 1825 | imp = moveimp; |
305c1fac | 1826 | cur = NULL; |
fb13c7ee | 1827 | goto assign; |
305c1fac | 1828 | } |
fb13c7ee | 1829 | |
88cca72c MG |
1830 | /* |
1831 | * Prefer swapping with a task moving to its preferred node over a | |
1832 | * task that is not. | |
1833 | */ | |
1834 | if (env->best_task && cur->numa_preferred_nid == env->src_nid && | |
1835 | env->best_task->numa_preferred_nid != env->src_nid) { | |
1836 | goto assign; | |
1837 | } | |
1838 | ||
6fd98e77 SD |
1839 | /* |
1840 | * If the NUMA importance is less than SMALLIMP, | |
1841 | * task migration might only result in ping pong | |
1842 | * of tasks and also hurt performance due to cache | |
1843 | * misses. | |
1844 | */ | |
1845 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
1846 | goto unlock; | |
1847 | ||
fb13c7ee MG |
1848 | /* |
1849 | * In the overloaded case, try and keep the load balanced. | |
1850 | */ | |
305c1fac SD |
1851 | load = task_h_load(env->p) - task_h_load(cur); |
1852 | if (!load) | |
1853 | goto assign; | |
1854 | ||
e720fff6 PZ |
1855 | dst_load = env->dst_stats.load + load; |
1856 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1857 | |
28a21745 | 1858 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1859 | goto unlock; |
1860 | ||
305c1fac | 1861 | assign: |
ff7db0bf | 1862 | /* Evaluate an idle CPU for a task numa move. */ |
10e2f1ac | 1863 | if (!cur) { |
ff7db0bf MG |
1864 | int cpu = env->dst_stats.idle_cpu; |
1865 | ||
1866 | /* Nothing cached so current CPU went idle since the search. */ | |
1867 | if (cpu < 0) | |
1868 | cpu = env->dst_cpu; | |
1869 | ||
10e2f1ac | 1870 | /* |
ff7db0bf MG |
1871 | * If the CPU is no longer truly idle and the previous best CPU |
1872 | * is, keep using it. | |
10e2f1ac | 1873 | */ |
ff7db0bf MG |
1874 | if (!idle_cpu(cpu) && env->best_cpu >= 0 && |
1875 | idle_cpu(env->best_cpu)) { | |
1876 | cpu = env->best_cpu; | |
1877 | } | |
1878 | ||
ff7db0bf | 1879 | env->dst_cpu = cpu; |
10e2f1ac | 1880 | } |
ba7e5a27 | 1881 | |
fb13c7ee | 1882 | task_numa_assign(env, cur, imp); |
a0f03b61 MG |
1883 | |
1884 | /* | |
1885 | * If a move to idle is allowed because there is capacity or load | |
1886 | * balance improves then stop the search. While a better swap | |
1887 | * candidate may exist, a search is not free. | |
1888 | */ | |
1889 | if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) | |
1890 | stopsearch = true; | |
1891 | ||
1892 | /* | |
1893 | * If a swap candidate must be identified and the current best task | |
1894 | * moves its preferred node then stop the search. | |
1895 | */ | |
1896 | if (!maymove && env->best_task && | |
1897 | env->best_task->numa_preferred_nid == env->src_nid) { | |
1898 | stopsearch = true; | |
1899 | } | |
fb13c7ee MG |
1900 | unlock: |
1901 | rcu_read_unlock(); | |
a0f03b61 MG |
1902 | |
1903 | return stopsearch; | |
fb13c7ee MG |
1904 | } |
1905 | ||
887c290e RR |
1906 | static void task_numa_find_cpu(struct task_numa_env *env, |
1907 | long taskimp, long groupimp) | |
2c8a50aa | 1908 | { |
305c1fac | 1909 | bool maymove = false; |
2c8a50aa MG |
1910 | int cpu; |
1911 | ||
305c1fac | 1912 | /* |
fb86f5b2 MG |
1913 | * If dst node has spare capacity, then check if there is an |
1914 | * imbalance that would be overruled by the load balancer. | |
305c1fac | 1915 | */ |
fb86f5b2 MG |
1916 | if (env->dst_stats.node_type == node_has_spare) { |
1917 | unsigned int imbalance; | |
1918 | int src_running, dst_running; | |
1919 | ||
1920 | /* | |
1921 | * Would movement cause an imbalance? Note that if src has | |
1922 | * more running tasks that the imbalance is ignored as the | |
1923 | * move improves the imbalance from the perspective of the | |
1924 | * CPU load balancer. | |
1925 | * */ | |
1926 | src_running = env->src_stats.nr_running - 1; | |
1927 | dst_running = env->dst_stats.nr_running + 1; | |
1928 | imbalance = max(0, dst_running - src_running); | |
7d2b5dd0 MG |
1929 | imbalance = adjust_numa_imbalance(imbalance, dst_running, |
1930 | env->dst_stats.weight); | |
fb86f5b2 MG |
1931 | |
1932 | /* Use idle CPU if there is no imbalance */ | |
ff7db0bf | 1933 | if (!imbalance) { |
fb86f5b2 | 1934 | maymove = true; |
ff7db0bf MG |
1935 | if (env->dst_stats.idle_cpu >= 0) { |
1936 | env->dst_cpu = env->dst_stats.idle_cpu; | |
1937 | task_numa_assign(env, NULL, 0); | |
1938 | return; | |
1939 | } | |
1940 | } | |
fb86f5b2 MG |
1941 | } else { |
1942 | long src_load, dst_load, load; | |
1943 | /* | |
1944 | * If the improvement from just moving env->p direction is better | |
1945 | * than swapping tasks around, check if a move is possible. | |
1946 | */ | |
1947 | load = task_h_load(env->p); | |
1948 | dst_load = env->dst_stats.load + load; | |
1949 | src_load = env->src_stats.load - load; | |
1950 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
1951 | } | |
305c1fac | 1952 | |
2c8a50aa MG |
1953 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1954 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 1955 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
1956 | continue; |
1957 | ||
1958 | env->dst_cpu = cpu; | |
a0f03b61 MG |
1959 | if (task_numa_compare(env, taskimp, groupimp, maymove)) |
1960 | break; | |
2c8a50aa MG |
1961 | } |
1962 | } | |
1963 | ||
58d081b5 MG |
1964 | static int task_numa_migrate(struct task_struct *p) |
1965 | { | |
58d081b5 MG |
1966 | struct task_numa_env env = { |
1967 | .p = p, | |
fb13c7ee | 1968 | |
58d081b5 | 1969 | .src_cpu = task_cpu(p), |
b32e86b4 | 1970 | .src_nid = task_node(p), |
fb13c7ee MG |
1971 | |
1972 | .imbalance_pct = 112, | |
1973 | ||
1974 | .best_task = NULL, | |
1975 | .best_imp = 0, | |
4142c3eb | 1976 | .best_cpu = -1, |
58d081b5 | 1977 | }; |
cb361d8c | 1978 | unsigned long taskweight, groupweight; |
58d081b5 | 1979 | struct sched_domain *sd; |
cb361d8c JH |
1980 | long taskimp, groupimp; |
1981 | struct numa_group *ng; | |
a4739eca | 1982 | struct rq *best_rq; |
7bd95320 | 1983 | int nid, ret, dist; |
e6628d5b | 1984 | |
58d081b5 | 1985 | /* |
fb13c7ee MG |
1986 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1987 | * imbalance and would be the first to start moving tasks about. | |
1988 | * | |
1989 | * And we want to avoid any moving of tasks about, as that would create | |
1990 | * random movement of tasks -- counter the numa conditions we're trying | |
1991 | * to satisfy here. | |
58d081b5 MG |
1992 | */ |
1993 | rcu_read_lock(); | |
fb13c7ee | 1994 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1995 | if (sd) |
1996 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1997 | rcu_read_unlock(); |
1998 | ||
46a73e8a RR |
1999 | /* |
2000 | * Cpusets can break the scheduler domain tree into smaller | |
2001 | * balance domains, some of which do not cross NUMA boundaries. | |
2002 | * Tasks that are "trapped" in such domains cannot be migrated | |
2003 | * elsewhere, so there is no point in (re)trying. | |
2004 | */ | |
2005 | if (unlikely(!sd)) { | |
8cd45eee | 2006 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
2007 | return -EINVAL; |
2008 | } | |
2009 | ||
2c8a50aa | 2010 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
2011 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
2012 | taskweight = task_weight(p, env.src_nid, dist); | |
2013 | groupweight = group_weight(p, env.src_nid, dist); | |
ff7db0bf | 2014 | update_numa_stats(&env, &env.src_stats, env.src_nid, false); |
7bd95320 RR |
2015 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
2016 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
ff7db0bf | 2017 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
58d081b5 | 2018 | |
a43455a1 | 2019 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 2020 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 2021 | |
9de05d48 RR |
2022 | /* |
2023 | * Look at other nodes in these cases: | |
2024 | * - there is no space available on the preferred_nid | |
2025 | * - the task is part of a numa_group that is interleaved across | |
2026 | * multiple NUMA nodes; in order to better consolidate the group, | |
2027 | * we need to check other locations. | |
2028 | */ | |
cb361d8c JH |
2029 | ng = deref_curr_numa_group(p); |
2030 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
2c8a50aa MG |
2031 | for_each_online_node(nid) { |
2032 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
2033 | continue; | |
58d081b5 | 2034 | |
7bd95320 | 2035 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
2036 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
2037 | dist != env.dist) { | |
2038 | taskweight = task_weight(p, env.src_nid, dist); | |
2039 | groupweight = group_weight(p, env.src_nid, dist); | |
2040 | } | |
7bd95320 | 2041 | |
83e1d2cd | 2042 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
2043 | taskimp = task_weight(p, nid, dist) - taskweight; |
2044 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 2045 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
2046 | continue; |
2047 | ||
7bd95320 | 2048 | env.dist = dist; |
2c8a50aa | 2049 | env.dst_nid = nid; |
ff7db0bf | 2050 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
2d4056fa | 2051 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
2052 | } |
2053 | } | |
2054 | ||
68d1b02a RR |
2055 | /* |
2056 | * If the task is part of a workload that spans multiple NUMA nodes, | |
2057 | * and is migrating into one of the workload's active nodes, remember | |
2058 | * this node as the task's preferred numa node, so the workload can | |
2059 | * settle down. | |
2060 | * A task that migrated to a second choice node will be better off | |
2061 | * trying for a better one later. Do not set the preferred node here. | |
2062 | */ | |
cb361d8c | 2063 | if (ng) { |
db015dae RR |
2064 | if (env.best_cpu == -1) |
2065 | nid = env.src_nid; | |
2066 | else | |
8cd45eee | 2067 | nid = cpu_to_node(env.best_cpu); |
db015dae | 2068 | |
8cd45eee SD |
2069 | if (nid != p->numa_preferred_nid) |
2070 | sched_setnuma(p, nid); | |
db015dae RR |
2071 | } |
2072 | ||
2073 | /* No better CPU than the current one was found. */ | |
f22aef4a | 2074 | if (env.best_cpu == -1) { |
b2b2042b | 2075 | trace_sched_stick_numa(p, env.src_cpu, NULL, -1); |
db015dae | 2076 | return -EAGAIN; |
f22aef4a | 2077 | } |
0ec8aa00 | 2078 | |
a4739eca | 2079 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 2080 | if (env.best_task == NULL) { |
286549dc | 2081 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 2082 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc | 2083 | if (ret != 0) |
b2b2042b | 2084 | trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); |
fb13c7ee MG |
2085 | return ret; |
2086 | } | |
2087 | ||
0ad4e3df | 2088 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 2089 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 2090 | |
286549dc | 2091 | if (ret != 0) |
b2b2042b | 2092 | trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); |
fb13c7ee MG |
2093 | put_task_struct(env.best_task); |
2094 | return ret; | |
e6628d5b MG |
2095 | } |
2096 | ||
6b9a7460 MG |
2097 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
2098 | static void numa_migrate_preferred(struct task_struct *p) | |
2099 | { | |
5085e2a3 RR |
2100 | unsigned long interval = HZ; |
2101 | ||
2739d3ee | 2102 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 2103 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
2104 | return; |
2105 | ||
2739d3ee | 2106 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 2107 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 2108 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
2109 | |
2110 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 2111 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
2112 | return; |
2113 | ||
2114 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 2115 | task_numa_migrate(p); |
6b9a7460 MG |
2116 | } |
2117 | ||
20e07dea | 2118 | /* |
4142c3eb | 2119 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
2120 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
2121 | * be different from the set of nodes where the workload's memory is currently | |
2122 | * located. | |
20e07dea | 2123 | */ |
4142c3eb | 2124 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
2125 | { |
2126 | unsigned long faults, max_faults = 0; | |
4142c3eb | 2127 | int nid, active_nodes = 0; |
20e07dea RR |
2128 | |
2129 | for_each_online_node(nid) { | |
2130 | faults = group_faults_cpu(numa_group, nid); | |
2131 | if (faults > max_faults) | |
2132 | max_faults = faults; | |
2133 | } | |
2134 | ||
2135 | for_each_online_node(nid) { | |
2136 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
2137 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
2138 | active_nodes++; | |
20e07dea | 2139 | } |
4142c3eb RR |
2140 | |
2141 | numa_group->max_faults_cpu = max_faults; | |
2142 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
2143 | } |
2144 | ||
04bb2f94 RR |
2145 | /* |
2146 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
2147 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
2148 | * period will be for the next scan window. If local/(local+remote) ratio is |
2149 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
2150 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
2151 | */ |
2152 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 2153 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
2154 | |
2155 | /* | |
2156 | * Increase the scan period (slow down scanning) if the majority of | |
2157 | * our memory is already on our local node, or if the majority of | |
2158 | * the page accesses are shared with other processes. | |
2159 | * Otherwise, decrease the scan period. | |
2160 | */ | |
2161 | static void update_task_scan_period(struct task_struct *p, | |
2162 | unsigned long shared, unsigned long private) | |
2163 | { | |
2164 | unsigned int period_slot; | |
37ec97de | 2165 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
2166 | int diff; |
2167 | ||
2168 | unsigned long remote = p->numa_faults_locality[0]; | |
2169 | unsigned long local = p->numa_faults_locality[1]; | |
2170 | ||
2171 | /* | |
2172 | * If there were no record hinting faults then either the task is | |
2173 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
2174 | * to automatic numa balancing. Related to that, if there were failed |
2175 | * migration then it implies we are migrating too quickly or the local | |
2176 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 2177 | */ |
074c2381 | 2178 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
2179 | p->numa_scan_period = min(p->numa_scan_period_max, |
2180 | p->numa_scan_period << 1); | |
2181 | ||
2182 | p->mm->numa_next_scan = jiffies + | |
2183 | msecs_to_jiffies(p->numa_scan_period); | |
2184 | ||
2185 | return; | |
2186 | } | |
2187 | ||
2188 | /* | |
2189 | * Prepare to scale scan period relative to the current period. | |
2190 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
2191 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
2192 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
2193 | */ | |
2194 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
2195 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
2196 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
2197 | ||
2198 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2199 | /* | |
2200 | * Most memory accesses are local. There is no need to | |
2201 | * do fast NUMA scanning, since memory is already local. | |
2202 | */ | |
2203 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
2204 | if (!slot) | |
2205 | slot = 1; | |
2206 | diff = slot * period_slot; | |
2207 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2208 | /* | |
2209 | * Most memory accesses are shared with other tasks. | |
2210 | * There is no point in continuing fast NUMA scanning, | |
2211 | * since other tasks may just move the memory elsewhere. | |
2212 | */ | |
2213 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
2214 | if (!slot) |
2215 | slot = 1; | |
2216 | diff = slot * period_slot; | |
2217 | } else { | |
04bb2f94 | 2218 | /* |
37ec97de RR |
2219 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
2220 | * yet they are not on the local NUMA node. Speed up | |
2221 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 2222 | */ |
37ec97de RR |
2223 | int ratio = max(lr_ratio, ps_ratio); |
2224 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
2225 | } |
2226 | ||
2227 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2228 | task_scan_min(p), task_scan_max(p)); | |
2229 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2230 | } | |
2231 | ||
7e2703e6 RR |
2232 | /* |
2233 | * Get the fraction of time the task has been running since the last | |
2234 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2235 | * decays those on a 32ms period, which is orders of magnitude off | |
2236 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2237 | * stats only if the task is so new there are no NUMA statistics yet. | |
2238 | */ | |
2239 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2240 | { | |
2241 | u64 runtime, delta, now; | |
2242 | /* Use the start of this time slice to avoid calculations. */ | |
2243 | now = p->se.exec_start; | |
2244 | runtime = p->se.sum_exec_runtime; | |
2245 | ||
2246 | if (p->last_task_numa_placement) { | |
2247 | delta = runtime - p->last_sum_exec_runtime; | |
2248 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2249 | |
2250 | /* Avoid time going backwards, prevent potential divide error: */ | |
2251 | if (unlikely((s64)*period < 0)) | |
2252 | *period = 0; | |
7e2703e6 | 2253 | } else { |
c7b50216 | 2254 | delta = p->se.avg.load_sum; |
9d89c257 | 2255 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2256 | } |
2257 | ||
2258 | p->last_sum_exec_runtime = runtime; | |
2259 | p->last_task_numa_placement = now; | |
2260 | ||
2261 | return delta; | |
2262 | } | |
2263 | ||
54009416 RR |
2264 | /* |
2265 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2266 | * be done in a way that produces consistent results with group_weight, | |
2267 | * otherwise workloads might not converge. | |
2268 | */ | |
2269 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2270 | { | |
2271 | nodemask_t nodes; | |
2272 | int dist; | |
2273 | ||
2274 | /* Direct connections between all NUMA nodes. */ | |
2275 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2276 | return nid; | |
2277 | ||
2278 | /* | |
2279 | * On a system with glueless mesh NUMA topology, group_weight | |
2280 | * scores nodes according to the number of NUMA hinting faults on | |
2281 | * both the node itself, and on nearby nodes. | |
2282 | */ | |
2283 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2284 | unsigned long score, max_score = 0; | |
2285 | int node, max_node = nid; | |
2286 | ||
2287 | dist = sched_max_numa_distance; | |
2288 | ||
2289 | for_each_online_node(node) { | |
2290 | score = group_weight(p, node, dist); | |
2291 | if (score > max_score) { | |
2292 | max_score = score; | |
2293 | max_node = node; | |
2294 | } | |
2295 | } | |
2296 | return max_node; | |
2297 | } | |
2298 | ||
2299 | /* | |
2300 | * Finding the preferred nid in a system with NUMA backplane | |
2301 | * interconnect topology is more involved. The goal is to locate | |
2302 | * tasks from numa_groups near each other in the system, and | |
2303 | * untangle workloads from different sides of the system. This requires | |
2304 | * searching down the hierarchy of node groups, recursively searching | |
2305 | * inside the highest scoring group of nodes. The nodemask tricks | |
2306 | * keep the complexity of the search down. | |
2307 | */ | |
2308 | nodes = node_online_map; | |
2309 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2310 | unsigned long max_faults = 0; | |
81907478 | 2311 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2312 | int a, b; |
2313 | ||
2314 | /* Are there nodes at this distance from each other? */ | |
2315 | if (!find_numa_distance(dist)) | |
2316 | continue; | |
2317 | ||
2318 | for_each_node_mask(a, nodes) { | |
2319 | unsigned long faults = 0; | |
2320 | nodemask_t this_group; | |
2321 | nodes_clear(this_group); | |
2322 | ||
2323 | /* Sum group's NUMA faults; includes a==b case. */ | |
2324 | for_each_node_mask(b, nodes) { | |
2325 | if (node_distance(a, b) < dist) { | |
2326 | faults += group_faults(p, b); | |
2327 | node_set(b, this_group); | |
2328 | node_clear(b, nodes); | |
2329 | } | |
2330 | } | |
2331 | ||
2332 | /* Remember the top group. */ | |
2333 | if (faults > max_faults) { | |
2334 | max_faults = faults; | |
2335 | max_group = this_group; | |
2336 | /* | |
2337 | * subtle: at the smallest distance there is | |
2338 | * just one node left in each "group", the | |
2339 | * winner is the preferred nid. | |
2340 | */ | |
2341 | nid = a; | |
2342 | } | |
2343 | } | |
2344 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2345 | if (!max_faults) |
2346 | break; | |
54009416 RR |
2347 | nodes = max_group; |
2348 | } | |
2349 | return nid; | |
2350 | } | |
2351 | ||
cbee9f88 PZ |
2352 | static void task_numa_placement(struct task_struct *p) |
2353 | { | |
98fa15f3 | 2354 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2355 | unsigned long max_faults = 0; |
04bb2f94 | 2356 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2357 | unsigned long total_faults; |
2358 | u64 runtime, period; | |
7dbd13ed | 2359 | spinlock_t *group_lock = NULL; |
cb361d8c | 2360 | struct numa_group *ng; |
cbee9f88 | 2361 | |
7e5a2c17 JL |
2362 | /* |
2363 | * The p->mm->numa_scan_seq field gets updated without | |
2364 | * exclusive access. Use READ_ONCE() here to ensure | |
2365 | * that the field is read in a single access: | |
2366 | */ | |
316c1608 | 2367 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2368 | if (p->numa_scan_seq == seq) |
2369 | return; | |
2370 | p->numa_scan_seq = seq; | |
598f0ec0 | 2371 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2372 | |
7e2703e6 RR |
2373 | total_faults = p->numa_faults_locality[0] + |
2374 | p->numa_faults_locality[1]; | |
2375 | runtime = numa_get_avg_runtime(p, &period); | |
2376 | ||
7dbd13ed | 2377 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2378 | ng = deref_curr_numa_group(p); |
2379 | if (ng) { | |
2380 | group_lock = &ng->lock; | |
60e69eed | 2381 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2382 | } |
2383 | ||
688b7585 MG |
2384 | /* Find the node with the highest number of faults */ |
2385 | for_each_online_node(nid) { | |
44dba3d5 IM |
2386 | /* Keep track of the offsets in numa_faults array */ |
2387 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2388 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2389 | int priv; |
745d6147 | 2390 | |
be1e4e76 | 2391 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2392 | long diff, f_diff, f_weight; |
8c8a743c | 2393 | |
44dba3d5 IM |
2394 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2395 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2396 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2397 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2398 | |
ac8e895b | 2399 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2400 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2401 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2402 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2403 | |
7e2703e6 RR |
2404 | /* |
2405 | * Normalize the faults_from, so all tasks in a group | |
2406 | * count according to CPU use, instead of by the raw | |
2407 | * number of faults. Tasks with little runtime have | |
2408 | * little over-all impact on throughput, and thus their | |
2409 | * faults are less important. | |
2410 | */ | |
2411 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2412 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2413 | (total_faults + 1); |
44dba3d5 IM |
2414 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2415 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2416 | |
44dba3d5 IM |
2417 | p->numa_faults[mem_idx] += diff; |
2418 | p->numa_faults[cpu_idx] += f_diff; | |
2419 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2420 | p->total_numa_faults += diff; |
cb361d8c | 2421 | if (ng) { |
44dba3d5 IM |
2422 | /* |
2423 | * safe because we can only change our own group | |
2424 | * | |
2425 | * mem_idx represents the offset for a given | |
2426 | * nid and priv in a specific region because it | |
2427 | * is at the beginning of the numa_faults array. | |
2428 | */ | |
cb361d8c JH |
2429 | ng->faults[mem_idx] += diff; |
2430 | ng->faults_cpu[mem_idx] += f_diff; | |
2431 | ng->total_faults += diff; | |
2432 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2433 | } |
ac8e895b MG |
2434 | } |
2435 | ||
cb361d8c | 2436 | if (!ng) { |
f03bb676 SD |
2437 | if (faults > max_faults) { |
2438 | max_faults = faults; | |
2439 | max_nid = nid; | |
2440 | } | |
2441 | } else if (group_faults > max_faults) { | |
2442 | max_faults = group_faults; | |
688b7585 MG |
2443 | max_nid = nid; |
2444 | } | |
83e1d2cd MG |
2445 | } |
2446 | ||
cb361d8c JH |
2447 | if (ng) { |
2448 | numa_group_count_active_nodes(ng); | |
60e69eed | 2449 | spin_unlock_irq(group_lock); |
f03bb676 | 2450 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2451 | } |
2452 | ||
bb97fc31 RR |
2453 | if (max_faults) { |
2454 | /* Set the new preferred node */ | |
2455 | if (max_nid != p->numa_preferred_nid) | |
2456 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2457 | } |
30619c89 SD |
2458 | |
2459 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2460 | } |
2461 | ||
8c8a743c PZ |
2462 | static inline int get_numa_group(struct numa_group *grp) |
2463 | { | |
c45a7795 | 2464 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2465 | } |
2466 | ||
2467 | static inline void put_numa_group(struct numa_group *grp) | |
2468 | { | |
c45a7795 | 2469 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2470 | kfree_rcu(grp, rcu); |
2471 | } | |
2472 | ||
3e6a9418 MG |
2473 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2474 | int *priv) | |
8c8a743c PZ |
2475 | { |
2476 | struct numa_group *grp, *my_grp; | |
2477 | struct task_struct *tsk; | |
2478 | bool join = false; | |
2479 | int cpu = cpupid_to_cpu(cpupid); | |
2480 | int i; | |
2481 | ||
cb361d8c | 2482 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2483 | unsigned int size = sizeof(struct numa_group) + |
50ec8a40 | 2484 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2485 | |
2486 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2487 | if (!grp) | |
2488 | return; | |
2489 | ||
c45a7795 | 2490 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2491 | grp->active_nodes = 1; |
2492 | grp->max_faults_cpu = 0; | |
8c8a743c | 2493 | spin_lock_init(&grp->lock); |
e29cf08b | 2494 | grp->gid = p->pid; |
50ec8a40 | 2495 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2496 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2497 | nr_node_ids; | |
8c8a743c | 2498 | |
be1e4e76 | 2499 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2500 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2501 | |
989348b5 | 2502 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2503 | |
8c8a743c PZ |
2504 | grp->nr_tasks++; |
2505 | rcu_assign_pointer(p->numa_group, grp); | |
2506 | } | |
2507 | ||
2508 | rcu_read_lock(); | |
316c1608 | 2509 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2510 | |
2511 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2512 | goto no_join; |
8c8a743c PZ |
2513 | |
2514 | grp = rcu_dereference(tsk->numa_group); | |
2515 | if (!grp) | |
3354781a | 2516 | goto no_join; |
8c8a743c | 2517 | |
cb361d8c | 2518 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2519 | if (grp == my_grp) |
3354781a | 2520 | goto no_join; |
8c8a743c PZ |
2521 | |
2522 | /* | |
2523 | * Only join the other group if its bigger; if we're the bigger group, | |
2524 | * the other task will join us. | |
2525 | */ | |
2526 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2527 | goto no_join; |
8c8a743c PZ |
2528 | |
2529 | /* | |
2530 | * Tie-break on the grp address. | |
2531 | */ | |
2532 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2533 | goto no_join; |
8c8a743c | 2534 | |
dabe1d99 RR |
2535 | /* Always join threads in the same process. */ |
2536 | if (tsk->mm == current->mm) | |
2537 | join = true; | |
2538 | ||
2539 | /* Simple filter to avoid false positives due to PID collisions */ | |
2540 | if (flags & TNF_SHARED) | |
2541 | join = true; | |
8c8a743c | 2542 | |
3e6a9418 MG |
2543 | /* Update priv based on whether false sharing was detected */ |
2544 | *priv = !join; | |
2545 | ||
dabe1d99 | 2546 | if (join && !get_numa_group(grp)) |
3354781a | 2547 | goto no_join; |
8c8a743c | 2548 | |
8c8a743c PZ |
2549 | rcu_read_unlock(); |
2550 | ||
2551 | if (!join) | |
2552 | return; | |
2553 | ||
60e69eed MG |
2554 | BUG_ON(irqs_disabled()); |
2555 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2556 | |
be1e4e76 | 2557 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2558 | my_grp->faults[i] -= p->numa_faults[i]; |
2559 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2560 | } |
989348b5 MG |
2561 | my_grp->total_faults -= p->total_numa_faults; |
2562 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2563 | |
8c8a743c PZ |
2564 | my_grp->nr_tasks--; |
2565 | grp->nr_tasks++; | |
2566 | ||
2567 | spin_unlock(&my_grp->lock); | |
60e69eed | 2568 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2569 | |
2570 | rcu_assign_pointer(p->numa_group, grp); | |
2571 | ||
2572 | put_numa_group(my_grp); | |
3354781a PZ |
2573 | return; |
2574 | ||
2575 | no_join: | |
2576 | rcu_read_unlock(); | |
2577 | return; | |
8c8a743c PZ |
2578 | } |
2579 | ||
16d51a59 | 2580 | /* |
3b03706f | 2581 | * Get rid of NUMA statistics associated with a task (either current or dead). |
16d51a59 JH |
2582 | * If @final is set, the task is dead and has reached refcount zero, so we can |
2583 | * safely free all relevant data structures. Otherwise, there might be | |
2584 | * concurrent reads from places like load balancing and procfs, and we should | |
2585 | * reset the data back to default state without freeing ->numa_faults. | |
2586 | */ | |
2587 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 2588 | { |
cb361d8c JH |
2589 | /* safe: p either is current or is being freed by current */ |
2590 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 2591 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
2592 | unsigned long flags; |
2593 | int i; | |
8c8a743c | 2594 | |
16d51a59 JH |
2595 | if (!numa_faults) |
2596 | return; | |
2597 | ||
8c8a743c | 2598 | if (grp) { |
e9dd685c | 2599 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2600 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2601 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2602 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2603 | |
8c8a743c | 2604 | grp->nr_tasks--; |
e9dd685c | 2605 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2606 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2607 | put_numa_group(grp); |
2608 | } | |
2609 | ||
16d51a59 JH |
2610 | if (final) { |
2611 | p->numa_faults = NULL; | |
2612 | kfree(numa_faults); | |
2613 | } else { | |
2614 | p->total_numa_faults = 0; | |
2615 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
2616 | numa_faults[i] = 0; | |
2617 | } | |
8c8a743c PZ |
2618 | } |
2619 | ||
cbee9f88 PZ |
2620 | /* |
2621 | * Got a PROT_NONE fault for a page on @node. | |
2622 | */ | |
58b46da3 | 2623 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2624 | { |
2625 | struct task_struct *p = current; | |
6688cc05 | 2626 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2627 | int cpu_node = task_node(current); |
792568ec | 2628 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2629 | struct numa_group *ng; |
ac8e895b | 2630 | int priv; |
cbee9f88 | 2631 | |
2a595721 | 2632 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2633 | return; |
2634 | ||
9ff1d9ff MG |
2635 | /* for example, ksmd faulting in a user's mm */ |
2636 | if (!p->mm) | |
2637 | return; | |
2638 | ||
f809ca9a | 2639 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2640 | if (unlikely(!p->numa_faults)) { |
2641 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2642 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2643 | |
44dba3d5 IM |
2644 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2645 | if (!p->numa_faults) | |
f809ca9a | 2646 | return; |
745d6147 | 2647 | |
83e1d2cd | 2648 | p->total_numa_faults = 0; |
04bb2f94 | 2649 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2650 | } |
cbee9f88 | 2651 | |
8c8a743c PZ |
2652 | /* |
2653 | * First accesses are treated as private, otherwise consider accesses | |
2654 | * to be private if the accessing pid has not changed | |
2655 | */ | |
2656 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2657 | priv = 1; | |
2658 | } else { | |
2659 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2660 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2661 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2662 | } |
2663 | ||
792568ec RR |
2664 | /* |
2665 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2666 | * occurs wholly within the set of nodes that the workload is | |
2667 | * actively using should be counted as local. This allows the | |
2668 | * scan rate to slow down when a workload has settled down. | |
2669 | */ | |
cb361d8c | 2670 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
2671 | if (!priv && !local && ng && ng->active_nodes > 1 && |
2672 | numa_is_active_node(cpu_node, ng) && | |
2673 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2674 | local = 1; |
2675 | ||
2739d3ee | 2676 | /* |
e1ff516a YW |
2677 | * Retry to migrate task to preferred node periodically, in case it |
2678 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2679 | */ |
b6a60cf3 SD |
2680 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2681 | task_numa_placement(p); | |
6b9a7460 | 2682 | numa_migrate_preferred(p); |
b6a60cf3 | 2683 | } |
6b9a7460 | 2684 | |
b32e86b4 IM |
2685 | if (migrated) |
2686 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2687 | if (flags & TNF_MIGRATE_FAIL) |
2688 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2689 | |
44dba3d5 IM |
2690 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2691 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2692 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2693 | } |
2694 | ||
6e5fb223 PZ |
2695 | static void reset_ptenuma_scan(struct task_struct *p) |
2696 | { | |
7e5a2c17 JL |
2697 | /* |
2698 | * We only did a read acquisition of the mmap sem, so | |
2699 | * p->mm->numa_scan_seq is written to without exclusive access | |
2700 | * and the update is not guaranteed to be atomic. That's not | |
2701 | * much of an issue though, since this is just used for | |
2702 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2703 | * expensive, to avoid any form of compiler optimizations: | |
2704 | */ | |
316c1608 | 2705 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2706 | p->mm->numa_scan_offset = 0; |
2707 | } | |
2708 | ||
cbee9f88 PZ |
2709 | /* |
2710 | * The expensive part of numa migration is done from task_work context. | |
2711 | * Triggered from task_tick_numa(). | |
2712 | */ | |
9434f9f5 | 2713 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
2714 | { |
2715 | unsigned long migrate, next_scan, now = jiffies; | |
2716 | struct task_struct *p = current; | |
2717 | struct mm_struct *mm = p->mm; | |
51170840 | 2718 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2719 | struct vm_area_struct *vma; |
9f40604c | 2720 | unsigned long start, end; |
598f0ec0 | 2721 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2722 | long pages, virtpages; |
cbee9f88 | 2723 | |
9148a3a1 | 2724 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 2725 | |
b34920d4 | 2726 | work->next = work; |
cbee9f88 PZ |
2727 | /* |
2728 | * Who cares about NUMA placement when they're dying. | |
2729 | * | |
2730 | * NOTE: make sure not to dereference p->mm before this check, | |
2731 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2732 | * without p->mm even though we still had it when we enqueued this | |
2733 | * work. | |
2734 | */ | |
2735 | if (p->flags & PF_EXITING) | |
2736 | return; | |
2737 | ||
930aa174 | 2738 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2739 | mm->numa_next_scan = now + |
2740 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2741 | } |
2742 | ||
cbee9f88 PZ |
2743 | /* |
2744 | * Enforce maximal scan/migration frequency.. | |
2745 | */ | |
2746 | migrate = mm->numa_next_scan; | |
2747 | if (time_before(now, migrate)) | |
2748 | return; | |
2749 | ||
598f0ec0 MG |
2750 | if (p->numa_scan_period == 0) { |
2751 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2752 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2753 | } |
cbee9f88 | 2754 | |
fb003b80 | 2755 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2756 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2757 | return; | |
2758 | ||
19a78d11 PZ |
2759 | /* |
2760 | * Delay this task enough that another task of this mm will likely win | |
2761 | * the next time around. | |
2762 | */ | |
2763 | p->node_stamp += 2 * TICK_NSEC; | |
2764 | ||
9f40604c MG |
2765 | start = mm->numa_scan_offset; |
2766 | pages = sysctl_numa_balancing_scan_size; | |
2767 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2768 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2769 | if (!pages) |
2770 | return; | |
cbee9f88 | 2771 | |
4620f8c1 | 2772 | |
d8ed45c5 | 2773 | if (!mmap_read_trylock(mm)) |
8655d549 | 2774 | return; |
9f40604c | 2775 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2776 | if (!vma) { |
2777 | reset_ptenuma_scan(p); | |
9f40604c | 2778 | start = 0; |
6e5fb223 PZ |
2779 | vma = mm->mmap; |
2780 | } | |
9f40604c | 2781 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2782 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2783 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2784 | continue; |
6b79c57b | 2785 | } |
6e5fb223 | 2786 | |
4591ce4f MG |
2787 | /* |
2788 | * Shared library pages mapped by multiple processes are not | |
2789 | * migrated as it is expected they are cache replicated. Avoid | |
2790 | * hinting faults in read-only file-backed mappings or the vdso | |
2791 | * as migrating the pages will be of marginal benefit. | |
2792 | */ | |
2793 | if (!vma->vm_mm || | |
2794 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2795 | continue; | |
2796 | ||
3c67f474 MG |
2797 | /* |
2798 | * Skip inaccessible VMAs to avoid any confusion between | |
2799 | * PROT_NONE and NUMA hinting ptes | |
2800 | */ | |
3122e80e | 2801 | if (!vma_is_accessible(vma)) |
3c67f474 | 2802 | continue; |
4591ce4f | 2803 | |
9f40604c MG |
2804 | do { |
2805 | start = max(start, vma->vm_start); | |
2806 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2807 | end = min(end, vma->vm_end); | |
4620f8c1 | 2808 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2809 | |
2810 | /* | |
4620f8c1 RR |
2811 | * Try to scan sysctl_numa_balancing_size worth of |
2812 | * hpages that have at least one present PTE that | |
2813 | * is not already pte-numa. If the VMA contains | |
2814 | * areas that are unused or already full of prot_numa | |
2815 | * PTEs, scan up to virtpages, to skip through those | |
2816 | * areas faster. | |
598f0ec0 MG |
2817 | */ |
2818 | if (nr_pte_updates) | |
2819 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2820 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2821 | |
9f40604c | 2822 | start = end; |
4620f8c1 | 2823 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2824 | goto out; |
3cf1962c RR |
2825 | |
2826 | cond_resched(); | |
9f40604c | 2827 | } while (end != vma->vm_end); |
cbee9f88 | 2828 | } |
6e5fb223 | 2829 | |
9f40604c | 2830 | out: |
6e5fb223 | 2831 | /* |
c69307d5 PZ |
2832 | * It is possible to reach the end of the VMA list but the last few |
2833 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2834 | * would find the !migratable VMA on the next scan but not reset the | |
2835 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2836 | */ |
2837 | if (vma) | |
9f40604c | 2838 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2839 | else |
2840 | reset_ptenuma_scan(p); | |
d8ed45c5 | 2841 | mmap_read_unlock(mm); |
51170840 RR |
2842 | |
2843 | /* | |
2844 | * Make sure tasks use at least 32x as much time to run other code | |
2845 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2846 | * Usually update_task_scan_period slows down scanning enough; on an | |
2847 | * overloaded system we need to limit overhead on a per task basis. | |
2848 | */ | |
2849 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2850 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2851 | p->node_stamp += 32 * diff; | |
2852 | } | |
cbee9f88 PZ |
2853 | } |
2854 | ||
d35927a1 VS |
2855 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
2856 | { | |
2857 | int mm_users = 0; | |
2858 | struct mm_struct *mm = p->mm; | |
2859 | ||
2860 | if (mm) { | |
2861 | mm_users = atomic_read(&mm->mm_users); | |
2862 | if (mm_users == 1) { | |
2863 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
2864 | mm->numa_scan_seq = 0; | |
2865 | } | |
2866 | } | |
2867 | p->node_stamp = 0; | |
2868 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
2869 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
b34920d4 | 2870 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
2871 | p->numa_work.next = &p->numa_work; |
2872 | p->numa_faults = NULL; | |
2873 | RCU_INIT_POINTER(p->numa_group, NULL); | |
2874 | p->last_task_numa_placement = 0; | |
2875 | p->last_sum_exec_runtime = 0; | |
2876 | ||
b34920d4 VS |
2877 | init_task_work(&p->numa_work, task_numa_work); |
2878 | ||
d35927a1 VS |
2879 | /* New address space, reset the preferred nid */ |
2880 | if (!(clone_flags & CLONE_VM)) { | |
2881 | p->numa_preferred_nid = NUMA_NO_NODE; | |
2882 | return; | |
2883 | } | |
2884 | ||
2885 | /* | |
2886 | * New thread, keep existing numa_preferred_nid which should be copied | |
2887 | * already by arch_dup_task_struct but stagger when scans start. | |
2888 | */ | |
2889 | if (mm) { | |
2890 | unsigned int delay; | |
2891 | ||
2892 | delay = min_t(unsigned int, task_scan_max(current), | |
2893 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
2894 | delay += 2 * TICK_NSEC; | |
2895 | p->node_stamp = delay; | |
2896 | } | |
2897 | } | |
2898 | ||
cbee9f88 PZ |
2899 | /* |
2900 | * Drive the periodic memory faults.. | |
2901 | */ | |
b1546edc | 2902 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
2903 | { |
2904 | struct callback_head *work = &curr->numa_work; | |
2905 | u64 period, now; | |
2906 | ||
2907 | /* | |
2908 | * We don't care about NUMA placement if we don't have memory. | |
2909 | */ | |
18f855e5 | 2910 | if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) |
cbee9f88 PZ |
2911 | return; |
2912 | ||
2913 | /* | |
2914 | * Using runtime rather than walltime has the dual advantage that | |
2915 | * we (mostly) drive the selection from busy threads and that the | |
2916 | * task needs to have done some actual work before we bother with | |
2917 | * NUMA placement. | |
2918 | */ | |
2919 | now = curr->se.sum_exec_runtime; | |
2920 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2921 | ||
25b3e5a3 | 2922 | if (now > curr->node_stamp + period) { |
4b96a29b | 2923 | if (!curr->node_stamp) |
b5dd77c8 | 2924 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2925 | curr->node_stamp += period; |
cbee9f88 | 2926 | |
b34920d4 | 2927 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
91989c70 | 2928 | task_work_add(curr, work, TWA_RESUME); |
cbee9f88 PZ |
2929 | } |
2930 | } | |
3fed382b | 2931 | |
3f9672ba SD |
2932 | static void update_scan_period(struct task_struct *p, int new_cpu) |
2933 | { | |
2934 | int src_nid = cpu_to_node(task_cpu(p)); | |
2935 | int dst_nid = cpu_to_node(new_cpu); | |
2936 | ||
05cbdf4f MG |
2937 | if (!static_branch_likely(&sched_numa_balancing)) |
2938 | return; | |
2939 | ||
3f9672ba SD |
2940 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
2941 | return; | |
2942 | ||
05cbdf4f MG |
2943 | if (src_nid == dst_nid) |
2944 | return; | |
2945 | ||
2946 | /* | |
2947 | * Allow resets if faults have been trapped before one scan | |
2948 | * has completed. This is most likely due to a new task that | |
2949 | * is pulled cross-node due to wakeups or load balancing. | |
2950 | */ | |
2951 | if (p->numa_scan_seq) { | |
2952 | /* | |
2953 | * Avoid scan adjustments if moving to the preferred | |
2954 | * node or if the task was not previously running on | |
2955 | * the preferred node. | |
2956 | */ | |
2957 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
2958 | (p->numa_preferred_nid != NUMA_NO_NODE && |
2959 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
2960 | return; |
2961 | } | |
2962 | ||
2963 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
2964 | } |
2965 | ||
cbee9f88 PZ |
2966 | #else |
2967 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2968 | { | |
2969 | } | |
0ec8aa00 PZ |
2970 | |
2971 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2972 | { | |
2973 | } | |
2974 | ||
2975 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2976 | { | |
2977 | } | |
3fed382b | 2978 | |
3f9672ba SD |
2979 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
2980 | { | |
2981 | } | |
2982 | ||
cbee9f88 PZ |
2983 | #endif /* CONFIG_NUMA_BALANCING */ |
2984 | ||
30cfdcfc DA |
2985 | static void |
2986 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2987 | { | |
2988 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 2989 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2990 | if (entity_is_task(se)) { |
2991 | struct rq *rq = rq_of(cfs_rq); | |
2992 | ||
2993 | account_numa_enqueue(rq, task_of(se)); | |
2994 | list_add(&se->group_node, &rq->cfs_tasks); | |
2995 | } | |
367456c7 | 2996 | #endif |
30cfdcfc | 2997 | cfs_rq->nr_running++; |
30cfdcfc DA |
2998 | } |
2999 | ||
3000 | static void | |
3001 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3002 | { | |
3003 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 3004 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3005 | if (entity_is_task(se)) { |
3006 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 3007 | list_del_init(&se->group_node); |
0ec8aa00 | 3008 | } |
bfdb198c | 3009 | #endif |
30cfdcfc | 3010 | cfs_rq->nr_running--; |
30cfdcfc DA |
3011 | } |
3012 | ||
8d5b9025 PZ |
3013 | /* |
3014 | * Signed add and clamp on underflow. | |
3015 | * | |
3016 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3017 | * memory. This allows lockless observations without ever seeing the negative | |
3018 | * values. | |
3019 | */ | |
3020 | #define add_positive(_ptr, _val) do { \ | |
3021 | typeof(_ptr) ptr = (_ptr); \ | |
3022 | typeof(_val) val = (_val); \ | |
3023 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3024 | \ | |
3025 | res = var + val; \ | |
3026 | \ | |
3027 | if (val < 0 && res > var) \ | |
3028 | res = 0; \ | |
3029 | \ | |
3030 | WRITE_ONCE(*ptr, res); \ | |
3031 | } while (0) | |
3032 | ||
3033 | /* | |
3034 | * Unsigned subtract and clamp on underflow. | |
3035 | * | |
3036 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3037 | * memory. This allows lockless observations without ever seeing the negative | |
3038 | * values. | |
3039 | */ | |
3040 | #define sub_positive(_ptr, _val) do { \ | |
3041 | typeof(_ptr) ptr = (_ptr); \ | |
3042 | typeof(*ptr) val = (_val); \ | |
3043 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3044 | res = var - val; \ | |
3045 | if (res > var) \ | |
3046 | res = 0; \ | |
3047 | WRITE_ONCE(*ptr, res); \ | |
3048 | } while (0) | |
3049 | ||
b5c0ce7b PB |
3050 | /* |
3051 | * Remove and clamp on negative, from a local variable. | |
3052 | * | |
3053 | * A variant of sub_positive(), which does not use explicit load-store | |
3054 | * and is thus optimized for local variable updates. | |
3055 | */ | |
3056 | #define lsub_positive(_ptr, _val) do { \ | |
3057 | typeof(_ptr) ptr = (_ptr); \ | |
3058 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
3059 | } while (0) | |
3060 | ||
8d5b9025 | 3061 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
3062 | static inline void |
3063 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3064 | { | |
3065 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3066 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
3067 | } | |
3068 | ||
3069 | static inline void | |
3070 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3071 | { | |
ceb6ba45 | 3072 | u32 divider = get_pelt_divider(&se->avg); |
8d5b9025 | 3073 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); |
ceb6ba45 | 3074 | cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider; |
8d5b9025 PZ |
3075 | } |
3076 | #else | |
3077 | static inline void | |
8d5b9025 PZ |
3078 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } |
3079 | static inline void | |
3080 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
3081 | #endif | |
3082 | ||
9059393e | 3083 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
0dacee1b | 3084 | unsigned long weight) |
9059393e VG |
3085 | { |
3086 | if (se->on_rq) { | |
3087 | /* commit outstanding execution time */ | |
3088 | if (cfs_rq->curr == se) | |
3089 | update_curr(cfs_rq); | |
1724b95b | 3090 | update_load_sub(&cfs_rq->load, se->load.weight); |
9059393e VG |
3091 | } |
3092 | dequeue_load_avg(cfs_rq, se); | |
3093 | ||
3094 | update_load_set(&se->load, weight); | |
3095 | ||
3096 | #ifdef CONFIG_SMP | |
1ea6c46a | 3097 | do { |
87e867b4 | 3098 | u32 divider = get_pelt_divider(&se->avg); |
1ea6c46a PZ |
3099 | |
3100 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
1ea6c46a | 3101 | } while (0); |
9059393e VG |
3102 | #endif |
3103 | ||
3104 | enqueue_load_avg(cfs_rq, se); | |
0dacee1b | 3105 | if (se->on_rq) |
1724b95b | 3106 | update_load_add(&cfs_rq->load, se->load.weight); |
0dacee1b | 3107 | |
9059393e VG |
3108 | } |
3109 | ||
3110 | void reweight_task(struct task_struct *p, int prio) | |
3111 | { | |
3112 | struct sched_entity *se = &p->se; | |
3113 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3114 | struct load_weight *load = &se->load; | |
3115 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
3116 | ||
0dacee1b | 3117 | reweight_entity(cfs_rq, se, weight); |
9059393e VG |
3118 | load->inv_weight = sched_prio_to_wmult[prio]; |
3119 | } | |
3120 | ||
3ff6dcac | 3121 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 3122 | #ifdef CONFIG_SMP |
cef27403 PZ |
3123 | /* |
3124 | * All this does is approximate the hierarchical proportion which includes that | |
3125 | * global sum we all love to hate. | |
3126 | * | |
3127 | * That is, the weight of a group entity, is the proportional share of the | |
3128 | * group weight based on the group runqueue weights. That is: | |
3129 | * | |
3130 | * tg->weight * grq->load.weight | |
3131 | * ge->load.weight = ----------------------------- (1) | |
08f7c2f4 | 3132 | * \Sum grq->load.weight |
cef27403 PZ |
3133 | * |
3134 | * Now, because computing that sum is prohibitively expensive to compute (been | |
3135 | * there, done that) we approximate it with this average stuff. The average | |
3136 | * moves slower and therefore the approximation is cheaper and more stable. | |
3137 | * | |
3138 | * So instead of the above, we substitute: | |
3139 | * | |
3140 | * grq->load.weight -> grq->avg.load_avg (2) | |
3141 | * | |
3142 | * which yields the following: | |
3143 | * | |
3144 | * tg->weight * grq->avg.load_avg | |
3145 | * ge->load.weight = ------------------------------ (3) | |
08f7c2f4 | 3146 | * tg->load_avg |
cef27403 PZ |
3147 | * |
3148 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
3149 | * | |
3150 | * That is shares_avg, and it is right (given the approximation (2)). | |
3151 | * | |
3152 | * The problem with it is that because the average is slow -- it was designed | |
3153 | * to be exactly that of course -- this leads to transients in boundary | |
3154 | * conditions. In specific, the case where the group was idle and we start the | |
3155 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
3156 | * yielding bad latency etc.. | |
3157 | * | |
3158 | * Now, in that special case (1) reduces to: | |
3159 | * | |
3160 | * tg->weight * grq->load.weight | |
17de4ee0 | 3161 | * ge->load.weight = ----------------------------- = tg->weight (4) |
08f7c2f4 | 3162 | * grp->load.weight |
cef27403 PZ |
3163 | * |
3164 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
3165 | * | |
3166 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
3167 | * UP case, like: | |
3168 | * | |
3169 | * ge->load.weight = | |
3170 | * | |
3171 | * tg->weight * grq->load.weight | |
3172 | * --------------------------------------------------- (5) | |
3173 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
3174 | * | |
17de4ee0 PZ |
3175 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
3176 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
3177 | * | |
3178 | * | |
3179 | * tg->weight * grq->load.weight | |
3180 | * ge->load.weight = ----------------------------- (6) | |
08f7c2f4 | 3181 | * tg_load_avg' |
17de4ee0 PZ |
3182 | * |
3183 | * Where: | |
3184 | * | |
3185 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
3186 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
3187 | * |
3188 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
3189 | * (4) while in the normal case it approaches (3). It consistently | |
3190 | * overestimates the ge->load.weight and therefore: | |
3191 | * | |
3192 | * \Sum ge->load.weight >= tg->weight | |
3193 | * | |
3194 | * hence icky! | |
3195 | */ | |
2c8e4dce | 3196 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 3197 | { |
7c80cfc9 PZ |
3198 | long tg_weight, tg_shares, load, shares; |
3199 | struct task_group *tg = cfs_rq->tg; | |
3200 | ||
3201 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 3202 | |
3d4b60d3 | 3203 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 3204 | |
ea1dc6fc | 3205 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 3206 | |
ea1dc6fc PZ |
3207 | /* Ensure tg_weight >= load */ |
3208 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
3209 | tg_weight += load; | |
3ff6dcac | 3210 | |
7c80cfc9 | 3211 | shares = (tg_shares * load); |
cf5f0acf PZ |
3212 | if (tg_weight) |
3213 | shares /= tg_weight; | |
3ff6dcac | 3214 | |
b8fd8423 DE |
3215 | /* |
3216 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
3217 | * of a group with small tg->shares value. It is a floor value which is | |
3218 | * assigned as a minimum load.weight to the sched_entity representing | |
3219 | * the group on a CPU. | |
3220 | * | |
3221 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
3222 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
3223 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
3224 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
3225 | * instead of 0. | |
3226 | */ | |
7c80cfc9 | 3227 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 3228 | } |
387f77cc | 3229 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 3230 | |
82958366 PT |
3231 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
3232 | ||
1ea6c46a PZ |
3233 | /* |
3234 | * Recomputes the group entity based on the current state of its group | |
3235 | * runqueue. | |
3236 | */ | |
3237 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 3238 | { |
1ea6c46a | 3239 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
0dacee1b | 3240 | long shares; |
2069dd75 | 3241 | |
1ea6c46a | 3242 | if (!gcfs_rq) |
89ee048f VG |
3243 | return; |
3244 | ||
1ea6c46a | 3245 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3246 | return; |
89ee048f | 3247 | |
3ff6dcac | 3248 | #ifndef CONFIG_SMP |
0dacee1b | 3249 | shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
3250 | |
3251 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 3252 | return; |
7c80cfc9 | 3253 | #else |
2c8e4dce | 3254 | shares = calc_group_shares(gcfs_rq); |
3ff6dcac | 3255 | #endif |
2069dd75 | 3256 | |
0dacee1b | 3257 | reweight_entity(cfs_rq_of(se), se, shares); |
2069dd75 | 3258 | } |
89ee048f | 3259 | |
2069dd75 | 3260 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3261 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3262 | { |
3263 | } | |
3264 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3265 | ||
ea14b57e | 3266 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3267 | { |
43964409 LT |
3268 | struct rq *rq = rq_of(cfs_rq); |
3269 | ||
a4f9a0e5 | 3270 | if (&rq->cfs == cfs_rq) { |
a030d738 VK |
3271 | /* |
3272 | * There are a few boundary cases this might miss but it should | |
3273 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3274 | * a real problem. |
a030d738 VK |
3275 | * |
3276 | * It will not get called when we go idle, because the idle | |
3277 | * thread is a different class (!fair), nor will the utilization | |
3278 | * number include things like RT tasks. | |
3279 | * | |
3280 | * As is, the util number is not freq-invariant (we'd have to | |
3281 | * implement arch_scale_freq_capacity() for that). | |
3282 | * | |
3283 | * See cpu_util(). | |
3284 | */ | |
ea14b57e | 3285 | cpufreq_update_util(rq, flags); |
a030d738 VK |
3286 | } |
3287 | } | |
3288 | ||
141965c7 | 3289 | #ifdef CONFIG_SMP |
c566e8e9 | 3290 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fdaba61e RR |
3291 | /* |
3292 | * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list | |
3293 | * immediately before a parent cfs_rq, and cfs_rqs are removed from the list | |
3294 | * bottom-up, we only have to test whether the cfs_rq before us on the list | |
3295 | * is our child. | |
3296 | * If cfs_rq is not on the list, test whether a child needs its to be added to | |
3297 | * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details). | |
3298 | */ | |
3299 | static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq) | |
3300 | { | |
3301 | struct cfs_rq *prev_cfs_rq; | |
3302 | struct list_head *prev; | |
3303 | ||
3304 | if (cfs_rq->on_list) { | |
3305 | prev = cfs_rq->leaf_cfs_rq_list.prev; | |
3306 | } else { | |
3307 | struct rq *rq = rq_of(cfs_rq); | |
3308 | ||
3309 | prev = rq->tmp_alone_branch; | |
3310 | } | |
3311 | ||
3312 | prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list); | |
3313 | ||
3314 | return (prev_cfs_rq->tg->parent == cfs_rq->tg); | |
3315 | } | |
a7b359fc OU |
3316 | |
3317 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) | |
3318 | { | |
3319 | if (cfs_rq->load.weight) | |
3320 | return false; | |
3321 | ||
3322 | if (cfs_rq->avg.load_sum) | |
3323 | return false; | |
3324 | ||
3325 | if (cfs_rq->avg.util_sum) | |
3326 | return false; | |
3327 | ||
3328 | if (cfs_rq->avg.runnable_sum) | |
3329 | return false; | |
3330 | ||
fdaba61e RR |
3331 | if (child_cfs_rq_on_list(cfs_rq)) |
3332 | return false; | |
3333 | ||
b2c0931a IM |
3334 | /* |
3335 | * _avg must be null when _sum are null because _avg = _sum / divider | |
3336 | * Make sure that rounding and/or propagation of PELT values never | |
3337 | * break this. | |
3338 | */ | |
3339 | SCHED_WARN_ON(cfs_rq->avg.load_avg || | |
3340 | cfs_rq->avg.util_avg || | |
3341 | cfs_rq->avg.runnable_avg); | |
3342 | ||
a7b359fc OU |
3343 | return true; |
3344 | } | |
3345 | ||
7c3edd2c PZ |
3346 | /** |
3347 | * update_tg_load_avg - update the tg's load avg | |
3348 | * @cfs_rq: the cfs_rq whose avg changed | |
7c3edd2c PZ |
3349 | * |
3350 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3351 | * However, because tg->load_avg is a global value there are performance | |
3352 | * considerations. | |
3353 | * | |
3354 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3355 | * differential update where we store the last value we propagated. This in | |
3356 | * turn allows skipping updates if the differential is 'small'. | |
3357 | * | |
815abf5a | 3358 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3359 | */ |
fe749158 | 3360 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) |
bb17f655 | 3361 | { |
9d89c257 | 3362 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3363 | |
aa0b7ae0 WL |
3364 | /* |
3365 | * No need to update load_avg for root_task_group as it is not used. | |
3366 | */ | |
3367 | if (cfs_rq->tg == &root_task_group) | |
3368 | return; | |
3369 | ||
fe749158 | 3370 | if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
9d89c257 YD |
3371 | atomic_long_add(delta, &cfs_rq->tg->load_avg); |
3372 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3373 | } |
8165e145 | 3374 | } |
f5f9739d | 3375 | |
ad936d86 | 3376 | /* |
97fb7a0a | 3377 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3378 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3379 | * including the state of rq->lock, should be made. | |
3380 | */ | |
3381 | void set_task_rq_fair(struct sched_entity *se, | |
3382 | struct cfs_rq *prev, struct cfs_rq *next) | |
3383 | { | |
0ccb977f PZ |
3384 | u64 p_last_update_time; |
3385 | u64 n_last_update_time; | |
3386 | ||
ad936d86 BP |
3387 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3388 | return; | |
3389 | ||
3390 | /* | |
3391 | * We are supposed to update the task to "current" time, then its up to | |
3392 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3393 | * getting what current time is, so simply throw away the out-of-date | |
3394 | * time. This will result in the wakee task is less decayed, but giving | |
3395 | * the wakee more load sounds not bad. | |
3396 | */ | |
0ccb977f PZ |
3397 | if (!(se->avg.last_update_time && prev)) |
3398 | return; | |
ad936d86 BP |
3399 | |
3400 | #ifndef CONFIG_64BIT | |
0ccb977f | 3401 | { |
ad936d86 BP |
3402 | u64 p_last_update_time_copy; |
3403 | u64 n_last_update_time_copy; | |
3404 | ||
3405 | do { | |
3406 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3407 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3408 | ||
3409 | smp_rmb(); | |
3410 | ||
3411 | p_last_update_time = prev->avg.last_update_time; | |
3412 | n_last_update_time = next->avg.last_update_time; | |
3413 | ||
3414 | } while (p_last_update_time != p_last_update_time_copy || | |
3415 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3416 | } |
ad936d86 | 3417 | #else |
0ccb977f PZ |
3418 | p_last_update_time = prev->avg.last_update_time; |
3419 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3420 | #endif |
23127296 | 3421 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 3422 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 3423 | } |
09a43ace | 3424 | |
0e2d2aaa PZ |
3425 | /* |
3426 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3427 | * propagate its contribution. The key to this propagation is the invariant | |
3428 | * that for each group: | |
3429 | * | |
3430 | * ge->avg == grq->avg (1) | |
3431 | * | |
3432 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3433 | * represent the very same entity, just at different points in the hierarchy. | |
3434 | * | |
9f683953 VG |
3435 | * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial |
3436 | * and simply copies the running/runnable sum over (but still wrong, because | |
3437 | * the group entity and group rq do not have their PELT windows aligned). | |
0e2d2aaa | 3438 | * |
0dacee1b | 3439 | * However, update_tg_cfs_load() is more complex. So we have: |
0e2d2aaa PZ |
3440 | * |
3441 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3442 | * | |
3443 | * And since, like util, the runnable part should be directly transferable, | |
3444 | * the following would _appear_ to be the straight forward approach: | |
3445 | * | |
a4c3c049 | 3446 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3447 | * |
3448 | * And per (1) we have: | |
3449 | * | |
a4c3c049 | 3450 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3451 | * |
3452 | * Which gives: | |
3453 | * | |
3454 | * ge->load.weight * grq->avg.load_avg | |
3455 | * ge->avg.load_avg = ----------------------------------- (4) | |
3456 | * grq->load.weight | |
3457 | * | |
3458 | * Except that is wrong! | |
3459 | * | |
3460 | * Because while for entities historical weight is not important and we | |
3461 | * really only care about our future and therefore can consider a pure | |
3462 | * runnable sum, runqueues can NOT do this. | |
3463 | * | |
3464 | * We specifically want runqueues to have a load_avg that includes | |
3465 | * historical weights. Those represent the blocked load, the load we expect | |
3466 | * to (shortly) return to us. This only works by keeping the weights as | |
3467 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3468 | * | |
a4c3c049 VG |
3469 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3470 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3471 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3472 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3473 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3474 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3475 | * |
a4c3c049 | 3476 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3477 | * |
a4c3c049 | 3478 | * Given the constraint: |
0e2d2aaa | 3479 | * |
a4c3c049 | 3480 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3481 | * |
a4c3c049 VG |
3482 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3483 | * overlap. | |
0e2d2aaa | 3484 | * |
a4c3c049 | 3485 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3486 | * |
a4c3c049 | 3487 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3488 | * |
a4c3c049 | 3489 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3490 | * |
0e2d2aaa | 3491 | */ |
09a43ace | 3492 | static inline void |
0e2d2aaa | 3493 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3494 | { |
09a43ace | 3495 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; |
87e867b4 | 3496 | u32 divider; |
09a43ace VG |
3497 | |
3498 | /* Nothing to update */ | |
3499 | if (!delta) | |
3500 | return; | |
3501 | ||
87e867b4 VG |
3502 | /* |
3503 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3504 | * See ___update_load_avg() for details. | |
3505 | */ | |
3506 | divider = get_pelt_divider(&cfs_rq->avg); | |
3507 | ||
09a43ace VG |
3508 | /* Set new sched_entity's utilization */ |
3509 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
95d68593 | 3510 | se->avg.util_sum = se->avg.util_avg * divider; |
09a43ace VG |
3511 | |
3512 | /* Update parent cfs_rq utilization */ | |
3513 | add_positive(&cfs_rq->avg.util_avg, delta); | |
95d68593 | 3514 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; |
09a43ace VG |
3515 | } |
3516 | ||
9f683953 VG |
3517 | static inline void |
3518 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) | |
3519 | { | |
3520 | long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; | |
87e867b4 | 3521 | u32 divider; |
9f683953 VG |
3522 | |
3523 | /* Nothing to update */ | |
3524 | if (!delta) | |
3525 | return; | |
3526 | ||
87e867b4 VG |
3527 | /* |
3528 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3529 | * See ___update_load_avg() for details. | |
3530 | */ | |
3531 | divider = get_pelt_divider(&cfs_rq->avg); | |
3532 | ||
9f683953 VG |
3533 | /* Set new sched_entity's runnable */ |
3534 | se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; | |
95d68593 | 3535 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
9f683953 VG |
3536 | |
3537 | /* Update parent cfs_rq runnable */ | |
3538 | add_positive(&cfs_rq->avg.runnable_avg, delta); | |
95d68593 | 3539 | cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; |
9f683953 VG |
3540 | } |
3541 | ||
09a43ace | 3542 | static inline void |
0dacee1b | 3543 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3544 | { |
7c7ad626 | 3545 | long delta, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
0dacee1b VG |
3546 | unsigned long load_avg; |
3547 | u64 load_sum = 0; | |
95d68593 | 3548 | u32 divider; |
09a43ace | 3549 | |
0e2d2aaa PZ |
3550 | if (!runnable_sum) |
3551 | return; | |
09a43ace | 3552 | |
0e2d2aaa | 3553 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3554 | |
95d68593 VG |
3555 | /* |
3556 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3557 | * See ___update_load_avg() for details. | |
3558 | */ | |
87e867b4 | 3559 | divider = get_pelt_divider(&cfs_rq->avg); |
95d68593 | 3560 | |
a4c3c049 VG |
3561 | if (runnable_sum >= 0) { |
3562 | /* | |
3563 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3564 | * the CPU is saturated running == runnable. | |
3565 | */ | |
3566 | runnable_sum += se->avg.load_sum; | |
95d68593 | 3567 | runnable_sum = min_t(long, runnable_sum, divider); |
a4c3c049 VG |
3568 | } else { |
3569 | /* | |
3570 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3571 | * assuming all tasks are equally runnable. | |
3572 | */ | |
3573 | if (scale_load_down(gcfs_rq->load.weight)) { | |
3574 | load_sum = div_s64(gcfs_rq->avg.load_sum, | |
3575 | scale_load_down(gcfs_rq->load.weight)); | |
3576 | } | |
3577 | ||
3578 | /* But make sure to not inflate se's runnable */ | |
3579 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3580 | } | |
3581 | ||
3582 | /* | |
3583 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
3584 | * Rescale running sum to be in the same range as runnable sum |
3585 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
3586 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 3587 | */ |
23127296 | 3588 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
3589 | runnable_sum = max(runnable_sum, running_sum); |
3590 | ||
0e2d2aaa | 3591 | load_sum = (s64)se_weight(se) * runnable_sum; |
95d68593 | 3592 | load_avg = div_s64(load_sum, divider); |
09a43ace | 3593 | |
83c5e9d5 DE |
3594 | se->avg.load_sum = runnable_sum; |
3595 | ||
7c7ad626 | 3596 | delta = load_avg - se->avg.load_avg; |
83c5e9d5 DE |
3597 | if (!delta) |
3598 | return; | |
09a43ace | 3599 | |
a4c3c049 | 3600 | se->avg.load_avg = load_avg; |
7c7ad626 VG |
3601 | |
3602 | add_positive(&cfs_rq->avg.load_avg, delta); | |
3603 | cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider; | |
09a43ace VG |
3604 | } |
3605 | ||
0e2d2aaa | 3606 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3607 | { |
0e2d2aaa PZ |
3608 | cfs_rq->propagate = 1; |
3609 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3610 | } |
3611 | ||
3612 | /* Update task and its cfs_rq load average */ | |
3613 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3614 | { | |
0e2d2aaa | 3615 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3616 | |
3617 | if (entity_is_task(se)) | |
3618 | return 0; | |
3619 | ||
0e2d2aaa PZ |
3620 | gcfs_rq = group_cfs_rq(se); |
3621 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3622 | return 0; |
3623 | ||
0e2d2aaa PZ |
3624 | gcfs_rq->propagate = 0; |
3625 | ||
09a43ace VG |
3626 | cfs_rq = cfs_rq_of(se); |
3627 | ||
0e2d2aaa | 3628 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3629 | |
0e2d2aaa | 3630 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
9f683953 | 3631 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); |
0dacee1b | 3632 | update_tg_cfs_load(cfs_rq, se, gcfs_rq); |
09a43ace | 3633 | |
ba19f51f | 3634 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 3635 | trace_pelt_se_tp(se); |
ba19f51f | 3636 | |
09a43ace VG |
3637 | return 1; |
3638 | } | |
3639 | ||
bc427898 VG |
3640 | /* |
3641 | * Check if we need to update the load and the utilization of a blocked | |
3642 | * group_entity: | |
3643 | */ | |
3644 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3645 | { | |
3646 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3647 | ||
3648 | /* | |
3649 | * If sched_entity still have not zero load or utilization, we have to | |
3650 | * decay it: | |
3651 | */ | |
3652 | if (se->avg.load_avg || se->avg.util_avg) | |
3653 | return false; | |
3654 | ||
3655 | /* | |
3656 | * If there is a pending propagation, we have to update the load and | |
3657 | * the utilization of the sched_entity: | |
3658 | */ | |
0e2d2aaa | 3659 | if (gcfs_rq->propagate) |
bc427898 VG |
3660 | return false; |
3661 | ||
3662 | /* | |
3663 | * Otherwise, the load and the utilization of the sched_entity is | |
3664 | * already zero and there is no pending propagation, so it will be a | |
3665 | * waste of time to try to decay it: | |
3666 | */ | |
3667 | return true; | |
3668 | } | |
3669 | ||
6e83125c | 3670 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3671 | |
fe749158 | 3672 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} |
09a43ace VG |
3673 | |
3674 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3675 | { | |
3676 | return 0; | |
3677 | } | |
3678 | ||
0e2d2aaa | 3679 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3680 | |
6e83125c | 3681 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3682 | |
3d30544f PZ |
3683 | /** |
3684 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 3685 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 3686 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
3687 | * |
3688 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3689 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3690 | * post_init_entity_util_avg(). | |
3691 | * | |
3692 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3693 | * | |
7c3edd2c PZ |
3694 | * Returns true if the load decayed or we removed load. |
3695 | * | |
3696 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3697 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3698 | */ |
a2c6c91f | 3699 | static inline int |
3a123bbb | 3700 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3701 | { |
9f683953 | 3702 | unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; |
9d89c257 | 3703 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3704 | int decayed = 0; |
2dac754e | 3705 | |
2a2f5d4e PZ |
3706 | if (cfs_rq->removed.nr) { |
3707 | unsigned long r; | |
87e867b4 | 3708 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
2a2f5d4e PZ |
3709 | |
3710 | raw_spin_lock(&cfs_rq->removed.lock); | |
3711 | swap(cfs_rq->removed.util_avg, removed_util); | |
3712 | swap(cfs_rq->removed.load_avg, removed_load); | |
9f683953 | 3713 | swap(cfs_rq->removed.runnable_avg, removed_runnable); |
2a2f5d4e PZ |
3714 | cfs_rq->removed.nr = 0; |
3715 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3716 | ||
2a2f5d4e | 3717 | r = removed_load; |
89741892 | 3718 | sub_positive(&sa->load_avg, r); |
1c35b07e | 3719 | sa->load_sum = sa->load_avg * divider; |
2dac754e | 3720 | |
2a2f5d4e | 3721 | r = removed_util; |
89741892 | 3722 | sub_positive(&sa->util_avg, r); |
943858a1 VG |
3723 | sub_positive(&sa->util_sum, r * divider); |
3724 | /* | |
3725 | * Because of rounding, se->util_sum might ends up being +1 more than | |
3726 | * cfs->util_sum. Although this is not a problem by itself, detaching | |
3727 | * a lot of tasks with the rounding problem between 2 updates of | |
3728 | * util_avg (~1ms) can make cfs->util_sum becoming null whereas | |
3729 | * cfs_util_avg is not. | |
3730 | * Check that util_sum is still above its lower bound for the new | |
3731 | * util_avg. Given that period_contrib might have moved since the last | |
3732 | * sync, we are only sure that util_sum must be above or equal to | |
3733 | * util_avg * minimum possible divider | |
3734 | */ | |
3735 | sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER); | |
2a2f5d4e | 3736 | |
9f683953 VG |
3737 | r = removed_runnable; |
3738 | sub_positive(&sa->runnable_avg, r); | |
1c35b07e | 3739 | sa->runnable_sum = sa->runnable_avg * divider; |
9f683953 VG |
3740 | |
3741 | /* | |
3742 | * removed_runnable is the unweighted version of removed_load so we | |
3743 | * can use it to estimate removed_load_sum. | |
3744 | */ | |
3745 | add_tg_cfs_propagate(cfs_rq, | |
3746 | -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); | |
2a2f5d4e PZ |
3747 | |
3748 | decayed = 1; | |
9d89c257 | 3749 | } |
36ee28e4 | 3750 | |
23127296 | 3751 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
36ee28e4 | 3752 | |
9d89c257 YD |
3753 | #ifndef CONFIG_64BIT |
3754 | smp_wmb(); | |
3755 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3756 | #endif | |
36ee28e4 | 3757 | |
2a2f5d4e | 3758 | return decayed; |
21e96f88 SM |
3759 | } |
3760 | ||
3d30544f PZ |
3761 | /** |
3762 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3763 | * @cfs_rq: cfs_rq to attach to | |
3764 | * @se: sched_entity to attach | |
3765 | * | |
3766 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3767 | * cfs_rq->avg.last_update_time being current. | |
3768 | */ | |
a4f9a0e5 | 3769 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
a05e8c51 | 3770 | { |
95d68593 VG |
3771 | /* |
3772 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3773 | * See ___update_load_avg() for details. | |
3774 | */ | |
87e867b4 | 3775 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
f207934f PZ |
3776 | |
3777 | /* | |
3778 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3779 | * window because without that, really weird and wonderful things can | |
3780 | * happen. | |
3781 | * | |
3782 | * XXX illustrate | |
3783 | */ | |
a05e8c51 | 3784 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3785 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3786 | ||
3787 | /* | |
3788 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3789 | * period_contrib. This isn't strictly correct, but since we're | |
3790 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3791 | * _sum a little. | |
3792 | */ | |
3793 | se->avg.util_sum = se->avg.util_avg * divider; | |
3794 | ||
9f683953 VG |
3795 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
3796 | ||
f207934f PZ |
3797 | se->avg.load_sum = divider; |
3798 | if (se_weight(se)) { | |
3799 | se->avg.load_sum = | |
3800 | div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); | |
3801 | } | |
3802 | ||
8d5b9025 | 3803 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3804 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3805 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
9f683953 VG |
3806 | cfs_rq->avg.runnable_avg += se->avg.runnable_avg; |
3807 | cfs_rq->avg.runnable_sum += se->avg.runnable_sum; | |
0e2d2aaa PZ |
3808 | |
3809 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 3810 | |
a4f9a0e5 | 3811 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3812 | |
3813 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3814 | } |
3815 | ||
3d30544f PZ |
3816 | /** |
3817 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3818 | * @cfs_rq: cfs_rq to detach from | |
3819 | * @se: sched_entity to detach | |
3820 | * | |
3821 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3822 | * cfs_rq->avg.last_update_time being current. | |
3823 | */ | |
a05e8c51 BP |
3824 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3825 | { | |
fcf6631f VG |
3826 | /* |
3827 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3828 | * See ___update_load_avg() for details. | |
3829 | */ | |
3830 | u32 divider = get_pelt_divider(&cfs_rq->avg); | |
3831 | ||
8d5b9025 | 3832 | dequeue_load_avg(cfs_rq, se); |
89741892 | 3833 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
fcf6631f | 3834 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; |
9f683953 | 3835 | sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); |
fcf6631f | 3836 | cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; |
0e2d2aaa PZ |
3837 | |
3838 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 3839 | |
ea14b57e | 3840 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3841 | |
3842 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3843 | } |
3844 | ||
b382a531 PZ |
3845 | /* |
3846 | * Optional action to be done while updating the load average | |
3847 | */ | |
3848 | #define UPDATE_TG 0x1 | |
3849 | #define SKIP_AGE_LOAD 0x2 | |
3850 | #define DO_ATTACH 0x4 | |
3851 | ||
3852 | /* Update task and its cfs_rq load average */ | |
3853 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
3854 | { | |
23127296 | 3855 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
3856 | int decayed; |
3857 | ||
3858 | /* | |
3859 | * Track task load average for carrying it to new CPU after migrated, and | |
3860 | * track group sched_entity load average for task_h_load calc in migration | |
3861 | */ | |
3862 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 3863 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
3864 | |
3865 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
3866 | decayed |= propagate_entity_load_avg(se); | |
3867 | ||
3868 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
3869 | ||
ea14b57e PZ |
3870 | /* |
3871 | * DO_ATTACH means we're here from enqueue_entity(). | |
3872 | * !last_update_time means we've passed through | |
3873 | * migrate_task_rq_fair() indicating we migrated. | |
3874 | * | |
3875 | * IOW we're enqueueing a task on a new CPU. | |
3876 | */ | |
a4f9a0e5 | 3877 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 3878 | update_tg_load_avg(cfs_rq); |
b382a531 | 3879 | |
bef69dd8 VG |
3880 | } else if (decayed) { |
3881 | cfs_rq_util_change(cfs_rq, 0); | |
3882 | ||
3883 | if (flags & UPDATE_TG) | |
fe749158 | 3884 | update_tg_load_avg(cfs_rq); |
bef69dd8 | 3885 | } |
b382a531 PZ |
3886 | } |
3887 | ||
9d89c257 | 3888 | #ifndef CONFIG_64BIT |
0905f04e YD |
3889 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3890 | { | |
9d89c257 | 3891 | u64 last_update_time_copy; |
0905f04e | 3892 | u64 last_update_time; |
9ee474f5 | 3893 | |
9d89c257 YD |
3894 | do { |
3895 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3896 | smp_rmb(); | |
3897 | last_update_time = cfs_rq->avg.last_update_time; | |
3898 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3899 | |
3900 | return last_update_time; | |
3901 | } | |
9d89c257 | 3902 | #else |
0905f04e YD |
3903 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3904 | { | |
3905 | return cfs_rq->avg.last_update_time; | |
3906 | } | |
9d89c257 YD |
3907 | #endif |
3908 | ||
104cb16d MR |
3909 | /* |
3910 | * Synchronize entity load avg of dequeued entity without locking | |
3911 | * the previous rq. | |
3912 | */ | |
71b47eaf | 3913 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
3914 | { |
3915 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3916 | u64 last_update_time; | |
3917 | ||
3918 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 3919 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
3920 | } |
3921 | ||
0905f04e YD |
3922 | /* |
3923 | * Task first catches up with cfs_rq, and then subtract | |
3924 | * itself from the cfs_rq (task must be off the queue now). | |
3925 | */ | |
71b47eaf | 3926 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
3927 | { |
3928 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 3929 | unsigned long flags; |
0905f04e YD |
3930 | |
3931 | /* | |
7dc603c9 PZ |
3932 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3933 | * post_init_entity_util_avg() which will have added things to the | |
3934 | * cfs_rq, so we can remove unconditionally. | |
0905f04e | 3935 | */ |
0905f04e | 3936 | |
104cb16d | 3937 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
3938 | |
3939 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
3940 | ++cfs_rq->removed.nr; | |
3941 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
3942 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
9f683953 | 3943 | cfs_rq->removed.runnable_avg += se->avg.runnable_avg; |
2a2f5d4e | 3944 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 3945 | } |
642dbc39 | 3946 | |
9f683953 VG |
3947 | static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) |
3948 | { | |
3949 | return cfs_rq->avg.runnable_avg; | |
3950 | } | |
3951 | ||
7ea241af YD |
3952 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) |
3953 | { | |
3954 | return cfs_rq->avg.load_avg; | |
3955 | } | |
3956 | ||
d91cecc1 CY |
3957 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); |
3958 | ||
7f65ea42 PB |
3959 | static inline unsigned long task_util(struct task_struct *p) |
3960 | { | |
3961 | return READ_ONCE(p->se.avg.util_avg); | |
3962 | } | |
3963 | ||
3964 | static inline unsigned long _task_util_est(struct task_struct *p) | |
3965 | { | |
3966 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
3967 | ||
68d7a190 | 3968 | return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED)); |
7f65ea42 PB |
3969 | } |
3970 | ||
3971 | static inline unsigned long task_util_est(struct task_struct *p) | |
3972 | { | |
3973 | return max(task_util(p), _task_util_est(p)); | |
3974 | } | |
3975 | ||
a7008c07 VS |
3976 | #ifdef CONFIG_UCLAMP_TASK |
3977 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
3978 | { | |
3979 | return clamp(task_util_est(p), | |
3980 | uclamp_eff_value(p, UCLAMP_MIN), | |
3981 | uclamp_eff_value(p, UCLAMP_MAX)); | |
3982 | } | |
3983 | #else | |
3984 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
3985 | { | |
3986 | return task_util_est(p); | |
3987 | } | |
3988 | #endif | |
3989 | ||
7f65ea42 PB |
3990 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, |
3991 | struct task_struct *p) | |
3992 | { | |
3993 | unsigned int enqueued; | |
3994 | ||
3995 | if (!sched_feat(UTIL_EST)) | |
3996 | return; | |
3997 | ||
3998 | /* Update root cfs_rq's estimated utilization */ | |
3999 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 4000 | enqueued += _task_util_est(p); |
7f65ea42 | 4001 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
4581bea8 VD |
4002 | |
4003 | trace_sched_util_est_cfs_tp(cfs_rq); | |
7f65ea42 PB |
4004 | } |
4005 | ||
8c1f560c XY |
4006 | static inline void util_est_dequeue(struct cfs_rq *cfs_rq, |
4007 | struct task_struct *p) | |
4008 | { | |
4009 | unsigned int enqueued; | |
4010 | ||
4011 | if (!sched_feat(UTIL_EST)) | |
4012 | return; | |
4013 | ||
4014 | /* Update root cfs_rq's estimated utilization */ | |
4015 | enqueued = cfs_rq->avg.util_est.enqueued; | |
4016 | enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); | |
4017 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); | |
4018 | ||
4019 | trace_sched_util_est_cfs_tp(cfs_rq); | |
4020 | } | |
4021 | ||
b89997aa VD |
4022 | #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100) |
4023 | ||
7f65ea42 PB |
4024 | /* |
4025 | * Check if a (signed) value is within a specified (unsigned) margin, | |
4026 | * based on the observation that: | |
4027 | * | |
4028 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
4029 | * | |
3b03706f | 4030 | * NOTE: this only works when value + margin < INT_MAX. |
7f65ea42 PB |
4031 | */ |
4032 | static inline bool within_margin(int value, int margin) | |
4033 | { | |
4034 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
4035 | } | |
4036 | ||
8c1f560c XY |
4037 | static inline void util_est_update(struct cfs_rq *cfs_rq, |
4038 | struct task_struct *p, | |
4039 | bool task_sleep) | |
7f65ea42 | 4040 | { |
b89997aa | 4041 | long last_ewma_diff, last_enqueued_diff; |
7f65ea42 PB |
4042 | struct util_est ue; |
4043 | ||
4044 | if (!sched_feat(UTIL_EST)) | |
4045 | return; | |
4046 | ||
7f65ea42 PB |
4047 | /* |
4048 | * Skip update of task's estimated utilization when the task has not | |
4049 | * yet completed an activation, e.g. being migrated. | |
4050 | */ | |
4051 | if (!task_sleep) | |
4052 | return; | |
4053 | ||
d519329f PB |
4054 | /* |
4055 | * If the PELT values haven't changed since enqueue time, | |
4056 | * skip the util_est update. | |
4057 | */ | |
4058 | ue = p->se.avg.util_est; | |
4059 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
4060 | return; | |
4061 | ||
b89997aa VD |
4062 | last_enqueued_diff = ue.enqueued; |
4063 | ||
b8c96361 PB |
4064 | /* |
4065 | * Reset EWMA on utilization increases, the moving average is used only | |
4066 | * to smooth utilization decreases. | |
4067 | */ | |
68d7a190 | 4068 | ue.enqueued = task_util(p); |
b8c96361 PB |
4069 | if (sched_feat(UTIL_EST_FASTUP)) { |
4070 | if (ue.ewma < ue.enqueued) { | |
4071 | ue.ewma = ue.enqueued; | |
4072 | goto done; | |
4073 | } | |
4074 | } | |
4075 | ||
7f65ea42 | 4076 | /* |
b89997aa | 4077 | * Skip update of task's estimated utilization when its members are |
7f65ea42 PB |
4078 | * already ~1% close to its last activation value. |
4079 | */ | |
7f65ea42 | 4080 | last_ewma_diff = ue.enqueued - ue.ewma; |
b89997aa VD |
4081 | last_enqueued_diff -= ue.enqueued; |
4082 | if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) { | |
4083 | if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN)) | |
4084 | goto done; | |
4085 | ||
7f65ea42 | 4086 | return; |
b89997aa | 4087 | } |
7f65ea42 | 4088 | |
10a35e68 VG |
4089 | /* |
4090 | * To avoid overestimation of actual task utilization, skip updates if | |
4091 | * we cannot grant there is idle time in this CPU. | |
4092 | */ | |
8c1f560c | 4093 | if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq)))) |
10a35e68 VG |
4094 | return; |
4095 | ||
7f65ea42 PB |
4096 | /* |
4097 | * Update Task's estimated utilization | |
4098 | * | |
4099 | * When *p completes an activation we can consolidate another sample | |
4100 | * of the task size. This is done by storing the current PELT value | |
4101 | * as ue.enqueued and by using this value to update the Exponential | |
4102 | * Weighted Moving Average (EWMA): | |
4103 | * | |
4104 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
4105 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
4106 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
4107 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
4108 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
4109 | * | |
4110 | * Where 'w' is the weight of new samples, which is configured to be | |
4111 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
4112 | */ | |
4113 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
4114 | ue.ewma += last_ewma_diff; | |
4115 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
b8c96361 | 4116 | done: |
68d7a190 | 4117 | ue.enqueued |= UTIL_AVG_UNCHANGED; |
7f65ea42 | 4118 | WRITE_ONCE(p->se.avg.util_est, ue); |
4581bea8 VD |
4119 | |
4120 | trace_sched_util_est_se_tp(&p->se); | |
7f65ea42 PB |
4121 | } |
4122 | ||
3b1baa64 MR |
4123 | static inline int task_fits_capacity(struct task_struct *p, long capacity) |
4124 | { | |
a7008c07 | 4125 | return fits_capacity(uclamp_task_util(p), capacity); |
3b1baa64 MR |
4126 | } |
4127 | ||
4128 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
4129 | { | |
4130 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) | |
4131 | return; | |
4132 | ||
0ae78eec | 4133 | if (!p || p->nr_cpus_allowed == 1) { |
3b1baa64 MR |
4134 | rq->misfit_task_load = 0; |
4135 | return; | |
4136 | } | |
4137 | ||
4138 | if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { | |
4139 | rq->misfit_task_load = 0; | |
4140 | return; | |
4141 | } | |
4142 | ||
01cfcde9 VG |
4143 | /* |
4144 | * Make sure that misfit_task_load will not be null even if | |
4145 | * task_h_load() returns 0. | |
4146 | */ | |
4147 | rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); | |
3b1baa64 MR |
4148 | } |
4149 | ||
38033c37 PZ |
4150 | #else /* CONFIG_SMP */ |
4151 | ||
a7b359fc OU |
4152 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
4153 | { | |
4154 | return true; | |
4155 | } | |
4156 | ||
d31b1a66 VG |
4157 | #define UPDATE_TG 0x0 |
4158 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 4159 | #define DO_ATTACH 0x0 |
d31b1a66 | 4160 | |
88c0616e | 4161 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 4162 | { |
ea14b57e | 4163 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
4164 | } |
4165 | ||
9d89c257 | 4166 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 4167 | |
a05e8c51 | 4168 | static inline void |
a4f9a0e5 | 4169 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} |
a05e8c51 BP |
4170 | static inline void |
4171 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
4172 | ||
d91cecc1 | 4173 | static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
4174 | { |
4175 | return 0; | |
4176 | } | |
4177 | ||
7f65ea42 PB |
4178 | static inline void |
4179 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
4180 | ||
4181 | static inline void | |
8c1f560c XY |
4182 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {} |
4183 | ||
4184 | static inline void | |
4185 | util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p, | |
4186 | bool task_sleep) {} | |
3b1baa64 | 4187 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 4188 | |
38033c37 | 4189 | #endif /* CONFIG_SMP */ |
9d85f21c | 4190 | |
ddc97297 PZ |
4191 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4192 | { | |
4193 | #ifdef CONFIG_SCHED_DEBUG | |
4194 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
4195 | ||
4196 | if (d < 0) | |
4197 | d = -d; | |
4198 | ||
4199 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 4200 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
4201 | #endif |
4202 | } | |
4203 | ||
aeb73b04 PZ |
4204 | static void |
4205 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
4206 | { | |
1af5f730 | 4207 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 4208 | |
2cb8600e PZ |
4209 | /* |
4210 | * The 'current' period is already promised to the current tasks, | |
4211 | * however the extra weight of the new task will slow them down a | |
4212 | * little, place the new task so that it fits in the slot that | |
4213 | * stays open at the end. | |
4214 | */ | |
94dfb5e7 | 4215 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 4216 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 4217 | |
a2e7a7eb | 4218 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 4219 | if (!initial) { |
a2e7a7eb | 4220 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 4221 | |
a2e7a7eb MG |
4222 | /* |
4223 | * Halve their sleep time's effect, to allow | |
4224 | * for a gentler effect of sleepers: | |
4225 | */ | |
4226 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
4227 | thresh >>= 1; | |
51e0304c | 4228 | |
a2e7a7eb | 4229 | vruntime -= thresh; |
aeb73b04 PZ |
4230 | } |
4231 | ||
b5d9d734 | 4232 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 4233 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
4234 | } |
4235 | ||
d3d9dc33 PT |
4236 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
4237 | ||
cb251765 MG |
4238 | static inline void check_schedstat_required(void) |
4239 | { | |
4240 | #ifdef CONFIG_SCHEDSTATS | |
4241 | if (schedstat_enabled()) | |
4242 | return; | |
4243 | ||
4244 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
4245 | if (trace_sched_stat_wait_enabled() || | |
4246 | trace_sched_stat_sleep_enabled() || | |
4247 | trace_sched_stat_iowait_enabled() || | |
4248 | trace_sched_stat_blocked_enabled() || | |
4249 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 4250 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 | 4251 | "stat_blocked and stat_runtime require the " |
f67abed5 | 4252 | "kernel parameter schedstats=enable or " |
cb251765 MG |
4253 | "kernel.sched_schedstats=1\n"); |
4254 | } | |
4255 | #endif | |
4256 | } | |
4257 | ||
fe61468b | 4258 | static inline bool cfs_bandwidth_used(void); |
b5179ac7 PZ |
4259 | |
4260 | /* | |
4261 | * MIGRATION | |
4262 | * | |
4263 | * dequeue | |
4264 | * update_curr() | |
4265 | * update_min_vruntime() | |
4266 | * vruntime -= min_vruntime | |
4267 | * | |
4268 | * enqueue | |
4269 | * update_curr() | |
4270 | * update_min_vruntime() | |
4271 | * vruntime += min_vruntime | |
4272 | * | |
4273 | * this way the vruntime transition between RQs is done when both | |
4274 | * min_vruntime are up-to-date. | |
4275 | * | |
4276 | * WAKEUP (remote) | |
4277 | * | |
59efa0ba | 4278 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
4279 | * vruntime -= min_vruntime |
4280 | * | |
4281 | * enqueue | |
4282 | * update_curr() | |
4283 | * update_min_vruntime() | |
4284 | * vruntime += min_vruntime | |
4285 | * | |
4286 | * this way we don't have the most up-to-date min_vruntime on the originating | |
4287 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
4288 | */ | |
4289 | ||
bf0f6f24 | 4290 | static void |
88ec22d3 | 4291 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4292 | { |
2f950354 PZ |
4293 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
4294 | bool curr = cfs_rq->curr == se; | |
4295 | ||
88ec22d3 | 4296 | /* |
2f950354 PZ |
4297 | * If we're the current task, we must renormalise before calling |
4298 | * update_curr(). | |
88ec22d3 | 4299 | */ |
2f950354 | 4300 | if (renorm && curr) |
88ec22d3 PZ |
4301 | se->vruntime += cfs_rq->min_vruntime; |
4302 | ||
2f950354 PZ |
4303 | update_curr(cfs_rq); |
4304 | ||
bf0f6f24 | 4305 | /* |
2f950354 PZ |
4306 | * Otherwise, renormalise after, such that we're placed at the current |
4307 | * moment in time, instead of some random moment in the past. Being | |
4308 | * placed in the past could significantly boost this task to the | |
4309 | * fairness detriment of existing tasks. | |
bf0f6f24 | 4310 | */ |
2f950354 PZ |
4311 | if (renorm && !curr) |
4312 | se->vruntime += cfs_rq->min_vruntime; | |
4313 | ||
89ee048f VG |
4314 | /* |
4315 | * When enqueuing a sched_entity, we must: | |
4316 | * - Update loads to have both entity and cfs_rq synced with now. | |
9f683953 | 4317 | * - Add its load to cfs_rq->runnable_avg |
89ee048f VG |
4318 | * - For group_entity, update its weight to reflect the new share of |
4319 | * its group cfs_rq | |
4320 | * - Add its new weight to cfs_rq->load.weight | |
4321 | */ | |
b382a531 | 4322 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
9f683953 | 4323 | se_update_runnable(se); |
1ea6c46a | 4324 | update_cfs_group(se); |
17bc14b7 | 4325 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 4326 | |
1a3d027c | 4327 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 4328 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 4329 | |
cb251765 | 4330 | check_schedstat_required(); |
4fa8d299 JP |
4331 | update_stats_enqueue(cfs_rq, se, flags); |
4332 | check_spread(cfs_rq, se); | |
2f950354 | 4333 | if (!curr) |
83b699ed | 4334 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4335 | se->on_rq = 1; |
3d4b47b4 | 4336 | |
fe61468b VG |
4337 | /* |
4338 | * When bandwidth control is enabled, cfs might have been removed | |
4339 | * because of a parent been throttled but cfs->nr_running > 1. Try to | |
3b03706f | 4340 | * add it unconditionally. |
fe61468b VG |
4341 | */ |
4342 | if (cfs_rq->nr_running == 1 || cfs_bandwidth_used()) | |
3d4b47b4 | 4343 | list_add_leaf_cfs_rq(cfs_rq); |
fe61468b VG |
4344 | |
4345 | if (cfs_rq->nr_running == 1) | |
d3d9dc33 | 4346 | check_enqueue_throttle(cfs_rq); |
bf0f6f24 IM |
4347 | } |
4348 | ||
2c13c919 | 4349 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4350 | { |
2c13c919 RR |
4351 | for_each_sched_entity(se) { |
4352 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4353 | if (cfs_rq->last != se) |
2c13c919 | 4354 | break; |
f1044799 PZ |
4355 | |
4356 | cfs_rq->last = NULL; | |
2c13c919 RR |
4357 | } |
4358 | } | |
2002c695 | 4359 | |
2c13c919 RR |
4360 | static void __clear_buddies_next(struct sched_entity *se) |
4361 | { | |
4362 | for_each_sched_entity(se) { | |
4363 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4364 | if (cfs_rq->next != se) |
2c13c919 | 4365 | break; |
f1044799 PZ |
4366 | |
4367 | cfs_rq->next = NULL; | |
2c13c919 | 4368 | } |
2002c695 PZ |
4369 | } |
4370 | ||
ac53db59 RR |
4371 | static void __clear_buddies_skip(struct sched_entity *se) |
4372 | { | |
4373 | for_each_sched_entity(se) { | |
4374 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4375 | if (cfs_rq->skip != se) |
ac53db59 | 4376 | break; |
f1044799 PZ |
4377 | |
4378 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4379 | } |
4380 | } | |
4381 | ||
a571bbea PZ |
4382 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4383 | { | |
2c13c919 RR |
4384 | if (cfs_rq->last == se) |
4385 | __clear_buddies_last(se); | |
4386 | ||
4387 | if (cfs_rq->next == se) | |
4388 | __clear_buddies_next(se); | |
ac53db59 RR |
4389 | |
4390 | if (cfs_rq->skip == se) | |
4391 | __clear_buddies_skip(se); | |
a571bbea PZ |
4392 | } |
4393 | ||
6c16a6dc | 4394 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4395 | |
bf0f6f24 | 4396 | static void |
371fd7e7 | 4397 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4398 | { |
a2a2d680 DA |
4399 | /* |
4400 | * Update run-time statistics of the 'current'. | |
4401 | */ | |
4402 | update_curr(cfs_rq); | |
89ee048f VG |
4403 | |
4404 | /* | |
4405 | * When dequeuing a sched_entity, we must: | |
4406 | * - Update loads to have both entity and cfs_rq synced with now. | |
9f683953 | 4407 | * - Subtract its load from the cfs_rq->runnable_avg. |
dfcb245e | 4408 | * - Subtract its previous weight from cfs_rq->load.weight. |
89ee048f VG |
4409 | * - For group entity, update its weight to reflect the new share |
4410 | * of its group cfs_rq. | |
4411 | */ | |
88c0616e | 4412 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 4413 | se_update_runnable(se); |
a2a2d680 | 4414 | |
4fa8d299 | 4415 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 4416 | |
2002c695 | 4417 | clear_buddies(cfs_rq, se); |
4793241b | 4418 | |
83b699ed | 4419 | if (se != cfs_rq->curr) |
30cfdcfc | 4420 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4421 | se->on_rq = 0; |
30cfdcfc | 4422 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4423 | |
4424 | /* | |
b60205c7 PZ |
4425 | * Normalize after update_curr(); which will also have moved |
4426 | * min_vruntime if @se is the one holding it back. But before doing | |
4427 | * update_min_vruntime() again, which will discount @se's position and | |
4428 | * can move min_vruntime forward still more. | |
88ec22d3 | 4429 | */ |
371fd7e7 | 4430 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4431 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4432 | |
d8b4986d PT |
4433 | /* return excess runtime on last dequeue */ |
4434 | return_cfs_rq_runtime(cfs_rq); | |
4435 | ||
1ea6c46a | 4436 | update_cfs_group(se); |
b60205c7 PZ |
4437 | |
4438 | /* | |
4439 | * Now advance min_vruntime if @se was the entity holding it back, | |
4440 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4441 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4442 | * further than we started -- ie. we'll be penalized. | |
4443 | */ | |
9845c49c | 4444 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4445 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
4446 | } |
4447 | ||
4448 | /* | |
4449 | * Preempt the current task with a newly woken task if needed: | |
4450 | */ | |
7c92e54f | 4451 | static void |
2e09bf55 | 4452 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4453 | { |
11697830 | 4454 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4455 | struct sched_entity *se; |
4456 | s64 delta; | |
11697830 | 4457 | |
6d0f0ebd | 4458 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4459 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4460 | if (delta_exec > ideal_runtime) { |
8875125e | 4461 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4462 | /* |
4463 | * The current task ran long enough, ensure it doesn't get | |
4464 | * re-elected due to buddy favours. | |
4465 | */ | |
4466 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4467 | return; |
4468 | } | |
4469 | ||
4470 | /* | |
4471 | * Ensure that a task that missed wakeup preemption by a | |
4472 | * narrow margin doesn't have to wait for a full slice. | |
4473 | * This also mitigates buddy induced latencies under load. | |
4474 | */ | |
f685ceac MG |
4475 | if (delta_exec < sysctl_sched_min_granularity) |
4476 | return; | |
4477 | ||
f4cfb33e WX |
4478 | se = __pick_first_entity(cfs_rq); |
4479 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4480 | |
f4cfb33e WX |
4481 | if (delta < 0) |
4482 | return; | |
d7d82944 | 4483 | |
f4cfb33e | 4484 | if (delta > ideal_runtime) |
8875125e | 4485 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4486 | } |
4487 | ||
83b699ed | 4488 | static void |
8494f412 | 4489 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4490 | { |
21f56ffe PZ |
4491 | clear_buddies(cfs_rq, se); |
4492 | ||
83b699ed SV |
4493 | /* 'current' is not kept within the tree. */ |
4494 | if (se->on_rq) { | |
4495 | /* | |
4496 | * Any task has to be enqueued before it get to execute on | |
4497 | * a CPU. So account for the time it spent waiting on the | |
4498 | * runqueue. | |
4499 | */ | |
4fa8d299 | 4500 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 4501 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4502 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4503 | } |
4504 | ||
79303e9e | 4505 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4506 | cfs_rq->curr = se; |
4fa8d299 | 4507 | |
eba1ed4b IM |
4508 | /* |
4509 | * Track our maximum slice length, if the CPU's load is at | |
4510 | * least twice that of our own weight (i.e. dont track it | |
4511 | * when there are only lesser-weight tasks around): | |
4512 | */ | |
f2bedc47 DE |
4513 | if (schedstat_enabled() && |
4514 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
4fa8d299 JP |
4515 | schedstat_set(se->statistics.slice_max, |
4516 | max((u64)schedstat_val(se->statistics.slice_max), | |
4517 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 4518 | } |
4fa8d299 | 4519 | |
4a55b450 | 4520 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4521 | } |
4522 | ||
3f3a4904 PZ |
4523 | static int |
4524 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4525 | ||
ac53db59 RR |
4526 | /* |
4527 | * Pick the next process, keeping these things in mind, in this order: | |
4528 | * 1) keep things fair between processes/task groups | |
4529 | * 2) pick the "next" process, since someone really wants that to run | |
4530 | * 3) pick the "last" process, for cache locality | |
4531 | * 4) do not run the "skip" process, if something else is available | |
4532 | */ | |
678d5718 PZ |
4533 | static struct sched_entity * |
4534 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4535 | { |
678d5718 PZ |
4536 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4537 | struct sched_entity *se; | |
4538 | ||
4539 | /* | |
4540 | * If curr is set we have to see if its left of the leftmost entity | |
4541 | * still in the tree, provided there was anything in the tree at all. | |
4542 | */ | |
4543 | if (!left || (curr && entity_before(curr, left))) | |
4544 | left = curr; | |
4545 | ||
4546 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4547 | |
ac53db59 RR |
4548 | /* |
4549 | * Avoid running the skip buddy, if running something else can | |
4550 | * be done without getting too unfair. | |
4551 | */ | |
21f56ffe | 4552 | if (cfs_rq->skip && cfs_rq->skip == se) { |
678d5718 PZ |
4553 | struct sched_entity *second; |
4554 | ||
4555 | if (se == curr) { | |
4556 | second = __pick_first_entity(cfs_rq); | |
4557 | } else { | |
4558 | second = __pick_next_entity(se); | |
4559 | if (!second || (curr && entity_before(curr, second))) | |
4560 | second = curr; | |
4561 | } | |
4562 | ||
ac53db59 RR |
4563 | if (second && wakeup_preempt_entity(second, left) < 1) |
4564 | se = second; | |
4565 | } | |
aa2ac252 | 4566 | |
9abb8973 PO |
4567 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { |
4568 | /* | |
4569 | * Someone really wants this to run. If it's not unfair, run it. | |
4570 | */ | |
ac53db59 | 4571 | se = cfs_rq->next; |
9abb8973 PO |
4572 | } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { |
4573 | /* | |
4574 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4575 | */ | |
4576 | se = cfs_rq->last; | |
4577 | } | |
ac53db59 | 4578 | |
4793241b | 4579 | return se; |
aa2ac252 PZ |
4580 | } |
4581 | ||
678d5718 | 4582 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4583 | |
ab6cde26 | 4584 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4585 | { |
4586 | /* | |
4587 | * If still on the runqueue then deactivate_task() | |
4588 | * was not called and update_curr() has to be done: | |
4589 | */ | |
4590 | if (prev->on_rq) | |
b7cc0896 | 4591 | update_curr(cfs_rq); |
bf0f6f24 | 4592 | |
d3d9dc33 PT |
4593 | /* throttle cfs_rqs exceeding runtime */ |
4594 | check_cfs_rq_runtime(cfs_rq); | |
4595 | ||
4fa8d299 | 4596 | check_spread(cfs_rq, prev); |
cb251765 | 4597 | |
30cfdcfc | 4598 | if (prev->on_rq) { |
4fa8d299 | 4599 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
4600 | /* Put 'current' back into the tree. */ |
4601 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4602 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4603 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4604 | } |
429d43bc | 4605 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4606 | } |
4607 | ||
8f4d37ec PZ |
4608 | static void |
4609 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4610 | { |
bf0f6f24 | 4611 | /* |
30cfdcfc | 4612 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4613 | */ |
30cfdcfc | 4614 | update_curr(cfs_rq); |
bf0f6f24 | 4615 | |
9d85f21c PT |
4616 | /* |
4617 | * Ensure that runnable average is periodically updated. | |
4618 | */ | |
88c0616e | 4619 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4620 | update_cfs_group(curr); |
9d85f21c | 4621 | |
8f4d37ec PZ |
4622 | #ifdef CONFIG_SCHED_HRTICK |
4623 | /* | |
4624 | * queued ticks are scheduled to match the slice, so don't bother | |
4625 | * validating it and just reschedule. | |
4626 | */ | |
983ed7a6 | 4627 | if (queued) { |
8875125e | 4628 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4629 | return; |
4630 | } | |
8f4d37ec PZ |
4631 | /* |
4632 | * don't let the period tick interfere with the hrtick preemption | |
4633 | */ | |
4634 | if (!sched_feat(DOUBLE_TICK) && | |
4635 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4636 | return; | |
4637 | #endif | |
4638 | ||
2c2efaed | 4639 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4640 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4641 | } |
4642 | ||
ab84d31e PT |
4643 | |
4644 | /************************************************** | |
4645 | * CFS bandwidth control machinery | |
4646 | */ | |
4647 | ||
4648 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 4649 | |
e9666d10 | 4650 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 4651 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4652 | |
4653 | static inline bool cfs_bandwidth_used(void) | |
4654 | { | |
c5905afb | 4655 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4656 | } |
4657 | ||
1ee14e6c | 4658 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4659 | { |
ce48c146 | 4660 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
4661 | } |
4662 | ||
4663 | void cfs_bandwidth_usage_dec(void) | |
4664 | { | |
ce48c146 | 4665 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 4666 | } |
e9666d10 | 4667 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
4668 | static bool cfs_bandwidth_used(void) |
4669 | { | |
4670 | return true; | |
4671 | } | |
4672 | ||
1ee14e6c BS |
4673 | void cfs_bandwidth_usage_inc(void) {} |
4674 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 4675 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 4676 | |
ab84d31e PT |
4677 | /* |
4678 | * default period for cfs group bandwidth. | |
4679 | * default: 0.1s, units: nanoseconds | |
4680 | */ | |
4681 | static inline u64 default_cfs_period(void) | |
4682 | { | |
4683 | return 100000000ULL; | |
4684 | } | |
ec12cb7f PT |
4685 | |
4686 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4687 | { | |
4688 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4689 | } | |
4690 | ||
a9cf55b2 | 4691 | /* |
763a9ec0 QC |
4692 | * Replenish runtime according to assigned quota. We use sched_clock_cpu |
4693 | * directly instead of rq->clock to avoid adding additional synchronization | |
4694 | * around rq->lock. | |
a9cf55b2 PT |
4695 | * |
4696 | * requires cfs_b->lock | |
4697 | */ | |
029632fb | 4698 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 | 4699 | { |
f4183717 HC |
4700 | if (unlikely(cfs_b->quota == RUNTIME_INF)) |
4701 | return; | |
4702 | ||
4703 | cfs_b->runtime += cfs_b->quota; | |
4704 | cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst); | |
a9cf55b2 PT |
4705 | } |
4706 | ||
029632fb PZ |
4707 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4708 | { | |
4709 | return &tg->cfs_bandwidth; | |
4710 | } | |
4711 | ||
85dac906 | 4712 | /* returns 0 on failure to allocate runtime */ |
e98fa02c PT |
4713 | static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, |
4714 | struct cfs_rq *cfs_rq, u64 target_runtime) | |
ec12cb7f | 4715 | { |
e98fa02c PT |
4716 | u64 min_amount, amount = 0; |
4717 | ||
4718 | lockdep_assert_held(&cfs_b->lock); | |
ec12cb7f PT |
4719 | |
4720 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
e98fa02c | 4721 | min_amount = target_runtime - cfs_rq->runtime_remaining; |
ec12cb7f | 4722 | |
ec12cb7f PT |
4723 | if (cfs_b->quota == RUNTIME_INF) |
4724 | amount = min_amount; | |
58088ad0 | 4725 | else { |
77a4d1a1 | 4726 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4727 | |
4728 | if (cfs_b->runtime > 0) { | |
4729 | amount = min(cfs_b->runtime, min_amount); | |
4730 | cfs_b->runtime -= amount; | |
4731 | cfs_b->idle = 0; | |
4732 | } | |
ec12cb7f | 4733 | } |
ec12cb7f PT |
4734 | |
4735 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
4736 | |
4737 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4738 | } |
4739 | ||
e98fa02c PT |
4740 | /* returns 0 on failure to allocate runtime */ |
4741 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4742 | { | |
4743 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4744 | int ret; | |
4745 | ||
4746 | raw_spin_lock(&cfs_b->lock); | |
4747 | ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); | |
4748 | raw_spin_unlock(&cfs_b->lock); | |
4749 | ||
4750 | return ret; | |
4751 | } | |
4752 | ||
9dbdb155 | 4753 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4754 | { |
4755 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4756 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4757 | |
4758 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4759 | return; |
4760 | ||
5e2d2cc2 L |
4761 | if (cfs_rq->throttled) |
4762 | return; | |
85dac906 PT |
4763 | /* |
4764 | * if we're unable to extend our runtime we resched so that the active | |
4765 | * hierarchy can be throttled | |
4766 | */ | |
4767 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4768 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4769 | } |
4770 | ||
6c16a6dc | 4771 | static __always_inline |
9dbdb155 | 4772 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4773 | { |
56f570e5 | 4774 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4775 | return; |
4776 | ||
4777 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4778 | } | |
4779 | ||
85dac906 PT |
4780 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4781 | { | |
56f570e5 | 4782 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4783 | } |
4784 | ||
64660c86 PT |
4785 | /* check whether cfs_rq, or any parent, is throttled */ |
4786 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4787 | { | |
56f570e5 | 4788 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4789 | } |
4790 | ||
4791 | /* | |
4792 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4793 | * dest_cpu are members of a throttled hierarchy when performing group | |
4794 | * load-balance operations. | |
4795 | */ | |
4796 | static inline int throttled_lb_pair(struct task_group *tg, | |
4797 | int src_cpu, int dest_cpu) | |
4798 | { | |
4799 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4800 | ||
4801 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4802 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4803 | ||
4804 | return throttled_hierarchy(src_cfs_rq) || | |
4805 | throttled_hierarchy(dest_cfs_rq); | |
4806 | } | |
4807 | ||
64660c86 PT |
4808 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
4809 | { | |
4810 | struct rq *rq = data; | |
4811 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4812 | ||
4813 | cfs_rq->throttle_count--; | |
64660c86 | 4814 | if (!cfs_rq->throttle_count) { |
78becc27 | 4815 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4816 | cfs_rq->throttled_clock_task; |
31bc6aea | 4817 | |
a7b359fc OU |
4818 | /* Add cfs_rq with load or one or more already running entities to the list */ |
4819 | if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running) | |
31bc6aea | 4820 | list_add_leaf_cfs_rq(cfs_rq); |
64660c86 | 4821 | } |
64660c86 PT |
4822 | |
4823 | return 0; | |
4824 | } | |
4825 | ||
4826 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4827 | { | |
4828 | struct rq *rq = data; | |
4829 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4830 | ||
82958366 | 4831 | /* group is entering throttled state, stop time */ |
31bc6aea | 4832 | if (!cfs_rq->throttle_count) { |
78becc27 | 4833 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
31bc6aea VG |
4834 | list_del_leaf_cfs_rq(cfs_rq); |
4835 | } | |
64660c86 PT |
4836 | cfs_rq->throttle_count++; |
4837 | ||
4838 | return 0; | |
4839 | } | |
4840 | ||
e98fa02c | 4841 | static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4842 | { |
4843 | struct rq *rq = rq_of(cfs_rq); | |
4844 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4845 | struct sched_entity *se; | |
43e9f7f2 | 4846 | long task_delta, idle_task_delta, dequeue = 1; |
e98fa02c PT |
4847 | |
4848 | raw_spin_lock(&cfs_b->lock); | |
4849 | /* This will start the period timer if necessary */ | |
4850 | if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { | |
4851 | /* | |
4852 | * We have raced with bandwidth becoming available, and if we | |
4853 | * actually throttled the timer might not unthrottle us for an | |
4854 | * entire period. We additionally needed to make sure that any | |
4855 | * subsequent check_cfs_rq_runtime calls agree not to throttle | |
4856 | * us, as we may commit to do cfs put_prev+pick_next, so we ask | |
4857 | * for 1ns of runtime rather than just check cfs_b. | |
4858 | */ | |
4859 | dequeue = 0; | |
4860 | } else { | |
4861 | list_add_tail_rcu(&cfs_rq->throttled_list, | |
4862 | &cfs_b->throttled_cfs_rq); | |
4863 | } | |
4864 | raw_spin_unlock(&cfs_b->lock); | |
4865 | ||
4866 | if (!dequeue) | |
4867 | return false; /* Throttle no longer required. */ | |
85dac906 PT |
4868 | |
4869 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4870 | ||
f1b17280 | 4871 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4872 | rcu_read_lock(); |
4873 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4874 | rcu_read_unlock(); | |
85dac906 PT |
4875 | |
4876 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4877 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
4878 | for_each_sched_entity(se) { |
4879 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4880 | /* throttled entity or throttle-on-deactivate */ | |
4881 | if (!se->on_rq) | |
b6d37a76 | 4882 | goto done; |
85dac906 | 4883 | |
b6d37a76 | 4884 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
6212437f | 4885 | |
30400039 JD |
4886 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
4887 | idle_task_delta = cfs_rq->h_nr_running; | |
4888 | ||
85dac906 | 4889 | qcfs_rq->h_nr_running -= task_delta; |
43e9f7f2 | 4890 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 | 4891 | |
b6d37a76 PW |
4892 | if (qcfs_rq->load.weight) { |
4893 | /* Avoid re-evaluating load for this entity: */ | |
4894 | se = parent_entity(se); | |
4895 | break; | |
4896 | } | |
4897 | } | |
4898 | ||
4899 | for_each_sched_entity(se) { | |
4900 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4901 | /* throttled entity or throttle-on-deactivate */ | |
4902 | if (!se->on_rq) | |
4903 | goto done; | |
4904 | ||
4905 | update_load_avg(qcfs_rq, se, 0); | |
4906 | se_update_runnable(se); | |
4907 | ||
30400039 JD |
4908 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
4909 | idle_task_delta = cfs_rq->h_nr_running; | |
4910 | ||
b6d37a76 PW |
4911 | qcfs_rq->h_nr_running -= task_delta; |
4912 | qcfs_rq->idle_h_nr_running -= idle_task_delta; | |
85dac906 PT |
4913 | } |
4914 | ||
b6d37a76 PW |
4915 | /* At this point se is NULL and we are at root level*/ |
4916 | sub_nr_running(rq, task_delta); | |
85dac906 | 4917 | |
b6d37a76 | 4918 | done: |
c06f04c7 | 4919 | /* |
e98fa02c PT |
4920 | * Note: distribution will already see us throttled via the |
4921 | * throttled-list. rq->lock protects completion. | |
c06f04c7 | 4922 | */ |
e98fa02c PT |
4923 | cfs_rq->throttled = 1; |
4924 | cfs_rq->throttled_clock = rq_clock(rq); | |
4925 | return true; | |
85dac906 PT |
4926 | } |
4927 | ||
029632fb | 4928 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4929 | { |
4930 | struct rq *rq = rq_of(cfs_rq); | |
4931 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4932 | struct sched_entity *se; | |
43e9f7f2 | 4933 | long task_delta, idle_task_delta; |
671fd9da | 4934 | |
22b958d8 | 4935 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4936 | |
4937 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4938 | |
4939 | update_rq_clock(rq); | |
4940 | ||
671fd9da | 4941 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4942 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4943 | list_del_rcu(&cfs_rq->throttled_list); |
4944 | raw_spin_unlock(&cfs_b->lock); | |
4945 | ||
64660c86 PT |
4946 | /* update hierarchical throttle state */ |
4947 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4948 | ||
2630cde2 MK |
4949 | /* Nothing to run but something to decay (on_list)? Complete the branch */ |
4950 | if (!cfs_rq->load.weight) { | |
4951 | if (cfs_rq->on_list) | |
4952 | goto unthrottle_throttle; | |
671fd9da | 4953 | return; |
2630cde2 | 4954 | } |
671fd9da PT |
4955 | |
4956 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4957 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da | 4958 | for_each_sched_entity(se) { |
30400039 JD |
4959 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
4960 | ||
671fd9da | 4961 | if (se->on_rq) |
39f23ce0 | 4962 | break; |
30400039 JD |
4963 | enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP); |
4964 | ||
4965 | if (cfs_rq_is_idle(group_cfs_rq(se))) | |
4966 | idle_task_delta = cfs_rq->h_nr_running; | |
39f23ce0 | 4967 | |
30400039 JD |
4968 | qcfs_rq->h_nr_running += task_delta; |
4969 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
4970 | |
4971 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 4972 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 VG |
4973 | goto unthrottle_throttle; |
4974 | } | |
671fd9da | 4975 | |
39f23ce0 | 4976 | for_each_sched_entity(se) { |
30400039 | 4977 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
39f23ce0 | 4978 | |
30400039 | 4979 | update_load_avg(qcfs_rq, se, UPDATE_TG); |
39f23ce0 | 4980 | se_update_runnable(se); |
6212437f | 4981 | |
30400039 JD |
4982 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
4983 | idle_task_delta = cfs_rq->h_nr_running; | |
671fd9da | 4984 | |
30400039 JD |
4985 | qcfs_rq->h_nr_running += task_delta; |
4986 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
4987 | |
4988 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 4989 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 VG |
4990 | goto unthrottle_throttle; |
4991 | ||
4992 | /* | |
4993 | * One parent has been throttled and cfs_rq removed from the | |
4994 | * list. Add it back to not break the leaf list. | |
4995 | */ | |
30400039 JD |
4996 | if (throttled_hierarchy(qcfs_rq)) |
4997 | list_add_leaf_cfs_rq(qcfs_rq); | |
671fd9da PT |
4998 | } |
4999 | ||
39f23ce0 VG |
5000 | /* At this point se is NULL and we are at root level*/ |
5001 | add_nr_running(rq, task_delta); | |
671fd9da | 5002 | |
39f23ce0 | 5003 | unthrottle_throttle: |
fe61468b VG |
5004 | /* |
5005 | * The cfs_rq_throttled() breaks in the above iteration can result in | |
5006 | * incomplete leaf list maintenance, resulting in triggering the | |
5007 | * assertion below. | |
5008 | */ | |
5009 | for_each_sched_entity(se) { | |
30400039 | 5010 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
fe61468b | 5011 | |
30400039 | 5012 | if (list_add_leaf_cfs_rq(qcfs_rq)) |
39f23ce0 | 5013 | break; |
fe61468b VG |
5014 | } |
5015 | ||
5016 | assert_list_leaf_cfs_rq(rq); | |
5017 | ||
97fb7a0a | 5018 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 5019 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 5020 | resched_curr(rq); |
671fd9da PT |
5021 | } |
5022 | ||
26a8b127 | 5023 | static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) |
671fd9da PT |
5024 | { |
5025 | struct cfs_rq *cfs_rq; | |
26a8b127 | 5026 | u64 runtime, remaining = 1; |
671fd9da PT |
5027 | |
5028 | rcu_read_lock(); | |
5029 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
5030 | throttled_list) { | |
5031 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 5032 | struct rq_flags rf; |
671fd9da | 5033 | |
c0ad4aa4 | 5034 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
5035 | if (!cfs_rq_throttled(cfs_rq)) |
5036 | goto next; | |
5037 | ||
5e2d2cc2 L |
5038 | /* By the above check, this should never be true */ |
5039 | SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); | |
5040 | ||
26a8b127 | 5041 | raw_spin_lock(&cfs_b->lock); |
671fd9da | 5042 | runtime = -cfs_rq->runtime_remaining + 1; |
26a8b127 HC |
5043 | if (runtime > cfs_b->runtime) |
5044 | runtime = cfs_b->runtime; | |
5045 | cfs_b->runtime -= runtime; | |
5046 | remaining = cfs_b->runtime; | |
5047 | raw_spin_unlock(&cfs_b->lock); | |
671fd9da PT |
5048 | |
5049 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
5050 | |
5051 | /* we check whether we're throttled above */ | |
5052 | if (cfs_rq->runtime_remaining > 0) | |
5053 | unthrottle_cfs_rq(cfs_rq); | |
5054 | ||
5055 | next: | |
c0ad4aa4 | 5056 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
5057 | |
5058 | if (!remaining) | |
5059 | break; | |
5060 | } | |
5061 | rcu_read_unlock(); | |
671fd9da PT |
5062 | } |
5063 | ||
58088ad0 PT |
5064 | /* |
5065 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
5066 | * cfs_rqs as appropriate. If there has been no activity within the last | |
5067 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
5068 | * used to track this state. | |
5069 | */ | |
c0ad4aa4 | 5070 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 5071 | { |
51f2176d | 5072 | int throttled; |
58088ad0 | 5073 | |
58088ad0 PT |
5074 | /* no need to continue the timer with no bandwidth constraint */ |
5075 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 5076 | goto out_deactivate; |
58088ad0 | 5077 | |
671fd9da | 5078 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 5079 | cfs_b->nr_periods += overrun; |
671fd9da | 5080 | |
f4183717 HC |
5081 | /* Refill extra burst quota even if cfs_b->idle */ |
5082 | __refill_cfs_bandwidth_runtime(cfs_b); | |
5083 | ||
51f2176d BS |
5084 | /* |
5085 | * idle depends on !throttled (for the case of a large deficit), and if | |
5086 | * we're going inactive then everything else can be deferred | |
5087 | */ | |
5088 | if (cfs_b->idle && !throttled) | |
5089 | goto out_deactivate; | |
a9cf55b2 | 5090 | |
671fd9da PT |
5091 | if (!throttled) { |
5092 | /* mark as potentially idle for the upcoming period */ | |
5093 | cfs_b->idle = 1; | |
51f2176d | 5094 | return 0; |
671fd9da PT |
5095 | } |
5096 | ||
e8da1b18 NR |
5097 | /* account preceding periods in which throttling occurred */ |
5098 | cfs_b->nr_throttled += overrun; | |
5099 | ||
671fd9da | 5100 | /* |
26a8b127 | 5101 | * This check is repeated as we release cfs_b->lock while we unthrottle. |
671fd9da | 5102 | */ |
ab93a4bc | 5103 | while (throttled && cfs_b->runtime > 0) { |
c0ad4aa4 | 5104 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 5105 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
26a8b127 | 5106 | distribute_cfs_runtime(cfs_b); |
c0ad4aa4 | 5107 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da PT |
5108 | |
5109 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
5110 | } | |
58088ad0 | 5111 | |
671fd9da PT |
5112 | /* |
5113 | * While we are ensured activity in the period following an | |
5114 | * unthrottle, this also covers the case in which the new bandwidth is | |
5115 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
5116 | * timer to remain active while there are any throttled entities.) | |
5117 | */ | |
5118 | cfs_b->idle = 0; | |
58088ad0 | 5119 | |
51f2176d BS |
5120 | return 0; |
5121 | ||
5122 | out_deactivate: | |
51f2176d | 5123 | return 1; |
58088ad0 | 5124 | } |
d3d9dc33 | 5125 | |
d8b4986d PT |
5126 | /* a cfs_rq won't donate quota below this amount */ |
5127 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
5128 | /* minimum remaining period time to redistribute slack quota */ | |
5129 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
5130 | /* how long we wait to gather additional slack before distributing */ | |
5131 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
5132 | ||
db06e78c BS |
5133 | /* |
5134 | * Are we near the end of the current quota period? | |
5135 | * | |
5136 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 5137 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
5138 | * migrate_hrtimers, base is never cleared, so we are fine. |
5139 | */ | |
d8b4986d PT |
5140 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
5141 | { | |
5142 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
72d0ad7c | 5143 | s64 remaining; |
d8b4986d PT |
5144 | |
5145 | /* if the call-back is running a quota refresh is already occurring */ | |
5146 | if (hrtimer_callback_running(refresh_timer)) | |
5147 | return 1; | |
5148 | ||
5149 | /* is a quota refresh about to occur? */ | |
5150 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
72d0ad7c | 5151 | if (remaining < (s64)min_expire) |
d8b4986d PT |
5152 | return 1; |
5153 | ||
5154 | return 0; | |
5155 | } | |
5156 | ||
5157 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
5158 | { | |
5159 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
5160 | ||
5161 | /* if there's a quota refresh soon don't bother with slack */ | |
5162 | if (runtime_refresh_within(cfs_b, min_left)) | |
5163 | return; | |
5164 | ||
66567fcb | 5165 | /* don't push forwards an existing deferred unthrottle */ |
5166 | if (cfs_b->slack_started) | |
5167 | return; | |
5168 | cfs_b->slack_started = true; | |
5169 | ||
4cfafd30 PZ |
5170 | hrtimer_start(&cfs_b->slack_timer, |
5171 | ns_to_ktime(cfs_bandwidth_slack_period), | |
5172 | HRTIMER_MODE_REL); | |
d8b4986d PT |
5173 | } |
5174 | ||
5175 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
5176 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5177 | { | |
5178 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5179 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
5180 | ||
5181 | if (slack_runtime <= 0) | |
5182 | return; | |
5183 | ||
5184 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 5185 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
5186 | cfs_b->runtime += slack_runtime; |
5187 | ||
5188 | /* we are under rq->lock, defer unthrottling using a timer */ | |
5189 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
5190 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
5191 | start_cfs_slack_bandwidth(cfs_b); | |
5192 | } | |
5193 | raw_spin_unlock(&cfs_b->lock); | |
5194 | ||
5195 | /* even if it's not valid for return we don't want to try again */ | |
5196 | cfs_rq->runtime_remaining -= slack_runtime; | |
5197 | } | |
5198 | ||
5199 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5200 | { | |
56f570e5 PT |
5201 | if (!cfs_bandwidth_used()) |
5202 | return; | |
5203 | ||
fccfdc6f | 5204 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
5205 | return; |
5206 | ||
5207 | __return_cfs_rq_runtime(cfs_rq); | |
5208 | } | |
5209 | ||
5210 | /* | |
5211 | * This is done with a timer (instead of inline with bandwidth return) since | |
5212 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
5213 | */ | |
5214 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
5215 | { | |
5216 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 5217 | unsigned long flags; |
d8b4986d PT |
5218 | |
5219 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 5220 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 5221 | cfs_b->slack_started = false; |
baa9be4f | 5222 | |
db06e78c | 5223 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 5224 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 5225 | return; |
db06e78c | 5226 | } |
d8b4986d | 5227 | |
c06f04c7 | 5228 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 5229 | runtime = cfs_b->runtime; |
c06f04c7 | 5230 | |
c0ad4aa4 | 5231 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
5232 | |
5233 | if (!runtime) | |
5234 | return; | |
5235 | ||
26a8b127 | 5236 | distribute_cfs_runtime(cfs_b); |
d8b4986d PT |
5237 | } |
5238 | ||
d3d9dc33 PT |
5239 | /* |
5240 | * When a group wakes up we want to make sure that its quota is not already | |
5241 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
c034f48e | 5242 | * runtime as update_curr() throttling can not trigger until it's on-rq. |
d3d9dc33 PT |
5243 | */ |
5244 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
5245 | { | |
56f570e5 PT |
5246 | if (!cfs_bandwidth_used()) |
5247 | return; | |
5248 | ||
d3d9dc33 PT |
5249 | /* an active group must be handled by the update_curr()->put() path */ |
5250 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
5251 | return; | |
5252 | ||
5253 | /* ensure the group is not already throttled */ | |
5254 | if (cfs_rq_throttled(cfs_rq)) | |
5255 | return; | |
5256 | ||
5257 | /* update runtime allocation */ | |
5258 | account_cfs_rq_runtime(cfs_rq, 0); | |
5259 | if (cfs_rq->runtime_remaining <= 0) | |
5260 | throttle_cfs_rq(cfs_rq); | |
5261 | } | |
5262 | ||
55e16d30 PZ |
5263 | static void sync_throttle(struct task_group *tg, int cpu) |
5264 | { | |
5265 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
5266 | ||
5267 | if (!cfs_bandwidth_used()) | |
5268 | return; | |
5269 | ||
5270 | if (!tg->parent) | |
5271 | return; | |
5272 | ||
5273 | cfs_rq = tg->cfs_rq[cpu]; | |
5274 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
5275 | ||
5276 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 5277 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
5278 | } |
5279 | ||
d3d9dc33 | 5280 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 5281 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 5282 | { |
56f570e5 | 5283 | if (!cfs_bandwidth_used()) |
678d5718 | 5284 | return false; |
56f570e5 | 5285 | |
d3d9dc33 | 5286 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 5287 | return false; |
d3d9dc33 PT |
5288 | |
5289 | /* | |
5290 | * it's possible for a throttled entity to be forced into a running | |
5291 | * state (e.g. set_curr_task), in this case we're finished. | |
5292 | */ | |
5293 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 5294 | return true; |
d3d9dc33 | 5295 | |
e98fa02c | 5296 | return throttle_cfs_rq(cfs_rq); |
d3d9dc33 | 5297 | } |
029632fb | 5298 | |
029632fb PZ |
5299 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
5300 | { | |
5301 | struct cfs_bandwidth *cfs_b = | |
5302 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 5303 | |
029632fb PZ |
5304 | do_sched_cfs_slack_timer(cfs_b); |
5305 | ||
5306 | return HRTIMER_NORESTART; | |
5307 | } | |
5308 | ||
2e8e1922 PA |
5309 | extern const u64 max_cfs_quota_period; |
5310 | ||
029632fb PZ |
5311 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
5312 | { | |
5313 | struct cfs_bandwidth *cfs_b = | |
5314 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 5315 | unsigned long flags; |
029632fb PZ |
5316 | int overrun; |
5317 | int idle = 0; | |
2e8e1922 | 5318 | int count = 0; |
029632fb | 5319 | |
c0ad4aa4 | 5320 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 5321 | for (;;) { |
77a4d1a1 | 5322 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
5323 | if (!overrun) |
5324 | break; | |
5325 | ||
5a6d6a6c HC |
5326 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
5327 | ||
2e8e1922 PA |
5328 | if (++count > 3) { |
5329 | u64 new, old = ktime_to_ns(cfs_b->period); | |
5330 | ||
4929a4e6 XZ |
5331 | /* |
5332 | * Grow period by a factor of 2 to avoid losing precision. | |
5333 | * Precision loss in the quota/period ratio can cause __cfs_schedulable | |
5334 | * to fail. | |
5335 | */ | |
5336 | new = old * 2; | |
5337 | if (new < max_cfs_quota_period) { | |
5338 | cfs_b->period = ns_to_ktime(new); | |
5339 | cfs_b->quota *= 2; | |
f4183717 | 5340 | cfs_b->burst *= 2; |
4929a4e6 XZ |
5341 | |
5342 | pr_warn_ratelimited( | |
5343 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5344 | smp_processor_id(), | |
5345 | div_u64(new, NSEC_PER_USEC), | |
5346 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5347 | } else { | |
5348 | pr_warn_ratelimited( | |
5349 | "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5350 | smp_processor_id(), | |
5351 | div_u64(old, NSEC_PER_USEC), | |
5352 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5353 | } | |
2e8e1922 PA |
5354 | |
5355 | /* reset count so we don't come right back in here */ | |
5356 | count = 0; | |
5357 | } | |
029632fb | 5358 | } |
4cfafd30 PZ |
5359 | if (idle) |
5360 | cfs_b->period_active = 0; | |
c0ad4aa4 | 5361 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
5362 | |
5363 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
5364 | } | |
5365 | ||
5366 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5367 | { | |
5368 | raw_spin_lock_init(&cfs_b->lock); | |
5369 | cfs_b->runtime = 0; | |
5370 | cfs_b->quota = RUNTIME_INF; | |
5371 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
f4183717 | 5372 | cfs_b->burst = 0; |
029632fb PZ |
5373 | |
5374 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 5375 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5376 | cfs_b->period_timer.function = sched_cfs_period_timer; |
5377 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5378 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
66567fcb | 5379 | cfs_b->slack_started = false; |
029632fb PZ |
5380 | } |
5381 | ||
5382 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5383 | { | |
5384 | cfs_rq->runtime_enabled = 0; | |
5385 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
5386 | } | |
5387 | ||
77a4d1a1 | 5388 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 5389 | { |
4cfafd30 | 5390 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 5391 | |
f1d1be8a XP |
5392 | if (cfs_b->period_active) |
5393 | return; | |
5394 | ||
5395 | cfs_b->period_active = 1; | |
763a9ec0 | 5396 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); |
f1d1be8a | 5397 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5398 | } |
5399 | ||
5400 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5401 | { | |
7f1a169b TH |
5402 | /* init_cfs_bandwidth() was not called */ |
5403 | if (!cfs_b->throttled_cfs_rq.next) | |
5404 | return; | |
5405 | ||
029632fb PZ |
5406 | hrtimer_cancel(&cfs_b->period_timer); |
5407 | hrtimer_cancel(&cfs_b->slack_timer); | |
5408 | } | |
5409 | ||
502ce005 | 5410 | /* |
97fb7a0a | 5411 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
5412 | * |
5413 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5414 | * bits doesn't do much. | |
5415 | */ | |
5416 | ||
3b03706f | 5417 | /* cpu online callback */ |
0e59bdae KT |
5418 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5419 | { | |
502ce005 | 5420 | struct task_group *tg; |
0e59bdae | 5421 | |
5cb9eaa3 | 5422 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
5423 | |
5424 | rcu_read_lock(); | |
5425 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5426 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5427 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5428 | |
5429 | raw_spin_lock(&cfs_b->lock); | |
5430 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5431 | raw_spin_unlock(&cfs_b->lock); | |
5432 | } | |
502ce005 | 5433 | rcu_read_unlock(); |
0e59bdae KT |
5434 | } |
5435 | ||
502ce005 | 5436 | /* cpu offline callback */ |
38dc3348 | 5437 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5438 | { |
502ce005 PZ |
5439 | struct task_group *tg; |
5440 | ||
5cb9eaa3 | 5441 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
5442 | |
5443 | rcu_read_lock(); | |
5444 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5445 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5446 | |
029632fb PZ |
5447 | if (!cfs_rq->runtime_enabled) |
5448 | continue; | |
5449 | ||
5450 | /* | |
5451 | * clock_task is not advancing so we just need to make sure | |
5452 | * there's some valid quota amount | |
5453 | */ | |
51f2176d | 5454 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 5455 | /* |
97fb7a0a | 5456 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
5457 | * in take_cpu_down(), so we prevent new cfs throttling here. |
5458 | */ | |
5459 | cfs_rq->runtime_enabled = 0; | |
5460 | ||
029632fb PZ |
5461 | if (cfs_rq_throttled(cfs_rq)) |
5462 | unthrottle_cfs_rq(cfs_rq); | |
5463 | } | |
502ce005 | 5464 | rcu_read_unlock(); |
029632fb PZ |
5465 | } |
5466 | ||
5467 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f6783319 VG |
5468 | |
5469 | static inline bool cfs_bandwidth_used(void) | |
5470 | { | |
5471 | return false; | |
5472 | } | |
5473 | ||
9dbdb155 | 5474 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5475 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5476 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5477 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5478 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5479 | |
5480 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5481 | { | |
5482 | return 0; | |
5483 | } | |
64660c86 PT |
5484 | |
5485 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5486 | { | |
5487 | return 0; | |
5488 | } | |
5489 | ||
5490 | static inline int throttled_lb_pair(struct task_group *tg, | |
5491 | int src_cpu, int dest_cpu) | |
5492 | { | |
5493 | return 0; | |
5494 | } | |
029632fb PZ |
5495 | |
5496 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5497 | ||
5498 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5499 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5500 | #endif |
5501 | ||
029632fb PZ |
5502 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5503 | { | |
5504 | return NULL; | |
5505 | } | |
5506 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5507 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5508 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5509 | |
5510 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5511 | ||
bf0f6f24 IM |
5512 | /************************************************** |
5513 | * CFS operations on tasks: | |
5514 | */ | |
5515 | ||
8f4d37ec PZ |
5516 | #ifdef CONFIG_SCHED_HRTICK |
5517 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5518 | { | |
8f4d37ec PZ |
5519 | struct sched_entity *se = &p->se; |
5520 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5521 | ||
9148a3a1 | 5522 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5523 | |
8bf46a39 | 5524 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5525 | u64 slice = sched_slice(cfs_rq, se); |
5526 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5527 | s64 delta = slice - ran; | |
5528 | ||
5529 | if (delta < 0) { | |
65bcf072 | 5530 | if (task_current(rq, p)) |
8875125e | 5531 | resched_curr(rq); |
8f4d37ec PZ |
5532 | return; |
5533 | } | |
31656519 | 5534 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5535 | } |
5536 | } | |
a4c2f00f PZ |
5537 | |
5538 | /* | |
5539 | * called from enqueue/dequeue and updates the hrtick when the | |
5540 | * current task is from our class and nr_running is low enough | |
5541 | * to matter. | |
5542 | */ | |
5543 | static void hrtick_update(struct rq *rq) | |
5544 | { | |
5545 | struct task_struct *curr = rq->curr; | |
5546 | ||
e0ee463c | 5547 | if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5548 | return; |
5549 | ||
5550 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5551 | hrtick_start_fair(rq, curr); | |
5552 | } | |
55e12e5e | 5553 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5554 | static inline void |
5555 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5556 | { | |
5557 | } | |
a4c2f00f PZ |
5558 | |
5559 | static inline void hrtick_update(struct rq *rq) | |
5560 | { | |
5561 | } | |
8f4d37ec PZ |
5562 | #endif |
5563 | ||
2802bf3c MR |
5564 | #ifdef CONFIG_SMP |
5565 | static inline unsigned long cpu_util(int cpu); | |
2802bf3c MR |
5566 | |
5567 | static inline bool cpu_overutilized(int cpu) | |
5568 | { | |
60e17f5c | 5569 | return !fits_capacity(cpu_util(cpu), capacity_of(cpu)); |
2802bf3c MR |
5570 | } |
5571 | ||
5572 | static inline void update_overutilized_status(struct rq *rq) | |
5573 | { | |
f9f240f9 | 5574 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { |
2802bf3c | 5575 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); |
f9f240f9 QY |
5576 | trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); |
5577 | } | |
2802bf3c MR |
5578 | } |
5579 | #else | |
5580 | static inline void update_overutilized_status(struct rq *rq) { } | |
5581 | #endif | |
5582 | ||
323af6de VK |
5583 | /* Runqueue only has SCHED_IDLE tasks enqueued */ |
5584 | static int sched_idle_rq(struct rq *rq) | |
5585 | { | |
5586 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
5587 | rq->nr_running); | |
5588 | } | |
5589 | ||
afa70d94 | 5590 | #ifdef CONFIG_SMP |
323af6de VK |
5591 | static int sched_idle_cpu(int cpu) |
5592 | { | |
5593 | return sched_idle_rq(cpu_rq(cpu)); | |
5594 | } | |
afa70d94 | 5595 | #endif |
323af6de | 5596 | |
bf0f6f24 IM |
5597 | /* |
5598 | * The enqueue_task method is called before nr_running is | |
5599 | * increased. Here we update the fair scheduling stats and | |
5600 | * then put the task into the rbtree: | |
5601 | */ | |
ea87bb78 | 5602 | static void |
371fd7e7 | 5603 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5604 | { |
5605 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5606 | struct sched_entity *se = &p->se; |
43e9f7f2 | 5607 | int idle_h_nr_running = task_has_idle_policy(p); |
8e1ac429 | 5608 | int task_new = !(flags & ENQUEUE_WAKEUP); |
bf0f6f24 | 5609 | |
2539fc82 PB |
5610 | /* |
5611 | * The code below (indirectly) updates schedutil which looks at | |
5612 | * the cfs_rq utilization to select a frequency. | |
5613 | * Let's add the task's estimated utilization to the cfs_rq's | |
5614 | * estimated utilization, before we update schedutil. | |
5615 | */ | |
5616 | util_est_enqueue(&rq->cfs, p); | |
5617 | ||
8c34ab19 RW |
5618 | /* |
5619 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5620 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5621 | * passed. | |
5622 | */ | |
5623 | if (p->in_iowait) | |
674e7541 | 5624 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5625 | |
bf0f6f24 | 5626 | for_each_sched_entity(se) { |
62fb1851 | 5627 | if (se->on_rq) |
bf0f6f24 IM |
5628 | break; |
5629 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5630 | enqueue_entity(cfs_rq, se, flags); |
85dac906 | 5631 | |
953bfcd1 | 5632 | cfs_rq->h_nr_running++; |
43e9f7f2 | 5633 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 5634 | |
30400039 JD |
5635 | if (cfs_rq_is_idle(cfs_rq)) |
5636 | idle_h_nr_running = 1; | |
5637 | ||
6d4d2246 VG |
5638 | /* end evaluation on encountering a throttled cfs_rq */ |
5639 | if (cfs_rq_throttled(cfs_rq)) | |
5640 | goto enqueue_throttle; | |
5641 | ||
88ec22d3 | 5642 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5643 | } |
8f4d37ec | 5644 | |
2069dd75 | 5645 | for_each_sched_entity(se) { |
0f317143 | 5646 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5647 | |
88c0616e | 5648 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5649 | se_update_runnable(se); |
1ea6c46a | 5650 | update_cfs_group(se); |
6d4d2246 VG |
5651 | |
5652 | cfs_rq->h_nr_running++; | |
5653 | cfs_rq->idle_h_nr_running += idle_h_nr_running; | |
5ab297ba | 5654 | |
30400039 JD |
5655 | if (cfs_rq_is_idle(cfs_rq)) |
5656 | idle_h_nr_running = 1; | |
5657 | ||
5ab297ba VG |
5658 | /* end evaluation on encountering a throttled cfs_rq */ |
5659 | if (cfs_rq_throttled(cfs_rq)) | |
5660 | goto enqueue_throttle; | |
b34cb07d PA |
5661 | |
5662 | /* | |
5663 | * One parent has been throttled and cfs_rq removed from the | |
5664 | * list. Add it back to not break the leaf list. | |
5665 | */ | |
5666 | if (throttled_hierarchy(cfs_rq)) | |
5667 | list_add_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
5668 | } |
5669 | ||
7d148be6 VG |
5670 | /* At this point se is NULL and we are at root level*/ |
5671 | add_nr_running(rq, 1); | |
2802bf3c | 5672 | |
7d148be6 VG |
5673 | /* |
5674 | * Since new tasks are assigned an initial util_avg equal to | |
5675 | * half of the spare capacity of their CPU, tiny tasks have the | |
5676 | * ability to cross the overutilized threshold, which will | |
5677 | * result in the load balancer ruining all the task placement | |
5678 | * done by EAS. As a way to mitigate that effect, do not account | |
5679 | * for the first enqueue operation of new tasks during the | |
5680 | * overutilized flag detection. | |
5681 | * | |
5682 | * A better way of solving this problem would be to wait for | |
5683 | * the PELT signals of tasks to converge before taking them | |
5684 | * into account, but that is not straightforward to implement, | |
5685 | * and the following generally works well enough in practice. | |
5686 | */ | |
8e1ac429 | 5687 | if (!task_new) |
7d148be6 | 5688 | update_overutilized_status(rq); |
cd126afe | 5689 | |
7d148be6 | 5690 | enqueue_throttle: |
f6783319 VG |
5691 | if (cfs_bandwidth_used()) { |
5692 | /* | |
5693 | * When bandwidth control is enabled; the cfs_rq_throttled() | |
5694 | * breaks in the above iteration can result in incomplete | |
5695 | * leaf list maintenance, resulting in triggering the assertion | |
5696 | * below. | |
5697 | */ | |
5698 | for_each_sched_entity(se) { | |
5699 | cfs_rq = cfs_rq_of(se); | |
5700 | ||
5701 | if (list_add_leaf_cfs_rq(cfs_rq)) | |
5702 | break; | |
5703 | } | |
5704 | } | |
5705 | ||
5d299eab PZ |
5706 | assert_list_leaf_cfs_rq(rq); |
5707 | ||
a4c2f00f | 5708 | hrtick_update(rq); |
bf0f6f24 IM |
5709 | } |
5710 | ||
2f36825b VP |
5711 | static void set_next_buddy(struct sched_entity *se); |
5712 | ||
bf0f6f24 IM |
5713 | /* |
5714 | * The dequeue_task method is called before nr_running is | |
5715 | * decreased. We remove the task from the rbtree and | |
5716 | * update the fair scheduling stats: | |
5717 | */ | |
371fd7e7 | 5718 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5719 | { |
5720 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5721 | struct sched_entity *se = &p->se; |
2f36825b | 5722 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 5723 | int idle_h_nr_running = task_has_idle_policy(p); |
323af6de | 5724 | bool was_sched_idle = sched_idle_rq(rq); |
bf0f6f24 | 5725 | |
8c1f560c XY |
5726 | util_est_dequeue(&rq->cfs, p); |
5727 | ||
bf0f6f24 IM |
5728 | for_each_sched_entity(se) { |
5729 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5730 | dequeue_entity(cfs_rq, se, flags); |
85dac906 | 5731 | |
953bfcd1 | 5732 | cfs_rq->h_nr_running--; |
43e9f7f2 | 5733 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 5734 | |
30400039 JD |
5735 | if (cfs_rq_is_idle(cfs_rq)) |
5736 | idle_h_nr_running = 1; | |
5737 | ||
6d4d2246 VG |
5738 | /* end evaluation on encountering a throttled cfs_rq */ |
5739 | if (cfs_rq_throttled(cfs_rq)) | |
5740 | goto dequeue_throttle; | |
5741 | ||
bf0f6f24 | 5742 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5743 | if (cfs_rq->load.weight) { |
754bd598 KK |
5744 | /* Avoid re-evaluating load for this entity: */ |
5745 | se = parent_entity(se); | |
2f36825b VP |
5746 | /* |
5747 | * Bias pick_next to pick a task from this cfs_rq, as | |
5748 | * p is sleeping when it is within its sched_slice. | |
5749 | */ | |
754bd598 KK |
5750 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5751 | set_next_buddy(se); | |
bf0f6f24 | 5752 | break; |
2f36825b | 5753 | } |
371fd7e7 | 5754 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5755 | } |
8f4d37ec | 5756 | |
2069dd75 | 5757 | for_each_sched_entity(se) { |
0f317143 | 5758 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5759 | |
88c0616e | 5760 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5761 | se_update_runnable(se); |
1ea6c46a | 5762 | update_cfs_group(se); |
6d4d2246 VG |
5763 | |
5764 | cfs_rq->h_nr_running--; | |
5765 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; | |
5ab297ba | 5766 | |
30400039 JD |
5767 | if (cfs_rq_is_idle(cfs_rq)) |
5768 | idle_h_nr_running = 1; | |
5769 | ||
5ab297ba VG |
5770 | /* end evaluation on encountering a throttled cfs_rq */ |
5771 | if (cfs_rq_throttled(cfs_rq)) | |
5772 | goto dequeue_throttle; | |
5773 | ||
2069dd75 PZ |
5774 | } |
5775 | ||
423d02e1 PW |
5776 | /* At this point se is NULL and we are at root level*/ |
5777 | sub_nr_running(rq, 1); | |
cd126afe | 5778 | |
323af6de VK |
5779 | /* balance early to pull high priority tasks */ |
5780 | if (unlikely(!was_sched_idle && sched_idle_rq(rq))) | |
5781 | rq->next_balance = jiffies; | |
5782 | ||
423d02e1 | 5783 | dequeue_throttle: |
8c1f560c | 5784 | util_est_update(&rq->cfs, p, task_sleep); |
a4c2f00f | 5785 | hrtick_update(rq); |
bf0f6f24 IM |
5786 | } |
5787 | ||
e7693a36 | 5788 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5789 | |
5790 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
5791 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5792 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5793 | ||
9fd81dd5 | 5794 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
5795 | |
5796 | static struct { | |
5797 | cpumask_var_t idle_cpus_mask; | |
5798 | atomic_t nr_cpus; | |
f643ea22 | 5799 | int has_blocked; /* Idle CPUS has blocked load */ |
e022e0d3 | 5800 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 5801 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
5802 | } nohz ____cacheline_aligned; |
5803 | ||
9fd81dd5 | 5804 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5805 | |
b0fb1eb4 VG |
5806 | static unsigned long cpu_load(struct rq *rq) |
5807 | { | |
5808 | return cfs_rq_load_avg(&rq->cfs); | |
5809 | } | |
5810 | ||
3318544b VG |
5811 | /* |
5812 | * cpu_load_without - compute CPU load without any contributions from *p | |
5813 | * @cpu: the CPU which load is requested | |
5814 | * @p: the task which load should be discounted | |
5815 | * | |
5816 | * The load of a CPU is defined by the load of tasks currently enqueued on that | |
5817 | * CPU as well as tasks which are currently sleeping after an execution on that | |
5818 | * CPU. | |
5819 | * | |
5820 | * This method returns the load of the specified CPU by discounting the load of | |
5821 | * the specified task, whenever the task is currently contributing to the CPU | |
5822 | * load. | |
5823 | */ | |
5824 | static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) | |
5825 | { | |
5826 | struct cfs_rq *cfs_rq; | |
5827 | unsigned int load; | |
5828 | ||
5829 | /* Task has no contribution or is new */ | |
5830 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5831 | return cpu_load(rq); | |
5832 | ||
5833 | cfs_rq = &rq->cfs; | |
5834 | load = READ_ONCE(cfs_rq->avg.load_avg); | |
5835 | ||
5836 | /* Discount task's util from CPU's util */ | |
5837 | lsub_positive(&load, task_h_load(p)); | |
5838 | ||
5839 | return load; | |
5840 | } | |
5841 | ||
9f683953 VG |
5842 | static unsigned long cpu_runnable(struct rq *rq) |
5843 | { | |
5844 | return cfs_rq_runnable_avg(&rq->cfs); | |
5845 | } | |
5846 | ||
070f5e86 VG |
5847 | static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) |
5848 | { | |
5849 | struct cfs_rq *cfs_rq; | |
5850 | unsigned int runnable; | |
5851 | ||
5852 | /* Task has no contribution or is new */ | |
5853 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5854 | return cpu_runnable(rq); | |
5855 | ||
5856 | cfs_rq = &rq->cfs; | |
5857 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
5858 | ||
5859 | /* Discount task's runnable from CPU's runnable */ | |
5860 | lsub_positive(&runnable, p->se.avg.runnable_avg); | |
5861 | ||
5862 | return runnable; | |
5863 | } | |
5864 | ||
ced549fa | 5865 | static unsigned long capacity_of(int cpu) |
029632fb | 5866 | { |
ced549fa | 5867 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5868 | } |
5869 | ||
c58d25f3 PZ |
5870 | static void record_wakee(struct task_struct *p) |
5871 | { | |
5872 | /* | |
5873 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5874 | * jiffy will not have built up many flips. | |
5875 | */ | |
5876 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5877 | current->wakee_flips >>= 1; | |
5878 | current->wakee_flip_decay_ts = jiffies; | |
5879 | } | |
5880 | ||
5881 | if (current->last_wakee != p) { | |
5882 | current->last_wakee = p; | |
5883 | current->wakee_flips++; | |
5884 | } | |
5885 | } | |
5886 | ||
63b0e9ed MG |
5887 | /* |
5888 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5889 | * |
63b0e9ed | 5890 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5891 | * at a frequency roughly N times higher than one of its wakees. |
5892 | * | |
5893 | * In order to determine whether we should let the load spread vs consolidating | |
5894 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5895 | * partner, and a factor of lls_size higher frequency in the other. | |
5896 | * | |
5897 | * With both conditions met, we can be relatively sure that the relationship is | |
5898 | * non-monogamous, with partner count exceeding socket size. | |
5899 | * | |
5900 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5901 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5902 | * socket size. | |
63b0e9ed | 5903 | */ |
62470419 MW |
5904 | static int wake_wide(struct task_struct *p) |
5905 | { | |
63b0e9ed MG |
5906 | unsigned int master = current->wakee_flips; |
5907 | unsigned int slave = p->wakee_flips; | |
17c891ab | 5908 | int factor = __this_cpu_read(sd_llc_size); |
62470419 | 5909 | |
63b0e9ed MG |
5910 | if (master < slave) |
5911 | swap(master, slave); | |
5912 | if (slave < factor || master < slave * factor) | |
5913 | return 0; | |
5914 | return 1; | |
62470419 MW |
5915 | } |
5916 | ||
90001d67 | 5917 | /* |
d153b153 PZ |
5918 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
5919 | * soonest. For the purpose of speed we only consider the waking and previous | |
5920 | * CPU. | |
90001d67 | 5921 | * |
7332dec0 MG |
5922 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
5923 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
5924 | * |
5925 | * wake_affine_weight() - considers the weight to reflect the average | |
5926 | * scheduling latency of the CPUs. This seems to work | |
5927 | * for the overloaded case. | |
90001d67 | 5928 | */ |
3b76c4a3 | 5929 | static int |
89a55f56 | 5930 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 5931 | { |
7332dec0 MG |
5932 | /* |
5933 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
5934 | * context. Only allow the move if cache is shared. Otherwise an | |
5935 | * interrupt intensive workload could force all tasks onto one | |
5936 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
5937 | * |
5938 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
5939 | * There is no guarantee that the cache hot data from an interrupt | |
5940 | * is more important than cache hot data on the prev_cpu and from | |
5941 | * a cpufreq perspective, it's better to have higher utilisation | |
5942 | * on one CPU. | |
7332dec0 | 5943 | */ |
943d355d RJ |
5944 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
5945 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 5946 | |
d153b153 | 5947 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 5948 | return this_cpu; |
90001d67 | 5949 | |
d8fcb81f JL |
5950 | if (available_idle_cpu(prev_cpu)) |
5951 | return prev_cpu; | |
5952 | ||
3b76c4a3 | 5953 | return nr_cpumask_bits; |
90001d67 PZ |
5954 | } |
5955 | ||
3b76c4a3 | 5956 | static int |
f2cdd9cc PZ |
5957 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
5958 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5959 | { |
90001d67 PZ |
5960 | s64 this_eff_load, prev_eff_load; |
5961 | unsigned long task_load; | |
5962 | ||
11f10e54 | 5963 | this_eff_load = cpu_load(cpu_rq(this_cpu)); |
90001d67 | 5964 | |
90001d67 PZ |
5965 | if (sync) { |
5966 | unsigned long current_load = task_h_load(current); | |
5967 | ||
f2cdd9cc | 5968 | if (current_load > this_eff_load) |
3b76c4a3 | 5969 | return this_cpu; |
90001d67 | 5970 | |
f2cdd9cc | 5971 | this_eff_load -= current_load; |
90001d67 PZ |
5972 | } |
5973 | ||
90001d67 PZ |
5974 | task_load = task_h_load(p); |
5975 | ||
f2cdd9cc PZ |
5976 | this_eff_load += task_load; |
5977 | if (sched_feat(WA_BIAS)) | |
5978 | this_eff_load *= 100; | |
5979 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 5980 | |
11f10e54 | 5981 | prev_eff_load = cpu_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
5982 | prev_eff_load -= task_load; |
5983 | if (sched_feat(WA_BIAS)) | |
5984 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
5985 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 5986 | |
082f764a MG |
5987 | /* |
5988 | * If sync, adjust the weight of prev_eff_load such that if | |
5989 | * prev_eff == this_eff that select_idle_sibling() will consider | |
5990 | * stacking the wakee on top of the waker if no other CPU is | |
5991 | * idle. | |
5992 | */ | |
5993 | if (sync) | |
5994 | prev_eff_load += 1; | |
5995 | ||
5996 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
5997 | } |
5998 | ||
772bd008 | 5999 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 6000 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 6001 | { |
3b76c4a3 | 6002 | int target = nr_cpumask_bits; |
098fb9db | 6003 | |
89a55f56 | 6004 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 6005 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 6006 | |
3b76c4a3 MG |
6007 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
6008 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 6009 | |
ae92882e | 6010 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
3b76c4a3 MG |
6011 | if (target == nr_cpumask_bits) |
6012 | return prev_cpu; | |
098fb9db | 6013 | |
3b76c4a3 MG |
6014 | schedstat_inc(sd->ttwu_move_affine); |
6015 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
6016 | return target; | |
098fb9db IM |
6017 | } |
6018 | ||
aaee1203 | 6019 | static struct sched_group * |
45da2773 | 6020 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); |
aaee1203 PZ |
6021 | |
6022 | /* | |
97fb7a0a | 6023 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
6024 | */ |
6025 | static int | |
18bd1b4b | 6026 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
6027 | { |
6028 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
6029 | unsigned int min_exit_latency = UINT_MAX; |
6030 | u64 latest_idle_timestamp = 0; | |
6031 | int least_loaded_cpu = this_cpu; | |
17346452 | 6032 | int shallowest_idle_cpu = -1; |
aaee1203 PZ |
6033 | int i; |
6034 | ||
eaecf41f MR |
6035 | /* Check if we have any choice: */ |
6036 | if (group->group_weight == 1) | |
ae4df9d6 | 6037 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 6038 | |
aaee1203 | 6039 | /* Traverse only the allowed CPUs */ |
3bd37062 | 6040 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
97886d9d AL |
6041 | struct rq *rq = cpu_rq(i); |
6042 | ||
6043 | if (!sched_core_cookie_match(rq, p)) | |
6044 | continue; | |
6045 | ||
17346452 VK |
6046 | if (sched_idle_cpu(i)) |
6047 | return i; | |
6048 | ||
943d355d | 6049 | if (available_idle_cpu(i)) { |
83a0a96a NP |
6050 | struct cpuidle_state *idle = idle_get_state(rq); |
6051 | if (idle && idle->exit_latency < min_exit_latency) { | |
6052 | /* | |
6053 | * We give priority to a CPU whose idle state | |
6054 | * has the smallest exit latency irrespective | |
6055 | * of any idle timestamp. | |
6056 | */ | |
6057 | min_exit_latency = idle->exit_latency; | |
6058 | latest_idle_timestamp = rq->idle_stamp; | |
6059 | shallowest_idle_cpu = i; | |
6060 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
6061 | rq->idle_stamp > latest_idle_timestamp) { | |
6062 | /* | |
6063 | * If equal or no active idle state, then | |
6064 | * the most recently idled CPU might have | |
6065 | * a warmer cache. | |
6066 | */ | |
6067 | latest_idle_timestamp = rq->idle_stamp; | |
6068 | shallowest_idle_cpu = i; | |
6069 | } | |
17346452 | 6070 | } else if (shallowest_idle_cpu == -1) { |
11f10e54 | 6071 | load = cpu_load(cpu_rq(i)); |
18cec7e0 | 6072 | if (load < min_load) { |
83a0a96a NP |
6073 | min_load = load; |
6074 | least_loaded_cpu = i; | |
6075 | } | |
e7693a36 GH |
6076 | } |
6077 | } | |
6078 | ||
17346452 | 6079 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 6080 | } |
e7693a36 | 6081 | |
18bd1b4b BJ |
6082 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
6083 | int cpu, int prev_cpu, int sd_flag) | |
6084 | { | |
93f50f90 | 6085 | int new_cpu = cpu; |
18bd1b4b | 6086 | |
3bd37062 | 6087 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
6088 | return prev_cpu; |
6089 | ||
c976a862 | 6090 | /* |
57abff06 | 6091 | * We need task's util for cpu_util_without, sync it up to |
c469933e | 6092 | * prev_cpu's last_update_time. |
c976a862 VK |
6093 | */ |
6094 | if (!(sd_flag & SD_BALANCE_FORK)) | |
6095 | sync_entity_load_avg(&p->se); | |
6096 | ||
18bd1b4b BJ |
6097 | while (sd) { |
6098 | struct sched_group *group; | |
6099 | struct sched_domain *tmp; | |
6100 | int weight; | |
6101 | ||
6102 | if (!(sd->flags & sd_flag)) { | |
6103 | sd = sd->child; | |
6104 | continue; | |
6105 | } | |
6106 | ||
45da2773 | 6107 | group = find_idlest_group(sd, p, cpu); |
18bd1b4b BJ |
6108 | if (!group) { |
6109 | sd = sd->child; | |
6110 | continue; | |
6111 | } | |
6112 | ||
6113 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 6114 | if (new_cpu == cpu) { |
97fb7a0a | 6115 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
6116 | sd = sd->child; |
6117 | continue; | |
6118 | } | |
6119 | ||
97fb7a0a | 6120 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
6121 | cpu = new_cpu; |
6122 | weight = sd->span_weight; | |
6123 | sd = NULL; | |
6124 | for_each_domain(cpu, tmp) { | |
6125 | if (weight <= tmp->span_weight) | |
6126 | break; | |
6127 | if (tmp->flags & sd_flag) | |
6128 | sd = tmp; | |
6129 | } | |
18bd1b4b BJ |
6130 | } |
6131 | ||
6132 | return new_cpu; | |
6133 | } | |
6134 | ||
97886d9d | 6135 | static inline int __select_idle_cpu(int cpu, struct task_struct *p) |
9fe1f127 | 6136 | { |
97886d9d AL |
6137 | if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) && |
6138 | sched_cpu_cookie_match(cpu_rq(cpu), p)) | |
9fe1f127 MG |
6139 | return cpu; |
6140 | ||
6141 | return -1; | |
6142 | } | |
6143 | ||
10e2f1ac | 6144 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 6145 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 6146 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
6147 | |
6148 | static inline void set_idle_cores(int cpu, int val) | |
6149 | { | |
6150 | struct sched_domain_shared *sds; | |
6151 | ||
6152 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
6153 | if (sds) | |
6154 | WRITE_ONCE(sds->has_idle_cores, val); | |
6155 | } | |
6156 | ||
6157 | static inline bool test_idle_cores(int cpu, bool def) | |
6158 | { | |
6159 | struct sched_domain_shared *sds; | |
6160 | ||
c722f35b RR |
6161 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
6162 | if (sds) | |
6163 | return READ_ONCE(sds->has_idle_cores); | |
10e2f1ac PZ |
6164 | |
6165 | return def; | |
6166 | } | |
6167 | ||
6168 | /* | |
6169 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
6170 | * information in sd_llc_shared->has_idle_cores. | |
6171 | * | |
6172 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
6173 | * state should be fairly cheap. | |
6174 | */ | |
1b568f0a | 6175 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
6176 | { |
6177 | int core = cpu_of(rq); | |
6178 | int cpu; | |
6179 | ||
6180 | rcu_read_lock(); | |
6181 | if (test_idle_cores(core, true)) | |
6182 | goto unlock; | |
6183 | ||
6184 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6185 | if (cpu == core) | |
6186 | continue; | |
6187 | ||
943d355d | 6188 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6189 | goto unlock; |
6190 | } | |
6191 | ||
6192 | set_idle_cores(core, 1); | |
6193 | unlock: | |
6194 | rcu_read_unlock(); | |
6195 | } | |
6196 | ||
6197 | /* | |
6198 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
6199 | * there are no idle cores left in the system; tracked through | |
6200 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
6201 | */ | |
9fe1f127 | 6202 | static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) |
10e2f1ac | 6203 | { |
9fe1f127 MG |
6204 | bool idle = true; |
6205 | int cpu; | |
10e2f1ac | 6206 | |
1b568f0a | 6207 | if (!static_branch_likely(&sched_smt_present)) |
97886d9d | 6208 | return __select_idle_cpu(core, p); |
10e2f1ac | 6209 | |
9fe1f127 MG |
6210 | for_each_cpu(cpu, cpu_smt_mask(core)) { |
6211 | if (!available_idle_cpu(cpu)) { | |
6212 | idle = false; | |
6213 | if (*idle_cpu == -1) { | |
6214 | if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) { | |
6215 | *idle_cpu = cpu; | |
6216 | break; | |
6217 | } | |
6218 | continue; | |
bec2860a | 6219 | } |
9fe1f127 | 6220 | break; |
10e2f1ac | 6221 | } |
9fe1f127 MG |
6222 | if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr)) |
6223 | *idle_cpu = cpu; | |
10e2f1ac PZ |
6224 | } |
6225 | ||
9fe1f127 MG |
6226 | if (idle) |
6227 | return core; | |
10e2f1ac | 6228 | |
9fe1f127 | 6229 | cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); |
10e2f1ac PZ |
6230 | return -1; |
6231 | } | |
6232 | ||
c722f35b RR |
6233 | /* |
6234 | * Scan the local SMT mask for idle CPUs. | |
6235 | */ | |
6236 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6237 | { | |
6238 | int cpu; | |
6239 | ||
6240 | for_each_cpu(cpu, cpu_smt_mask(target)) { | |
6241 | if (!cpumask_test_cpu(cpu, p->cpus_ptr) || | |
6242 | !cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
6243 | continue; | |
6244 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) | |
6245 | return cpu; | |
6246 | } | |
6247 | ||
6248 | return -1; | |
6249 | } | |
6250 | ||
10e2f1ac PZ |
6251 | #else /* CONFIG_SCHED_SMT */ |
6252 | ||
9fe1f127 | 6253 | static inline void set_idle_cores(int cpu, int val) |
10e2f1ac | 6254 | { |
9fe1f127 MG |
6255 | } |
6256 | ||
6257 | static inline bool test_idle_cores(int cpu, bool def) | |
6258 | { | |
6259 | return def; | |
6260 | } | |
6261 | ||
6262 | static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) | |
6263 | { | |
97886d9d | 6264 | return __select_idle_cpu(core, p); |
10e2f1ac PZ |
6265 | } |
6266 | ||
c722f35b RR |
6267 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) |
6268 | { | |
6269 | return -1; | |
6270 | } | |
6271 | ||
10e2f1ac PZ |
6272 | #endif /* CONFIG_SCHED_SMT */ |
6273 | ||
6274 | /* | |
6275 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6276 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6277 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6278 | */ |
c722f35b | 6279 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target) |
10e2f1ac | 6280 | { |
60588bfa | 6281 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); |
9fe1f127 | 6282 | int i, cpu, idle_cpu = -1, nr = INT_MAX; |
94aafc3e | 6283 | struct rq *this_rq = this_rq(); |
9fe1f127 | 6284 | int this = smp_processor_id(); |
9cfb38a7 | 6285 | struct sched_domain *this_sd; |
94aafc3e | 6286 | u64 time = 0; |
10e2f1ac | 6287 | |
9cfb38a7 WL |
6288 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6289 | if (!this_sd) | |
6290 | return -1; | |
6291 | ||
bae4ec13 MG |
6292 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6293 | ||
c722f35b | 6294 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
e6e0dc2d | 6295 | u64 avg_cost, avg_idle, span_avg; |
94aafc3e | 6296 | unsigned long now = jiffies; |
1ad3aaf3 | 6297 | |
e6e0dc2d | 6298 | /* |
94aafc3e PZ |
6299 | * If we're busy, the assumption that the last idle period |
6300 | * predicts the future is flawed; age away the remaining | |
6301 | * predicted idle time. | |
e6e0dc2d | 6302 | */ |
94aafc3e PZ |
6303 | if (unlikely(this_rq->wake_stamp < now)) { |
6304 | while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) { | |
6305 | this_rq->wake_stamp++; | |
6306 | this_rq->wake_avg_idle >>= 1; | |
6307 | } | |
6308 | } | |
6309 | ||
6310 | avg_idle = this_rq->wake_avg_idle; | |
e6e0dc2d | 6311 | avg_cost = this_sd->avg_scan_cost + 1; |
10e2f1ac | 6312 | |
e6e0dc2d | 6313 | span_avg = sd->span_weight * avg_idle; |
1ad3aaf3 PZ |
6314 | if (span_avg > 4*avg_cost) |
6315 | nr = div_u64(span_avg, avg_cost); | |
6316 | else | |
6317 | nr = 4; | |
10e2f1ac | 6318 | |
bae4ec13 MG |
6319 | time = cpu_clock(this); |
6320 | } | |
60588bfa | 6321 | |
56498cfb | 6322 | for_each_cpu_wrap(cpu, cpus, target + 1) { |
c722f35b | 6323 | if (has_idle_core) { |
9fe1f127 MG |
6324 | i = select_idle_core(p, cpu, cpus, &idle_cpu); |
6325 | if ((unsigned int)i < nr_cpumask_bits) | |
6326 | return i; | |
6327 | ||
6328 | } else { | |
6329 | if (!--nr) | |
6330 | return -1; | |
97886d9d | 6331 | idle_cpu = __select_idle_cpu(cpu, p); |
9fe1f127 MG |
6332 | if ((unsigned int)idle_cpu < nr_cpumask_bits) |
6333 | break; | |
6334 | } | |
10e2f1ac PZ |
6335 | } |
6336 | ||
c722f35b | 6337 | if (has_idle_core) |
02dbb724 | 6338 | set_idle_cores(target, false); |
9fe1f127 | 6339 | |
c722f35b | 6340 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
bae4ec13 | 6341 | time = cpu_clock(this) - time; |
94aafc3e PZ |
6342 | |
6343 | /* | |
6344 | * Account for the scan cost of wakeups against the average | |
6345 | * idle time. | |
6346 | */ | |
6347 | this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time); | |
6348 | ||
bae4ec13 MG |
6349 | update_avg(&this_sd->avg_scan_cost, time); |
6350 | } | |
10e2f1ac | 6351 | |
9fe1f127 | 6352 | return idle_cpu; |
10e2f1ac PZ |
6353 | } |
6354 | ||
b7a33161 MR |
6355 | /* |
6356 | * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which | |
6357 | * the task fits. If no CPU is big enough, but there are idle ones, try to | |
6358 | * maximize capacity. | |
6359 | */ | |
6360 | static int | |
6361 | select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) | |
6362 | { | |
b4c9c9f1 | 6363 | unsigned long task_util, best_cap = 0; |
b7a33161 MR |
6364 | int cpu, best_cpu = -1; |
6365 | struct cpumask *cpus; | |
6366 | ||
b7a33161 MR |
6367 | cpus = this_cpu_cpumask_var_ptr(select_idle_mask); |
6368 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); | |
6369 | ||
b4c9c9f1 VG |
6370 | task_util = uclamp_task_util(p); |
6371 | ||
b7a33161 MR |
6372 | for_each_cpu_wrap(cpu, cpus, target) { |
6373 | unsigned long cpu_cap = capacity_of(cpu); | |
6374 | ||
6375 | if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) | |
6376 | continue; | |
b4c9c9f1 | 6377 | if (fits_capacity(task_util, cpu_cap)) |
b7a33161 MR |
6378 | return cpu; |
6379 | ||
6380 | if (cpu_cap > best_cap) { | |
6381 | best_cap = cpu_cap; | |
6382 | best_cpu = cpu; | |
6383 | } | |
6384 | } | |
6385 | ||
6386 | return best_cpu; | |
6387 | } | |
6388 | ||
b4c9c9f1 VG |
6389 | static inline bool asym_fits_capacity(int task_util, int cpu) |
6390 | { | |
6391 | if (static_branch_unlikely(&sched_asym_cpucapacity)) | |
6392 | return fits_capacity(task_util, capacity_of(cpu)); | |
6393 | ||
6394 | return true; | |
6395 | } | |
6396 | ||
10e2f1ac PZ |
6397 | /* |
6398 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6399 | */ |
772bd008 | 6400 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6401 | { |
c722f35b | 6402 | bool has_idle_core = false; |
99bd5e2f | 6403 | struct sched_domain *sd; |
b4c9c9f1 | 6404 | unsigned long task_util; |
32e839dd | 6405 | int i, recent_used_cpu; |
a50bde51 | 6406 | |
b7a33161 | 6407 | /* |
b4c9c9f1 VG |
6408 | * On asymmetric system, update task utilization because we will check |
6409 | * that the task fits with cpu's capacity. | |
b7a33161 MR |
6410 | */ |
6411 | if (static_branch_unlikely(&sched_asym_cpucapacity)) { | |
b4c9c9f1 VG |
6412 | sync_entity_load_avg(&p->se); |
6413 | task_util = uclamp_task_util(p); | |
b7a33161 MR |
6414 | } |
6415 | ||
9099a147 PZ |
6416 | /* |
6417 | * per-cpu select_idle_mask usage | |
6418 | */ | |
6419 | lockdep_assert_irqs_disabled(); | |
6420 | ||
b4c9c9f1 VG |
6421 | if ((available_idle_cpu(target) || sched_idle_cpu(target)) && |
6422 | asym_fits_capacity(task_util, target)) | |
e0a79f52 | 6423 | return target; |
99bd5e2f SS |
6424 | |
6425 | /* | |
97fb7a0a | 6426 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6427 | */ |
3c29e651 | 6428 | if (prev != target && cpus_share_cache(prev, target) && |
b4c9c9f1 VG |
6429 | (available_idle_cpu(prev) || sched_idle_cpu(prev)) && |
6430 | asym_fits_capacity(task_util, prev)) | |
772bd008 | 6431 | return prev; |
a50bde51 | 6432 | |
52262ee5 MG |
6433 | /* |
6434 | * Allow a per-cpu kthread to stack with the wakee if the | |
6435 | * kworker thread and the tasks previous CPUs are the same. | |
6436 | * The assumption is that the wakee queued work for the | |
6437 | * per-cpu kthread that is now complete and the wakeup is | |
6438 | * essentially a sync wakeup. An obvious example of this | |
6439 | * pattern is IO completions. | |
6440 | */ | |
6441 | if (is_per_cpu_kthread(current) && | |
d492a48a | 6442 | in_task() && |
52262ee5 | 6443 | prev == smp_processor_id() && |
5a069c4c VD |
6444 | this_rq()->nr_running <= 1 && |
6445 | asym_fits_capacity(task_util, prev)) { | |
52262ee5 MG |
6446 | return prev; |
6447 | } | |
6448 | ||
97fb7a0a | 6449 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd | 6450 | recent_used_cpu = p->recent_used_cpu; |
89aafd67 | 6451 | p->recent_used_cpu = prev; |
32e839dd MG |
6452 | if (recent_used_cpu != prev && |
6453 | recent_used_cpu != target && | |
6454 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 6455 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
b4c9c9f1 VG |
6456 | cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) && |
6457 | asym_fits_capacity(task_util, recent_used_cpu)) { | |
32e839dd MG |
6458 | /* |
6459 | * Replace recent_used_cpu with prev as it is a potential | |
97fb7a0a | 6460 | * candidate for the next wake: |
32e839dd MG |
6461 | */ |
6462 | p->recent_used_cpu = prev; | |
6463 | return recent_used_cpu; | |
6464 | } | |
6465 | ||
b4c9c9f1 VG |
6466 | /* |
6467 | * For asymmetric CPU capacity systems, our domain of interest is | |
6468 | * sd_asym_cpucapacity rather than sd_llc. | |
6469 | */ | |
6470 | if (static_branch_unlikely(&sched_asym_cpucapacity)) { | |
6471 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); | |
6472 | /* | |
6473 | * On an asymmetric CPU capacity system where an exclusive | |
6474 | * cpuset defines a symmetric island (i.e. one unique | |
6475 | * capacity_orig value through the cpuset), the key will be set | |
6476 | * but the CPUs within that cpuset will not have a domain with | |
6477 | * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric | |
6478 | * capacity path. | |
6479 | */ | |
6480 | if (sd) { | |
6481 | i = select_idle_capacity(p, sd, target); | |
6482 | return ((unsigned)i < nr_cpumask_bits) ? i : target; | |
6483 | } | |
6484 | } | |
6485 | ||
518cd623 | 6486 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6487 | if (!sd) |
6488 | return target; | |
772bd008 | 6489 | |
c722f35b RR |
6490 | if (sched_smt_active()) { |
6491 | has_idle_core = test_idle_cores(target, false); | |
6492 | ||
6493 | if (!has_idle_core && cpus_share_cache(prev, target)) { | |
6494 | i = select_idle_smt(p, sd, prev); | |
6495 | if ((unsigned int)i < nr_cpumask_bits) | |
6496 | return i; | |
6497 | } | |
6498 | } | |
6499 | ||
6500 | i = select_idle_cpu(p, sd, has_idle_core, target); | |
10e2f1ac PZ |
6501 | if ((unsigned)i < nr_cpumask_bits) |
6502 | return i; | |
6503 | ||
a50bde51 PZ |
6504 | return target; |
6505 | } | |
231678b7 | 6506 | |
f9be3e59 | 6507 | /** |
59a74b15 | 6508 | * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks. |
f9be3e59 PB |
6509 | * @cpu: the CPU to get the utilization of |
6510 | * | |
6511 | * The unit of the return value must be the one of capacity so we can compare | |
6512 | * the utilization with the capacity of the CPU that is available for CFS task | |
6513 | * (ie cpu_capacity). | |
231678b7 DE |
6514 | * |
6515 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
6516 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
6517 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
6518 | * capacity_orig is the cpu_capacity available at the highest frequency | |
6519 | * (arch_scale_freq_capacity()). | |
6520 | * The utilization of a CPU converges towards a sum equal to or less than the | |
6521 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
6522 | * the running time on this CPU scaled by capacity_curr. | |
6523 | * | |
f9be3e59 PB |
6524 | * The estimated utilization of a CPU is defined to be the maximum between its |
6525 | * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks | |
6526 | * currently RUNNABLE on that CPU. | |
6527 | * This allows to properly represent the expected utilization of a CPU which | |
6528 | * has just got a big task running since a long sleep period. At the same time | |
6529 | * however it preserves the benefits of the "blocked utilization" in | |
6530 | * describing the potential for other tasks waking up on the same CPU. | |
6531 | * | |
231678b7 DE |
6532 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even |
6533 | * higher than capacity_orig because of unfortunate rounding in | |
6534 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
6535 | * the average stabilizes with the new running time. We need to check that the | |
6536 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
6537 | * necessary. Without utilization capping, a group could be seen as overloaded | |
6538 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
6539 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
6540 | * capacity_orig) as it useful for predicting the capacity required after task | |
6541 | * migrations (scheduler-driven DVFS). | |
f9be3e59 PB |
6542 | * |
6543 | * Return: the (estimated) utilization for the specified CPU | |
8bb5b00c | 6544 | */ |
f9be3e59 | 6545 | static inline unsigned long cpu_util(int cpu) |
8bb5b00c | 6546 | { |
f9be3e59 PB |
6547 | struct cfs_rq *cfs_rq; |
6548 | unsigned int util; | |
6549 | ||
6550 | cfs_rq = &cpu_rq(cpu)->cfs; | |
6551 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6552 | ||
6553 | if (sched_feat(UTIL_EST)) | |
6554 | util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); | |
8bb5b00c | 6555 | |
f9be3e59 | 6556 | return min_t(unsigned long, util, capacity_orig_of(cpu)); |
8bb5b00c | 6557 | } |
a50bde51 | 6558 | |
104cb16d | 6559 | /* |
c469933e PB |
6560 | * cpu_util_without: compute cpu utilization without any contributions from *p |
6561 | * @cpu: the CPU which utilization is requested | |
6562 | * @p: the task which utilization should be discounted | |
6563 | * | |
6564 | * The utilization of a CPU is defined by the utilization of tasks currently | |
6565 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
6566 | * execution on that CPU. | |
6567 | * | |
6568 | * This method returns the utilization of the specified CPU by discounting the | |
6569 | * utilization of the specified task, whenever the task is currently | |
6570 | * contributing to the CPU utilization. | |
104cb16d | 6571 | */ |
c469933e | 6572 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) |
104cb16d | 6573 | { |
f9be3e59 PB |
6574 | struct cfs_rq *cfs_rq; |
6575 | unsigned int util; | |
104cb16d MR |
6576 | |
6577 | /* Task has no contribution or is new */ | |
f9be3e59 | 6578 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) |
104cb16d MR |
6579 | return cpu_util(cpu); |
6580 | ||
f9be3e59 PB |
6581 | cfs_rq = &cpu_rq(cpu)->cfs; |
6582 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6583 | ||
c469933e | 6584 | /* Discount task's util from CPU's util */ |
b5c0ce7b | 6585 | lsub_positive(&util, task_util(p)); |
104cb16d | 6586 | |
f9be3e59 PB |
6587 | /* |
6588 | * Covered cases: | |
6589 | * | |
6590 | * a) if *p is the only task sleeping on this CPU, then: | |
6591 | * cpu_util (== task_util) > util_est (== 0) | |
6592 | * and thus we return: | |
c469933e | 6593 | * cpu_util_without = (cpu_util - task_util) = 0 |
f9be3e59 PB |
6594 | * |
6595 | * b) if other tasks are SLEEPING on this CPU, which is now exiting | |
6596 | * IDLE, then: | |
6597 | * cpu_util >= task_util | |
6598 | * cpu_util > util_est (== 0) | |
6599 | * and thus we discount *p's blocked utilization to return: | |
c469933e | 6600 | * cpu_util_without = (cpu_util - task_util) >= 0 |
f9be3e59 PB |
6601 | * |
6602 | * c) if other tasks are RUNNABLE on that CPU and | |
6603 | * util_est > cpu_util | |
6604 | * then we use util_est since it returns a more restrictive | |
6605 | * estimation of the spare capacity on that CPU, by just | |
6606 | * considering the expected utilization of tasks already | |
6607 | * runnable on that CPU. | |
6608 | * | |
6609 | * Cases a) and b) are covered by the above code, while case c) is | |
6610 | * covered by the following code when estimated utilization is | |
6611 | * enabled. | |
6612 | */ | |
c469933e PB |
6613 | if (sched_feat(UTIL_EST)) { |
6614 | unsigned int estimated = | |
6615 | READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6616 | ||
6617 | /* | |
6618 | * Despite the following checks we still have a small window | |
6619 | * for a possible race, when an execl's select_task_rq_fair() | |
6620 | * races with LB's detach_task(): | |
6621 | * | |
6622 | * detach_task() | |
6623 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
6624 | * ---------------------------------- A | |
6625 | * deactivate_task() \ | |
6626 | * dequeue_task() + RaceTime | |
6627 | * util_est_dequeue() / | |
6628 | * ---------------------------------- B | |
6629 | * | |
6630 | * The additional check on "current == p" it's required to | |
6631 | * properly fix the execl regression and it helps in further | |
6632 | * reducing the chances for the above race. | |
6633 | */ | |
b5c0ce7b PB |
6634 | if (unlikely(task_on_rq_queued(p) || current == p)) |
6635 | lsub_positive(&estimated, _task_util_est(p)); | |
6636 | ||
c469933e PB |
6637 | util = max(util, estimated); |
6638 | } | |
f9be3e59 PB |
6639 | |
6640 | /* | |
6641 | * Utilization (estimated) can exceed the CPU capacity, thus let's | |
6642 | * clamp to the maximum CPU capacity to ensure consistency with | |
6643 | * the cpu_util call. | |
6644 | */ | |
6645 | return min_t(unsigned long, util, capacity_orig_of(cpu)); | |
104cb16d MR |
6646 | } |
6647 | ||
390031e4 QP |
6648 | /* |
6649 | * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) | |
6650 | * to @dst_cpu. | |
6651 | */ | |
6652 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
6653 | { | |
6654 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
6655 | unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); | |
6656 | ||
6657 | /* | |
6658 | * If @p migrates from @cpu to another, remove its contribution. Or, | |
6659 | * if @p migrates from another CPU to @cpu, add its contribution. In | |
6660 | * the other cases, @cpu is not impacted by the migration, so the | |
6661 | * util_avg should already be correct. | |
6662 | */ | |
6663 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
736cc6b3 | 6664 | lsub_positive(&util, task_util(p)); |
390031e4 QP |
6665 | else if (task_cpu(p) != cpu && dst_cpu == cpu) |
6666 | util += task_util(p); | |
6667 | ||
6668 | if (sched_feat(UTIL_EST)) { | |
6669 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6670 | ||
6671 | /* | |
6672 | * During wake-up, the task isn't enqueued yet and doesn't | |
6673 | * appear in the cfs_rq->avg.util_est.enqueued of any rq, | |
6674 | * so just add it (if needed) to "simulate" what will be | |
6675 | * cpu_util() after the task has been enqueued. | |
6676 | */ | |
6677 | if (dst_cpu == cpu) | |
6678 | util_est += _task_util_est(p); | |
6679 | ||
6680 | util = max(util, util_est); | |
6681 | } | |
6682 | ||
6683 | return min(util, capacity_orig_of(cpu)); | |
6684 | } | |
6685 | ||
6686 | /* | |
eb92692b | 6687 | * compute_energy(): Estimates the energy that @pd would consume if @p was |
390031e4 | 6688 | * migrated to @dst_cpu. compute_energy() predicts what will be the utilization |
eb92692b | 6689 | * landscape of @pd's CPUs after the task migration, and uses the Energy Model |
390031e4 QP |
6690 | * to compute what would be the energy if we decided to actually migrate that |
6691 | * task. | |
6692 | */ | |
6693 | static long | |
6694 | compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) | |
6695 | { | |
eb92692b QP |
6696 | struct cpumask *pd_mask = perf_domain_span(pd); |
6697 | unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); | |
6698 | unsigned long max_util = 0, sum_util = 0; | |
489f1645 | 6699 | unsigned long _cpu_cap = cpu_cap; |
390031e4 QP |
6700 | int cpu; |
6701 | ||
489f1645 LL |
6702 | _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask)); |
6703 | ||
eb92692b QP |
6704 | /* |
6705 | * The capacity state of CPUs of the current rd can be driven by CPUs | |
6706 | * of another rd if they belong to the same pd. So, account for the | |
6707 | * utilization of these CPUs too by masking pd with cpu_online_mask | |
6708 | * instead of the rd span. | |
6709 | * | |
6710 | * If an entire pd is outside of the current rd, it will not appear in | |
6711 | * its pd list and will not be accounted by compute_energy(). | |
6712 | */ | |
6713 | for_each_cpu_and(cpu, pd_mask, cpu_online_mask) { | |
0372e1cf VD |
6714 | unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu); |
6715 | unsigned long cpu_util, util_running = util_freq; | |
6716 | struct task_struct *tsk = NULL; | |
6717 | ||
6718 | /* | |
6719 | * When @p is placed on @cpu: | |
6720 | * | |
6721 | * util_running = max(cpu_util, cpu_util_est) + | |
6722 | * max(task_util, _task_util_est) | |
6723 | * | |
6724 | * while cpu_util_next is: max(cpu_util + task_util, | |
6725 | * cpu_util_est + _task_util_est) | |
6726 | */ | |
6727 | if (cpu == dst_cpu) { | |
6728 | tsk = p; | |
6729 | util_running = | |
6730 | cpu_util_next(cpu, p, -1) + task_util_est(p); | |
6731 | } | |
af24bde8 PB |
6732 | |
6733 | /* | |
eb92692b QP |
6734 | * Busy time computation: utilization clamping is not |
6735 | * required since the ratio (sum_util / cpu_capacity) | |
6736 | * is already enough to scale the EM reported power | |
6737 | * consumption at the (eventually clamped) cpu_capacity. | |
af24bde8 | 6738 | */ |
489f1645 LL |
6739 | cpu_util = effective_cpu_util(cpu, util_running, cpu_cap, |
6740 | ENERGY_UTIL, NULL); | |
6741 | ||
6742 | sum_util += min(cpu_util, _cpu_cap); | |
af24bde8 | 6743 | |
390031e4 | 6744 | /* |
eb92692b QP |
6745 | * Performance domain frequency: utilization clamping |
6746 | * must be considered since it affects the selection | |
6747 | * of the performance domain frequency. | |
6748 | * NOTE: in case RT tasks are running, by default the | |
6749 | * FREQUENCY_UTIL's utilization can be max OPP. | |
390031e4 | 6750 | */ |
0372e1cf | 6751 | cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap, |
eb92692b | 6752 | FREQUENCY_UTIL, tsk); |
489f1645 | 6753 | max_util = max(max_util, min(cpu_util, _cpu_cap)); |
390031e4 QP |
6754 | } |
6755 | ||
8f1b971b | 6756 | return em_cpu_energy(pd->em_pd, max_util, sum_util, _cpu_cap); |
390031e4 QP |
6757 | } |
6758 | ||
732cd75b QP |
6759 | /* |
6760 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
6761 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
6762 | * spare capacity in each performance domain and uses it as a potential | |
6763 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
6764 | * out which of the CPU candidates is the most energy-efficient. | |
6765 | * | |
6766 | * The rationale for this heuristic is as follows. In a performance domain, | |
6767 | * all the most energy efficient CPU candidates (according to the Energy | |
6768 | * Model) are those for which we'll request a low frequency. When there are | |
6769 | * several CPUs for which the frequency request will be the same, we don't | |
6770 | * have enough data to break the tie between them, because the Energy Model | |
6771 | * only includes active power costs. With this model, if we assume that | |
6772 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
6773 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
6774 | * the best candidates of the performance domain. | |
6775 | * | |
6776 | * In practice, it could be preferable from an energy standpoint to pack | |
6777 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
6778 | * but that could also hurt our chances to go cluster idle, and we have no | |
6779 | * ways to tell with the current Energy Model if this is actually a good | |
6780 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
6781 | * cluster-packing, and spreading inside a cluster. That should at least be | |
6782 | * a good thing for latency, and this is consistent with the idea that most | |
6783 | * of the energy savings of EAS come from the asymmetry of the system, and | |
6784 | * not so much from breaking the tie between identical CPUs. That's also the | |
6785 | * reason why EAS is enabled in the topology code only for systems where | |
6786 | * SD_ASYM_CPUCAPACITY is set. | |
6787 | * | |
6788 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
6789 | * they don't have any useful utilization data yet and it's not possible to | |
6790 | * forecast their impact on energy consumption. Consequently, they will be | |
6791 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
6792 | * to be energy-inefficient in some use-cases. The alternative would be to | |
6793 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
6794 | * their util_avg from the parent task, but those heuristics could hurt | |
6795 | * other use-cases too. So, until someone finds a better way to solve this, | |
6796 | * let's keep things simple by re-using the existing slow path. | |
6797 | */ | |
732cd75b QP |
6798 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) |
6799 | { | |
eb92692b | 6800 | unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; |
732cd75b | 6801 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; |
619e090c | 6802 | int cpu, best_energy_cpu = prev_cpu, target = -1; |
eb92692b | 6803 | unsigned long cpu_cap, util, base_energy = 0; |
732cd75b | 6804 | struct sched_domain *sd; |
eb92692b | 6805 | struct perf_domain *pd; |
732cd75b QP |
6806 | |
6807 | rcu_read_lock(); | |
6808 | pd = rcu_dereference(rd->pd); | |
6809 | if (!pd || READ_ONCE(rd->overutilized)) | |
619e090c | 6810 | goto unlock; |
732cd75b QP |
6811 | |
6812 | /* | |
6813 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
6814 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
6815 | */ | |
6816 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
6817 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
6818 | sd = sd->parent; | |
6819 | if (!sd) | |
619e090c PG |
6820 | goto unlock; |
6821 | ||
6822 | target = prev_cpu; | |
732cd75b QP |
6823 | |
6824 | sync_entity_load_avg(&p->se); | |
6825 | if (!task_util_est(p)) | |
6826 | goto unlock; | |
6827 | ||
6828 | for (; pd; pd = pd->next) { | |
eb92692b | 6829 | unsigned long cur_delta, spare_cap, max_spare_cap = 0; |
8d4c97c1 | 6830 | bool compute_prev_delta = false; |
eb92692b | 6831 | unsigned long base_energy_pd; |
732cd75b QP |
6832 | int max_spare_cap_cpu = -1; |
6833 | ||
6834 | for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { | |
3bd37062 | 6835 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
6836 | continue; |
6837 | ||
732cd75b QP |
6838 | util = cpu_util_next(cpu, p, cpu); |
6839 | cpu_cap = capacity_of(cpu); | |
da0777d3 LL |
6840 | spare_cap = cpu_cap; |
6841 | lsub_positive(&spare_cap, util); | |
1d42509e VS |
6842 | |
6843 | /* | |
6844 | * Skip CPUs that cannot satisfy the capacity request. | |
6845 | * IOW, placing the task there would make the CPU | |
6846 | * overutilized. Take uclamp into account to see how | |
6847 | * much capacity we can get out of the CPU; this is | |
a5418be9 | 6848 | * aligned with sched_cpu_util(). |
1d42509e VS |
6849 | */ |
6850 | util = uclamp_rq_util_with(cpu_rq(cpu), util, p); | |
60e17f5c | 6851 | if (!fits_capacity(util, cpu_cap)) |
732cd75b QP |
6852 | continue; |
6853 | ||
732cd75b | 6854 | if (cpu == prev_cpu) { |
8d4c97c1 PG |
6855 | /* Always use prev_cpu as a candidate. */ |
6856 | compute_prev_delta = true; | |
6857 | } else if (spare_cap > max_spare_cap) { | |
6858 | /* | |
6859 | * Find the CPU with the maximum spare capacity | |
6860 | * in the performance domain. | |
6861 | */ | |
732cd75b QP |
6862 | max_spare_cap = spare_cap; |
6863 | max_spare_cap_cpu = cpu; | |
6864 | } | |
6865 | } | |
6866 | ||
8d4c97c1 PG |
6867 | if (max_spare_cap_cpu < 0 && !compute_prev_delta) |
6868 | continue; | |
6869 | ||
6870 | /* Compute the 'base' energy of the pd, without @p */ | |
6871 | base_energy_pd = compute_energy(p, -1, pd); | |
6872 | base_energy += base_energy_pd; | |
6873 | ||
6874 | /* Evaluate the energy impact of using prev_cpu. */ | |
6875 | if (compute_prev_delta) { | |
6876 | prev_delta = compute_energy(p, prev_cpu, pd); | |
619e090c PG |
6877 | if (prev_delta < base_energy_pd) |
6878 | goto unlock; | |
8d4c97c1 PG |
6879 | prev_delta -= base_energy_pd; |
6880 | best_delta = min(best_delta, prev_delta); | |
6881 | } | |
6882 | ||
6883 | /* Evaluate the energy impact of using max_spare_cap_cpu. */ | |
6884 | if (max_spare_cap_cpu >= 0) { | |
eb92692b | 6885 | cur_delta = compute_energy(p, max_spare_cap_cpu, pd); |
619e090c PG |
6886 | if (cur_delta < base_energy_pd) |
6887 | goto unlock; | |
eb92692b QP |
6888 | cur_delta -= base_energy_pd; |
6889 | if (cur_delta < best_delta) { | |
6890 | best_delta = cur_delta; | |
732cd75b QP |
6891 | best_energy_cpu = max_spare_cap_cpu; |
6892 | } | |
6893 | } | |
6894 | } | |
732cd75b QP |
6895 | rcu_read_unlock(); |
6896 | ||
6897 | /* | |
6898 | * Pick the best CPU if prev_cpu cannot be used, or if it saves at | |
6899 | * least 6% of the energy used by prev_cpu. | |
6900 | */ | |
619e090c PG |
6901 | if ((prev_delta == ULONG_MAX) || |
6902 | (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) | |
6903 | target = best_energy_cpu; | |
732cd75b | 6904 | |
619e090c | 6905 | return target; |
732cd75b | 6906 | |
619e090c | 6907 | unlock: |
732cd75b QP |
6908 | rcu_read_unlock(); |
6909 | ||
619e090c | 6910 | return target; |
732cd75b QP |
6911 | } |
6912 | ||
aaee1203 | 6913 | /* |
de91b9cb | 6914 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
3aef1551 | 6915 | * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE, |
de91b9cb | 6916 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. |
aaee1203 | 6917 | * |
97fb7a0a IM |
6918 | * Balances load by selecting the idlest CPU in the idlest group, or under |
6919 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 6920 | * |
97fb7a0a | 6921 | * Returns the target CPU number. |
aaee1203 | 6922 | */ |
0017d735 | 6923 | static int |
3aef1551 | 6924 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags) |
aaee1203 | 6925 | { |
3aef1551 | 6926 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
f1d88b44 | 6927 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 6928 | int cpu = smp_processor_id(); |
63b0e9ed | 6929 | int new_cpu = prev_cpu; |
99bd5e2f | 6930 | int want_affine = 0; |
3aef1551 VS |
6931 | /* SD_flags and WF_flags share the first nibble */ |
6932 | int sd_flag = wake_flags & 0xF; | |
c88d5910 | 6933 | |
9099a147 PZ |
6934 | /* |
6935 | * required for stable ->cpus_allowed | |
6936 | */ | |
6937 | lockdep_assert_held(&p->pi_lock); | |
dc824eb8 | 6938 | if (wake_flags & WF_TTWU) { |
c58d25f3 | 6939 | record_wakee(p); |
732cd75b | 6940 | |
f8a696f2 | 6941 | if (sched_energy_enabled()) { |
732cd75b QP |
6942 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
6943 | if (new_cpu >= 0) | |
6944 | return new_cpu; | |
6945 | new_cpu = prev_cpu; | |
6946 | } | |
6947 | ||
00061968 | 6948 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 6949 | } |
aaee1203 | 6950 | |
dce840a0 | 6951 | rcu_read_lock(); |
aaee1203 | 6952 | for_each_domain(cpu, tmp) { |
fe3bcfe1 | 6953 | /* |
97fb7a0a | 6954 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 6955 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 6956 | */ |
99bd5e2f SS |
6957 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6958 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
6959 | if (cpu != prev_cpu) |
6960 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
6961 | ||
6962 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 6963 | break; |
f03542a7 | 6964 | } |
29cd8bae | 6965 | |
f03542a7 | 6966 | if (tmp->flags & sd_flag) |
29cd8bae | 6967 | sd = tmp; |
63b0e9ed MG |
6968 | else if (!want_affine) |
6969 | break; | |
29cd8bae PZ |
6970 | } |
6971 | ||
f1d88b44 VK |
6972 | if (unlikely(sd)) { |
6973 | /* Slow path */ | |
18bd1b4b | 6974 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
dc824eb8 | 6975 | } else if (wake_flags & WF_TTWU) { /* XXX always ? */ |
f1d88b44 | 6976 | /* Fast path */ |
f1d88b44 | 6977 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
e7693a36 | 6978 | } |
dce840a0 | 6979 | rcu_read_unlock(); |
e7693a36 | 6980 | |
c88d5910 | 6981 | return new_cpu; |
e7693a36 | 6982 | } |
0a74bef8 | 6983 | |
144d8487 PZ |
6984 | static void detach_entity_cfs_rq(struct sched_entity *se); |
6985 | ||
0a74bef8 | 6986 | /* |
97fb7a0a | 6987 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 6988 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 6989 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6990 | */ |
3f9672ba | 6991 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 6992 | { |
59efa0ba PZ |
6993 | /* |
6994 | * As blocked tasks retain absolute vruntime the migration needs to | |
6995 | * deal with this by subtracting the old and adding the new | |
6996 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6997 | * the task on the new runqueue. | |
6998 | */ | |
2f064a59 | 6999 | if (READ_ONCE(p->__state) == TASK_WAKING) { |
59efa0ba PZ |
7000 | struct sched_entity *se = &p->se; |
7001 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
7002 | u64 min_vruntime; | |
7003 | ||
7004 | #ifndef CONFIG_64BIT | |
7005 | u64 min_vruntime_copy; | |
7006 | ||
7007 | do { | |
7008 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
7009 | smp_rmb(); | |
7010 | min_vruntime = cfs_rq->min_vruntime; | |
7011 | } while (min_vruntime != min_vruntime_copy); | |
7012 | #else | |
7013 | min_vruntime = cfs_rq->min_vruntime; | |
7014 | #endif | |
7015 | ||
7016 | se->vruntime -= min_vruntime; | |
7017 | } | |
7018 | ||
144d8487 PZ |
7019 | if (p->on_rq == TASK_ON_RQ_MIGRATING) { |
7020 | /* | |
7021 | * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' | |
7022 | * rq->lock and can modify state directly. | |
7023 | */ | |
5cb9eaa3 | 7024 | lockdep_assert_rq_held(task_rq(p)); |
144d8487 PZ |
7025 | detach_entity_cfs_rq(&p->se); |
7026 | ||
7027 | } else { | |
7028 | /* | |
7029 | * We are supposed to update the task to "current" time, then | |
7030 | * its up to date and ready to go to new CPU/cfs_rq. But we | |
7031 | * have difficulty in getting what current time is, so simply | |
7032 | * throw away the out-of-date time. This will result in the | |
7033 | * wakee task is less decayed, but giving the wakee more load | |
7034 | * sounds not bad. | |
7035 | */ | |
7036 | remove_entity_load_avg(&p->se); | |
7037 | } | |
9d89c257 YD |
7038 | |
7039 | /* Tell new CPU we are migrated */ | |
7040 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
7041 | |
7042 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 7043 | p->se.exec_start = 0; |
3f9672ba SD |
7044 | |
7045 | update_scan_period(p, new_cpu); | |
0a74bef8 | 7046 | } |
12695578 YD |
7047 | |
7048 | static void task_dead_fair(struct task_struct *p) | |
7049 | { | |
7050 | remove_entity_load_avg(&p->se); | |
7051 | } | |
6e2df058 PZ |
7052 | |
7053 | static int | |
7054 | balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
7055 | { | |
7056 | if (rq->nr_running) | |
7057 | return 1; | |
7058 | ||
7059 | return newidle_balance(rq, rf) != 0; | |
7060 | } | |
e7693a36 GH |
7061 | #endif /* CONFIG_SMP */ |
7062 | ||
a555e9d8 | 7063 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
7064 | { |
7065 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
7066 | ||
7067 | /* | |
e52fb7c0 PZ |
7068 | * Since its curr running now, convert the gran from real-time |
7069 | * to virtual-time in his units. | |
13814d42 MG |
7070 | * |
7071 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
7072 | * they get preempted easier. That is, if 'se' < 'curr' then | |
7073 | * the resulting gran will be larger, therefore penalizing the | |
7074 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
7075 | * be smaller, again penalizing the lighter task. | |
7076 | * | |
7077 | * This is especially important for buddies when the leftmost | |
7078 | * task is higher priority than the buddy. | |
0bbd3336 | 7079 | */ |
f4ad9bd2 | 7080 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
7081 | } |
7082 | ||
464b7527 PZ |
7083 | /* |
7084 | * Should 'se' preempt 'curr'. | |
7085 | * | |
7086 | * |s1 | |
7087 | * |s2 | |
7088 | * |s3 | |
7089 | * g | |
7090 | * |<--->|c | |
7091 | * | |
7092 | * w(c, s1) = -1 | |
7093 | * w(c, s2) = 0 | |
7094 | * w(c, s3) = 1 | |
7095 | * | |
7096 | */ | |
7097 | static int | |
7098 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
7099 | { | |
7100 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
7101 | ||
7102 | if (vdiff <= 0) | |
7103 | return -1; | |
7104 | ||
a555e9d8 | 7105 | gran = wakeup_gran(se); |
464b7527 PZ |
7106 | if (vdiff > gran) |
7107 | return 1; | |
7108 | ||
7109 | return 0; | |
7110 | } | |
7111 | ||
02479099 PZ |
7112 | static void set_last_buddy(struct sched_entity *se) |
7113 | { | |
c5ae366e DA |
7114 | for_each_sched_entity(se) { |
7115 | if (SCHED_WARN_ON(!se->on_rq)) | |
7116 | return; | |
30400039 JD |
7117 | if (se_is_idle(se)) |
7118 | return; | |
69c80f3e | 7119 | cfs_rq_of(se)->last = se; |
c5ae366e | 7120 | } |
02479099 PZ |
7121 | } |
7122 | ||
7123 | static void set_next_buddy(struct sched_entity *se) | |
7124 | { | |
c5ae366e DA |
7125 | for_each_sched_entity(se) { |
7126 | if (SCHED_WARN_ON(!se->on_rq)) | |
7127 | return; | |
30400039 JD |
7128 | if (se_is_idle(se)) |
7129 | return; | |
69c80f3e | 7130 | cfs_rq_of(se)->next = se; |
c5ae366e | 7131 | } |
02479099 PZ |
7132 | } |
7133 | ||
ac53db59 RR |
7134 | static void set_skip_buddy(struct sched_entity *se) |
7135 | { | |
69c80f3e VP |
7136 | for_each_sched_entity(se) |
7137 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
7138 | } |
7139 | ||
bf0f6f24 IM |
7140 | /* |
7141 | * Preempt the current task with a newly woken task if needed: | |
7142 | */ | |
5a9b86f6 | 7143 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
7144 | { |
7145 | struct task_struct *curr = rq->curr; | |
8651a86c | 7146 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 7147 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 7148 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 7149 | int next_buddy_marked = 0; |
30400039 | 7150 | int cse_is_idle, pse_is_idle; |
bf0f6f24 | 7151 | |
4ae7d5ce IM |
7152 | if (unlikely(se == pse)) |
7153 | return; | |
7154 | ||
5238cdd3 | 7155 | /* |
163122b7 | 7156 | * This is possible from callers such as attach_tasks(), in which we |
3b03706f | 7157 | * unconditionally check_preempt_curr() after an enqueue (which may have |
5238cdd3 PT |
7158 | * lead to a throttle). This both saves work and prevents false |
7159 | * next-buddy nomination below. | |
7160 | */ | |
7161 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
7162 | return; | |
7163 | ||
2f36825b | 7164 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 7165 | set_next_buddy(pse); |
2f36825b VP |
7166 | next_buddy_marked = 1; |
7167 | } | |
57fdc26d | 7168 | |
aec0a514 BR |
7169 | /* |
7170 | * We can come here with TIF_NEED_RESCHED already set from new task | |
7171 | * wake up path. | |
5238cdd3 PT |
7172 | * |
7173 | * Note: this also catches the edge-case of curr being in a throttled | |
7174 | * group (e.g. via set_curr_task), since update_curr() (in the | |
7175 | * enqueue of curr) will have resulted in resched being set. This | |
7176 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
7177 | * below. | |
aec0a514 BR |
7178 | */ |
7179 | if (test_tsk_need_resched(curr)) | |
7180 | return; | |
7181 | ||
a2f5c9ab | 7182 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
7183 | if (unlikely(task_has_idle_policy(curr)) && |
7184 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
7185 | goto preempt; |
7186 | ||
91c234b4 | 7187 | /* |
a2f5c9ab DH |
7188 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
7189 | * is driven by the tick): | |
91c234b4 | 7190 | */ |
8ed92e51 | 7191 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 7192 | return; |
bf0f6f24 | 7193 | |
464b7527 | 7194 | find_matching_se(&se, &pse); |
002f128b | 7195 | BUG_ON(!pse); |
30400039 JD |
7196 | |
7197 | cse_is_idle = se_is_idle(se); | |
7198 | pse_is_idle = se_is_idle(pse); | |
7199 | ||
7200 | /* | |
7201 | * Preempt an idle group in favor of a non-idle group (and don't preempt | |
7202 | * in the inverse case). | |
7203 | */ | |
7204 | if (cse_is_idle && !pse_is_idle) | |
7205 | goto preempt; | |
7206 | if (cse_is_idle != pse_is_idle) | |
7207 | return; | |
7208 | ||
7209 | update_curr(cfs_rq_of(se)); | |
2f36825b VP |
7210 | if (wakeup_preempt_entity(se, pse) == 1) { |
7211 | /* | |
7212 | * Bias pick_next to pick the sched entity that is | |
7213 | * triggering this preemption. | |
7214 | */ | |
7215 | if (!next_buddy_marked) | |
7216 | set_next_buddy(pse); | |
3a7e73a2 | 7217 | goto preempt; |
2f36825b | 7218 | } |
464b7527 | 7219 | |
3a7e73a2 | 7220 | return; |
a65ac745 | 7221 | |
3a7e73a2 | 7222 | preempt: |
8875125e | 7223 | resched_curr(rq); |
3a7e73a2 PZ |
7224 | /* |
7225 | * Only set the backward buddy when the current task is still | |
7226 | * on the rq. This can happen when a wakeup gets interleaved | |
7227 | * with schedule on the ->pre_schedule() or idle_balance() | |
7228 | * point, either of which can * drop the rq lock. | |
7229 | * | |
7230 | * Also, during early boot the idle thread is in the fair class, | |
7231 | * for obvious reasons its a bad idea to schedule back to it. | |
7232 | */ | |
7233 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
7234 | return; | |
7235 | ||
7236 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
7237 | set_last_buddy(se); | |
bf0f6f24 IM |
7238 | } |
7239 | ||
21f56ffe PZ |
7240 | #ifdef CONFIG_SMP |
7241 | static struct task_struct *pick_task_fair(struct rq *rq) | |
7242 | { | |
7243 | struct sched_entity *se; | |
7244 | struct cfs_rq *cfs_rq; | |
7245 | ||
7246 | again: | |
7247 | cfs_rq = &rq->cfs; | |
7248 | if (!cfs_rq->nr_running) | |
7249 | return NULL; | |
7250 | ||
7251 | do { | |
7252 | struct sched_entity *curr = cfs_rq->curr; | |
7253 | ||
7254 | /* When we pick for a remote RQ, we'll not have done put_prev_entity() */ | |
7255 | if (curr) { | |
7256 | if (curr->on_rq) | |
7257 | update_curr(cfs_rq); | |
7258 | else | |
7259 | curr = NULL; | |
7260 | ||
7261 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
7262 | goto again; | |
7263 | } | |
7264 | ||
7265 | se = pick_next_entity(cfs_rq, curr); | |
7266 | cfs_rq = group_cfs_rq(se); | |
7267 | } while (cfs_rq); | |
7268 | ||
7269 | return task_of(se); | |
7270 | } | |
7271 | #endif | |
7272 | ||
5d7d6056 | 7273 | struct task_struct * |
d8ac8971 | 7274 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
7275 | { |
7276 | struct cfs_rq *cfs_rq = &rq->cfs; | |
7277 | struct sched_entity *se; | |
678d5718 | 7278 | struct task_struct *p; |
37e117c0 | 7279 | int new_tasks; |
678d5718 | 7280 | |
6e83125c | 7281 | again: |
6e2df058 | 7282 | if (!sched_fair_runnable(rq)) |
38033c37 | 7283 | goto idle; |
678d5718 | 7284 | |
9674f5ca | 7285 | #ifdef CONFIG_FAIR_GROUP_SCHED |
67692435 | 7286 | if (!prev || prev->sched_class != &fair_sched_class) |
678d5718 PZ |
7287 | goto simple; |
7288 | ||
7289 | /* | |
7290 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
7291 | * likely that a next task is from the same cgroup as the current. | |
7292 | * | |
7293 | * Therefore attempt to avoid putting and setting the entire cgroup | |
7294 | * hierarchy, only change the part that actually changes. | |
7295 | */ | |
7296 | ||
7297 | do { | |
7298 | struct sched_entity *curr = cfs_rq->curr; | |
7299 | ||
7300 | /* | |
7301 | * Since we got here without doing put_prev_entity() we also | |
7302 | * have to consider cfs_rq->curr. If it is still a runnable | |
7303 | * entity, update_curr() will update its vruntime, otherwise | |
7304 | * forget we've ever seen it. | |
7305 | */ | |
54d27365 BS |
7306 | if (curr) { |
7307 | if (curr->on_rq) | |
7308 | update_curr(cfs_rq); | |
7309 | else | |
7310 | curr = NULL; | |
678d5718 | 7311 | |
54d27365 BS |
7312 | /* |
7313 | * This call to check_cfs_rq_runtime() will do the | |
7314 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 7315 | * Therefore the nr_running test will indeed |
54d27365 BS |
7316 | * be correct. |
7317 | */ | |
9674f5ca VK |
7318 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
7319 | cfs_rq = &rq->cfs; | |
7320 | ||
7321 | if (!cfs_rq->nr_running) | |
7322 | goto idle; | |
7323 | ||
54d27365 | 7324 | goto simple; |
9674f5ca | 7325 | } |
54d27365 | 7326 | } |
678d5718 PZ |
7327 | |
7328 | se = pick_next_entity(cfs_rq, curr); | |
7329 | cfs_rq = group_cfs_rq(se); | |
7330 | } while (cfs_rq); | |
7331 | ||
7332 | p = task_of(se); | |
7333 | ||
7334 | /* | |
7335 | * Since we haven't yet done put_prev_entity and if the selected task | |
7336 | * is a different task than we started out with, try and touch the | |
7337 | * least amount of cfs_rqs. | |
7338 | */ | |
7339 | if (prev != p) { | |
7340 | struct sched_entity *pse = &prev->se; | |
7341 | ||
7342 | while (!(cfs_rq = is_same_group(se, pse))) { | |
7343 | int se_depth = se->depth; | |
7344 | int pse_depth = pse->depth; | |
7345 | ||
7346 | if (se_depth <= pse_depth) { | |
7347 | put_prev_entity(cfs_rq_of(pse), pse); | |
7348 | pse = parent_entity(pse); | |
7349 | } | |
7350 | if (se_depth >= pse_depth) { | |
7351 | set_next_entity(cfs_rq_of(se), se); | |
7352 | se = parent_entity(se); | |
7353 | } | |
7354 | } | |
7355 | ||
7356 | put_prev_entity(cfs_rq, pse); | |
7357 | set_next_entity(cfs_rq, se); | |
7358 | } | |
7359 | ||
93824900 | 7360 | goto done; |
678d5718 | 7361 | simple: |
678d5718 | 7362 | #endif |
67692435 PZ |
7363 | if (prev) |
7364 | put_prev_task(rq, prev); | |
606dba2e | 7365 | |
bf0f6f24 | 7366 | do { |
678d5718 | 7367 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 7368 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
7369 | cfs_rq = group_cfs_rq(se); |
7370 | } while (cfs_rq); | |
7371 | ||
8f4d37ec | 7372 | p = task_of(se); |
678d5718 | 7373 | |
13a453c2 | 7374 | done: __maybe_unused; |
93824900 UR |
7375 | #ifdef CONFIG_SMP |
7376 | /* | |
7377 | * Move the next running task to the front of | |
7378 | * the list, so our cfs_tasks list becomes MRU | |
7379 | * one. | |
7380 | */ | |
7381 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
7382 | #endif | |
7383 | ||
e0ee463c | 7384 | if (hrtick_enabled_fair(rq)) |
b39e66ea | 7385 | hrtick_start_fair(rq, p); |
8f4d37ec | 7386 | |
3b1baa64 MR |
7387 | update_misfit_status(p, rq); |
7388 | ||
8f4d37ec | 7389 | return p; |
38033c37 PZ |
7390 | |
7391 | idle: | |
67692435 PZ |
7392 | if (!rf) |
7393 | return NULL; | |
7394 | ||
5ba553ef | 7395 | new_tasks = newidle_balance(rq, rf); |
46f69fa3 | 7396 | |
37e117c0 | 7397 | /* |
5ba553ef | 7398 | * Because newidle_balance() releases (and re-acquires) rq->lock, it is |
37e117c0 PZ |
7399 | * possible for any higher priority task to appear. In that case we |
7400 | * must re-start the pick_next_entity() loop. | |
7401 | */ | |
e4aa358b | 7402 | if (new_tasks < 0) |
37e117c0 PZ |
7403 | return RETRY_TASK; |
7404 | ||
e4aa358b | 7405 | if (new_tasks > 0) |
38033c37 | 7406 | goto again; |
38033c37 | 7407 | |
23127296 VG |
7408 | /* |
7409 | * rq is about to be idle, check if we need to update the | |
7410 | * lost_idle_time of clock_pelt | |
7411 | */ | |
7412 | update_idle_rq_clock_pelt(rq); | |
7413 | ||
38033c37 | 7414 | return NULL; |
bf0f6f24 IM |
7415 | } |
7416 | ||
98c2f700 PZ |
7417 | static struct task_struct *__pick_next_task_fair(struct rq *rq) |
7418 | { | |
7419 | return pick_next_task_fair(rq, NULL, NULL); | |
7420 | } | |
7421 | ||
bf0f6f24 IM |
7422 | /* |
7423 | * Account for a descheduled task: | |
7424 | */ | |
6e2df058 | 7425 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
7426 | { |
7427 | struct sched_entity *se = &prev->se; | |
7428 | struct cfs_rq *cfs_rq; | |
7429 | ||
7430 | for_each_sched_entity(se) { | |
7431 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 7432 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
7433 | } |
7434 | } | |
7435 | ||
ac53db59 RR |
7436 | /* |
7437 | * sched_yield() is very simple | |
7438 | * | |
7439 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
7440 | */ | |
7441 | static void yield_task_fair(struct rq *rq) | |
7442 | { | |
7443 | struct task_struct *curr = rq->curr; | |
7444 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
7445 | struct sched_entity *se = &curr->se; | |
7446 | ||
7447 | /* | |
7448 | * Are we the only task in the tree? | |
7449 | */ | |
7450 | if (unlikely(rq->nr_running == 1)) | |
7451 | return; | |
7452 | ||
7453 | clear_buddies(cfs_rq, se); | |
7454 | ||
7455 | if (curr->policy != SCHED_BATCH) { | |
7456 | update_rq_clock(rq); | |
7457 | /* | |
7458 | * Update run-time statistics of the 'current'. | |
7459 | */ | |
7460 | update_curr(cfs_rq); | |
916671c0 MG |
7461 | /* |
7462 | * Tell update_rq_clock() that we've just updated, | |
7463 | * so we don't do microscopic update in schedule() | |
7464 | * and double the fastpath cost. | |
7465 | */ | |
adcc8da8 | 7466 | rq_clock_skip_update(rq); |
ac53db59 RR |
7467 | } |
7468 | ||
7469 | set_skip_buddy(se); | |
7470 | } | |
7471 | ||
0900acf2 | 7472 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) |
d95f4122 MG |
7473 | { |
7474 | struct sched_entity *se = &p->se; | |
7475 | ||
5238cdd3 PT |
7476 | /* throttled hierarchies are not runnable */ |
7477 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
7478 | return false; |
7479 | ||
7480 | /* Tell the scheduler that we'd really like pse to run next. */ | |
7481 | set_next_buddy(se); | |
7482 | ||
d95f4122 MG |
7483 | yield_task_fair(rq); |
7484 | ||
7485 | return true; | |
7486 | } | |
7487 | ||
681f3e68 | 7488 | #ifdef CONFIG_SMP |
bf0f6f24 | 7489 | /************************************************** |
e9c84cb8 PZ |
7490 | * Fair scheduling class load-balancing methods. |
7491 | * | |
7492 | * BASICS | |
7493 | * | |
7494 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 7495 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
7496 | * time to each task. This is expressed in the following equation: |
7497 | * | |
7498 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
7499 | * | |
97fb7a0a | 7500 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
7501 | * W_i,0 is defined as: |
7502 | * | |
7503 | * W_i,0 = \Sum_j w_i,j (2) | |
7504 | * | |
97fb7a0a | 7505 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 7506 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
7507 | * |
7508 | * The weight average is an exponential decay average of the instantaneous | |
7509 | * weight: | |
7510 | * | |
7511 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
7512 | * | |
97fb7a0a | 7513 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
7514 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
7515 | * can also include other factors [XXX]. | |
7516 | * | |
7517 | * To achieve this balance we define a measure of imbalance which follows | |
7518 | * directly from (1): | |
7519 | * | |
ced549fa | 7520 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
7521 | * |
7522 | * We them move tasks around to minimize the imbalance. In the continuous | |
7523 | * function space it is obvious this converges, in the discrete case we get | |
7524 | * a few fun cases generally called infeasible weight scenarios. | |
7525 | * | |
7526 | * [XXX expand on: | |
7527 | * - infeasible weights; | |
7528 | * - local vs global optima in the discrete case. ] | |
7529 | * | |
7530 | * | |
7531 | * SCHED DOMAINS | |
7532 | * | |
7533 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 7534 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 7535 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 7536 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 7537 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 7538 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
7539 | * the groups. |
7540 | * | |
7541 | * This yields: | |
7542 | * | |
7543 | * log_2 n 1 n | |
7544 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
7545 | * i = 0 2^i 2^i | |
7546 | * `- size of each group | |
97fb7a0a | 7547 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
7548 | * | `- freq |
7549 | * `- sum over all levels | |
7550 | * | |
7551 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
7552 | * this makes (5) the runtime complexity of the balancer. | |
7553 | * | |
7554 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 7555 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
7556 | * |
7557 | * The adjacency matrix of the resulting graph is given by: | |
7558 | * | |
97a7142f | 7559 | * log_2 n |
e9c84cb8 PZ |
7560 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
7561 | * k = 0 | |
7562 | * | |
7563 | * And you'll find that: | |
7564 | * | |
7565 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
7566 | * | |
97fb7a0a | 7567 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
7568 | * The task movement gives a factor of O(m), giving a convergence complexity |
7569 | * of: | |
7570 | * | |
7571 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
7572 | * | |
7573 | * | |
7574 | * WORK CONSERVING | |
7575 | * | |
7576 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 7577 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
7578 | * tree itself instead of relying on other CPUs to bring it work. |
7579 | * | |
7580 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
7581 | * time. | |
7582 | * | |
7583 | * [XXX more?] | |
7584 | * | |
7585 | * | |
7586 | * CGROUPS | |
7587 | * | |
7588 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
7589 | * | |
7590 | * s_k,i | |
7591 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
7592 | * S_k | |
7593 | * | |
7594 | * Where | |
7595 | * | |
7596 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
7597 | * | |
97fb7a0a | 7598 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
7599 | * |
7600 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
7601 | * property. | |
7602 | * | |
7603 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
7604 | * rewrite all of this once again.] | |
97a7142f | 7605 | */ |
bf0f6f24 | 7606 | |
ed387b78 HS |
7607 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
7608 | ||
0ec8aa00 PZ |
7609 | enum fbq_type { regular, remote, all }; |
7610 | ||
0b0695f2 | 7611 | /* |
a9723389 VG |
7612 | * 'group_type' describes the group of CPUs at the moment of load balancing. |
7613 | * | |
0b0695f2 | 7614 | * The enum is ordered by pulling priority, with the group with lowest priority |
a9723389 VG |
7615 | * first so the group_type can simply be compared when selecting the busiest |
7616 | * group. See update_sd_pick_busiest(). | |
0b0695f2 | 7617 | */ |
3b1baa64 | 7618 | enum group_type { |
a9723389 | 7619 | /* The group has spare capacity that can be used to run more tasks. */ |
0b0695f2 | 7620 | group_has_spare = 0, |
a9723389 VG |
7621 | /* |
7622 | * The group is fully used and the tasks don't compete for more CPU | |
7623 | * cycles. Nevertheless, some tasks might wait before running. | |
7624 | */ | |
0b0695f2 | 7625 | group_fully_busy, |
a9723389 VG |
7626 | /* |
7627 | * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity | |
7628 | * and must be migrated to a more powerful CPU. | |
7629 | */ | |
3b1baa64 | 7630 | group_misfit_task, |
a9723389 VG |
7631 | /* |
7632 | * SD_ASYM_PACKING only: One local CPU with higher capacity is available, | |
7633 | * and the task should be migrated to it instead of running on the | |
7634 | * current CPU. | |
7635 | */ | |
0b0695f2 | 7636 | group_asym_packing, |
a9723389 VG |
7637 | /* |
7638 | * The tasks' affinity constraints previously prevented the scheduler | |
7639 | * from balancing the load across the system. | |
7640 | */ | |
3b1baa64 | 7641 | group_imbalanced, |
a9723389 VG |
7642 | /* |
7643 | * The CPU is overloaded and can't provide expected CPU cycles to all | |
7644 | * tasks. | |
7645 | */ | |
0b0695f2 VG |
7646 | group_overloaded |
7647 | }; | |
7648 | ||
7649 | enum migration_type { | |
7650 | migrate_load = 0, | |
7651 | migrate_util, | |
7652 | migrate_task, | |
7653 | migrate_misfit | |
3b1baa64 MR |
7654 | }; |
7655 | ||
ddcdf6e7 | 7656 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 7657 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
7658 | #define LBF_DST_PINNED 0x04 |
7659 | #define LBF_SOME_PINNED 0x08 | |
23fb06d9 | 7660 | #define LBF_ACTIVE_LB 0x10 |
ddcdf6e7 PZ |
7661 | |
7662 | struct lb_env { | |
7663 | struct sched_domain *sd; | |
7664 | ||
ddcdf6e7 | 7665 | struct rq *src_rq; |
85c1e7da | 7666 | int src_cpu; |
ddcdf6e7 PZ |
7667 | |
7668 | int dst_cpu; | |
7669 | struct rq *dst_rq; | |
7670 | ||
88b8dac0 SV |
7671 | struct cpumask *dst_grpmask; |
7672 | int new_dst_cpu; | |
ddcdf6e7 | 7673 | enum cpu_idle_type idle; |
bd939f45 | 7674 | long imbalance; |
b9403130 MW |
7675 | /* The set of CPUs under consideration for load-balancing */ |
7676 | struct cpumask *cpus; | |
7677 | ||
ddcdf6e7 | 7678 | unsigned int flags; |
367456c7 PZ |
7679 | |
7680 | unsigned int loop; | |
7681 | unsigned int loop_break; | |
7682 | unsigned int loop_max; | |
0ec8aa00 PZ |
7683 | |
7684 | enum fbq_type fbq_type; | |
0b0695f2 | 7685 | enum migration_type migration_type; |
163122b7 | 7686 | struct list_head tasks; |
ddcdf6e7 PZ |
7687 | }; |
7688 | ||
029632fb PZ |
7689 | /* |
7690 | * Is this task likely cache-hot: | |
7691 | */ | |
5d5e2b1b | 7692 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
7693 | { |
7694 | s64 delta; | |
7695 | ||
5cb9eaa3 | 7696 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 7697 | |
029632fb PZ |
7698 | if (p->sched_class != &fair_sched_class) |
7699 | return 0; | |
7700 | ||
1da1843f | 7701 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
7702 | return 0; |
7703 | ||
ec73240b JD |
7704 | /* SMT siblings share cache */ |
7705 | if (env->sd->flags & SD_SHARE_CPUCAPACITY) | |
7706 | return 0; | |
7707 | ||
029632fb PZ |
7708 | /* |
7709 | * Buddy candidates are cache hot: | |
7710 | */ | |
5d5e2b1b | 7711 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
7712 | (&p->se == cfs_rq_of(&p->se)->next || |
7713 | &p->se == cfs_rq_of(&p->se)->last)) | |
7714 | return 1; | |
7715 | ||
7716 | if (sysctl_sched_migration_cost == -1) | |
7717 | return 1; | |
97886d9d AL |
7718 | |
7719 | /* | |
7720 | * Don't migrate task if the task's cookie does not match | |
7721 | * with the destination CPU's core cookie. | |
7722 | */ | |
7723 | if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p)) | |
7724 | return 1; | |
7725 | ||
029632fb PZ |
7726 | if (sysctl_sched_migration_cost == 0) |
7727 | return 0; | |
7728 | ||
5d5e2b1b | 7729 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
7730 | |
7731 | return delta < (s64)sysctl_sched_migration_cost; | |
7732 | } | |
7733 | ||
3a7053b3 | 7734 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 7735 | /* |
2a1ed24c SD |
7736 | * Returns 1, if task migration degrades locality |
7737 | * Returns 0, if task migration improves locality i.e migration preferred. | |
7738 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 7739 | */ |
2a1ed24c | 7740 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 7741 | { |
b1ad065e | 7742 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
7743 | unsigned long src_weight, dst_weight; |
7744 | int src_nid, dst_nid, dist; | |
3a7053b3 | 7745 | |
2a595721 | 7746 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
7747 | return -1; |
7748 | ||
c3b9bc5b | 7749 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 7750 | return -1; |
7a0f3083 MG |
7751 | |
7752 | src_nid = cpu_to_node(env->src_cpu); | |
7753 | dst_nid = cpu_to_node(env->dst_cpu); | |
7754 | ||
83e1d2cd | 7755 | if (src_nid == dst_nid) |
2a1ed24c | 7756 | return -1; |
7a0f3083 | 7757 | |
2a1ed24c SD |
7758 | /* Migrating away from the preferred node is always bad. */ |
7759 | if (src_nid == p->numa_preferred_nid) { | |
7760 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
7761 | return 1; | |
7762 | else | |
7763 | return -1; | |
7764 | } | |
b1ad065e | 7765 | |
c1ceac62 RR |
7766 | /* Encourage migration to the preferred node. */ |
7767 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 7768 | return 0; |
b1ad065e | 7769 | |
739294fb | 7770 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 7771 | if (env->idle == CPU_IDLE) |
739294fb RR |
7772 | return -1; |
7773 | ||
f35678b6 | 7774 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 7775 | if (numa_group) { |
f35678b6 SD |
7776 | src_weight = group_weight(p, src_nid, dist); |
7777 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 7778 | } else { |
f35678b6 SD |
7779 | src_weight = task_weight(p, src_nid, dist); |
7780 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
7781 | } |
7782 | ||
f35678b6 | 7783 | return dst_weight < src_weight; |
7a0f3083 MG |
7784 | } |
7785 | ||
3a7053b3 | 7786 | #else |
2a1ed24c | 7787 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
7788 | struct lb_env *env) |
7789 | { | |
2a1ed24c | 7790 | return -1; |
7a0f3083 | 7791 | } |
3a7053b3 MG |
7792 | #endif |
7793 | ||
1e3c88bd PZ |
7794 | /* |
7795 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
7796 | */ | |
7797 | static | |
8e45cb54 | 7798 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 7799 | { |
2a1ed24c | 7800 | int tsk_cache_hot; |
e5673f28 | 7801 | |
5cb9eaa3 | 7802 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 7803 | |
1e3c88bd PZ |
7804 | /* |
7805 | * We do not migrate tasks that are: | |
d3198084 | 7806 | * 1) throttled_lb_pair, or |
3bd37062 | 7807 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
7808 | * 3) running (obviously), or |
7809 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7810 | */ |
d3198084 JK |
7811 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7812 | return 0; | |
7813 | ||
9bcb959d | 7814 | /* Disregard pcpu kthreads; they are where they need to be. */ |
3a7956e2 | 7815 | if (kthread_is_per_cpu(p)) |
9bcb959d LC |
7816 | return 0; |
7817 | ||
3bd37062 | 7818 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 7819 | int cpu; |
88b8dac0 | 7820 | |
ae92882e | 7821 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 7822 | |
6263322c PZ |
7823 | env->flags |= LBF_SOME_PINNED; |
7824 | ||
88b8dac0 | 7825 | /* |
97fb7a0a | 7826 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
7827 | * our sched_group. We may want to revisit it if we couldn't |
7828 | * meet load balance goals by pulling other tasks on src_cpu. | |
7829 | * | |
23fb06d9 VS |
7830 | * Avoid computing new_dst_cpu |
7831 | * - for NEWLY_IDLE | |
7832 | * - if we have already computed one in current iteration | |
7833 | * - if it's an active balance | |
88b8dac0 | 7834 | */ |
23fb06d9 VS |
7835 | if (env->idle == CPU_NEWLY_IDLE || |
7836 | env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) | |
88b8dac0 SV |
7837 | return 0; |
7838 | ||
97fb7a0a | 7839 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 7840 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 7841 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 7842 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7843 | env->new_dst_cpu = cpu; |
7844 | break; | |
7845 | } | |
88b8dac0 | 7846 | } |
e02e60c1 | 7847 | |
1e3c88bd PZ |
7848 | return 0; |
7849 | } | |
88b8dac0 | 7850 | |
3b03706f | 7851 | /* Record that we found at least one task that could run on dst_cpu */ |
8e45cb54 | 7852 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7853 | |
ddcdf6e7 | 7854 | if (task_running(env->src_rq, p)) { |
ae92882e | 7855 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
7856 | return 0; |
7857 | } | |
7858 | ||
7859 | /* | |
7860 | * Aggressive migration if: | |
23fb06d9 VS |
7861 | * 1) active balance |
7862 | * 2) destination numa is preferred | |
7863 | * 3) task is cache cold, or | |
7864 | * 4) too many balance attempts have failed. | |
1e3c88bd | 7865 | */ |
23fb06d9 VS |
7866 | if (env->flags & LBF_ACTIVE_LB) |
7867 | return 1; | |
7868 | ||
2a1ed24c SD |
7869 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7870 | if (tsk_cache_hot == -1) | |
7871 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7872 | |
2a1ed24c | 7873 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7874 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7875 | if (tsk_cache_hot == 1) { |
ae92882e JP |
7876 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
7877 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 7878 | } |
1e3c88bd PZ |
7879 | return 1; |
7880 | } | |
7881 | ||
ae92882e | 7882 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 7883 | return 0; |
1e3c88bd PZ |
7884 | } |
7885 | ||
897c395f | 7886 | /* |
163122b7 KT |
7887 | * detach_task() -- detach the task for the migration specified in env |
7888 | */ | |
7889 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7890 | { | |
5cb9eaa3 | 7891 | lockdep_assert_rq_held(env->src_rq); |
163122b7 | 7892 | |
5704ac0a | 7893 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7894 | set_task_cpu(p, env->dst_cpu); |
7895 | } | |
7896 | ||
897c395f | 7897 | /* |
e5673f28 | 7898 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7899 | * part of active balancing operations within "domain". |
897c395f | 7900 | * |
e5673f28 | 7901 | * Returns a task if successful and NULL otherwise. |
897c395f | 7902 | */ |
e5673f28 | 7903 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7904 | { |
93824900 | 7905 | struct task_struct *p; |
897c395f | 7906 | |
5cb9eaa3 | 7907 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 7908 | |
93824900 UR |
7909 | list_for_each_entry_reverse(p, |
7910 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7911 | if (!can_migrate_task(p, env)) |
7912 | continue; | |
897c395f | 7913 | |
163122b7 | 7914 | detach_task(p, env); |
e5673f28 | 7915 | |
367456c7 | 7916 | /* |
e5673f28 | 7917 | * Right now, this is only the second place where |
163122b7 | 7918 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 7919 | * so we can safely collect stats here rather than |
163122b7 | 7920 | * inside detach_tasks(). |
367456c7 | 7921 | */ |
ae92882e | 7922 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 7923 | return p; |
897c395f | 7924 | } |
e5673f28 | 7925 | return NULL; |
897c395f PZ |
7926 | } |
7927 | ||
eb95308e PZ |
7928 | static const unsigned int sched_nr_migrate_break = 32; |
7929 | ||
5d6523eb | 7930 | /* |
0b0695f2 | 7931 | * detach_tasks() -- tries to detach up to imbalance load/util/tasks from |
163122b7 | 7932 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 7933 | * |
163122b7 | 7934 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 7935 | */ |
163122b7 | 7936 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 7937 | { |
5d6523eb | 7938 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
0b0695f2 | 7939 | unsigned long util, load; |
5d6523eb | 7940 | struct task_struct *p; |
163122b7 KT |
7941 | int detached = 0; |
7942 | ||
5cb9eaa3 | 7943 | lockdep_assert_rq_held(env->src_rq); |
1e3c88bd | 7944 | |
acb4decc AL |
7945 | /* |
7946 | * Source run queue has been emptied by another CPU, clear | |
7947 | * LBF_ALL_PINNED flag as we will not test any task. | |
7948 | */ | |
7949 | if (env->src_rq->nr_running <= 1) { | |
7950 | env->flags &= ~LBF_ALL_PINNED; | |
7951 | return 0; | |
7952 | } | |
7953 | ||
bd939f45 | 7954 | if (env->imbalance <= 0) |
5d6523eb | 7955 | return 0; |
1e3c88bd | 7956 | |
5d6523eb | 7957 | while (!list_empty(tasks)) { |
985d3a4c YD |
7958 | /* |
7959 | * We don't want to steal all, otherwise we may be treated likewise, | |
7960 | * which could at worst lead to a livelock crash. | |
7961 | */ | |
7962 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
7963 | break; | |
7964 | ||
93824900 | 7965 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 7966 | |
367456c7 PZ |
7967 | env->loop++; |
7968 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 7969 | if (env->loop > env->loop_max) |
367456c7 | 7970 | break; |
5d6523eb PZ |
7971 | |
7972 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 7973 | if (env->loop > env->loop_break) { |
eb95308e | 7974 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 7975 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 7976 | break; |
a195f004 | 7977 | } |
1e3c88bd | 7978 | |
d3198084 | 7979 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
7980 | goto next; |
7981 | ||
0b0695f2 VG |
7982 | switch (env->migration_type) { |
7983 | case migrate_load: | |
01cfcde9 VG |
7984 | /* |
7985 | * Depending of the number of CPUs and tasks and the | |
7986 | * cgroup hierarchy, task_h_load() can return a null | |
7987 | * value. Make sure that env->imbalance decreases | |
7988 | * otherwise detach_tasks() will stop only after | |
7989 | * detaching up to loop_max tasks. | |
7990 | */ | |
7991 | load = max_t(unsigned long, task_h_load(p), 1); | |
5d6523eb | 7992 | |
0b0695f2 VG |
7993 | if (sched_feat(LB_MIN) && |
7994 | load < 16 && !env->sd->nr_balance_failed) | |
7995 | goto next; | |
367456c7 | 7996 | |
6cf82d55 VG |
7997 | /* |
7998 | * Make sure that we don't migrate too much load. | |
7999 | * Nevertheless, let relax the constraint if | |
8000 | * scheduler fails to find a good waiting task to | |
8001 | * migrate. | |
8002 | */ | |
39a2a6eb | 8003 | if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) |
0b0695f2 VG |
8004 | goto next; |
8005 | ||
8006 | env->imbalance -= load; | |
8007 | break; | |
8008 | ||
8009 | case migrate_util: | |
8010 | util = task_util_est(p); | |
8011 | ||
8012 | if (util > env->imbalance) | |
8013 | goto next; | |
8014 | ||
8015 | env->imbalance -= util; | |
8016 | break; | |
8017 | ||
8018 | case migrate_task: | |
8019 | env->imbalance--; | |
8020 | break; | |
8021 | ||
8022 | case migrate_misfit: | |
c63be7be VG |
8023 | /* This is not a misfit task */ |
8024 | if (task_fits_capacity(p, capacity_of(env->src_cpu))) | |
0b0695f2 VG |
8025 | goto next; |
8026 | ||
8027 | env->imbalance = 0; | |
8028 | break; | |
8029 | } | |
1e3c88bd | 8030 | |
163122b7 KT |
8031 | detach_task(p, env); |
8032 | list_add(&p->se.group_node, &env->tasks); | |
8033 | ||
8034 | detached++; | |
1e3c88bd | 8035 | |
c1a280b6 | 8036 | #ifdef CONFIG_PREEMPTION |
ee00e66f PZ |
8037 | /* |
8038 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 8039 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
8040 | * the critical section. |
8041 | */ | |
5d6523eb | 8042 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 8043 | break; |
1e3c88bd PZ |
8044 | #endif |
8045 | ||
ee00e66f PZ |
8046 | /* |
8047 | * We only want to steal up to the prescribed amount of | |
0b0695f2 | 8048 | * load/util/tasks. |
ee00e66f | 8049 | */ |
bd939f45 | 8050 | if (env->imbalance <= 0) |
ee00e66f | 8051 | break; |
367456c7 PZ |
8052 | |
8053 | continue; | |
8054 | next: | |
93824900 | 8055 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 8056 | } |
5d6523eb | 8057 | |
1e3c88bd | 8058 | /* |
163122b7 KT |
8059 | * Right now, this is one of only two places we collect this stat |
8060 | * so we can safely collect detach_one_task() stats here rather | |
8061 | * than inside detach_one_task(). | |
1e3c88bd | 8062 | */ |
ae92882e | 8063 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 8064 | |
163122b7 KT |
8065 | return detached; |
8066 | } | |
8067 | ||
8068 | /* | |
8069 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
8070 | */ | |
8071 | static void attach_task(struct rq *rq, struct task_struct *p) | |
8072 | { | |
5cb9eaa3 | 8073 | lockdep_assert_rq_held(rq); |
163122b7 KT |
8074 | |
8075 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 8076 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
163122b7 KT |
8077 | check_preempt_curr(rq, p, 0); |
8078 | } | |
8079 | ||
8080 | /* | |
8081 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
8082 | * its new rq. | |
8083 | */ | |
8084 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
8085 | { | |
8a8c69c3 PZ |
8086 | struct rq_flags rf; |
8087 | ||
8088 | rq_lock(rq, &rf); | |
5704ac0a | 8089 | update_rq_clock(rq); |
163122b7 | 8090 | attach_task(rq, p); |
8a8c69c3 | 8091 | rq_unlock(rq, &rf); |
163122b7 KT |
8092 | } |
8093 | ||
8094 | /* | |
8095 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
8096 | * new rq. | |
8097 | */ | |
8098 | static void attach_tasks(struct lb_env *env) | |
8099 | { | |
8100 | struct list_head *tasks = &env->tasks; | |
8101 | struct task_struct *p; | |
8a8c69c3 | 8102 | struct rq_flags rf; |
163122b7 | 8103 | |
8a8c69c3 | 8104 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 8105 | update_rq_clock(env->dst_rq); |
163122b7 KT |
8106 | |
8107 | while (!list_empty(tasks)) { | |
8108 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
8109 | list_del_init(&p->se.group_node); | |
1e3c88bd | 8110 | |
163122b7 KT |
8111 | attach_task(env->dst_rq, p); |
8112 | } | |
8113 | ||
8a8c69c3 | 8114 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
8115 | } |
8116 | ||
b0c79224 | 8117 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
8118 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
8119 | { | |
8120 | if (cfs_rq->avg.load_avg) | |
8121 | return true; | |
8122 | ||
8123 | if (cfs_rq->avg.util_avg) | |
8124 | return true; | |
8125 | ||
8126 | return false; | |
8127 | } | |
8128 | ||
91c27493 | 8129 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
8130 | { |
8131 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
8132 | return true; | |
8133 | ||
3727e0e1 VG |
8134 | if (READ_ONCE(rq->avg_dl.util_avg)) |
8135 | return true; | |
8136 | ||
b4eccf5f TG |
8137 | if (thermal_load_avg(rq)) |
8138 | return true; | |
8139 | ||
11d4afd4 | 8140 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
8141 | if (READ_ONCE(rq->avg_irq.util_avg)) |
8142 | return true; | |
8143 | #endif | |
8144 | ||
371bf427 VG |
8145 | return false; |
8146 | } | |
8147 | ||
39b6a429 | 8148 | static inline void update_blocked_load_tick(struct rq *rq) |
b0c79224 | 8149 | { |
39b6a429 VG |
8150 | WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); |
8151 | } | |
b0c79224 | 8152 | |
39b6a429 VG |
8153 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
8154 | { | |
b0c79224 VS |
8155 | if (!has_blocked) |
8156 | rq->has_blocked_load = 0; | |
8157 | } | |
8158 | #else | |
8159 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
8160 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
39b6a429 | 8161 | static inline void update_blocked_load_tick(struct rq *rq) {} |
b0c79224 VS |
8162 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} |
8163 | #endif | |
8164 | ||
bef69dd8 VG |
8165 | static bool __update_blocked_others(struct rq *rq, bool *done) |
8166 | { | |
8167 | const struct sched_class *curr_class; | |
8168 | u64 now = rq_clock_pelt(rq); | |
b4eccf5f | 8169 | unsigned long thermal_pressure; |
bef69dd8 VG |
8170 | bool decayed; |
8171 | ||
8172 | /* | |
8173 | * update_load_avg() can call cpufreq_update_util(). Make sure that RT, | |
8174 | * DL and IRQ signals have been updated before updating CFS. | |
8175 | */ | |
8176 | curr_class = rq->curr->sched_class; | |
8177 | ||
b4eccf5f TG |
8178 | thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); |
8179 | ||
bef69dd8 VG |
8180 | decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | |
8181 | update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | | |
05289b90 | 8182 | update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | |
bef69dd8 VG |
8183 | update_irq_load_avg(rq, 0); |
8184 | ||
8185 | if (others_have_blocked(rq)) | |
8186 | *done = false; | |
8187 | ||
8188 | return decayed; | |
8189 | } | |
8190 | ||
1936c53c VG |
8191 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8192 | ||
bef69dd8 | 8193 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8194 | { |
039ae8bc | 8195 | struct cfs_rq *cfs_rq, *pos; |
bef69dd8 VG |
8196 | bool decayed = false; |
8197 | int cpu = cpu_of(rq); | |
b90f7c9d | 8198 | |
9763b67f PZ |
8199 | /* |
8200 | * Iterates the task_group tree in a bottom up fashion, see | |
8201 | * list_add_leaf_cfs_rq() for details. | |
8202 | */ | |
039ae8bc | 8203 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
8204 | struct sched_entity *se; |
8205 | ||
bef69dd8 | 8206 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { |
fe749158 | 8207 | update_tg_load_avg(cfs_rq); |
4e516076 | 8208 | |
bef69dd8 VG |
8209 | if (cfs_rq == &rq->cfs) |
8210 | decayed = true; | |
8211 | } | |
8212 | ||
bc427898 VG |
8213 | /* Propagate pending load changes to the parent, if any: */ |
8214 | se = cfs_rq->tg->se[cpu]; | |
8215 | if (se && !skip_blocked_update(se)) | |
02da26ad | 8216 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
a9e7f654 | 8217 | |
039ae8bc VG |
8218 | /* |
8219 | * There can be a lot of idle CPU cgroups. Don't let fully | |
8220 | * decayed cfs_rqs linger on the list. | |
8221 | */ | |
8222 | if (cfs_rq_is_decayed(cfs_rq)) | |
8223 | list_del_leaf_cfs_rq(cfs_rq); | |
8224 | ||
1936c53c VG |
8225 | /* Don't need periodic decay once load/util_avg are null */ |
8226 | if (cfs_rq_has_blocked(cfs_rq)) | |
bef69dd8 | 8227 | *done = false; |
9d89c257 | 8228 | } |
12b04875 | 8229 | |
bef69dd8 | 8230 | return decayed; |
9e3081ca PZ |
8231 | } |
8232 | ||
9763b67f | 8233 | /* |
68520796 | 8234 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
8235 | * This needs to be done in a top-down fashion because the load of a child |
8236 | * group is a fraction of its parents load. | |
8237 | */ | |
68520796 | 8238 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 8239 | { |
68520796 VD |
8240 | struct rq *rq = rq_of(cfs_rq); |
8241 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 8242 | unsigned long now = jiffies; |
68520796 | 8243 | unsigned long load; |
a35b6466 | 8244 | |
68520796 | 8245 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
8246 | return; |
8247 | ||
0e9f0245 | 8248 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
8249 | for_each_sched_entity(se) { |
8250 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 8251 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
8252 | if (cfs_rq->last_h_load_update == now) |
8253 | break; | |
8254 | } | |
a35b6466 | 8255 | |
68520796 | 8256 | if (!se) { |
7ea241af | 8257 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
8258 | cfs_rq->last_h_load_update = now; |
8259 | } | |
8260 | ||
0e9f0245 | 8261 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 8262 | load = cfs_rq->h_load; |
7ea241af YD |
8263 | load = div64_ul(load * se->avg.load_avg, |
8264 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
8265 | cfs_rq = group_cfs_rq(se); |
8266 | cfs_rq->h_load = load; | |
8267 | cfs_rq->last_h_load_update = now; | |
8268 | } | |
9763b67f PZ |
8269 | } |
8270 | ||
367456c7 | 8271 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 8272 | { |
367456c7 | 8273 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 8274 | |
68520796 | 8275 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 8276 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 8277 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
8278 | } |
8279 | #else | |
bef69dd8 | 8280 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8281 | { |
6c1d47c0 | 8282 | struct cfs_rq *cfs_rq = &rq->cfs; |
bef69dd8 | 8283 | bool decayed; |
b90f7c9d | 8284 | |
bef69dd8 VG |
8285 | decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
8286 | if (cfs_rq_has_blocked(cfs_rq)) | |
8287 | *done = false; | |
b90f7c9d | 8288 | |
bef69dd8 | 8289 | return decayed; |
9e3081ca PZ |
8290 | } |
8291 | ||
367456c7 | 8292 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 8293 | { |
9d89c257 | 8294 | return p->se.avg.load_avg; |
1e3c88bd | 8295 | } |
230059de | 8296 | #endif |
1e3c88bd | 8297 | |
bef69dd8 VG |
8298 | static void update_blocked_averages(int cpu) |
8299 | { | |
8300 | bool decayed = false, done = true; | |
8301 | struct rq *rq = cpu_rq(cpu); | |
8302 | struct rq_flags rf; | |
8303 | ||
8304 | rq_lock_irqsave(rq, &rf); | |
39b6a429 | 8305 | update_blocked_load_tick(rq); |
bef69dd8 VG |
8306 | update_rq_clock(rq); |
8307 | ||
8308 | decayed |= __update_blocked_others(rq, &done); | |
8309 | decayed |= __update_blocked_fair(rq, &done); | |
8310 | ||
8311 | update_blocked_load_status(rq, !done); | |
8312 | if (decayed) | |
8313 | cpufreq_update_util(rq, 0); | |
8314 | rq_unlock_irqrestore(rq, &rf); | |
8315 | } | |
8316 | ||
1e3c88bd | 8317 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 8318 | |
1e3c88bd PZ |
8319 | /* |
8320 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
8321 | */ | |
8322 | struct sg_lb_stats { | |
8323 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
8324 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
63b2ca30 | 8325 | unsigned long group_capacity; |
070f5e86 VG |
8326 | unsigned long group_util; /* Total utilization over the CPUs of the group */ |
8327 | unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ | |
5e23e474 | 8328 | unsigned int sum_nr_running; /* Nr of tasks running in the group */ |
a3498347 | 8329 | unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ |
147c5fc2 PZ |
8330 | unsigned int idle_cpus; |
8331 | unsigned int group_weight; | |
caeb178c | 8332 | enum group_type group_type; |
490ba971 | 8333 | unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ |
3b1baa64 | 8334 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
8335 | #ifdef CONFIG_NUMA_BALANCING |
8336 | unsigned int nr_numa_running; | |
8337 | unsigned int nr_preferred_running; | |
8338 | #endif | |
1e3c88bd PZ |
8339 | }; |
8340 | ||
56cf515b JK |
8341 | /* |
8342 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
8343 | * during load balancing. | |
8344 | */ | |
8345 | struct sd_lb_stats { | |
8346 | struct sched_group *busiest; /* Busiest group in this sd */ | |
8347 | struct sched_group *local; /* Local group in this sd */ | |
8348 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 8349 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b | 8350 | unsigned long avg_load; /* Average load across all groups in sd */ |
0b0695f2 | 8351 | unsigned int prefer_sibling; /* tasks should go to sibling first */ |
56cf515b | 8352 | |
56cf515b | 8353 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 8354 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
8355 | }; |
8356 | ||
147c5fc2 PZ |
8357 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
8358 | { | |
8359 | /* | |
8360 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
8361 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
0b0695f2 VG |
8362 | * We must however set busiest_stat::group_type and |
8363 | * busiest_stat::idle_cpus to the worst busiest group because | |
8364 | * update_sd_pick_busiest() reads these before assignment. | |
147c5fc2 PZ |
8365 | */ |
8366 | *sds = (struct sd_lb_stats){ | |
8367 | .busiest = NULL, | |
8368 | .local = NULL, | |
8369 | .total_load = 0UL, | |
63b2ca30 | 8370 | .total_capacity = 0UL, |
147c5fc2 | 8371 | .busiest_stat = { |
0b0695f2 VG |
8372 | .idle_cpus = UINT_MAX, |
8373 | .group_type = group_has_spare, | |
147c5fc2 PZ |
8374 | }, |
8375 | }; | |
8376 | } | |
8377 | ||
1ca2034e | 8378 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
8379 | { |
8380 | struct rq *rq = cpu_rq(cpu); | |
8ec59c0f | 8381 | unsigned long max = arch_scale_cpu_capacity(cpu); |
523e979d | 8382 | unsigned long used, free; |
523e979d | 8383 | unsigned long irq; |
b654f7de | 8384 | |
2e62c474 | 8385 | irq = cpu_util_irq(rq); |
cadefd3d | 8386 | |
523e979d VG |
8387 | if (unlikely(irq >= max)) |
8388 | return 1; | |
aa483808 | 8389 | |
467b7d01 TG |
8390 | /* |
8391 | * avg_rt.util_avg and avg_dl.util_avg track binary signals | |
8392 | * (running and not running) with weights 0 and 1024 respectively. | |
8393 | * avg_thermal.load_avg tracks thermal pressure and the weighted | |
8394 | * average uses the actual delta max capacity(load). | |
8395 | */ | |
523e979d VG |
8396 | used = READ_ONCE(rq->avg_rt.util_avg); |
8397 | used += READ_ONCE(rq->avg_dl.util_avg); | |
467b7d01 | 8398 | used += thermal_load_avg(rq); |
1e3c88bd | 8399 | |
523e979d VG |
8400 | if (unlikely(used >= max)) |
8401 | return 1; | |
1e3c88bd | 8402 | |
523e979d | 8403 | free = max - used; |
2e62c474 VG |
8404 | |
8405 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
8406 | } |
8407 | ||
ced549fa | 8408 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 8409 | { |
1ca2034e | 8410 | unsigned long capacity = scale_rt_capacity(cpu); |
1e3c88bd PZ |
8411 | struct sched_group *sdg = sd->groups; |
8412 | ||
8ec59c0f | 8413 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); |
1e3c88bd | 8414 | |
ced549fa NP |
8415 | if (!capacity) |
8416 | capacity = 1; | |
1e3c88bd | 8417 | |
ced549fa | 8418 | cpu_rq(cpu)->cpu_capacity = capacity; |
51cf18c9 VD |
8419 | trace_sched_cpu_capacity_tp(cpu_rq(cpu)); |
8420 | ||
ced549fa | 8421 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8422 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 8423 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
8424 | } |
8425 | ||
63b2ca30 | 8426 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
8427 | { |
8428 | struct sched_domain *child = sd->child; | |
8429 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 8430 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
8431 | unsigned long interval; |
8432 | ||
8433 | interval = msecs_to_jiffies(sd->balance_interval); | |
8434 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 8435 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
8436 | |
8437 | if (!child) { | |
ced549fa | 8438 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
8439 | return; |
8440 | } | |
8441 | ||
dc7ff76e | 8442 | capacity = 0; |
bf475ce0 | 8443 | min_capacity = ULONG_MAX; |
e3d6d0cb | 8444 | max_capacity = 0; |
1e3c88bd | 8445 | |
74a5ce20 PZ |
8446 | if (child->flags & SD_OVERLAP) { |
8447 | /* | |
8448 | * SD_OVERLAP domains cannot assume that child groups | |
8449 | * span the current group. | |
8450 | */ | |
8451 | ||
ae4df9d6 | 8452 | for_each_cpu(cpu, sched_group_span(sdg)) { |
4c58f57f | 8453 | unsigned long cpu_cap = capacity_of(cpu); |
863bffc8 | 8454 | |
4c58f57f PL |
8455 | capacity += cpu_cap; |
8456 | min_capacity = min(cpu_cap, min_capacity); | |
8457 | max_capacity = max(cpu_cap, max_capacity); | |
863bffc8 | 8458 | } |
74a5ce20 PZ |
8459 | } else { |
8460 | /* | |
8461 | * !SD_OVERLAP domains can assume that child groups | |
8462 | * span the current group. | |
97a7142f | 8463 | */ |
74a5ce20 PZ |
8464 | |
8465 | group = child->groups; | |
8466 | do { | |
bf475ce0 MR |
8467 | struct sched_group_capacity *sgc = group->sgc; |
8468 | ||
8469 | capacity += sgc->capacity; | |
8470 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 8471 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
8472 | group = group->next; |
8473 | } while (group != child->groups); | |
8474 | } | |
1e3c88bd | 8475 | |
63b2ca30 | 8476 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8477 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 8478 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
8479 | } |
8480 | ||
9d5efe05 | 8481 | /* |
ea67821b VG |
8482 | * Check whether the capacity of the rq has been noticeably reduced by side |
8483 | * activity. The imbalance_pct is used for the threshold. | |
8484 | * Return true is the capacity is reduced | |
9d5efe05 SV |
8485 | */ |
8486 | static inline int | |
ea67821b | 8487 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 8488 | { |
ea67821b VG |
8489 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
8490 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
8491 | } |
8492 | ||
a0fe2cf0 VS |
8493 | /* |
8494 | * Check whether a rq has a misfit task and if it looks like we can actually | |
8495 | * help that task: we can migrate the task to a CPU of higher capacity, or | |
8496 | * the task's current CPU is heavily pressured. | |
8497 | */ | |
8498 | static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) | |
8499 | { | |
8500 | return rq->misfit_task_load && | |
8501 | (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || | |
8502 | check_cpu_capacity(rq, sd)); | |
8503 | } | |
8504 | ||
30ce5dab PZ |
8505 | /* |
8506 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 8507 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 8508 | * |
97fb7a0a IM |
8509 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
8510 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
8511 | * Something like: |
8512 | * | |
2b4d5b25 IM |
8513 | * { 0 1 2 3 } { 4 5 6 7 } |
8514 | * * * * * | |
30ce5dab PZ |
8515 | * |
8516 | * If we were to balance group-wise we'd place two tasks in the first group and | |
8517 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 8518 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
8519 | * |
8520 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
8521 | * by noticing the lower domain failed to reach balance and had difficulty |
8522 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
8523 | * |
8524 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 8525 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 8526 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
8527 | * to create an effective group imbalance. |
8528 | * | |
8529 | * This is a somewhat tricky proposition since the next run might not find the | |
8530 | * group imbalance and decide the groups need to be balanced again. A most | |
8531 | * subtle and fragile situation. | |
8532 | */ | |
8533 | ||
6263322c | 8534 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 8535 | { |
63b2ca30 | 8536 | return group->sgc->imbalance; |
30ce5dab PZ |
8537 | } |
8538 | ||
b37d9316 | 8539 | /* |
ea67821b VG |
8540 | * group_has_capacity returns true if the group has spare capacity that could |
8541 | * be used by some tasks. | |
8542 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
8543 | * smaller than the number of CPUs or if the utilization is lower than the |
8544 | * available capacity for CFS tasks. | |
ea67821b VG |
8545 | * For the latter, we use a threshold to stabilize the state, to take into |
8546 | * account the variance of the tasks' load and to return true if the available | |
8547 | * capacity in meaningful for the load balancer. | |
8548 | * As an example, an available capacity of 1% can appear but it doesn't make | |
8549 | * any benefit for the load balance. | |
b37d9316 | 8550 | */ |
ea67821b | 8551 | static inline bool |
57abff06 | 8552 | group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
b37d9316 | 8553 | { |
5e23e474 | 8554 | if (sgs->sum_nr_running < sgs->group_weight) |
ea67821b | 8555 | return true; |
c61037e9 | 8556 | |
070f5e86 VG |
8557 | if ((sgs->group_capacity * imbalance_pct) < |
8558 | (sgs->group_runnable * 100)) | |
8559 | return false; | |
8560 | ||
ea67821b | 8561 | if ((sgs->group_capacity * 100) > |
57abff06 | 8562 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8563 | return true; |
b37d9316 | 8564 | |
ea67821b VG |
8565 | return false; |
8566 | } | |
8567 | ||
8568 | /* | |
8569 | * group_is_overloaded returns true if the group has more tasks than it can | |
8570 | * handle. | |
8571 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
8572 | * with the exact right number of tasks, has no more spare capacity but is not | |
8573 | * overloaded so both group_has_capacity and group_is_overloaded return | |
8574 | * false. | |
8575 | */ | |
8576 | static inline bool | |
57abff06 | 8577 | group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
ea67821b | 8578 | { |
5e23e474 | 8579 | if (sgs->sum_nr_running <= sgs->group_weight) |
ea67821b | 8580 | return false; |
b37d9316 | 8581 | |
ea67821b | 8582 | if ((sgs->group_capacity * 100) < |
57abff06 | 8583 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8584 | return true; |
b37d9316 | 8585 | |
070f5e86 VG |
8586 | if ((sgs->group_capacity * imbalance_pct) < |
8587 | (sgs->group_runnable * 100)) | |
8588 | return true; | |
8589 | ||
ea67821b | 8590 | return false; |
b37d9316 PZ |
8591 | } |
8592 | ||
79a89f92 | 8593 | static inline enum |
57abff06 | 8594 | group_type group_classify(unsigned int imbalance_pct, |
0b0695f2 | 8595 | struct sched_group *group, |
79a89f92 | 8596 | struct sg_lb_stats *sgs) |
caeb178c | 8597 | { |
57abff06 | 8598 | if (group_is_overloaded(imbalance_pct, sgs)) |
caeb178c RR |
8599 | return group_overloaded; |
8600 | ||
8601 | if (sg_imbalanced(group)) | |
8602 | return group_imbalanced; | |
8603 | ||
0b0695f2 VG |
8604 | if (sgs->group_asym_packing) |
8605 | return group_asym_packing; | |
8606 | ||
3b1baa64 MR |
8607 | if (sgs->group_misfit_task_load) |
8608 | return group_misfit_task; | |
8609 | ||
57abff06 | 8610 | if (!group_has_capacity(imbalance_pct, sgs)) |
0b0695f2 VG |
8611 | return group_fully_busy; |
8612 | ||
8613 | return group_has_spare; | |
caeb178c RR |
8614 | } |
8615 | ||
512a6f55 RN |
8616 | static inline bool |
8617 | sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs, | |
8618 | struct sched_group *group) | |
8619 | { | |
8620 | return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu); | |
8621 | } | |
8622 | ||
1e3c88bd PZ |
8623 | /** |
8624 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 8625 | * @env: The load balancing environment. |
1e3c88bd | 8626 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 8627 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 8628 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 8629 | */ |
bd939f45 | 8630 | static inline void update_sg_lb_stats(struct lb_env *env, |
4304cb29 | 8631 | struct sd_lb_stats *sds, |
630246a0 QP |
8632 | struct sched_group *group, |
8633 | struct sg_lb_stats *sgs, | |
8634 | int *sg_status) | |
1e3c88bd | 8635 | { |
0b0695f2 | 8636 | int i, nr_running, local_group; |
1e3c88bd | 8637 | |
b72ff13c PZ |
8638 | memset(sgs, 0, sizeof(*sgs)); |
8639 | ||
4304cb29 | 8640 | local_group = group == sds->local; |
0b0695f2 | 8641 | |
ae4df9d6 | 8642 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd PZ |
8643 | struct rq *rq = cpu_rq(i); |
8644 | ||
b0fb1eb4 | 8645 | sgs->group_load += cpu_load(rq); |
9e91d61d | 8646 | sgs->group_util += cpu_util(i); |
070f5e86 | 8647 | sgs->group_runnable += cpu_runnable(rq); |
a3498347 | 8648 | sgs->sum_h_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 8649 | |
a426f99c | 8650 | nr_running = rq->nr_running; |
5e23e474 VG |
8651 | sgs->sum_nr_running += nr_running; |
8652 | ||
a426f99c | 8653 | if (nr_running > 1) |
630246a0 | 8654 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 8655 | |
2802bf3c MR |
8656 | if (cpu_overutilized(i)) |
8657 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 8658 | |
0ec8aa00 PZ |
8659 | #ifdef CONFIG_NUMA_BALANCING |
8660 | sgs->nr_numa_running += rq->nr_numa_running; | |
8661 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
8662 | #endif | |
a426f99c WL |
8663 | /* |
8664 | * No need to call idle_cpu() if nr_running is not 0 | |
8665 | */ | |
0b0695f2 | 8666 | if (!nr_running && idle_cpu(i)) { |
aae6d3dd | 8667 | sgs->idle_cpus++; |
0b0695f2 VG |
8668 | /* Idle cpu can't have misfit task */ |
8669 | continue; | |
8670 | } | |
8671 | ||
8672 | if (local_group) | |
8673 | continue; | |
3b1baa64 | 8674 | |
0b0695f2 | 8675 | /* Check for a misfit task on the cpu */ |
3b1baa64 | 8676 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && |
757ffdd7 | 8677 | sgs->group_misfit_task_load < rq->misfit_task_load) { |
3b1baa64 | 8678 | sgs->group_misfit_task_load = rq->misfit_task_load; |
630246a0 | 8679 | *sg_status |= SG_OVERLOAD; |
757ffdd7 | 8680 | } |
1e3c88bd PZ |
8681 | } |
8682 | ||
512a6f55 RN |
8683 | sgs->group_capacity = group->sgc->capacity; |
8684 | ||
8685 | sgs->group_weight = group->group_weight; | |
8686 | ||
0b0695f2 | 8687 | /* Check if dst CPU is idle and preferred to this group */ |
ddcc40c2 | 8688 | if (!local_group && env->sd->flags & SD_ASYM_PACKING && |
512a6f55 RN |
8689 | env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running && |
8690 | sched_asym(env, sds, sgs, group)) { | |
0b0695f2 VG |
8691 | sgs->group_asym_packing = 1; |
8692 | } | |
8693 | ||
57abff06 | 8694 | sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); |
0b0695f2 VG |
8695 | |
8696 | /* Computing avg_load makes sense only when group is overloaded */ | |
8697 | if (sgs->group_type == group_overloaded) | |
8698 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / | |
8699 | sgs->group_capacity; | |
1e3c88bd PZ |
8700 | } |
8701 | ||
532cb4c4 MN |
8702 | /** |
8703 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 8704 | * @env: The load balancing environment. |
532cb4c4 MN |
8705 | * @sds: sched_domain statistics |
8706 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 8707 | * @sgs: sched_group statistics |
532cb4c4 MN |
8708 | * |
8709 | * Determine if @sg is a busier group than the previously selected | |
8710 | * busiest group. | |
e69f6186 YB |
8711 | * |
8712 | * Return: %true if @sg is a busier group than the previously selected | |
8713 | * busiest group. %false otherwise. | |
532cb4c4 | 8714 | */ |
bd939f45 | 8715 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
8716 | struct sd_lb_stats *sds, |
8717 | struct sched_group *sg, | |
bd939f45 | 8718 | struct sg_lb_stats *sgs) |
532cb4c4 | 8719 | { |
caeb178c | 8720 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 8721 | |
0b0695f2 VG |
8722 | /* Make sure that there is at least one task to pull */ |
8723 | if (!sgs->sum_h_nr_running) | |
8724 | return false; | |
8725 | ||
cad68e55 MR |
8726 | /* |
8727 | * Don't try to pull misfit tasks we can't help. | |
8728 | * We can use max_capacity here as reduction in capacity on some | |
8729 | * CPUs in the group should either be possible to resolve | |
8730 | * internally or be covered by avg_load imbalance (eventually). | |
8731 | */ | |
8732 | if (sgs->group_type == group_misfit_task && | |
4aed8aa4 | 8733 | (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || |
0b0695f2 | 8734 | sds->local_stat.group_type != group_has_spare)) |
cad68e55 MR |
8735 | return false; |
8736 | ||
caeb178c | 8737 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
8738 | return true; |
8739 | ||
caeb178c RR |
8740 | if (sgs->group_type < busiest->group_type) |
8741 | return false; | |
8742 | ||
9e0994c0 | 8743 | /* |
0b0695f2 VG |
8744 | * The candidate and the current busiest group are the same type of |
8745 | * group. Let check which one is the busiest according to the type. | |
9e0994c0 | 8746 | */ |
9e0994c0 | 8747 | |
0b0695f2 VG |
8748 | switch (sgs->group_type) { |
8749 | case group_overloaded: | |
8750 | /* Select the overloaded group with highest avg_load. */ | |
8751 | if (sgs->avg_load <= busiest->avg_load) | |
8752 | return false; | |
8753 | break; | |
8754 | ||
8755 | case group_imbalanced: | |
8756 | /* | |
8757 | * Select the 1st imbalanced group as we don't have any way to | |
8758 | * choose one more than another. | |
8759 | */ | |
9e0994c0 MR |
8760 | return false; |
8761 | ||
0b0695f2 VG |
8762 | case group_asym_packing: |
8763 | /* Prefer to move from lowest priority CPU's work */ | |
8764 | if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) | |
8765 | return false; | |
8766 | break; | |
532cb4c4 | 8767 | |
0b0695f2 VG |
8768 | case group_misfit_task: |
8769 | /* | |
8770 | * If we have more than one misfit sg go with the biggest | |
8771 | * misfit. | |
8772 | */ | |
8773 | if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
8774 | return false; | |
8775 | break; | |
532cb4c4 | 8776 | |
0b0695f2 VG |
8777 | case group_fully_busy: |
8778 | /* | |
8779 | * Select the fully busy group with highest avg_load. In | |
8780 | * theory, there is no need to pull task from such kind of | |
8781 | * group because tasks have all compute capacity that they need | |
8782 | * but we can still improve the overall throughput by reducing | |
8783 | * contention when accessing shared HW resources. | |
8784 | * | |
8785 | * XXX for now avg_load is not computed and always 0 so we | |
8786 | * select the 1st one. | |
8787 | */ | |
8788 | if (sgs->avg_load <= busiest->avg_load) | |
8789 | return false; | |
8790 | break; | |
8791 | ||
8792 | case group_has_spare: | |
8793 | /* | |
5f68eb19 VG |
8794 | * Select not overloaded group with lowest number of idle cpus |
8795 | * and highest number of running tasks. We could also compare | |
8796 | * the spare capacity which is more stable but it can end up | |
8797 | * that the group has less spare capacity but finally more idle | |
0b0695f2 VG |
8798 | * CPUs which means less opportunity to pull tasks. |
8799 | */ | |
5f68eb19 | 8800 | if (sgs->idle_cpus > busiest->idle_cpus) |
0b0695f2 | 8801 | return false; |
5f68eb19 VG |
8802 | else if ((sgs->idle_cpus == busiest->idle_cpus) && |
8803 | (sgs->sum_nr_running <= busiest->sum_nr_running)) | |
8804 | return false; | |
8805 | ||
0b0695f2 | 8806 | break; |
532cb4c4 MN |
8807 | } |
8808 | ||
0b0695f2 VG |
8809 | /* |
8810 | * Candidate sg has no more than one task per CPU and has higher | |
8811 | * per-CPU capacity. Migrating tasks to less capable CPUs may harm | |
8812 | * throughput. Maximize throughput, power/energy consequences are not | |
8813 | * considered. | |
8814 | */ | |
8815 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && | |
8816 | (sgs->group_type <= group_fully_busy) && | |
4aed8aa4 | 8817 | (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) |
0b0695f2 VG |
8818 | return false; |
8819 | ||
8820 | return true; | |
532cb4c4 MN |
8821 | } |
8822 | ||
0ec8aa00 PZ |
8823 | #ifdef CONFIG_NUMA_BALANCING |
8824 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8825 | { | |
a3498347 | 8826 | if (sgs->sum_h_nr_running > sgs->nr_numa_running) |
0ec8aa00 | 8827 | return regular; |
a3498347 | 8828 | if (sgs->sum_h_nr_running > sgs->nr_preferred_running) |
0ec8aa00 PZ |
8829 | return remote; |
8830 | return all; | |
8831 | } | |
8832 | ||
8833 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8834 | { | |
8835 | if (rq->nr_running > rq->nr_numa_running) | |
8836 | return regular; | |
8837 | if (rq->nr_running > rq->nr_preferred_running) | |
8838 | return remote; | |
8839 | return all; | |
8840 | } | |
8841 | #else | |
8842 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8843 | { | |
8844 | return all; | |
8845 | } | |
8846 | ||
8847 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8848 | { | |
8849 | return regular; | |
8850 | } | |
8851 | #endif /* CONFIG_NUMA_BALANCING */ | |
8852 | ||
57abff06 VG |
8853 | |
8854 | struct sg_lb_stats; | |
8855 | ||
3318544b VG |
8856 | /* |
8857 | * task_running_on_cpu - return 1 if @p is running on @cpu. | |
8858 | */ | |
8859 | ||
8860 | static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) | |
8861 | { | |
8862 | /* Task has no contribution or is new */ | |
8863 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
8864 | return 0; | |
8865 | ||
8866 | if (task_on_rq_queued(p)) | |
8867 | return 1; | |
8868 | ||
8869 | return 0; | |
8870 | } | |
8871 | ||
8872 | /** | |
8873 | * idle_cpu_without - would a given CPU be idle without p ? | |
8874 | * @cpu: the processor on which idleness is tested. | |
8875 | * @p: task which should be ignored. | |
8876 | * | |
8877 | * Return: 1 if the CPU would be idle. 0 otherwise. | |
8878 | */ | |
8879 | static int idle_cpu_without(int cpu, struct task_struct *p) | |
8880 | { | |
8881 | struct rq *rq = cpu_rq(cpu); | |
8882 | ||
8883 | if (rq->curr != rq->idle && rq->curr != p) | |
8884 | return 0; | |
8885 | ||
8886 | /* | |
8887 | * rq->nr_running can't be used but an updated version without the | |
8888 | * impact of p on cpu must be used instead. The updated nr_running | |
8889 | * be computed and tested before calling idle_cpu_without(). | |
8890 | */ | |
8891 | ||
8892 | #ifdef CONFIG_SMP | |
126c2092 | 8893 | if (rq->ttwu_pending) |
3318544b VG |
8894 | return 0; |
8895 | #endif | |
8896 | ||
8897 | return 1; | |
8898 | } | |
8899 | ||
57abff06 VG |
8900 | /* |
8901 | * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. | |
3318544b | 8902 | * @sd: The sched_domain level to look for idlest group. |
57abff06 VG |
8903 | * @group: sched_group whose statistics are to be updated. |
8904 | * @sgs: variable to hold the statistics for this group. | |
3318544b | 8905 | * @p: The task for which we look for the idlest group/CPU. |
57abff06 VG |
8906 | */ |
8907 | static inline void update_sg_wakeup_stats(struct sched_domain *sd, | |
8908 | struct sched_group *group, | |
8909 | struct sg_lb_stats *sgs, | |
8910 | struct task_struct *p) | |
8911 | { | |
8912 | int i, nr_running; | |
8913 | ||
8914 | memset(sgs, 0, sizeof(*sgs)); | |
8915 | ||
8916 | for_each_cpu(i, sched_group_span(group)) { | |
8917 | struct rq *rq = cpu_rq(i); | |
3318544b | 8918 | unsigned int local; |
57abff06 | 8919 | |
3318544b | 8920 | sgs->group_load += cpu_load_without(rq, p); |
57abff06 | 8921 | sgs->group_util += cpu_util_without(i, p); |
070f5e86 | 8922 | sgs->group_runnable += cpu_runnable_without(rq, p); |
3318544b VG |
8923 | local = task_running_on_cpu(i, p); |
8924 | sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; | |
57abff06 | 8925 | |
3318544b | 8926 | nr_running = rq->nr_running - local; |
57abff06 VG |
8927 | sgs->sum_nr_running += nr_running; |
8928 | ||
8929 | /* | |
3318544b | 8930 | * No need to call idle_cpu_without() if nr_running is not 0 |
57abff06 | 8931 | */ |
3318544b | 8932 | if (!nr_running && idle_cpu_without(i, p)) |
57abff06 VG |
8933 | sgs->idle_cpus++; |
8934 | ||
57abff06 VG |
8935 | } |
8936 | ||
8937 | /* Check if task fits in the group */ | |
8938 | if (sd->flags & SD_ASYM_CPUCAPACITY && | |
8939 | !task_fits_capacity(p, group->sgc->max_capacity)) { | |
8940 | sgs->group_misfit_task_load = 1; | |
8941 | } | |
8942 | ||
8943 | sgs->group_capacity = group->sgc->capacity; | |
8944 | ||
289de359 VG |
8945 | sgs->group_weight = group->group_weight; |
8946 | ||
57abff06 VG |
8947 | sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); |
8948 | ||
8949 | /* | |
8950 | * Computing avg_load makes sense only when group is fully busy or | |
8951 | * overloaded | |
8952 | */ | |
6c8116c9 TZ |
8953 | if (sgs->group_type == group_fully_busy || |
8954 | sgs->group_type == group_overloaded) | |
57abff06 VG |
8955 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / |
8956 | sgs->group_capacity; | |
8957 | } | |
8958 | ||
8959 | static bool update_pick_idlest(struct sched_group *idlest, | |
8960 | struct sg_lb_stats *idlest_sgs, | |
8961 | struct sched_group *group, | |
8962 | struct sg_lb_stats *sgs) | |
8963 | { | |
8964 | if (sgs->group_type < idlest_sgs->group_type) | |
8965 | return true; | |
8966 | ||
8967 | if (sgs->group_type > idlest_sgs->group_type) | |
8968 | return false; | |
8969 | ||
8970 | /* | |
8971 | * The candidate and the current idlest group are the same type of | |
8972 | * group. Let check which one is the idlest according to the type. | |
8973 | */ | |
8974 | ||
8975 | switch (sgs->group_type) { | |
8976 | case group_overloaded: | |
8977 | case group_fully_busy: | |
8978 | /* Select the group with lowest avg_load. */ | |
8979 | if (idlest_sgs->avg_load <= sgs->avg_load) | |
8980 | return false; | |
8981 | break; | |
8982 | ||
8983 | case group_imbalanced: | |
8984 | case group_asym_packing: | |
8985 | /* Those types are not used in the slow wakeup path */ | |
8986 | return false; | |
8987 | ||
8988 | case group_misfit_task: | |
8989 | /* Select group with the highest max capacity */ | |
8990 | if (idlest->sgc->max_capacity >= group->sgc->max_capacity) | |
8991 | return false; | |
8992 | break; | |
8993 | ||
8994 | case group_has_spare: | |
8995 | /* Select group with most idle CPUs */ | |
3edecfef | 8996 | if (idlest_sgs->idle_cpus > sgs->idle_cpus) |
57abff06 | 8997 | return false; |
3edecfef PP |
8998 | |
8999 | /* Select group with lowest group_util */ | |
9000 | if (idlest_sgs->idle_cpus == sgs->idle_cpus && | |
9001 | idlest_sgs->group_util <= sgs->group_util) | |
9002 | return false; | |
9003 | ||
57abff06 VG |
9004 | break; |
9005 | } | |
9006 | ||
9007 | return true; | |
9008 | } | |
9009 | ||
23e6082a MG |
9010 | /* |
9011 | * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain. | |
9012 | * This is an approximation as the number of running tasks may not be | |
9013 | * related to the number of busy CPUs due to sched_setaffinity. | |
9014 | */ | |
9015 | static inline bool allow_numa_imbalance(int dst_running, int dst_weight) | |
9016 | { | |
9017 | return (dst_running < (dst_weight >> 2)); | |
9018 | } | |
9019 | ||
57abff06 VG |
9020 | /* |
9021 | * find_idlest_group() finds and returns the least busy CPU group within the | |
9022 | * domain. | |
9023 | * | |
9024 | * Assumes p is allowed on at least one CPU in sd. | |
9025 | */ | |
9026 | static struct sched_group * | |
45da2773 | 9027 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
57abff06 VG |
9028 | { |
9029 | struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; | |
9030 | struct sg_lb_stats local_sgs, tmp_sgs; | |
9031 | struct sg_lb_stats *sgs; | |
9032 | unsigned long imbalance; | |
9033 | struct sg_lb_stats idlest_sgs = { | |
9034 | .avg_load = UINT_MAX, | |
9035 | .group_type = group_overloaded, | |
9036 | }; | |
9037 | ||
57abff06 VG |
9038 | do { |
9039 | int local_group; | |
9040 | ||
9041 | /* Skip over this group if it has no CPUs allowed */ | |
9042 | if (!cpumask_intersects(sched_group_span(group), | |
9043 | p->cpus_ptr)) | |
9044 | continue; | |
9045 | ||
97886d9d AL |
9046 | /* Skip over this group if no cookie matched */ |
9047 | if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group)) | |
9048 | continue; | |
9049 | ||
57abff06 VG |
9050 | local_group = cpumask_test_cpu(this_cpu, |
9051 | sched_group_span(group)); | |
9052 | ||
9053 | if (local_group) { | |
9054 | sgs = &local_sgs; | |
9055 | local = group; | |
9056 | } else { | |
9057 | sgs = &tmp_sgs; | |
9058 | } | |
9059 | ||
9060 | update_sg_wakeup_stats(sd, group, sgs, p); | |
9061 | ||
9062 | if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { | |
9063 | idlest = group; | |
9064 | idlest_sgs = *sgs; | |
9065 | } | |
9066 | ||
9067 | } while (group = group->next, group != sd->groups); | |
9068 | ||
9069 | ||
9070 | /* There is no idlest group to push tasks to */ | |
9071 | if (!idlest) | |
9072 | return NULL; | |
9073 | ||
7ed735c3 VG |
9074 | /* The local group has been skipped because of CPU affinity */ |
9075 | if (!local) | |
9076 | return idlest; | |
9077 | ||
57abff06 VG |
9078 | /* |
9079 | * If the local group is idler than the selected idlest group | |
9080 | * don't try and push the task. | |
9081 | */ | |
9082 | if (local_sgs.group_type < idlest_sgs.group_type) | |
9083 | return NULL; | |
9084 | ||
9085 | /* | |
9086 | * If the local group is busier than the selected idlest group | |
9087 | * try and push the task. | |
9088 | */ | |
9089 | if (local_sgs.group_type > idlest_sgs.group_type) | |
9090 | return idlest; | |
9091 | ||
9092 | switch (local_sgs.group_type) { | |
9093 | case group_overloaded: | |
9094 | case group_fully_busy: | |
5c339005 MG |
9095 | |
9096 | /* Calculate allowed imbalance based on load */ | |
9097 | imbalance = scale_load_down(NICE_0_LOAD) * | |
9098 | (sd->imbalance_pct-100) / 100; | |
9099 | ||
57abff06 VG |
9100 | /* |
9101 | * When comparing groups across NUMA domains, it's possible for | |
9102 | * the local domain to be very lightly loaded relative to the | |
9103 | * remote domains but "imbalance" skews the comparison making | |
9104 | * remote CPUs look much more favourable. When considering | |
9105 | * cross-domain, add imbalance to the load on the remote node | |
9106 | * and consider staying local. | |
9107 | */ | |
9108 | ||
9109 | if ((sd->flags & SD_NUMA) && | |
9110 | ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) | |
9111 | return NULL; | |
9112 | ||
9113 | /* | |
9114 | * If the local group is less loaded than the selected | |
9115 | * idlest group don't try and push any tasks. | |
9116 | */ | |
9117 | if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) | |
9118 | return NULL; | |
9119 | ||
9120 | if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) | |
9121 | return NULL; | |
9122 | break; | |
9123 | ||
9124 | case group_imbalanced: | |
9125 | case group_asym_packing: | |
9126 | /* Those type are not used in the slow wakeup path */ | |
9127 | return NULL; | |
9128 | ||
9129 | case group_misfit_task: | |
9130 | /* Select group with the highest max capacity */ | |
9131 | if (local->sgc->max_capacity >= idlest->sgc->max_capacity) | |
9132 | return NULL; | |
9133 | break; | |
9134 | ||
9135 | case group_has_spare: | |
9136 | if (sd->flags & SD_NUMA) { | |
9137 | #ifdef CONFIG_NUMA_BALANCING | |
9138 | int idlest_cpu; | |
9139 | /* | |
9140 | * If there is spare capacity at NUMA, try to select | |
9141 | * the preferred node | |
9142 | */ | |
9143 | if (cpu_to_node(this_cpu) == p->numa_preferred_nid) | |
9144 | return NULL; | |
9145 | ||
9146 | idlest_cpu = cpumask_first(sched_group_span(idlest)); | |
9147 | if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) | |
9148 | return idlest; | |
9149 | #endif | |
9150 | /* | |
9151 | * Otherwise, keep the task on this node to stay close | |
9152 | * its wakeup source and improve locality. If there is | |
9153 | * a real need of migration, periodic load balance will | |
9154 | * take care of it. | |
9155 | */ | |
23e6082a | 9156 | if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight)) |
57abff06 VG |
9157 | return NULL; |
9158 | } | |
9159 | ||
9160 | /* | |
9161 | * Select group with highest number of idle CPUs. We could also | |
9162 | * compare the utilization which is more stable but it can end | |
9163 | * up that the group has less spare capacity but finally more | |
9164 | * idle CPUs which means more opportunity to run task. | |
9165 | */ | |
9166 | if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) | |
9167 | return NULL; | |
9168 | break; | |
9169 | } | |
9170 | ||
9171 | return idlest; | |
9172 | } | |
9173 | ||
1e3c88bd | 9174 | /** |
461819ac | 9175 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 9176 | * @env: The load balancing environment. |
1e3c88bd PZ |
9177 | * @sds: variable to hold the statistics for this sched_domain. |
9178 | */ | |
0b0695f2 | 9179 | |
0ec8aa00 | 9180 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9181 | { |
bd939f45 PZ |
9182 | struct sched_domain *child = env->sd->child; |
9183 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 9184 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 9185 | struct sg_lb_stats tmp_sgs; |
630246a0 | 9186 | int sg_status = 0; |
1e3c88bd | 9187 | |
1e3c88bd | 9188 | do { |
56cf515b | 9189 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
9190 | int local_group; |
9191 | ||
ae4df9d6 | 9192 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
9193 | if (local_group) { |
9194 | sds->local = sg; | |
05b40e05 | 9195 | sgs = local; |
b72ff13c PZ |
9196 | |
9197 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
9198 | time_after_eq(jiffies, sg->sgc->next_update)) |
9199 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 9200 | } |
1e3c88bd | 9201 | |
4304cb29 | 9202 | update_sg_lb_stats(env, sds, sg, sgs, &sg_status); |
1e3c88bd | 9203 | |
b72ff13c PZ |
9204 | if (local_group) |
9205 | goto next_group; | |
9206 | ||
1e3c88bd | 9207 | |
b72ff13c | 9208 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 9209 | sds->busiest = sg; |
56cf515b | 9210 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
9211 | } |
9212 | ||
b72ff13c PZ |
9213 | next_group: |
9214 | /* Now, start updating sd_lb_stats */ | |
9215 | sds->total_load += sgs->group_load; | |
63b2ca30 | 9216 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 9217 | |
532cb4c4 | 9218 | sg = sg->next; |
bd939f45 | 9219 | } while (sg != env->sd->groups); |
0ec8aa00 | 9220 | |
0b0695f2 VG |
9221 | /* Tag domain that child domain prefers tasks go to siblings first */ |
9222 | sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; | |
9223 | ||
f643ea22 | 9224 | |
0ec8aa00 PZ |
9225 | if (env->sd->flags & SD_NUMA) |
9226 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
9227 | |
9228 | if (!env->sd->parent) { | |
2802bf3c MR |
9229 | struct root_domain *rd = env->dst_rq->rd; |
9230 | ||
4486edd1 | 9231 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
9232 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
9233 | ||
9234 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
9235 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
f9f240f9 | 9236 | trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); |
2802bf3c | 9237 | } else if (sg_status & SG_OVERUTILIZED) { |
f9f240f9 QY |
9238 | struct root_domain *rd = env->dst_rq->rd; |
9239 | ||
9240 | WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); | |
9241 | trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); | |
4486edd1 | 9242 | } |
532cb4c4 MN |
9243 | } |
9244 | ||
abeae76a MG |
9245 | #define NUMA_IMBALANCE_MIN 2 |
9246 | ||
7d2b5dd0 MG |
9247 | static inline long adjust_numa_imbalance(int imbalance, |
9248 | int dst_running, int dst_weight) | |
fb86f5b2 | 9249 | { |
23e6082a MG |
9250 | if (!allow_numa_imbalance(dst_running, dst_weight)) |
9251 | return imbalance; | |
9252 | ||
fb86f5b2 MG |
9253 | /* |
9254 | * Allow a small imbalance based on a simple pair of communicating | |
7d2b5dd0 | 9255 | * tasks that remain local when the destination is lightly loaded. |
fb86f5b2 | 9256 | */ |
23e6082a | 9257 | if (imbalance <= NUMA_IMBALANCE_MIN) |
fb86f5b2 MG |
9258 | return 0; |
9259 | ||
9260 | return imbalance; | |
9261 | } | |
9262 | ||
1e3c88bd PZ |
9263 | /** |
9264 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
9265 | * groups of a given sched_domain during load balance. | |
bd939f45 | 9266 | * @env: load balance environment |
1e3c88bd | 9267 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 9268 | */ |
bd939f45 | 9269 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9270 | { |
56cf515b JK |
9271 | struct sg_lb_stats *local, *busiest; |
9272 | ||
9273 | local = &sds->local_stat; | |
56cf515b | 9274 | busiest = &sds->busiest_stat; |
dd5feea1 | 9275 | |
0b0695f2 VG |
9276 | if (busiest->group_type == group_misfit_task) { |
9277 | /* Set imbalance to allow misfit tasks to be balanced. */ | |
9278 | env->migration_type = migrate_misfit; | |
c63be7be | 9279 | env->imbalance = 1; |
0b0695f2 VG |
9280 | return; |
9281 | } | |
9282 | ||
9283 | if (busiest->group_type == group_asym_packing) { | |
9284 | /* | |
9285 | * In case of asym capacity, we will try to migrate all load to | |
9286 | * the preferred CPU. | |
9287 | */ | |
9288 | env->migration_type = migrate_task; | |
9289 | env->imbalance = busiest->sum_h_nr_running; | |
9290 | return; | |
9291 | } | |
9292 | ||
9293 | if (busiest->group_type == group_imbalanced) { | |
9294 | /* | |
9295 | * In the group_imb case we cannot rely on group-wide averages | |
9296 | * to ensure CPU-load equilibrium, try to move any task to fix | |
9297 | * the imbalance. The next load balance will take care of | |
9298 | * balancing back the system. | |
9299 | */ | |
9300 | env->migration_type = migrate_task; | |
9301 | env->imbalance = 1; | |
490ba971 VG |
9302 | return; |
9303 | } | |
9304 | ||
1e3c88bd | 9305 | /* |
0b0695f2 | 9306 | * Try to use spare capacity of local group without overloading it or |
a9723389 | 9307 | * emptying busiest. |
1e3c88bd | 9308 | */ |
0b0695f2 | 9309 | if (local->group_type == group_has_spare) { |
16b0a7a1 VG |
9310 | if ((busiest->group_type > group_fully_busy) && |
9311 | !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { | |
0b0695f2 VG |
9312 | /* |
9313 | * If busiest is overloaded, try to fill spare | |
9314 | * capacity. This might end up creating spare capacity | |
9315 | * in busiest or busiest still being overloaded but | |
9316 | * there is no simple way to directly compute the | |
9317 | * amount of load to migrate in order to balance the | |
9318 | * system. | |
9319 | */ | |
9320 | env->migration_type = migrate_util; | |
9321 | env->imbalance = max(local->group_capacity, local->group_util) - | |
9322 | local->group_util; | |
9323 | ||
9324 | /* | |
9325 | * In some cases, the group's utilization is max or even | |
9326 | * higher than capacity because of migrations but the | |
9327 | * local CPU is (newly) idle. There is at least one | |
9328 | * waiting task in this overloaded busiest group. Let's | |
9329 | * try to pull it. | |
9330 | */ | |
9331 | if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { | |
9332 | env->migration_type = migrate_task; | |
9333 | env->imbalance = 1; | |
9334 | } | |
9335 | ||
9336 | return; | |
9337 | } | |
9338 | ||
9339 | if (busiest->group_weight == 1 || sds->prefer_sibling) { | |
5e23e474 | 9340 | unsigned int nr_diff = busiest->sum_nr_running; |
0b0695f2 VG |
9341 | /* |
9342 | * When prefer sibling, evenly spread running tasks on | |
9343 | * groups. | |
9344 | */ | |
9345 | env->migration_type = migrate_task; | |
5e23e474 | 9346 | lsub_positive(&nr_diff, local->sum_nr_running); |
0b0695f2 | 9347 | env->imbalance = nr_diff >> 1; |
b396f523 | 9348 | } else { |
0b0695f2 | 9349 | |
b396f523 MG |
9350 | /* |
9351 | * If there is no overload, we just want to even the number of | |
9352 | * idle cpus. | |
9353 | */ | |
9354 | env->migration_type = migrate_task; | |
9355 | env->imbalance = max_t(long, 0, (local->idle_cpus - | |
0b0695f2 | 9356 | busiest->idle_cpus) >> 1); |
b396f523 MG |
9357 | } |
9358 | ||
9359 | /* Consider allowing a small imbalance between NUMA groups */ | |
7d2b5dd0 | 9360 | if (env->sd->flags & SD_NUMA) { |
fb86f5b2 | 9361 | env->imbalance = adjust_numa_imbalance(env->imbalance, |
7d2b5dd0 MG |
9362 | busiest->sum_nr_running, busiest->group_weight); |
9363 | } | |
b396f523 | 9364 | |
fcf0553d | 9365 | return; |
1e3c88bd PZ |
9366 | } |
9367 | ||
9a5d9ba6 | 9368 | /* |
0b0695f2 VG |
9369 | * Local is fully busy but has to take more load to relieve the |
9370 | * busiest group | |
9a5d9ba6 | 9371 | */ |
0b0695f2 VG |
9372 | if (local->group_type < group_overloaded) { |
9373 | /* | |
9374 | * Local will become overloaded so the avg_load metrics are | |
9375 | * finally needed. | |
9376 | */ | |
9377 | ||
9378 | local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / | |
9379 | local->group_capacity; | |
9380 | ||
9381 | sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / | |
9382 | sds->total_capacity; | |
111688ca AL |
9383 | /* |
9384 | * If the local group is more loaded than the selected | |
9385 | * busiest group don't try to pull any tasks. | |
9386 | */ | |
9387 | if (local->avg_load >= busiest->avg_load) { | |
9388 | env->imbalance = 0; | |
9389 | return; | |
9390 | } | |
dd5feea1 SS |
9391 | } |
9392 | ||
9393 | /* | |
0b0695f2 VG |
9394 | * Both group are or will become overloaded and we're trying to get all |
9395 | * the CPUs to the average_load, so we don't want to push ourselves | |
9396 | * above the average load, nor do we wish to reduce the max loaded CPU | |
9397 | * below the average load. At the same time, we also don't want to | |
9398 | * reduce the group load below the group capacity. Thus we look for | |
9399 | * the minimum possible imbalance. | |
dd5feea1 | 9400 | */ |
0b0695f2 | 9401 | env->migration_type = migrate_load; |
56cf515b | 9402 | env->imbalance = min( |
0b0695f2 | 9403 | (busiest->avg_load - sds->avg_load) * busiest->group_capacity, |
63b2ca30 | 9404 | (sds->avg_load - local->avg_load) * local->group_capacity |
ca8ce3d0 | 9405 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 9406 | } |
fab47622 | 9407 | |
1e3c88bd PZ |
9408 | /******* find_busiest_group() helpers end here *********************/ |
9409 | ||
0b0695f2 VG |
9410 | /* |
9411 | * Decision matrix according to the local and busiest group type: | |
9412 | * | |
9413 | * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded | |
9414 | * has_spare nr_idle balanced N/A N/A balanced balanced | |
9415 | * fully_busy nr_idle nr_idle N/A N/A balanced balanced | |
9416 | * misfit_task force N/A N/A N/A force force | |
9417 | * asym_packing force force N/A N/A force force | |
9418 | * imbalanced force force N/A N/A force force | |
9419 | * overloaded force force N/A N/A force avg_load | |
9420 | * | |
9421 | * N/A : Not Applicable because already filtered while updating | |
9422 | * statistics. | |
9423 | * balanced : The system is balanced for these 2 groups. | |
9424 | * force : Calculate the imbalance as load migration is probably needed. | |
9425 | * avg_load : Only if imbalance is significant enough. | |
9426 | * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite | |
9427 | * different in groups. | |
9428 | */ | |
9429 | ||
1e3c88bd PZ |
9430 | /** |
9431 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 9432 | * if there is an imbalance. |
1e3c88bd | 9433 | * |
a3df0679 | 9434 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
9435 | * to restore balance. |
9436 | * | |
cd96891d | 9437 | * @env: The load balancing environment. |
1e3c88bd | 9438 | * |
e69f6186 | 9439 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 9440 | */ |
56cf515b | 9441 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 9442 | { |
56cf515b | 9443 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
9444 | struct sd_lb_stats sds; |
9445 | ||
147c5fc2 | 9446 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
9447 | |
9448 | /* | |
b0fb1eb4 | 9449 | * Compute the various statistics relevant for load balancing at |
1e3c88bd PZ |
9450 | * this level. |
9451 | */ | |
23f0d209 | 9452 | update_sd_lb_stats(env, &sds); |
2802bf3c | 9453 | |
f8a696f2 | 9454 | if (sched_energy_enabled()) { |
2802bf3c MR |
9455 | struct root_domain *rd = env->dst_rq->rd; |
9456 | ||
9457 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
9458 | goto out_balanced; | |
9459 | } | |
9460 | ||
56cf515b JK |
9461 | local = &sds.local_stat; |
9462 | busiest = &sds.busiest_stat; | |
1e3c88bd | 9463 | |
cc57aa8f | 9464 | /* There is no busy sibling group to pull tasks from */ |
0b0695f2 | 9465 | if (!sds.busiest) |
1e3c88bd PZ |
9466 | goto out_balanced; |
9467 | ||
0b0695f2 VG |
9468 | /* Misfit tasks should be dealt with regardless of the avg load */ |
9469 | if (busiest->group_type == group_misfit_task) | |
9470 | goto force_balance; | |
9471 | ||
9472 | /* ASYM feature bypasses nice load balance check */ | |
9473 | if (busiest->group_type == group_asym_packing) | |
9474 | goto force_balance; | |
b0432d8f | 9475 | |
866ab43e PZ |
9476 | /* |
9477 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 9478 | * work because they assume all things are equal, which typically |
3bd37062 | 9479 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 9480 | */ |
caeb178c | 9481 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
9482 | goto force_balance; |
9483 | ||
cc57aa8f | 9484 | /* |
9c58c79a | 9485 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
9486 | * don't try and pull any tasks. |
9487 | */ | |
0b0695f2 | 9488 | if (local->group_type > busiest->group_type) |
1e3c88bd PZ |
9489 | goto out_balanced; |
9490 | ||
cc57aa8f | 9491 | /* |
0b0695f2 VG |
9492 | * When groups are overloaded, use the avg_load to ensure fairness |
9493 | * between tasks. | |
cc57aa8f | 9494 | */ |
0b0695f2 VG |
9495 | if (local->group_type == group_overloaded) { |
9496 | /* | |
9497 | * If the local group is more loaded than the selected | |
9498 | * busiest group don't try to pull any tasks. | |
9499 | */ | |
9500 | if (local->avg_load >= busiest->avg_load) | |
9501 | goto out_balanced; | |
9502 | ||
9503 | /* XXX broken for overlapping NUMA groups */ | |
9504 | sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / | |
9505 | sds.total_capacity; | |
1e3c88bd | 9506 | |
aae6d3dd | 9507 | /* |
0b0695f2 VG |
9508 | * Don't pull any tasks if this group is already above the |
9509 | * domain average load. | |
aae6d3dd | 9510 | */ |
0b0695f2 | 9511 | if (local->avg_load >= sds.avg_load) |
aae6d3dd | 9512 | goto out_balanced; |
0b0695f2 | 9513 | |
c186fafe | 9514 | /* |
0b0695f2 VG |
9515 | * If the busiest group is more loaded, use imbalance_pct to be |
9516 | * conservative. | |
c186fafe | 9517 | */ |
56cf515b JK |
9518 | if (100 * busiest->avg_load <= |
9519 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 9520 | goto out_balanced; |
aae6d3dd | 9521 | } |
1e3c88bd | 9522 | |
0b0695f2 VG |
9523 | /* Try to move all excess tasks to child's sibling domain */ |
9524 | if (sds.prefer_sibling && local->group_type == group_has_spare && | |
5e23e474 | 9525 | busiest->sum_nr_running > local->sum_nr_running + 1) |
0b0695f2 VG |
9526 | goto force_balance; |
9527 | ||
2ab4092f VG |
9528 | if (busiest->group_type != group_overloaded) { |
9529 | if (env->idle == CPU_NOT_IDLE) | |
9530 | /* | |
9531 | * If the busiest group is not overloaded (and as a | |
9532 | * result the local one too) but this CPU is already | |
9533 | * busy, let another idle CPU try to pull task. | |
9534 | */ | |
9535 | goto out_balanced; | |
9536 | ||
9537 | if (busiest->group_weight > 1 && | |
9538 | local->idle_cpus <= (busiest->idle_cpus + 1)) | |
9539 | /* | |
9540 | * If the busiest group is not overloaded | |
9541 | * and there is no imbalance between this and busiest | |
9542 | * group wrt idle CPUs, it is balanced. The imbalance | |
9543 | * becomes significant if the diff is greater than 1 | |
9544 | * otherwise we might end up to just move the imbalance | |
9545 | * on another group. Of course this applies only if | |
9546 | * there is more than 1 CPU per group. | |
9547 | */ | |
9548 | goto out_balanced; | |
9549 | ||
9550 | if (busiest->sum_h_nr_running == 1) | |
9551 | /* | |
9552 | * busiest doesn't have any tasks waiting to run | |
9553 | */ | |
9554 | goto out_balanced; | |
9555 | } | |
0b0695f2 | 9556 | |
fab47622 | 9557 | force_balance: |
1e3c88bd | 9558 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 9559 | calculate_imbalance(env, &sds); |
bb3485c8 | 9560 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
9561 | |
9562 | out_balanced: | |
bd939f45 | 9563 | env->imbalance = 0; |
1e3c88bd PZ |
9564 | return NULL; |
9565 | } | |
9566 | ||
9567 | /* | |
97fb7a0a | 9568 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 9569 | */ |
bd939f45 | 9570 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 9571 | struct sched_group *group) |
1e3c88bd PZ |
9572 | { |
9573 | struct rq *busiest = NULL, *rq; | |
0b0695f2 VG |
9574 | unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; |
9575 | unsigned int busiest_nr = 0; | |
1e3c88bd PZ |
9576 | int i; |
9577 | ||
ae4df9d6 | 9578 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
0b0695f2 VG |
9579 | unsigned long capacity, load, util; |
9580 | unsigned int nr_running; | |
0ec8aa00 PZ |
9581 | enum fbq_type rt; |
9582 | ||
9583 | rq = cpu_rq(i); | |
9584 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 9585 | |
0ec8aa00 PZ |
9586 | /* |
9587 | * We classify groups/runqueues into three groups: | |
9588 | * - regular: there are !numa tasks | |
9589 | * - remote: there are numa tasks that run on the 'wrong' node | |
9590 | * - all: there is no distinction | |
9591 | * | |
9592 | * In order to avoid migrating ideally placed numa tasks, | |
9593 | * ignore those when there's better options. | |
9594 | * | |
9595 | * If we ignore the actual busiest queue to migrate another | |
9596 | * task, the next balance pass can still reduce the busiest | |
9597 | * queue by moving tasks around inside the node. | |
9598 | * | |
9599 | * If we cannot move enough load due to this classification | |
9600 | * the next pass will adjust the group classification and | |
9601 | * allow migration of more tasks. | |
9602 | * | |
9603 | * Both cases only affect the total convergence complexity. | |
9604 | */ | |
9605 | if (rt > env->fbq_type) | |
9606 | continue; | |
9607 | ||
0b0695f2 | 9608 | nr_running = rq->cfs.h_nr_running; |
fc488ffd VG |
9609 | if (!nr_running) |
9610 | continue; | |
9611 | ||
9612 | capacity = capacity_of(i); | |
9d5efe05 | 9613 | |
4ad3831a CR |
9614 | /* |
9615 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
9616 | * eventually lead to active_balancing high->low capacity. | |
9617 | * Higher per-CPU capacity is considered better than balancing | |
9618 | * average load. | |
9619 | */ | |
9620 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
4aed8aa4 | 9621 | !capacity_greater(capacity_of(env->dst_cpu), capacity) && |
0b0695f2 | 9622 | nr_running == 1) |
4ad3831a CR |
9623 | continue; |
9624 | ||
0b0695f2 VG |
9625 | switch (env->migration_type) { |
9626 | case migrate_load: | |
9627 | /* | |
b0fb1eb4 VG |
9628 | * When comparing with load imbalance, use cpu_load() |
9629 | * which is not scaled with the CPU capacity. | |
0b0695f2 | 9630 | */ |
b0fb1eb4 | 9631 | load = cpu_load(rq); |
1e3c88bd | 9632 | |
0b0695f2 VG |
9633 | if (nr_running == 1 && load > env->imbalance && |
9634 | !check_cpu_capacity(rq, env->sd)) | |
9635 | break; | |
ea67821b | 9636 | |
0b0695f2 VG |
9637 | /* |
9638 | * For the load comparisons with the other CPUs, | |
b0fb1eb4 VG |
9639 | * consider the cpu_load() scaled with the CPU |
9640 | * capacity, so that the load can be moved away | |
9641 | * from the CPU that is potentially running at a | |
9642 | * lower capacity. | |
0b0695f2 VG |
9643 | * |
9644 | * Thus we're looking for max(load_i / capacity_i), | |
9645 | * crosswise multiplication to rid ourselves of the | |
9646 | * division works out to: | |
9647 | * load_i * capacity_j > load_j * capacity_i; | |
9648 | * where j is our previous maximum. | |
9649 | */ | |
9650 | if (load * busiest_capacity > busiest_load * capacity) { | |
9651 | busiest_load = load; | |
9652 | busiest_capacity = capacity; | |
9653 | busiest = rq; | |
9654 | } | |
9655 | break; | |
9656 | ||
9657 | case migrate_util: | |
9658 | util = cpu_util(cpu_of(rq)); | |
9659 | ||
c32b4308 VG |
9660 | /* |
9661 | * Don't try to pull utilization from a CPU with one | |
9662 | * running task. Whatever its utilization, we will fail | |
9663 | * detach the task. | |
9664 | */ | |
9665 | if (nr_running <= 1) | |
9666 | continue; | |
9667 | ||
0b0695f2 VG |
9668 | if (busiest_util < util) { |
9669 | busiest_util = util; | |
9670 | busiest = rq; | |
9671 | } | |
9672 | break; | |
9673 | ||
9674 | case migrate_task: | |
9675 | if (busiest_nr < nr_running) { | |
9676 | busiest_nr = nr_running; | |
9677 | busiest = rq; | |
9678 | } | |
9679 | break; | |
9680 | ||
9681 | case migrate_misfit: | |
9682 | /* | |
9683 | * For ASYM_CPUCAPACITY domains with misfit tasks we | |
9684 | * simply seek the "biggest" misfit task. | |
9685 | */ | |
9686 | if (rq->misfit_task_load > busiest_load) { | |
9687 | busiest_load = rq->misfit_task_load; | |
9688 | busiest = rq; | |
9689 | } | |
9690 | ||
9691 | break; | |
1e3c88bd | 9692 | |
1e3c88bd PZ |
9693 | } |
9694 | } | |
9695 | ||
9696 | return busiest; | |
9697 | } | |
9698 | ||
9699 | /* | |
9700 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
9701 | * so long as it is large enough. | |
9702 | */ | |
9703 | #define MAX_PINNED_INTERVAL 512 | |
9704 | ||
46a745d9 VG |
9705 | static inline bool |
9706 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 9707 | { |
46a745d9 VG |
9708 | /* |
9709 | * ASYM_PACKING needs to force migrate tasks from busy but | |
9710 | * lower priority CPUs in order to pack all tasks in the | |
9711 | * highest priority CPUs. | |
9712 | */ | |
9713 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
9714 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
9715 | } | |
bd939f45 | 9716 | |
46a745d9 | 9717 | static inline bool |
e9b9734b VG |
9718 | imbalanced_active_balance(struct lb_env *env) |
9719 | { | |
9720 | struct sched_domain *sd = env->sd; | |
9721 | ||
9722 | /* | |
9723 | * The imbalanced case includes the case of pinned tasks preventing a fair | |
9724 | * distribution of the load on the system but also the even distribution of the | |
9725 | * threads on a system with spare capacity | |
9726 | */ | |
9727 | if ((env->migration_type == migrate_task) && | |
9728 | (sd->nr_balance_failed > sd->cache_nice_tries+2)) | |
9729 | return 1; | |
9730 | ||
9731 | return 0; | |
9732 | } | |
9733 | ||
9734 | static int need_active_balance(struct lb_env *env) | |
46a745d9 VG |
9735 | { |
9736 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 9737 | |
46a745d9 VG |
9738 | if (asym_active_balance(env)) |
9739 | return 1; | |
1af3ed3d | 9740 | |
e9b9734b VG |
9741 | if (imbalanced_active_balance(env)) |
9742 | return 1; | |
9743 | ||
1aaf90a4 VG |
9744 | /* |
9745 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
9746 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
9747 | * because of other sched_class or IRQs if more capacity stays | |
9748 | * available on dst_cpu. | |
9749 | */ | |
9750 | if ((env->idle != CPU_NOT_IDLE) && | |
9751 | (env->src_rq->cfs.h_nr_running == 1)) { | |
9752 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
9753 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
9754 | return 1; | |
9755 | } | |
9756 | ||
0b0695f2 | 9757 | if (env->migration_type == migrate_misfit) |
cad68e55 MR |
9758 | return 1; |
9759 | ||
46a745d9 VG |
9760 | return 0; |
9761 | } | |
9762 | ||
969c7921 TH |
9763 | static int active_load_balance_cpu_stop(void *data); |
9764 | ||
23f0d209 JK |
9765 | static int should_we_balance(struct lb_env *env) |
9766 | { | |
9767 | struct sched_group *sg = env->sd->groups; | |
64297f2b | 9768 | int cpu; |
23f0d209 | 9769 | |
024c9d2f PZ |
9770 | /* |
9771 | * Ensure the balancing environment is consistent; can happen | |
9772 | * when the softirq triggers 'during' hotplug. | |
9773 | */ | |
9774 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
9775 | return 0; | |
9776 | ||
23f0d209 | 9777 | /* |
97fb7a0a | 9778 | * In the newly idle case, we will allow all the CPUs |
23f0d209 JK |
9779 | * to do the newly idle load balance. |
9780 | */ | |
9781 | if (env->idle == CPU_NEWLY_IDLE) | |
9782 | return 1; | |
9783 | ||
97fb7a0a | 9784 | /* Try to find first idle CPU */ |
e5c14b1f | 9785 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 9786 | if (!idle_cpu(cpu)) |
23f0d209 JK |
9787 | continue; |
9788 | ||
64297f2b PW |
9789 | /* Are we the first idle CPU? */ |
9790 | return cpu == env->dst_cpu; | |
23f0d209 JK |
9791 | } |
9792 | ||
64297f2b PW |
9793 | /* Are we the first CPU of this group ? */ |
9794 | return group_balance_cpu(sg) == env->dst_cpu; | |
23f0d209 JK |
9795 | } |
9796 | ||
1e3c88bd PZ |
9797 | /* |
9798 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
9799 | * tasks if there is an imbalance. | |
9800 | */ | |
9801 | static int load_balance(int this_cpu, struct rq *this_rq, | |
9802 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 9803 | int *continue_balancing) |
1e3c88bd | 9804 | { |
88b8dac0 | 9805 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 9806 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 9807 | struct sched_group *group; |
1e3c88bd | 9808 | struct rq *busiest; |
8a8c69c3 | 9809 | struct rq_flags rf; |
4ba29684 | 9810 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 9811 | |
8e45cb54 PZ |
9812 | struct lb_env env = { |
9813 | .sd = sd, | |
ddcdf6e7 PZ |
9814 | .dst_cpu = this_cpu, |
9815 | .dst_rq = this_rq, | |
ae4df9d6 | 9816 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 9817 | .idle = idle, |
eb95308e | 9818 | .loop_break = sched_nr_migrate_break, |
b9403130 | 9819 | .cpus = cpus, |
0ec8aa00 | 9820 | .fbq_type = all, |
163122b7 | 9821 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
9822 | }; |
9823 | ||
65a4433a | 9824 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 9825 | |
ae92882e | 9826 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
9827 | |
9828 | redo: | |
23f0d209 JK |
9829 | if (!should_we_balance(&env)) { |
9830 | *continue_balancing = 0; | |
1e3c88bd | 9831 | goto out_balanced; |
23f0d209 | 9832 | } |
1e3c88bd | 9833 | |
23f0d209 | 9834 | group = find_busiest_group(&env); |
1e3c88bd | 9835 | if (!group) { |
ae92882e | 9836 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
9837 | goto out_balanced; |
9838 | } | |
9839 | ||
b9403130 | 9840 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 9841 | if (!busiest) { |
ae92882e | 9842 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
9843 | goto out_balanced; |
9844 | } | |
9845 | ||
78feefc5 | 9846 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 9847 | |
ae92882e | 9848 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 9849 | |
1aaf90a4 VG |
9850 | env.src_cpu = busiest->cpu; |
9851 | env.src_rq = busiest; | |
9852 | ||
1e3c88bd | 9853 | ld_moved = 0; |
8a41dfcd VG |
9854 | /* Clear this flag as soon as we find a pullable task */ |
9855 | env.flags |= LBF_ALL_PINNED; | |
1e3c88bd PZ |
9856 | if (busiest->nr_running > 1) { |
9857 | /* | |
9858 | * Attempt to move tasks. If find_busiest_group has found | |
9859 | * an imbalance but busiest->nr_running <= 1, the group is | |
9860 | * still unbalanced. ld_moved simply stays zero, so it is | |
9861 | * correctly treated as an imbalance. | |
9862 | */ | |
c82513e5 | 9863 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 9864 | |
5d6523eb | 9865 | more_balance: |
8a8c69c3 | 9866 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 9867 | update_rq_clock(busiest); |
88b8dac0 SV |
9868 | |
9869 | /* | |
9870 | * cur_ld_moved - load moved in current iteration | |
9871 | * ld_moved - cumulative load moved across iterations | |
9872 | */ | |
163122b7 | 9873 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
9874 | |
9875 | /* | |
163122b7 KT |
9876 | * We've detached some tasks from busiest_rq. Every |
9877 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
9878 | * unlock busiest->lock, and we are able to be sure | |
9879 | * that nobody can manipulate the tasks in parallel. | |
9880 | * See task_rq_lock() family for the details. | |
1e3c88bd | 9881 | */ |
163122b7 | 9882 | |
8a8c69c3 | 9883 | rq_unlock(busiest, &rf); |
163122b7 KT |
9884 | |
9885 | if (cur_ld_moved) { | |
9886 | attach_tasks(&env); | |
9887 | ld_moved += cur_ld_moved; | |
9888 | } | |
9889 | ||
8a8c69c3 | 9890 | local_irq_restore(rf.flags); |
88b8dac0 | 9891 | |
f1cd0858 JK |
9892 | if (env.flags & LBF_NEED_BREAK) { |
9893 | env.flags &= ~LBF_NEED_BREAK; | |
9894 | goto more_balance; | |
9895 | } | |
9896 | ||
88b8dac0 SV |
9897 | /* |
9898 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
9899 | * us and move them to an alternate dst_cpu in our sched_group | |
9900 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 9901 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
9902 | * sched_group. |
9903 | * | |
9904 | * This changes load balance semantics a bit on who can move | |
9905 | * load to a given_cpu. In addition to the given_cpu itself | |
9906 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
9907 | * nohz-idle), we now have balance_cpu in a position to move | |
9908 | * load to given_cpu. In rare situations, this may cause | |
9909 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
9910 | * _independently_ and at _same_ time to move some load to | |
3b03706f | 9911 | * given_cpu) causing excess load to be moved to given_cpu. |
88b8dac0 SV |
9912 | * This however should not happen so much in practice and |
9913 | * moreover subsequent load balance cycles should correct the | |
9914 | * excess load moved. | |
9915 | */ | |
6263322c | 9916 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 9917 | |
97fb7a0a | 9918 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 9919 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 9920 | |
78feefc5 | 9921 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 9922 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 9923 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
9924 | env.loop = 0; |
9925 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 9926 | |
88b8dac0 SV |
9927 | /* |
9928 | * Go back to "more_balance" rather than "redo" since we | |
9929 | * need to continue with same src_cpu. | |
9930 | */ | |
9931 | goto more_balance; | |
9932 | } | |
1e3c88bd | 9933 | |
6263322c PZ |
9934 | /* |
9935 | * We failed to reach balance because of affinity. | |
9936 | */ | |
9937 | if (sd_parent) { | |
63b2ca30 | 9938 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 9939 | |
afdeee05 | 9940 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 9941 | *group_imbalance = 1; |
6263322c PZ |
9942 | } |
9943 | ||
1e3c88bd | 9944 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 9945 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 9946 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
9947 | /* |
9948 | * Attempting to continue load balancing at the current | |
9949 | * sched_domain level only makes sense if there are | |
9950 | * active CPUs remaining as possible busiest CPUs to | |
9951 | * pull load from which are not contained within the | |
9952 | * destination group that is receiving any migrated | |
9953 | * load. | |
9954 | */ | |
9955 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
9956 | env.loop = 0; |
9957 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 9958 | goto redo; |
bbf18b19 | 9959 | } |
afdeee05 | 9960 | goto out_all_pinned; |
1e3c88bd PZ |
9961 | } |
9962 | } | |
9963 | ||
9964 | if (!ld_moved) { | |
ae92882e | 9965 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
9966 | /* |
9967 | * Increment the failure counter only on periodic balance. | |
9968 | * We do not want newidle balance, which can be very | |
9969 | * frequent, pollute the failure counter causing | |
9970 | * excessive cache_hot migrations and active balances. | |
9971 | */ | |
9972 | if (idle != CPU_NEWLY_IDLE) | |
9973 | sd->nr_balance_failed++; | |
1e3c88bd | 9974 | |
bd939f45 | 9975 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
9976 | unsigned long flags; |
9977 | ||
5cb9eaa3 | 9978 | raw_spin_rq_lock_irqsave(busiest, flags); |
1e3c88bd | 9979 | |
97fb7a0a IM |
9980 | /* |
9981 | * Don't kick the active_load_balance_cpu_stop, | |
9982 | * if the curr task on busiest CPU can't be | |
9983 | * moved to this_cpu: | |
1e3c88bd | 9984 | */ |
3bd37062 | 9985 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
5cb9eaa3 | 9986 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
1e3c88bd PZ |
9987 | goto out_one_pinned; |
9988 | } | |
9989 | ||
8a41dfcd VG |
9990 | /* Record that we found at least one task that could run on this_cpu */ |
9991 | env.flags &= ~LBF_ALL_PINNED; | |
9992 | ||
969c7921 TH |
9993 | /* |
9994 | * ->active_balance synchronizes accesses to | |
9995 | * ->active_balance_work. Once set, it's cleared | |
9996 | * only after active load balance is finished. | |
9997 | */ | |
1e3c88bd PZ |
9998 | if (!busiest->active_balance) { |
9999 | busiest->active_balance = 1; | |
10000 | busiest->push_cpu = this_cpu; | |
10001 | active_balance = 1; | |
10002 | } | |
5cb9eaa3 | 10003 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
969c7921 | 10004 | |
bd939f45 | 10005 | if (active_balance) { |
969c7921 TH |
10006 | stop_one_cpu_nowait(cpu_of(busiest), |
10007 | active_load_balance_cpu_stop, busiest, | |
10008 | &busiest->active_balance_work); | |
bd939f45 | 10009 | } |
1e3c88bd | 10010 | } |
e9b9734b | 10011 | } else { |
1e3c88bd | 10012 | sd->nr_balance_failed = 0; |
e9b9734b | 10013 | } |
1e3c88bd | 10014 | |
e9b9734b | 10015 | if (likely(!active_balance) || need_active_balance(&env)) { |
1e3c88bd PZ |
10016 | /* We were unbalanced, so reset the balancing interval */ |
10017 | sd->balance_interval = sd->min_interval; | |
1e3c88bd PZ |
10018 | } |
10019 | ||
1e3c88bd PZ |
10020 | goto out; |
10021 | ||
10022 | out_balanced: | |
afdeee05 VG |
10023 | /* |
10024 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
10025 | * constraints. Clear the imbalance flag only if other tasks got |
10026 | * a chance to move and fix the imbalance. | |
afdeee05 | 10027 | */ |
f6cad8df | 10028 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
10029 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
10030 | ||
10031 | if (*group_imbalance) | |
10032 | *group_imbalance = 0; | |
10033 | } | |
10034 | ||
10035 | out_all_pinned: | |
10036 | /* | |
10037 | * We reach balance because all tasks are pinned at this level so | |
10038 | * we can't migrate them. Let the imbalance flag set so parent level | |
10039 | * can try to migrate them. | |
10040 | */ | |
ae92882e | 10041 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
10042 | |
10043 | sd->nr_balance_failed = 0; | |
10044 | ||
10045 | out_one_pinned: | |
3f130a37 VS |
10046 | ld_moved = 0; |
10047 | ||
10048 | /* | |
5ba553ef PZ |
10049 | * newidle_balance() disregards balance intervals, so we could |
10050 | * repeatedly reach this code, which would lead to balance_interval | |
3b03706f | 10051 | * skyrocketing in a short amount of time. Skip the balance_interval |
5ba553ef | 10052 | * increase logic to avoid that. |
3f130a37 VS |
10053 | */ |
10054 | if (env.idle == CPU_NEWLY_IDLE) | |
10055 | goto out; | |
10056 | ||
1e3c88bd | 10057 | /* tune up the balancing interval */ |
47b7aee1 VS |
10058 | if ((env.flags & LBF_ALL_PINNED && |
10059 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
10060 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 10061 | sd->balance_interval *= 2; |
1e3c88bd | 10062 | out: |
1e3c88bd PZ |
10063 | return ld_moved; |
10064 | } | |
10065 | ||
52a08ef1 JL |
10066 | static inline unsigned long |
10067 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
10068 | { | |
10069 | unsigned long interval = sd->balance_interval; | |
10070 | ||
10071 | if (cpu_busy) | |
10072 | interval *= sd->busy_factor; | |
10073 | ||
10074 | /* scale ms to jiffies */ | |
10075 | interval = msecs_to_jiffies(interval); | |
e4d32e4d VG |
10076 | |
10077 | /* | |
10078 | * Reduce likelihood of busy balancing at higher domains racing with | |
10079 | * balancing at lower domains by preventing their balancing periods | |
10080 | * from being multiples of each other. | |
10081 | */ | |
10082 | if (cpu_busy) | |
10083 | interval -= 1; | |
10084 | ||
52a08ef1 JL |
10085 | interval = clamp(interval, 1UL, max_load_balance_interval); |
10086 | ||
10087 | return interval; | |
10088 | } | |
10089 | ||
10090 | static inline void | |
31851a98 | 10091 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
10092 | { |
10093 | unsigned long interval, next; | |
10094 | ||
31851a98 LY |
10095 | /* used by idle balance, so cpu_busy = 0 */ |
10096 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
10097 | next = sd->last_balance + interval; |
10098 | ||
10099 | if (time_after(*next_balance, next)) | |
10100 | *next_balance = next; | |
10101 | } | |
10102 | ||
1e3c88bd | 10103 | /* |
97fb7a0a | 10104 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
10105 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
10106 | * least 1 task to be running on each physical CPU where possible, and | |
10107 | * avoids physical / logical imbalances. | |
1e3c88bd | 10108 | */ |
969c7921 | 10109 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 10110 | { |
969c7921 TH |
10111 | struct rq *busiest_rq = data; |
10112 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 10113 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 10114 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 10115 | struct sched_domain *sd; |
e5673f28 | 10116 | struct task_struct *p = NULL; |
8a8c69c3 | 10117 | struct rq_flags rf; |
969c7921 | 10118 | |
8a8c69c3 | 10119 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
10120 | /* |
10121 | * Between queueing the stop-work and running it is a hole in which | |
10122 | * CPUs can become inactive. We should not move tasks from or to | |
10123 | * inactive CPUs. | |
10124 | */ | |
10125 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
10126 | goto out_unlock; | |
969c7921 | 10127 | |
97fb7a0a | 10128 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
10129 | if (unlikely(busiest_cpu != smp_processor_id() || |
10130 | !busiest_rq->active_balance)) | |
10131 | goto out_unlock; | |
1e3c88bd PZ |
10132 | |
10133 | /* Is there any task to move? */ | |
10134 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 10135 | goto out_unlock; |
1e3c88bd PZ |
10136 | |
10137 | /* | |
10138 | * This condition is "impossible", if it occurs | |
10139 | * we need to fix it. Originally reported by | |
97fb7a0a | 10140 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd PZ |
10141 | */ |
10142 | BUG_ON(busiest_rq == target_rq); | |
10143 | ||
1e3c88bd | 10144 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 10145 | rcu_read_lock(); |
1e3c88bd | 10146 | for_each_domain(target_cpu, sd) { |
e669ac8a VS |
10147 | if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
10148 | break; | |
1e3c88bd PZ |
10149 | } |
10150 | ||
10151 | if (likely(sd)) { | |
8e45cb54 PZ |
10152 | struct lb_env env = { |
10153 | .sd = sd, | |
ddcdf6e7 PZ |
10154 | .dst_cpu = target_cpu, |
10155 | .dst_rq = target_rq, | |
10156 | .src_cpu = busiest_rq->cpu, | |
10157 | .src_rq = busiest_rq, | |
8e45cb54 | 10158 | .idle = CPU_IDLE, |
23fb06d9 | 10159 | .flags = LBF_ACTIVE_LB, |
8e45cb54 PZ |
10160 | }; |
10161 | ||
ae92882e | 10162 | schedstat_inc(sd->alb_count); |
3bed5e21 | 10163 | update_rq_clock(busiest_rq); |
1e3c88bd | 10164 | |
e5673f28 | 10165 | p = detach_one_task(&env); |
d02c0711 | 10166 | if (p) { |
ae92882e | 10167 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
10168 | /* Active balancing done, reset the failure counter. */ |
10169 | sd->nr_balance_failed = 0; | |
10170 | } else { | |
ae92882e | 10171 | schedstat_inc(sd->alb_failed); |
d02c0711 | 10172 | } |
1e3c88bd | 10173 | } |
dce840a0 | 10174 | rcu_read_unlock(); |
969c7921 TH |
10175 | out_unlock: |
10176 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 10177 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
10178 | |
10179 | if (p) | |
10180 | attach_one_task(target_rq, p); | |
10181 | ||
10182 | local_irq_enable(); | |
10183 | ||
969c7921 | 10184 | return 0; |
1e3c88bd PZ |
10185 | } |
10186 | ||
af3fe03c PZ |
10187 | static DEFINE_SPINLOCK(balancing); |
10188 | ||
10189 | /* | |
10190 | * Scale the max load_balance interval with the number of CPUs in the system. | |
10191 | * This trades load-balance latency on larger machines for less cross talk. | |
10192 | */ | |
10193 | void update_max_interval(void) | |
10194 | { | |
10195 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
10196 | } | |
10197 | ||
10198 | /* | |
10199 | * It checks each scheduling domain to see if it is due to be balanced, | |
10200 | * and initiates a balancing operation if so. | |
10201 | * | |
10202 | * Balancing parameters are set up in init_sched_domains. | |
10203 | */ | |
10204 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
10205 | { | |
10206 | int continue_balancing = 1; | |
10207 | int cpu = rq->cpu; | |
323af6de | 10208 | int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10209 | unsigned long interval; |
10210 | struct sched_domain *sd; | |
10211 | /* Earliest time when we have to do rebalance again */ | |
10212 | unsigned long next_balance = jiffies + 60*HZ; | |
10213 | int update_next_balance = 0; | |
10214 | int need_serialize, need_decay = 0; | |
10215 | u64 max_cost = 0; | |
10216 | ||
10217 | rcu_read_lock(); | |
10218 | for_each_domain(cpu, sd) { | |
10219 | /* | |
10220 | * Decay the newidle max times here because this is a regular | |
10221 | * visit to all the domains. Decay ~1% per second. | |
10222 | */ | |
10223 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
10224 | sd->max_newidle_lb_cost = | |
10225 | (sd->max_newidle_lb_cost * 253) / 256; | |
10226 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
10227 | need_decay = 1; | |
10228 | } | |
10229 | max_cost += sd->max_newidle_lb_cost; | |
10230 | ||
af3fe03c PZ |
10231 | /* |
10232 | * Stop the load balance at this level. There is another | |
10233 | * CPU in our sched group which is doing load balancing more | |
10234 | * actively. | |
10235 | */ | |
10236 | if (!continue_balancing) { | |
10237 | if (need_decay) | |
10238 | continue; | |
10239 | break; | |
10240 | } | |
10241 | ||
323af6de | 10242 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
10243 | |
10244 | need_serialize = sd->flags & SD_SERIALIZE; | |
10245 | if (need_serialize) { | |
10246 | if (!spin_trylock(&balancing)) | |
10247 | goto out; | |
10248 | } | |
10249 | ||
10250 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
10251 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
10252 | /* | |
10253 | * The LBF_DST_PINNED logic could have changed | |
10254 | * env->dst_cpu, so we can't know our idle | |
10255 | * state even if we migrated tasks. Update it. | |
10256 | */ | |
10257 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
323af6de | 10258 | busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10259 | } |
10260 | sd->last_balance = jiffies; | |
323af6de | 10261 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
10262 | } |
10263 | if (need_serialize) | |
10264 | spin_unlock(&balancing); | |
10265 | out: | |
10266 | if (time_after(next_balance, sd->last_balance + interval)) { | |
10267 | next_balance = sd->last_balance + interval; | |
10268 | update_next_balance = 1; | |
10269 | } | |
10270 | } | |
10271 | if (need_decay) { | |
10272 | /* | |
10273 | * Ensure the rq-wide value also decays but keep it at a | |
10274 | * reasonable floor to avoid funnies with rq->avg_idle. | |
10275 | */ | |
10276 | rq->max_idle_balance_cost = | |
10277 | max((u64)sysctl_sched_migration_cost, max_cost); | |
10278 | } | |
10279 | rcu_read_unlock(); | |
10280 | ||
10281 | /* | |
10282 | * next_balance will be updated only when there is a need. | |
10283 | * When the cpu is attached to null domain for ex, it will not be | |
10284 | * updated. | |
10285 | */ | |
7a82e5f5 | 10286 | if (likely(update_next_balance)) |
af3fe03c PZ |
10287 | rq->next_balance = next_balance; |
10288 | ||
af3fe03c PZ |
10289 | } |
10290 | ||
d987fc7f MG |
10291 | static inline int on_null_domain(struct rq *rq) |
10292 | { | |
10293 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
10294 | } | |
10295 | ||
3451d024 | 10296 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
10297 | /* |
10298 | * idle load balancing details | |
83cd4fe2 VP |
10299 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
10300 | * needed, they will kick the idle load balancer, which then does idle | |
10301 | * load balancing for all the idle CPUs. | |
9b019acb NP |
10302 | * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set |
10303 | * anywhere yet. | |
83cd4fe2 | 10304 | */ |
1e3c88bd | 10305 | |
3dd0337d | 10306 | static inline int find_new_ilb(void) |
1e3c88bd | 10307 | { |
9b019acb | 10308 | int ilb; |
031e3bd8 | 10309 | const struct cpumask *hk_mask; |
1e3c88bd | 10310 | |
031e3bd8 | 10311 | hk_mask = housekeeping_cpumask(HK_FLAG_MISC); |
1e3c88bd | 10312 | |
031e3bd8 | 10313 | for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) { |
45da7a2b PZ |
10314 | |
10315 | if (ilb == smp_processor_id()) | |
10316 | continue; | |
10317 | ||
9b019acb NP |
10318 | if (idle_cpu(ilb)) |
10319 | return ilb; | |
10320 | } | |
786d6dc7 SS |
10321 | |
10322 | return nr_cpu_ids; | |
1e3c88bd | 10323 | } |
1e3c88bd | 10324 | |
83cd4fe2 | 10325 | /* |
9b019acb NP |
10326 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick any |
10327 | * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one). | |
83cd4fe2 | 10328 | */ |
a4064fb6 | 10329 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
10330 | { |
10331 | int ilb_cpu; | |
10332 | ||
3ea2f097 VG |
10333 | /* |
10334 | * Increase nohz.next_balance only when if full ilb is triggered but | |
10335 | * not if we only update stats. | |
10336 | */ | |
10337 | if (flags & NOHZ_BALANCE_KICK) | |
10338 | nohz.next_balance = jiffies+1; | |
83cd4fe2 | 10339 | |
3dd0337d | 10340 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 10341 | |
0b005cf5 SS |
10342 | if (ilb_cpu >= nr_cpu_ids) |
10343 | return; | |
83cd4fe2 | 10344 | |
19a1f5ec PZ |
10345 | /* |
10346 | * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets | |
10347 | * the first flag owns it; cleared by nohz_csd_func(). | |
10348 | */ | |
a4064fb6 | 10349 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 10350 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 10351 | return; |
4550487a | 10352 | |
1c792db7 | 10353 | /* |
90b5363a | 10354 | * This way we generate an IPI on the target CPU which |
1c792db7 SS |
10355 | * is idle. And the softirq performing nohz idle load balance |
10356 | * will be run before returning from the IPI. | |
10357 | */ | |
90b5363a | 10358 | smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); |
4550487a PZ |
10359 | } |
10360 | ||
10361 | /* | |
9f132742 VS |
10362 | * Current decision point for kicking the idle load balancer in the presence |
10363 | * of idle CPUs in the system. | |
4550487a PZ |
10364 | */ |
10365 | static void nohz_balancer_kick(struct rq *rq) | |
10366 | { | |
10367 | unsigned long now = jiffies; | |
10368 | struct sched_domain_shared *sds; | |
10369 | struct sched_domain *sd; | |
10370 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 10371 | unsigned int flags = 0; |
4550487a PZ |
10372 | |
10373 | if (unlikely(rq->idle_balance)) | |
10374 | return; | |
10375 | ||
10376 | /* | |
10377 | * We may be recently in ticked or tickless idle mode. At the first | |
10378 | * busy tick after returning from idle, we will update the busy stats. | |
10379 | */ | |
00357f5e | 10380 | nohz_balance_exit_idle(rq); |
4550487a PZ |
10381 | |
10382 | /* | |
10383 | * None are in tickless mode and hence no need for NOHZ idle load | |
10384 | * balancing. | |
10385 | */ | |
10386 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
10387 | return; | |
10388 | ||
f643ea22 VG |
10389 | if (READ_ONCE(nohz.has_blocked) && |
10390 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
10391 | flags = NOHZ_STATS_KICK; |
10392 | ||
4550487a | 10393 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 10394 | goto out; |
4550487a | 10395 | |
a0fe2cf0 | 10396 | if (rq->nr_running >= 2) { |
a4064fb6 | 10397 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10398 | goto out; |
10399 | } | |
10400 | ||
10401 | rcu_read_lock(); | |
4550487a PZ |
10402 | |
10403 | sd = rcu_dereference(rq->sd); | |
10404 | if (sd) { | |
e25a7a94 VS |
10405 | /* |
10406 | * If there's a CFS task and the current CPU has reduced | |
10407 | * capacity; kick the ILB to see if there's a better CPU to run | |
10408 | * on. | |
10409 | */ | |
10410 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
a4064fb6 | 10411 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10412 | goto unlock; |
10413 | } | |
10414 | } | |
10415 | ||
011b27bb | 10416 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 10417 | if (sd) { |
b9a7b883 VS |
10418 | /* |
10419 | * When ASYM_PACKING; see if there's a more preferred CPU | |
10420 | * currently idle; in which case, kick the ILB to move tasks | |
10421 | * around. | |
10422 | */ | |
7edab78d | 10423 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
4550487a | 10424 | if (sched_asym_prefer(i, cpu)) { |
a4064fb6 | 10425 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10426 | goto unlock; |
10427 | } | |
10428 | } | |
10429 | } | |
b9a7b883 | 10430 | |
a0fe2cf0 VS |
10431 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
10432 | if (sd) { | |
10433 | /* | |
10434 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
10435 | * to run the misfit task on. | |
10436 | */ | |
10437 | if (check_misfit_status(rq, sd)) { | |
10438 | flags = NOHZ_KICK_MASK; | |
10439 | goto unlock; | |
10440 | } | |
b9a7b883 VS |
10441 | |
10442 | /* | |
10443 | * For asymmetric systems, we do not want to nicely balance | |
10444 | * cache use, instead we want to embrace asymmetry and only | |
10445 | * ensure tasks have enough CPU capacity. | |
10446 | * | |
10447 | * Skip the LLC logic because it's not relevant in that case. | |
10448 | */ | |
10449 | goto unlock; | |
a0fe2cf0 VS |
10450 | } |
10451 | ||
b9a7b883 VS |
10452 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
10453 | if (sds) { | |
e25a7a94 | 10454 | /* |
b9a7b883 VS |
10455 | * If there is an imbalance between LLC domains (IOW we could |
10456 | * increase the overall cache use), we need some less-loaded LLC | |
10457 | * domain to pull some load. Likewise, we may need to spread | |
10458 | * load within the current LLC domain (e.g. packed SMT cores but | |
10459 | * other CPUs are idle). We can't really know from here how busy | |
10460 | * the others are - so just get a nohz balance going if it looks | |
10461 | * like this LLC domain has tasks we could move. | |
e25a7a94 | 10462 | */ |
b9a7b883 VS |
10463 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
10464 | if (nr_busy > 1) { | |
10465 | flags = NOHZ_KICK_MASK; | |
10466 | goto unlock; | |
4550487a PZ |
10467 | } |
10468 | } | |
10469 | unlock: | |
10470 | rcu_read_unlock(); | |
10471 | out: | |
a4064fb6 PZ |
10472 | if (flags) |
10473 | kick_ilb(flags); | |
83cd4fe2 VP |
10474 | } |
10475 | ||
00357f5e | 10476 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 10477 | { |
00357f5e | 10478 | struct sched_domain *sd; |
a22e47a4 | 10479 | |
00357f5e PZ |
10480 | rcu_read_lock(); |
10481 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 10482 | |
00357f5e PZ |
10483 | if (!sd || !sd->nohz_idle) |
10484 | goto unlock; | |
10485 | sd->nohz_idle = 0; | |
10486 | ||
10487 | atomic_inc(&sd->shared->nr_busy_cpus); | |
10488 | unlock: | |
10489 | rcu_read_unlock(); | |
71325960 SS |
10490 | } |
10491 | ||
00357f5e PZ |
10492 | void nohz_balance_exit_idle(struct rq *rq) |
10493 | { | |
10494 | SCHED_WARN_ON(rq != this_rq()); | |
10495 | ||
10496 | if (likely(!rq->nohz_tick_stopped)) | |
10497 | return; | |
10498 | ||
10499 | rq->nohz_tick_stopped = 0; | |
10500 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
10501 | atomic_dec(&nohz.nr_cpus); | |
10502 | ||
10503 | set_cpu_sd_state_busy(rq->cpu); | |
10504 | } | |
10505 | ||
10506 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
10507 | { |
10508 | struct sched_domain *sd; | |
69e1e811 | 10509 | |
69e1e811 | 10510 | rcu_read_lock(); |
0e369d75 | 10511 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
10512 | |
10513 | if (!sd || sd->nohz_idle) | |
10514 | goto unlock; | |
10515 | sd->nohz_idle = 1; | |
10516 | ||
0e369d75 | 10517 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 10518 | unlock: |
69e1e811 SS |
10519 | rcu_read_unlock(); |
10520 | } | |
10521 | ||
1e3c88bd | 10522 | /* |
97fb7a0a | 10523 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 10524 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 10525 | */ |
c1cc017c | 10526 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 10527 | { |
00357f5e PZ |
10528 | struct rq *rq = cpu_rq(cpu); |
10529 | ||
10530 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
10531 | ||
97fb7a0a | 10532 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
10533 | if (!cpu_active(cpu)) |
10534 | return; | |
10535 | ||
387bc8b5 | 10536 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
de201559 | 10537 | if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) |
387bc8b5 FW |
10538 | return; |
10539 | ||
f643ea22 VG |
10540 | /* |
10541 | * Can be set safely without rq->lock held | |
10542 | * If a clear happens, it will have evaluated last additions because | |
10543 | * rq->lock is held during the check and the clear | |
10544 | */ | |
10545 | rq->has_blocked_load = 1; | |
10546 | ||
10547 | /* | |
10548 | * The tick is still stopped but load could have been added in the | |
10549 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
10550 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
10551 | * of nohz.has_blocked can only happen after checking the new load | |
10552 | */ | |
00357f5e | 10553 | if (rq->nohz_tick_stopped) |
f643ea22 | 10554 | goto out; |
1e3c88bd | 10555 | |
97fb7a0a | 10556 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 10557 | if (on_null_domain(rq)) |
d987fc7f MG |
10558 | return; |
10559 | ||
00357f5e PZ |
10560 | rq->nohz_tick_stopped = 1; |
10561 | ||
c1cc017c AS |
10562 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
10563 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 10564 | |
f643ea22 VG |
10565 | /* |
10566 | * Ensures that if nohz_idle_balance() fails to observe our | |
10567 | * @idle_cpus_mask store, it must observe the @has_blocked | |
10568 | * store. | |
10569 | */ | |
10570 | smp_mb__after_atomic(); | |
10571 | ||
00357f5e | 10572 | set_cpu_sd_state_idle(cpu); |
f643ea22 VG |
10573 | |
10574 | out: | |
10575 | /* | |
10576 | * Each time a cpu enter idle, we assume that it has blocked load and | |
10577 | * enable the periodic update of the load of idle cpus | |
10578 | */ | |
10579 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 10580 | } |
1e3c88bd | 10581 | |
3f5ad914 Y |
10582 | static bool update_nohz_stats(struct rq *rq) |
10583 | { | |
10584 | unsigned int cpu = rq->cpu; | |
10585 | ||
10586 | if (!rq->has_blocked_load) | |
10587 | return false; | |
10588 | ||
10589 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) | |
10590 | return false; | |
10591 | ||
10592 | if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) | |
10593 | return true; | |
10594 | ||
10595 | update_blocked_averages(cpu); | |
10596 | ||
10597 | return rq->has_blocked_load; | |
10598 | } | |
10599 | ||
1e3c88bd | 10600 | /* |
31e77c93 VG |
10601 | * Internal function that runs load balance for all idle cpus. The load balance |
10602 | * can be a simple update of blocked load or a complete load balance with | |
10603 | * tasks movement depending of flags. | |
1e3c88bd | 10604 | */ |
ab2dde5e | 10605 | static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags, |
31e77c93 | 10606 | enum cpu_idle_type idle) |
83cd4fe2 | 10607 | { |
c5afb6a8 | 10608 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
10609 | unsigned long now = jiffies; |
10610 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 10611 | bool has_blocked_load = false; |
c5afb6a8 | 10612 | int update_next_balance = 0; |
b7031a02 | 10613 | int this_cpu = this_rq->cpu; |
b7031a02 PZ |
10614 | int balance_cpu; |
10615 | struct rq *rq; | |
83cd4fe2 | 10616 | |
b7031a02 | 10617 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 10618 | |
f643ea22 VG |
10619 | /* |
10620 | * We assume there will be no idle load after this update and clear | |
10621 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
10622 | * set the has_blocked flag and trig another update of idle load. | |
10623 | * Because a cpu that becomes idle, is added to idle_cpus_mask before | |
10624 | * setting the flag, we are sure to not clear the state and not | |
10625 | * check the load of an idle cpu. | |
10626 | */ | |
10627 | WRITE_ONCE(nohz.has_blocked, 0); | |
10628 | ||
10629 | /* | |
10630 | * Ensures that if we miss the CPU, we must see the has_blocked | |
10631 | * store from nohz_balance_enter_idle(). | |
10632 | */ | |
10633 | smp_mb(); | |
10634 | ||
7a82e5f5 VG |
10635 | /* |
10636 | * Start with the next CPU after this_cpu so we will end with this_cpu and let a | |
10637 | * chance for other idle cpu to pull load. | |
10638 | */ | |
10639 | for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) { | |
10640 | if (!idle_cpu(balance_cpu)) | |
83cd4fe2 VP |
10641 | continue; |
10642 | ||
10643 | /* | |
97fb7a0a IM |
10644 | * If this CPU gets work to do, stop the load balancing |
10645 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
10646 | * balancing owner will pick it up. |
10647 | */ | |
f643ea22 VG |
10648 | if (need_resched()) { |
10649 | has_blocked_load = true; | |
10650 | goto abort; | |
10651 | } | |
83cd4fe2 | 10652 | |
5ed4f1d9 VG |
10653 | rq = cpu_rq(balance_cpu); |
10654 | ||
64f84f27 | 10655 | has_blocked_load |= update_nohz_stats(rq); |
f643ea22 | 10656 | |
ed61bbc6 TC |
10657 | /* |
10658 | * If time for next balance is due, | |
10659 | * do the balance. | |
10660 | */ | |
10661 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
10662 | struct rq_flags rf; |
10663 | ||
31e77c93 | 10664 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 10665 | update_rq_clock(rq); |
31e77c93 | 10666 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 10667 | |
b7031a02 PZ |
10668 | if (flags & NOHZ_BALANCE_KICK) |
10669 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 10670 | } |
83cd4fe2 | 10671 | |
c5afb6a8 VG |
10672 | if (time_after(next_balance, rq->next_balance)) { |
10673 | next_balance = rq->next_balance; | |
10674 | update_next_balance = 1; | |
10675 | } | |
83cd4fe2 | 10676 | } |
c5afb6a8 | 10677 | |
3ea2f097 VG |
10678 | /* |
10679 | * next_balance will be updated only when there is a need. | |
10680 | * When the CPU is attached to null domain for ex, it will not be | |
10681 | * updated. | |
10682 | */ | |
10683 | if (likely(update_next_balance)) | |
10684 | nohz.next_balance = next_balance; | |
10685 | ||
f643ea22 VG |
10686 | WRITE_ONCE(nohz.next_blocked, |
10687 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
10688 | ||
10689 | abort: | |
10690 | /* There is still blocked load, enable periodic update */ | |
10691 | if (has_blocked_load) | |
10692 | WRITE_ONCE(nohz.has_blocked, 1); | |
31e77c93 VG |
10693 | } |
10694 | ||
10695 | /* | |
10696 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
10697 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
10698 | */ | |
10699 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
10700 | { | |
19a1f5ec | 10701 | unsigned int flags = this_rq->nohz_idle_balance; |
31e77c93 | 10702 | |
19a1f5ec | 10703 | if (!flags) |
31e77c93 VG |
10704 | return false; |
10705 | ||
19a1f5ec | 10706 | this_rq->nohz_idle_balance = 0; |
31e77c93 | 10707 | |
19a1f5ec | 10708 | if (idle != CPU_IDLE) |
31e77c93 VG |
10709 | return false; |
10710 | ||
10711 | _nohz_idle_balance(this_rq, flags, idle); | |
10712 | ||
b7031a02 | 10713 | return true; |
83cd4fe2 | 10714 | } |
31e77c93 | 10715 | |
c6f88654 VG |
10716 | /* |
10717 | * Check if we need to run the ILB for updating blocked load before entering | |
10718 | * idle state. | |
10719 | */ | |
10720 | void nohz_run_idle_balance(int cpu) | |
10721 | { | |
10722 | unsigned int flags; | |
10723 | ||
10724 | flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu)); | |
10725 | ||
10726 | /* | |
10727 | * Update the blocked load only if no SCHED_SOFTIRQ is about to happen | |
10728 | * (ie NOHZ_STATS_KICK set) and will do the same. | |
10729 | */ | |
10730 | if ((flags == NOHZ_NEWILB_KICK) && !need_resched()) | |
10731 | _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE); | |
10732 | } | |
10733 | ||
31e77c93 VG |
10734 | static void nohz_newidle_balance(struct rq *this_rq) |
10735 | { | |
10736 | int this_cpu = this_rq->cpu; | |
10737 | ||
10738 | /* | |
10739 | * This CPU doesn't want to be disturbed by scheduler | |
10740 | * housekeeping | |
10741 | */ | |
10742 | if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) | |
10743 | return; | |
10744 | ||
10745 | /* Will wake up very soon. No time for doing anything else*/ | |
10746 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
10747 | return; | |
10748 | ||
10749 | /* Don't need to update blocked load of idle CPUs*/ | |
10750 | if (!READ_ONCE(nohz.has_blocked) || | |
10751 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
10752 | return; | |
10753 | ||
31e77c93 | 10754 | /* |
c6f88654 VG |
10755 | * Set the need to trigger ILB in order to update blocked load |
10756 | * before entering idle state. | |
31e77c93 | 10757 | */ |
c6f88654 | 10758 | atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu)); |
31e77c93 VG |
10759 | } |
10760 | ||
dd707247 PZ |
10761 | #else /* !CONFIG_NO_HZ_COMMON */ |
10762 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
10763 | ||
31e77c93 | 10764 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
10765 | { |
10766 | return false; | |
10767 | } | |
31e77c93 VG |
10768 | |
10769 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 10770 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 10771 | |
47ea5412 | 10772 | /* |
5b78f2dc | 10773 | * newidle_balance is called by schedule() if this_cpu is about to become |
47ea5412 | 10774 | * idle. Attempts to pull tasks from other CPUs. |
7277a34c PZ |
10775 | * |
10776 | * Returns: | |
10777 | * < 0 - we released the lock and there are !fair tasks present | |
10778 | * 0 - failed, no new tasks | |
10779 | * > 0 - success, new (fair) tasks present | |
47ea5412 | 10780 | */ |
d91cecc1 | 10781 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) |
47ea5412 PZ |
10782 | { |
10783 | unsigned long next_balance = jiffies + HZ; | |
10784 | int this_cpu = this_rq->cpu; | |
10785 | struct sched_domain *sd; | |
10786 | int pulled_task = 0; | |
10787 | u64 curr_cost = 0; | |
10788 | ||
5ba553ef | 10789 | update_misfit_status(NULL, this_rq); |
e5e678e4 RR |
10790 | |
10791 | /* | |
10792 | * There is a task waiting to run. No need to search for one. | |
10793 | * Return 0; the task will be enqueued when switching to idle. | |
10794 | */ | |
10795 | if (this_rq->ttwu_pending) | |
10796 | return 0; | |
10797 | ||
47ea5412 PZ |
10798 | /* |
10799 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
10800 | * measure the duration of idle_balance() as idle time. | |
10801 | */ | |
10802 | this_rq->idle_stamp = rq_clock(this_rq); | |
10803 | ||
10804 | /* | |
10805 | * Do not pull tasks towards !active CPUs... | |
10806 | */ | |
10807 | if (!cpu_active(this_cpu)) | |
10808 | return 0; | |
10809 | ||
10810 | /* | |
10811 | * This is OK, because current is on_cpu, which avoids it being picked | |
10812 | * for load-balance and preemption/IRQs are still disabled avoiding | |
10813 | * further scheduler activity on it and we're being very careful to | |
10814 | * re-start the picking loop. | |
10815 | */ | |
10816 | rq_unpin_lock(this_rq, rf); | |
10817 | ||
10818 | if (this_rq->avg_idle < sysctl_sched_migration_cost || | |
e90c8fe1 | 10819 | !READ_ONCE(this_rq->rd->overload)) { |
31e77c93 | 10820 | |
47ea5412 PZ |
10821 | rcu_read_lock(); |
10822 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
10823 | if (sd) | |
10824 | update_next_balance(sd, &next_balance); | |
10825 | rcu_read_unlock(); | |
10826 | ||
10827 | goto out; | |
10828 | } | |
10829 | ||
5cb9eaa3 | 10830 | raw_spin_rq_unlock(this_rq); |
47ea5412 PZ |
10831 | |
10832 | update_blocked_averages(this_cpu); | |
10833 | rcu_read_lock(); | |
10834 | for_each_domain(this_cpu, sd) { | |
10835 | int continue_balancing = 1; | |
10836 | u64 t0, domain_cost; | |
10837 | ||
47ea5412 PZ |
10838 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
10839 | update_next_balance(sd, &next_balance); | |
10840 | break; | |
10841 | } | |
10842 | ||
10843 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
10844 | t0 = sched_clock_cpu(this_cpu); | |
10845 | ||
10846 | pulled_task = load_balance(this_cpu, this_rq, | |
10847 | sd, CPU_NEWLY_IDLE, | |
10848 | &continue_balancing); | |
10849 | ||
10850 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
10851 | if (domain_cost > sd->max_newidle_lb_cost) | |
10852 | sd->max_newidle_lb_cost = domain_cost; | |
10853 | ||
10854 | curr_cost += domain_cost; | |
10855 | } | |
10856 | ||
10857 | update_next_balance(sd, &next_balance); | |
10858 | ||
10859 | /* | |
10860 | * Stop searching for tasks to pull if there are | |
10861 | * now runnable tasks on this rq. | |
10862 | */ | |
e5e678e4 RR |
10863 | if (pulled_task || this_rq->nr_running > 0 || |
10864 | this_rq->ttwu_pending) | |
47ea5412 PZ |
10865 | break; |
10866 | } | |
10867 | rcu_read_unlock(); | |
10868 | ||
5cb9eaa3 | 10869 | raw_spin_rq_lock(this_rq); |
47ea5412 PZ |
10870 | |
10871 | if (curr_cost > this_rq->max_idle_balance_cost) | |
10872 | this_rq->max_idle_balance_cost = curr_cost; | |
10873 | ||
10874 | /* | |
10875 | * While browsing the domains, we released the rq lock, a task could | |
10876 | * have been enqueued in the meantime. Since we're not going idle, | |
10877 | * pretend we pulled a task. | |
10878 | */ | |
10879 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
10880 | pulled_task = 1; | |
10881 | ||
47ea5412 PZ |
10882 | /* Is there a task of a high priority class? */ |
10883 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
10884 | pulled_task = -1; | |
10885 | ||
6553fc18 VG |
10886 | out: |
10887 | /* Move the next balance forward */ | |
10888 | if (time_after(this_rq->next_balance, next_balance)) | |
10889 | this_rq->next_balance = next_balance; | |
10890 | ||
47ea5412 PZ |
10891 | if (pulled_task) |
10892 | this_rq->idle_stamp = 0; | |
0826530d VG |
10893 | else |
10894 | nohz_newidle_balance(this_rq); | |
47ea5412 PZ |
10895 | |
10896 | rq_repin_lock(this_rq, rf); | |
10897 | ||
10898 | return pulled_task; | |
10899 | } | |
10900 | ||
83cd4fe2 VP |
10901 | /* |
10902 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
10903 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
10904 | */ | |
0766f788 | 10905 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 10906 | { |
208cb16b | 10907 | struct rq *this_rq = this_rq(); |
6eb57e0d | 10908 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
10909 | CPU_IDLE : CPU_NOT_IDLE; |
10910 | ||
1e3c88bd | 10911 | /* |
97fb7a0a IM |
10912 | * If this CPU has a pending nohz_balance_kick, then do the |
10913 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 10914 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 10915 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
10916 | * load balance only within the local sched_domain hierarchy |
10917 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 10918 | */ |
b7031a02 PZ |
10919 | if (nohz_idle_balance(this_rq, idle)) |
10920 | return; | |
10921 | ||
10922 | /* normal load balance */ | |
10923 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 10924 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
10925 | } |
10926 | ||
1e3c88bd PZ |
10927 | /* |
10928 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 10929 | */ |
7caff66f | 10930 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 10931 | { |
e0b257c3 AMB |
10932 | /* |
10933 | * Don't need to rebalance while attached to NULL domain or | |
10934 | * runqueue CPU is not active | |
10935 | */ | |
10936 | if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq)))) | |
c726099e DL |
10937 | return; |
10938 | ||
10939 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 10940 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
10941 | |
10942 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
10943 | } |
10944 | ||
0bcdcf28 CE |
10945 | static void rq_online_fair(struct rq *rq) |
10946 | { | |
10947 | update_sysctl(); | |
0e59bdae KT |
10948 | |
10949 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
10950 | } |
10951 | ||
10952 | static void rq_offline_fair(struct rq *rq) | |
10953 | { | |
10954 | update_sysctl(); | |
a4c96ae3 PB |
10955 | |
10956 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
10957 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
10958 | } |
10959 | ||
55e12e5e | 10960 | #endif /* CONFIG_SMP */ |
e1d1484f | 10961 | |
8039e96f VP |
10962 | #ifdef CONFIG_SCHED_CORE |
10963 | static inline bool | |
10964 | __entity_slice_used(struct sched_entity *se, int min_nr_tasks) | |
10965 | { | |
10966 | u64 slice = sched_slice(cfs_rq_of(se), se); | |
10967 | u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
10968 | ||
10969 | return (rtime * min_nr_tasks > slice); | |
10970 | } | |
10971 | ||
10972 | #define MIN_NR_TASKS_DURING_FORCEIDLE 2 | |
10973 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) | |
10974 | { | |
10975 | if (!sched_core_enabled(rq)) | |
10976 | return; | |
10977 | ||
10978 | /* | |
10979 | * If runqueue has only one task which used up its slice and | |
10980 | * if the sibling is forced idle, then trigger schedule to | |
10981 | * give forced idle task a chance. | |
10982 | * | |
10983 | * sched_slice() considers only this active rq and it gets the | |
10984 | * whole slice. But during force idle, we have siblings acting | |
10985 | * like a single runqueue and hence we need to consider runnable | |
cc00c198 | 10986 | * tasks on this CPU and the forced idle CPU. Ideally, we should |
8039e96f | 10987 | * go through the forced idle rq, but that would be a perf hit. |
cc00c198 | 10988 | * We can assume that the forced idle CPU has at least |
8039e96f | 10989 | * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check |
cc00c198 | 10990 | * if we need to give up the CPU. |
8039e96f VP |
10991 | */ |
10992 | if (rq->core->core_forceidle && rq->cfs.nr_running == 1 && | |
10993 | __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE)) | |
10994 | resched_curr(rq); | |
10995 | } | |
c6047c2e JFG |
10996 | |
10997 | /* | |
10998 | * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed. | |
10999 | */ | |
11000 | static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle) | |
11001 | { | |
11002 | for_each_sched_entity(se) { | |
11003 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11004 | ||
11005 | if (forceidle) { | |
11006 | if (cfs_rq->forceidle_seq == fi_seq) | |
11007 | break; | |
11008 | cfs_rq->forceidle_seq = fi_seq; | |
11009 | } | |
11010 | ||
11011 | cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime; | |
11012 | } | |
11013 | } | |
11014 | ||
11015 | void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi) | |
11016 | { | |
11017 | struct sched_entity *se = &p->se; | |
11018 | ||
11019 | if (p->sched_class != &fair_sched_class) | |
11020 | return; | |
11021 | ||
11022 | se_fi_update(se, rq->core->core_forceidle_seq, in_fi); | |
11023 | } | |
11024 | ||
11025 | bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi) | |
11026 | { | |
11027 | struct rq *rq = task_rq(a); | |
11028 | struct sched_entity *sea = &a->se; | |
11029 | struct sched_entity *seb = &b->se; | |
11030 | struct cfs_rq *cfs_rqa; | |
11031 | struct cfs_rq *cfs_rqb; | |
11032 | s64 delta; | |
11033 | ||
11034 | SCHED_WARN_ON(task_rq(b)->core != rq->core); | |
11035 | ||
11036 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
11037 | /* | |
11038 | * Find an se in the hierarchy for tasks a and b, such that the se's | |
11039 | * are immediate siblings. | |
11040 | */ | |
11041 | while (sea->cfs_rq->tg != seb->cfs_rq->tg) { | |
11042 | int sea_depth = sea->depth; | |
11043 | int seb_depth = seb->depth; | |
11044 | ||
11045 | if (sea_depth >= seb_depth) | |
11046 | sea = parent_entity(sea); | |
11047 | if (sea_depth <= seb_depth) | |
11048 | seb = parent_entity(seb); | |
11049 | } | |
11050 | ||
11051 | se_fi_update(sea, rq->core->core_forceidle_seq, in_fi); | |
11052 | se_fi_update(seb, rq->core->core_forceidle_seq, in_fi); | |
11053 | ||
11054 | cfs_rqa = sea->cfs_rq; | |
11055 | cfs_rqb = seb->cfs_rq; | |
11056 | #else | |
11057 | cfs_rqa = &task_rq(a)->cfs; | |
11058 | cfs_rqb = &task_rq(b)->cfs; | |
11059 | #endif | |
11060 | ||
11061 | /* | |
11062 | * Find delta after normalizing se's vruntime with its cfs_rq's | |
11063 | * min_vruntime_fi, which would have been updated in prior calls | |
11064 | * to se_fi_update(). | |
11065 | */ | |
11066 | delta = (s64)(sea->vruntime - seb->vruntime) + | |
11067 | (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi); | |
11068 | ||
11069 | return delta > 0; | |
11070 | } | |
8039e96f VP |
11071 | #else |
11072 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {} | |
11073 | #endif | |
11074 | ||
bf0f6f24 | 11075 | /* |
d84b3131 FW |
11076 | * scheduler tick hitting a task of our scheduling class. |
11077 | * | |
11078 | * NOTE: This function can be called remotely by the tick offload that | |
11079 | * goes along full dynticks. Therefore no local assumption can be made | |
11080 | * and everything must be accessed through the @rq and @curr passed in | |
11081 | * parameters. | |
bf0f6f24 | 11082 | */ |
8f4d37ec | 11083 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
11084 | { |
11085 | struct cfs_rq *cfs_rq; | |
11086 | struct sched_entity *se = &curr->se; | |
11087 | ||
11088 | for_each_sched_entity(se) { | |
11089 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 11090 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 11091 | } |
18bf2805 | 11092 | |
b52da86e | 11093 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 11094 | task_tick_numa(rq, curr); |
3b1baa64 MR |
11095 | |
11096 | update_misfit_status(curr, rq); | |
2802bf3c | 11097 | update_overutilized_status(task_rq(curr)); |
8039e96f VP |
11098 | |
11099 | task_tick_core(rq, curr); | |
bf0f6f24 IM |
11100 | } |
11101 | ||
11102 | /* | |
cd29fe6f PZ |
11103 | * called on fork with the child task as argument from the parent's context |
11104 | * - child not yet on the tasklist | |
11105 | * - preemption disabled | |
bf0f6f24 | 11106 | */ |
cd29fe6f | 11107 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 11108 | { |
4fc420c9 DN |
11109 | struct cfs_rq *cfs_rq; |
11110 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 11111 | struct rq *rq = this_rq(); |
8a8c69c3 | 11112 | struct rq_flags rf; |
bf0f6f24 | 11113 | |
8a8c69c3 | 11114 | rq_lock(rq, &rf); |
861d034e PZ |
11115 | update_rq_clock(rq); |
11116 | ||
4fc420c9 DN |
11117 | cfs_rq = task_cfs_rq(current); |
11118 | curr = cfs_rq->curr; | |
e210bffd PZ |
11119 | if (curr) { |
11120 | update_curr(cfs_rq); | |
b5d9d734 | 11121 | se->vruntime = curr->vruntime; |
e210bffd | 11122 | } |
aeb73b04 | 11123 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 11124 | |
cd29fe6f | 11125 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 11126 | /* |
edcb60a3 IM |
11127 | * Upon rescheduling, sched_class::put_prev_task() will place |
11128 | * 'current' within the tree based on its new key value. | |
11129 | */ | |
4d78e7b6 | 11130 | swap(curr->vruntime, se->vruntime); |
8875125e | 11131 | resched_curr(rq); |
4d78e7b6 | 11132 | } |
bf0f6f24 | 11133 | |
88ec22d3 | 11134 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 11135 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
11136 | } |
11137 | ||
cb469845 SR |
11138 | /* |
11139 | * Priority of the task has changed. Check to see if we preempt | |
11140 | * the current task. | |
11141 | */ | |
da7a735e PZ |
11142 | static void |
11143 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 11144 | { |
da0c1e65 | 11145 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
11146 | return; |
11147 | ||
7c2e8bbd FW |
11148 | if (rq->cfs.nr_running == 1) |
11149 | return; | |
11150 | ||
cb469845 SR |
11151 | /* |
11152 | * Reschedule if we are currently running on this runqueue and | |
11153 | * our priority decreased, or if we are not currently running on | |
11154 | * this runqueue and our priority is higher than the current's | |
11155 | */ | |
65bcf072 | 11156 | if (task_current(rq, p)) { |
cb469845 | 11157 | if (p->prio > oldprio) |
8875125e | 11158 | resched_curr(rq); |
cb469845 | 11159 | } else |
15afe09b | 11160 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
11161 | } |
11162 | ||
daa59407 | 11163 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
11164 | { |
11165 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
11166 | |
11167 | /* | |
daa59407 BP |
11168 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
11169 | * the dequeue_entity(.flags=0) will already have normalized the | |
11170 | * vruntime. | |
11171 | */ | |
11172 | if (p->on_rq) | |
11173 | return true; | |
11174 | ||
11175 | /* | |
11176 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
11177 | * But there are some cases where it has already been normalized: | |
da7a735e | 11178 | * |
daa59407 BP |
11179 | * - A forked child which is waiting for being woken up by |
11180 | * wake_up_new_task(). | |
11181 | * - A task which has been woken up by try_to_wake_up() and | |
11182 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 11183 | */ |
d0cdb3ce | 11184 | if (!se->sum_exec_runtime || |
2f064a59 | 11185 | (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup)) |
daa59407 BP |
11186 | return true; |
11187 | ||
11188 | return false; | |
11189 | } | |
11190 | ||
09a43ace VG |
11191 | #ifdef CONFIG_FAIR_GROUP_SCHED |
11192 | /* | |
11193 | * Propagate the changes of the sched_entity across the tg tree to make it | |
11194 | * visible to the root | |
11195 | */ | |
11196 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
11197 | { | |
11198 | struct cfs_rq *cfs_rq; | |
11199 | ||
0258bdfa OU |
11200 | list_add_leaf_cfs_rq(cfs_rq_of(se)); |
11201 | ||
09a43ace VG |
11202 | /* Start to propagate at parent */ |
11203 | se = se->parent; | |
11204 | ||
11205 | for_each_sched_entity(se) { | |
11206 | cfs_rq = cfs_rq_of(se); | |
11207 | ||
0258bdfa OU |
11208 | if (!cfs_rq_throttled(cfs_rq)){ |
11209 | update_load_avg(cfs_rq, se, UPDATE_TG); | |
11210 | list_add_leaf_cfs_rq(cfs_rq); | |
11211 | continue; | |
11212 | } | |
09a43ace | 11213 | |
0258bdfa OU |
11214 | if (list_add_leaf_cfs_rq(cfs_rq)) |
11215 | break; | |
09a43ace VG |
11216 | } |
11217 | } | |
11218 | #else | |
11219 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
11220 | #endif | |
11221 | ||
df217913 | 11222 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 11223 | { |
daa59407 BP |
11224 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
11225 | ||
9d89c257 | 11226 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 11227 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 11228 | detach_entity_load_avg(cfs_rq, se); |
fe749158 | 11229 | update_tg_load_avg(cfs_rq); |
09a43ace | 11230 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
11231 | } |
11232 | ||
df217913 | 11233 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 11234 | { |
daa59407 | 11235 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
11236 | |
11237 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
11238 | /* |
11239 | * Since the real-depth could have been changed (only FAIR | |
11240 | * class maintain depth value), reset depth properly. | |
11241 | */ | |
11242 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
11243 | #endif | |
7855a35a | 11244 | |
df217913 | 11245 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 11246 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
a4f9a0e5 | 11247 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 11248 | update_tg_load_avg(cfs_rq); |
09a43ace | 11249 | propagate_entity_cfs_rq(se); |
df217913 VG |
11250 | } |
11251 | ||
11252 | static void detach_task_cfs_rq(struct task_struct *p) | |
11253 | { | |
11254 | struct sched_entity *se = &p->se; | |
11255 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11256 | ||
11257 | if (!vruntime_normalized(p)) { | |
11258 | /* | |
11259 | * Fix up our vruntime so that the current sleep doesn't | |
11260 | * cause 'unlimited' sleep bonus. | |
11261 | */ | |
11262 | place_entity(cfs_rq, se, 0); | |
11263 | se->vruntime -= cfs_rq->min_vruntime; | |
11264 | } | |
11265 | ||
11266 | detach_entity_cfs_rq(se); | |
11267 | } | |
11268 | ||
11269 | static void attach_task_cfs_rq(struct task_struct *p) | |
11270 | { | |
11271 | struct sched_entity *se = &p->se; | |
11272 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11273 | ||
11274 | attach_entity_cfs_rq(se); | |
daa59407 BP |
11275 | |
11276 | if (!vruntime_normalized(p)) | |
11277 | se->vruntime += cfs_rq->min_vruntime; | |
11278 | } | |
6efdb105 | 11279 | |
daa59407 BP |
11280 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
11281 | { | |
11282 | detach_task_cfs_rq(p); | |
11283 | } | |
11284 | ||
11285 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
11286 | { | |
11287 | attach_task_cfs_rq(p); | |
7855a35a | 11288 | |
daa59407 | 11289 | if (task_on_rq_queued(p)) { |
7855a35a | 11290 | /* |
daa59407 BP |
11291 | * We were most likely switched from sched_rt, so |
11292 | * kick off the schedule if running, otherwise just see | |
11293 | * if we can still preempt the current task. | |
7855a35a | 11294 | */ |
65bcf072 | 11295 | if (task_current(rq, p)) |
daa59407 BP |
11296 | resched_curr(rq); |
11297 | else | |
11298 | check_preempt_curr(rq, p, 0); | |
7855a35a | 11299 | } |
cb469845 SR |
11300 | } |
11301 | ||
83b699ed SV |
11302 | /* Account for a task changing its policy or group. |
11303 | * | |
11304 | * This routine is mostly called to set cfs_rq->curr field when a task | |
11305 | * migrates between groups/classes. | |
11306 | */ | |
a0e813f2 | 11307 | static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) |
83b699ed | 11308 | { |
03b7fad1 PZ |
11309 | struct sched_entity *se = &p->se; |
11310 | ||
11311 | #ifdef CONFIG_SMP | |
11312 | if (task_on_rq_queued(p)) { | |
11313 | /* | |
11314 | * Move the next running task to the front of the list, so our | |
11315 | * cfs_tasks list becomes MRU one. | |
11316 | */ | |
11317 | list_move(&se->group_node, &rq->cfs_tasks); | |
11318 | } | |
11319 | #endif | |
83b699ed | 11320 | |
ec12cb7f PT |
11321 | for_each_sched_entity(se) { |
11322 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11323 | ||
11324 | set_next_entity(cfs_rq, se); | |
11325 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
11326 | account_cfs_rq_runtime(cfs_rq, 0); | |
11327 | } | |
83b699ed SV |
11328 | } |
11329 | ||
029632fb PZ |
11330 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
11331 | { | |
bfb06889 | 11332 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
029632fb PZ |
11333 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
11334 | #ifndef CONFIG_64BIT | |
11335 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
11336 | #endif | |
141965c7 | 11337 | #ifdef CONFIG_SMP |
2a2f5d4e | 11338 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 11339 | #endif |
029632fb PZ |
11340 | } |
11341 | ||
810b3817 | 11342 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
11343 | static void task_set_group_fair(struct task_struct *p) |
11344 | { | |
11345 | struct sched_entity *se = &p->se; | |
11346 | ||
11347 | set_task_rq(p, task_cpu(p)); | |
11348 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
11349 | } | |
11350 | ||
bc54da21 | 11351 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 11352 | { |
daa59407 | 11353 | detach_task_cfs_rq(p); |
b2b5ce02 | 11354 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
11355 | |
11356 | #ifdef CONFIG_SMP | |
11357 | /* Tell se's cfs_rq has been changed -- migrated */ | |
11358 | p->se.avg.last_update_time = 0; | |
11359 | #endif | |
daa59407 | 11360 | attach_task_cfs_rq(p); |
810b3817 | 11361 | } |
029632fb | 11362 | |
ea86cb4b VG |
11363 | static void task_change_group_fair(struct task_struct *p, int type) |
11364 | { | |
11365 | switch (type) { | |
11366 | case TASK_SET_GROUP: | |
11367 | task_set_group_fair(p); | |
11368 | break; | |
11369 | ||
11370 | case TASK_MOVE_GROUP: | |
11371 | task_move_group_fair(p); | |
11372 | break; | |
11373 | } | |
11374 | } | |
11375 | ||
029632fb PZ |
11376 | void free_fair_sched_group(struct task_group *tg) |
11377 | { | |
11378 | int i; | |
11379 | ||
029632fb PZ |
11380 | for_each_possible_cpu(i) { |
11381 | if (tg->cfs_rq) | |
11382 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 11383 | if (tg->se) |
029632fb PZ |
11384 | kfree(tg->se[i]); |
11385 | } | |
11386 | ||
11387 | kfree(tg->cfs_rq); | |
11388 | kfree(tg->se); | |
11389 | } | |
11390 | ||
11391 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11392 | { | |
029632fb | 11393 | struct sched_entity *se; |
b7fa30c9 | 11394 | struct cfs_rq *cfs_rq; |
029632fb PZ |
11395 | int i; |
11396 | ||
6396bb22 | 11397 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
11398 | if (!tg->cfs_rq) |
11399 | goto err; | |
6396bb22 | 11400 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
11401 | if (!tg->se) |
11402 | goto err; | |
11403 | ||
11404 | tg->shares = NICE_0_LOAD; | |
11405 | ||
11406 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
11407 | ||
11408 | for_each_possible_cpu(i) { | |
11409 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
11410 | GFP_KERNEL, cpu_to_node(i)); | |
11411 | if (!cfs_rq) | |
11412 | goto err; | |
11413 | ||
11414 | se = kzalloc_node(sizeof(struct sched_entity), | |
11415 | GFP_KERNEL, cpu_to_node(i)); | |
11416 | if (!se) | |
11417 | goto err_free_rq; | |
11418 | ||
11419 | init_cfs_rq(cfs_rq); | |
11420 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 11421 | init_entity_runnable_average(se); |
029632fb PZ |
11422 | } |
11423 | ||
11424 | return 1; | |
11425 | ||
11426 | err_free_rq: | |
11427 | kfree(cfs_rq); | |
11428 | err: | |
11429 | return 0; | |
11430 | } | |
11431 | ||
8663e24d PZ |
11432 | void online_fair_sched_group(struct task_group *tg) |
11433 | { | |
11434 | struct sched_entity *se; | |
a46d14ec | 11435 | struct rq_flags rf; |
8663e24d PZ |
11436 | struct rq *rq; |
11437 | int i; | |
11438 | ||
11439 | for_each_possible_cpu(i) { | |
11440 | rq = cpu_rq(i); | |
11441 | se = tg->se[i]; | |
a46d14ec | 11442 | rq_lock_irq(rq, &rf); |
4126bad6 | 11443 | update_rq_clock(rq); |
d0326691 | 11444 | attach_entity_cfs_rq(se); |
55e16d30 | 11445 | sync_throttle(tg, i); |
a46d14ec | 11446 | rq_unlock_irq(rq, &rf); |
8663e24d PZ |
11447 | } |
11448 | } | |
11449 | ||
6fe1f348 | 11450 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 11451 | { |
029632fb | 11452 | unsigned long flags; |
6fe1f348 PZ |
11453 | struct rq *rq; |
11454 | int cpu; | |
029632fb | 11455 | |
b4b2d765 MK |
11456 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); |
11457 | ||
6fe1f348 PZ |
11458 | for_each_possible_cpu(cpu) { |
11459 | if (tg->se[cpu]) | |
11460 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 11461 | |
6fe1f348 PZ |
11462 | /* |
11463 | * Only empty task groups can be destroyed; so we can speculatively | |
11464 | * check on_list without danger of it being re-added. | |
11465 | */ | |
11466 | if (!tg->cfs_rq[cpu]->on_list) | |
11467 | continue; | |
11468 | ||
11469 | rq = cpu_rq(cpu); | |
11470 | ||
5cb9eaa3 | 11471 | raw_spin_rq_lock_irqsave(rq, flags); |
6fe1f348 | 11472 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); |
5cb9eaa3 | 11473 | raw_spin_rq_unlock_irqrestore(rq, flags); |
6fe1f348 | 11474 | } |
029632fb PZ |
11475 | } |
11476 | ||
11477 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
11478 | struct sched_entity *se, int cpu, | |
11479 | struct sched_entity *parent) | |
11480 | { | |
11481 | struct rq *rq = cpu_rq(cpu); | |
11482 | ||
11483 | cfs_rq->tg = tg; | |
11484 | cfs_rq->rq = rq; | |
029632fb PZ |
11485 | init_cfs_rq_runtime(cfs_rq); |
11486 | ||
11487 | tg->cfs_rq[cpu] = cfs_rq; | |
11488 | tg->se[cpu] = se; | |
11489 | ||
11490 | /* se could be NULL for root_task_group */ | |
11491 | if (!se) | |
11492 | return; | |
11493 | ||
fed14d45 | 11494 | if (!parent) { |
029632fb | 11495 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
11496 | se->depth = 0; |
11497 | } else { | |
029632fb | 11498 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
11499 | se->depth = parent->depth + 1; |
11500 | } | |
029632fb PZ |
11501 | |
11502 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
11503 | /* guarantee group entities always have weight */ |
11504 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
11505 | se->parent = parent; |
11506 | } | |
11507 | ||
11508 | static DEFINE_MUTEX(shares_mutex); | |
11509 | ||
30400039 | 11510 | static int __sched_group_set_shares(struct task_group *tg, unsigned long shares) |
029632fb PZ |
11511 | { |
11512 | int i; | |
029632fb | 11513 | |
30400039 JD |
11514 | lockdep_assert_held(&shares_mutex); |
11515 | ||
029632fb PZ |
11516 | /* |
11517 | * We can't change the weight of the root cgroup. | |
11518 | */ | |
11519 | if (!tg->se[0]) | |
11520 | return -EINVAL; | |
11521 | ||
11522 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
11523 | ||
029632fb | 11524 | if (tg->shares == shares) |
30400039 | 11525 | return 0; |
029632fb PZ |
11526 | |
11527 | tg->shares = shares; | |
11528 | for_each_possible_cpu(i) { | |
11529 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
11530 | struct sched_entity *se = tg->se[i]; |
11531 | struct rq_flags rf; | |
029632fb | 11532 | |
029632fb | 11533 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 11534 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 11535 | update_rq_clock(rq); |
89ee048f | 11536 | for_each_sched_entity(se) { |
88c0616e | 11537 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 11538 | update_cfs_group(se); |
89ee048f | 11539 | } |
8a8c69c3 | 11540 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
11541 | } |
11542 | ||
30400039 JD |
11543 | return 0; |
11544 | } | |
11545 | ||
11546 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
11547 | { | |
11548 | int ret; | |
11549 | ||
11550 | mutex_lock(&shares_mutex); | |
11551 | if (tg_is_idle(tg)) | |
11552 | ret = -EINVAL; | |
11553 | else | |
11554 | ret = __sched_group_set_shares(tg, shares); | |
11555 | mutex_unlock(&shares_mutex); | |
11556 | ||
11557 | return ret; | |
11558 | } | |
11559 | ||
11560 | int sched_group_set_idle(struct task_group *tg, long idle) | |
11561 | { | |
11562 | int i; | |
11563 | ||
11564 | if (tg == &root_task_group) | |
11565 | return -EINVAL; | |
11566 | ||
11567 | if (idle < 0 || idle > 1) | |
11568 | return -EINVAL; | |
11569 | ||
11570 | mutex_lock(&shares_mutex); | |
11571 | ||
11572 | if (tg->idle == idle) { | |
11573 | mutex_unlock(&shares_mutex); | |
11574 | return 0; | |
11575 | } | |
11576 | ||
11577 | tg->idle = idle; | |
11578 | ||
11579 | for_each_possible_cpu(i) { | |
11580 | struct rq *rq = cpu_rq(i); | |
11581 | struct sched_entity *se = tg->se[i]; | |
11582 | struct cfs_rq *grp_cfs_rq = tg->cfs_rq[i]; | |
11583 | bool was_idle = cfs_rq_is_idle(grp_cfs_rq); | |
11584 | long idle_task_delta; | |
11585 | struct rq_flags rf; | |
11586 | ||
11587 | rq_lock_irqsave(rq, &rf); | |
11588 | ||
11589 | grp_cfs_rq->idle = idle; | |
11590 | if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq))) | |
11591 | goto next_cpu; | |
11592 | ||
11593 | idle_task_delta = grp_cfs_rq->h_nr_running - | |
11594 | grp_cfs_rq->idle_h_nr_running; | |
11595 | if (!cfs_rq_is_idle(grp_cfs_rq)) | |
11596 | idle_task_delta *= -1; | |
11597 | ||
11598 | for_each_sched_entity(se) { | |
11599 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11600 | ||
11601 | if (!se->on_rq) | |
11602 | break; | |
11603 | ||
11604 | cfs_rq->idle_h_nr_running += idle_task_delta; | |
11605 | ||
11606 | /* Already accounted at parent level and above. */ | |
11607 | if (cfs_rq_is_idle(cfs_rq)) | |
11608 | break; | |
11609 | } | |
11610 | ||
11611 | next_cpu: | |
11612 | rq_unlock_irqrestore(rq, &rf); | |
11613 | } | |
11614 | ||
11615 | /* Idle groups have minimum weight. */ | |
11616 | if (tg_is_idle(tg)) | |
11617 | __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO)); | |
11618 | else | |
11619 | __sched_group_set_shares(tg, NICE_0_LOAD); | |
11620 | ||
029632fb PZ |
11621 | mutex_unlock(&shares_mutex); |
11622 | return 0; | |
11623 | } | |
30400039 | 11624 | |
029632fb PZ |
11625 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
11626 | ||
11627 | void free_fair_sched_group(struct task_group *tg) { } | |
11628 | ||
11629 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11630 | { | |
11631 | return 1; | |
11632 | } | |
11633 | ||
8663e24d PZ |
11634 | void online_fair_sched_group(struct task_group *tg) { } |
11635 | ||
6fe1f348 | 11636 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
11637 | |
11638 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
11639 | ||
810b3817 | 11640 | |
6d686f45 | 11641 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
11642 | { |
11643 | struct sched_entity *se = &task->se; | |
0d721cea PW |
11644 | unsigned int rr_interval = 0; |
11645 | ||
11646 | /* | |
11647 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
11648 | * idle runqueue: | |
11649 | */ | |
0d721cea | 11650 | if (rq->cfs.load.weight) |
a59f4e07 | 11651 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
11652 | |
11653 | return rr_interval; | |
11654 | } | |
11655 | ||
bf0f6f24 IM |
11656 | /* |
11657 | * All the scheduling class methods: | |
11658 | */ | |
43c31ac0 PZ |
11659 | DEFINE_SCHED_CLASS(fair) = { |
11660 | ||
bf0f6f24 IM |
11661 | .enqueue_task = enqueue_task_fair, |
11662 | .dequeue_task = dequeue_task_fair, | |
11663 | .yield_task = yield_task_fair, | |
d95f4122 | 11664 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 11665 | |
2e09bf55 | 11666 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 | 11667 | |
98c2f700 | 11668 | .pick_next_task = __pick_next_task_fair, |
bf0f6f24 | 11669 | .put_prev_task = put_prev_task_fair, |
03b7fad1 | 11670 | .set_next_task = set_next_task_fair, |
bf0f6f24 | 11671 | |
681f3e68 | 11672 | #ifdef CONFIG_SMP |
6e2df058 | 11673 | .balance = balance_fair, |
21f56ffe | 11674 | .pick_task = pick_task_fair, |
4ce72a2c | 11675 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 11676 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 11677 | |
0bcdcf28 CE |
11678 | .rq_online = rq_online_fair, |
11679 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 11680 | |
12695578 | 11681 | .task_dead = task_dead_fair, |
c5b28038 | 11682 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 11683 | #endif |
bf0f6f24 | 11684 | |
bf0f6f24 | 11685 | .task_tick = task_tick_fair, |
cd29fe6f | 11686 | .task_fork = task_fork_fair, |
cb469845 SR |
11687 | |
11688 | .prio_changed = prio_changed_fair, | |
da7a735e | 11689 | .switched_from = switched_from_fair, |
cb469845 | 11690 | .switched_to = switched_to_fair, |
810b3817 | 11691 | |
0d721cea PW |
11692 | .get_rr_interval = get_rr_interval_fair, |
11693 | ||
6e998916 SG |
11694 | .update_curr = update_curr_fair, |
11695 | ||
810b3817 | 11696 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 11697 | .task_change_group = task_change_group_fair, |
810b3817 | 11698 | #endif |
982d9cdc PB |
11699 | |
11700 | #ifdef CONFIG_UCLAMP_TASK | |
11701 | .uclamp_enabled = 1, | |
11702 | #endif | |
bf0f6f24 IM |
11703 | }; |
11704 | ||
11705 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 11706 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 11707 | { |
039ae8bc | 11708 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 11709 | |
5973e5b9 | 11710 | rcu_read_lock(); |
039ae8bc | 11711 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 11712 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 11713 | rcu_read_unlock(); |
bf0f6f24 | 11714 | } |
397f2378 SD |
11715 | |
11716 | #ifdef CONFIG_NUMA_BALANCING | |
11717 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
11718 | { | |
11719 | int node; | |
11720 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 11721 | struct numa_group *ng; |
397f2378 | 11722 | |
cb361d8c JH |
11723 | rcu_read_lock(); |
11724 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
11725 | for_each_online_node(node) { |
11726 | if (p->numa_faults) { | |
11727 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
11728 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
11729 | } | |
cb361d8c JH |
11730 | if (ng) { |
11731 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
11732 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
11733 | } |
11734 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
11735 | } | |
cb361d8c | 11736 | rcu_read_unlock(); |
397f2378 SD |
11737 | } |
11738 | #endif /* CONFIG_NUMA_BALANCING */ | |
11739 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
11740 | |
11741 | __init void init_sched_fair_class(void) | |
11742 | { | |
11743 | #ifdef CONFIG_SMP | |
11744 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
11745 | ||
3451d024 | 11746 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 11747 | nohz.next_balance = jiffies; |
f643ea22 | 11748 | nohz.next_blocked = jiffies; |
029632fb | 11749 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
11750 | #endif |
11751 | #endif /* SMP */ | |
11752 | ||
11753 | } | |
3c93a0c0 QY |
11754 | |
11755 | /* | |
11756 | * Helper functions to facilitate extracting info from tracepoints. | |
11757 | */ | |
11758 | ||
11759 | const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq) | |
11760 | { | |
11761 | #ifdef CONFIG_SMP | |
11762 | return cfs_rq ? &cfs_rq->avg : NULL; | |
11763 | #else | |
11764 | return NULL; | |
11765 | #endif | |
11766 | } | |
11767 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg); | |
11768 | ||
11769 | char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len) | |
11770 | { | |
11771 | if (!cfs_rq) { | |
11772 | if (str) | |
11773 | strlcpy(str, "(null)", len); | |
11774 | else | |
11775 | return NULL; | |
11776 | } | |
11777 | ||
11778 | cfs_rq_tg_path(cfs_rq, str, len); | |
11779 | return str; | |
11780 | } | |
11781 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path); | |
11782 | ||
11783 | int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq) | |
11784 | { | |
11785 | return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; | |
11786 | } | |
11787 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu); | |
11788 | ||
11789 | const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq) | |
11790 | { | |
11791 | #ifdef CONFIG_SMP | |
11792 | return rq ? &rq->avg_rt : NULL; | |
11793 | #else | |
11794 | return NULL; | |
11795 | #endif | |
11796 | } | |
11797 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt); | |
11798 | ||
11799 | const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq) | |
11800 | { | |
11801 | #ifdef CONFIG_SMP | |
11802 | return rq ? &rq->avg_dl : NULL; | |
11803 | #else | |
11804 | return NULL; | |
11805 | #endif | |
11806 | } | |
11807 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl); | |
11808 | ||
11809 | const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq) | |
11810 | { | |
11811 | #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ) | |
11812 | return rq ? &rq->avg_irq : NULL; | |
11813 | #else | |
11814 | return NULL; | |
11815 | #endif | |
11816 | } | |
11817 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq); | |
11818 | ||
11819 | int sched_trace_rq_cpu(struct rq *rq) | |
11820 | { | |
11821 | return rq ? cpu_of(rq) : -1; | |
11822 | } | |
11823 | EXPORT_SYMBOL_GPL(sched_trace_rq_cpu); | |
11824 | ||
51cf18c9 VD |
11825 | int sched_trace_rq_cpu_capacity(struct rq *rq) |
11826 | { | |
11827 | return rq ? | |
11828 | #ifdef CONFIG_SMP | |
11829 | rq->cpu_capacity | |
11830 | #else | |
11831 | SCHED_CAPACITY_SCALE | |
11832 | #endif | |
11833 | : -1; | |
11834 | } | |
11835 | EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity); | |
11836 | ||
3c93a0c0 QY |
11837 | const struct cpumask *sched_trace_rd_span(struct root_domain *rd) |
11838 | { | |
11839 | #ifdef CONFIG_SMP | |
11840 | return rd ? rd->span : NULL; | |
11841 | #else | |
11842 | return NULL; | |
11843 | #endif | |
11844 | } | |
11845 | EXPORT_SYMBOL_GPL(sched_trace_rd_span); | |
9d246053 PA |
11846 | |
11847 | int sched_trace_rq_nr_running(struct rq *rq) | |
11848 | { | |
11849 | return rq ? rq->nr_running : -1; | |
11850 | } | |
11851 | EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running); |