]>
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 PZ |
24 | |
25 | #include <trace/events/sched.h> | |
26 | ||
bf0f6f24 | 27 | /* |
21805085 | 28 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 29 | * |
21805085 | 30 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
31 | * 'timeslice length' - timeslices in CFS are of variable length |
32 | * and have no persistent notion like in traditional, time-slice | |
33 | * based scheduling concepts. | |
bf0f6f24 | 34 | * |
d274a4ce IM |
35 | * (to see the precise effective timeslice length of your workload, |
36 | * run vmstat and monitor the context-switches (cs) field) | |
2b4d5b25 IM |
37 | * |
38 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 39 | */ |
2b4d5b25 | 40 | unsigned int sysctl_sched_latency = 6000000ULL; |
ed8885a1 | 41 | static unsigned int normalized_sysctl_sched_latency = 6000000ULL; |
2bd8e6d4 | 42 | |
1983a922 CE |
43 | /* |
44 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
45 | * |
46 | * Options are: | |
2b4d5b25 IM |
47 | * |
48 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
49 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
50 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
51 | * | |
52 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 53 | */ |
2b4d5b25 | 54 | enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 55 | |
2bd8e6d4 | 56 | /* |
b2be5e96 | 57 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 58 | * |
864616ee | 59 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 60 | */ |
ed8885a1 MS |
61 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
62 | static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
63 | |
64 | /* | |
2b4d5b25 | 65 | * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity |
b2be5e96 | 66 | */ |
0bf377bb | 67 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
68 | |
69 | /* | |
2bba22c5 | 70 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 71 | * parent will (try to) run first. |
21805085 | 72 | */ |
2bba22c5 | 73 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 74 | |
bf0f6f24 IM |
75 | /* |
76 | * SCHED_OTHER wake-up granularity. | |
bf0f6f24 IM |
77 | * |
78 | * This option delays the preemption effects of decoupled workloads | |
79 | * and reduces their over-scheduling. Synchronous workloads will still | |
80 | * have immediate wakeup/sleep latencies. | |
2b4d5b25 IM |
81 | * |
82 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 83 | */ |
ed8885a1 MS |
84 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
85 | static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
bf0f6f24 | 86 | |
2b4d5b25 | 87 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 88 | |
05289b90 TG |
89 | int sched_thermal_decay_shift; |
90 | static int __init setup_sched_thermal_decay_shift(char *str) | |
91 | { | |
92 | int _shift = 0; | |
93 | ||
94 | if (kstrtoint(str, 0, &_shift)) | |
95 | pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n"); | |
96 | ||
97 | sched_thermal_decay_shift = clamp(_shift, 0, 10); | |
98 | return 1; | |
99 | } | |
100 | __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift); | |
101 | ||
afe06efd TC |
102 | #ifdef CONFIG_SMP |
103 | /* | |
97fb7a0a | 104 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
105 | */ |
106 | int __weak arch_asym_cpu_priority(int cpu) | |
107 | { | |
108 | return -cpu; | |
109 | } | |
6d101ba6 OJ |
110 | |
111 | /* | |
60e17f5c | 112 | * The margin used when comparing utilization with CPU capacity. |
6d101ba6 OJ |
113 | * |
114 | * (default: ~20%) | |
115 | */ | |
60e17f5c VK |
116 | #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) |
117 | ||
afe06efd TC |
118 | #endif |
119 | ||
ec12cb7f PT |
120 | #ifdef CONFIG_CFS_BANDWIDTH |
121 | /* | |
122 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
123 | * each time a cfs_rq requests quota. | |
124 | * | |
125 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
126 | * to consumption or the quota being specified to be smaller than the slice) | |
127 | * we will always only issue the remaining available time. | |
128 | * | |
2b4d5b25 IM |
129 | * (default: 5 msec, units: microseconds) |
130 | */ | |
131 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
ec12cb7f PT |
132 | #endif |
133 | ||
8527632d PG |
134 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
135 | { | |
136 | lw->weight += inc; | |
137 | lw->inv_weight = 0; | |
138 | } | |
139 | ||
140 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
141 | { | |
142 | lw->weight -= dec; | |
143 | lw->inv_weight = 0; | |
144 | } | |
145 | ||
146 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
147 | { | |
148 | lw->weight = w; | |
149 | lw->inv_weight = 0; | |
150 | } | |
151 | ||
029632fb PZ |
152 | /* |
153 | * Increase the granularity value when there are more CPUs, | |
154 | * because with more CPUs the 'effective latency' as visible | |
155 | * to users decreases. But the relationship is not linear, | |
156 | * so pick a second-best guess by going with the log2 of the | |
157 | * number of CPUs. | |
158 | * | |
159 | * This idea comes from the SD scheduler of Con Kolivas: | |
160 | */ | |
58ac93e4 | 161 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 162 | { |
58ac93e4 | 163 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
164 | unsigned int factor; |
165 | ||
166 | switch (sysctl_sched_tunable_scaling) { | |
167 | case SCHED_TUNABLESCALING_NONE: | |
168 | factor = 1; | |
169 | break; | |
170 | case SCHED_TUNABLESCALING_LINEAR: | |
171 | factor = cpus; | |
172 | break; | |
173 | case SCHED_TUNABLESCALING_LOG: | |
174 | default: | |
175 | factor = 1 + ilog2(cpus); | |
176 | break; | |
177 | } | |
178 | ||
179 | return factor; | |
180 | } | |
181 | ||
182 | static void update_sysctl(void) | |
183 | { | |
184 | unsigned int factor = get_update_sysctl_factor(); | |
185 | ||
186 | #define SET_SYSCTL(name) \ | |
187 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
188 | SET_SYSCTL(sched_min_granularity); | |
189 | SET_SYSCTL(sched_latency); | |
190 | SET_SYSCTL(sched_wakeup_granularity); | |
191 | #undef SET_SYSCTL | |
192 | } | |
193 | ||
f38f12d1 | 194 | void __init sched_init_granularity(void) |
029632fb PZ |
195 | { |
196 | update_sysctl(); | |
197 | } | |
198 | ||
9dbdb155 | 199 | #define WMULT_CONST (~0U) |
029632fb PZ |
200 | #define WMULT_SHIFT 32 |
201 | ||
9dbdb155 PZ |
202 | static void __update_inv_weight(struct load_weight *lw) |
203 | { | |
204 | unsigned long w; | |
205 | ||
206 | if (likely(lw->inv_weight)) | |
207 | return; | |
208 | ||
209 | w = scale_load_down(lw->weight); | |
210 | ||
211 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
212 | lw->inv_weight = 1; | |
213 | else if (unlikely(!w)) | |
214 | lw->inv_weight = WMULT_CONST; | |
215 | else | |
216 | lw->inv_weight = WMULT_CONST / w; | |
217 | } | |
029632fb PZ |
218 | |
219 | /* | |
9dbdb155 PZ |
220 | * delta_exec * weight / lw.weight |
221 | * OR | |
222 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
223 | * | |
1c3de5e1 | 224 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
225 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
226 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
227 | * | |
228 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
229 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 230 | */ |
9dbdb155 | 231 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 232 | { |
9dbdb155 PZ |
233 | u64 fact = scale_load_down(weight); |
234 | int shift = WMULT_SHIFT; | |
029632fb | 235 | |
9dbdb155 | 236 | __update_inv_weight(lw); |
029632fb | 237 | |
9dbdb155 PZ |
238 | if (unlikely(fact >> 32)) { |
239 | while (fact >> 32) { | |
240 | fact >>= 1; | |
241 | shift--; | |
242 | } | |
029632fb PZ |
243 | } |
244 | ||
2eeb01a2 | 245 | fact = mul_u32_u32(fact, lw->inv_weight); |
029632fb | 246 | |
9dbdb155 PZ |
247 | while (fact >> 32) { |
248 | fact >>= 1; | |
249 | shift--; | |
250 | } | |
029632fb | 251 | |
9dbdb155 | 252 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
253 | } |
254 | ||
255 | ||
256 | const struct sched_class fair_sched_class; | |
a4c2f00f | 257 | |
bf0f6f24 IM |
258 | /************************************************************** |
259 | * CFS operations on generic schedulable entities: | |
260 | */ | |
261 | ||
62160e3f | 262 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8f48894f PZ |
263 | static inline struct task_struct *task_of(struct sched_entity *se) |
264 | { | |
9148a3a1 | 265 | SCHED_WARN_ON(!entity_is_task(se)); |
8f48894f PZ |
266 | return container_of(se, struct task_struct, se); |
267 | } | |
268 | ||
b758149c PZ |
269 | /* Walk up scheduling entities hierarchy */ |
270 | #define for_each_sched_entity(se) \ | |
271 | for (; se; se = se->parent) | |
272 | ||
273 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
274 | { | |
275 | return p->se.cfs_rq; | |
276 | } | |
277 | ||
278 | /* runqueue on which this entity is (to be) queued */ | |
279 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
280 | { | |
281 | return se->cfs_rq; | |
282 | } | |
283 | ||
284 | /* runqueue "owned" by this group */ | |
285 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
286 | { | |
287 | return grp->my_q; | |
288 | } | |
289 | ||
3c93a0c0 QY |
290 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
291 | { | |
292 | if (!path) | |
293 | return; | |
294 | ||
295 | if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) | |
296 | autogroup_path(cfs_rq->tg, path, len); | |
297 | else if (cfs_rq && cfs_rq->tg->css.cgroup) | |
298 | cgroup_path(cfs_rq->tg->css.cgroup, path, len); | |
299 | else | |
300 | strlcpy(path, "(null)", len); | |
301 | } | |
302 | ||
f6783319 | 303 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 304 | { |
5d299eab PZ |
305 | struct rq *rq = rq_of(cfs_rq); |
306 | int cpu = cpu_of(rq); | |
307 | ||
308 | if (cfs_rq->on_list) | |
f6783319 | 309 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
310 | |
311 | cfs_rq->on_list = 1; | |
312 | ||
313 | /* | |
314 | * Ensure we either appear before our parent (if already | |
315 | * enqueued) or force our parent to appear after us when it is | |
316 | * enqueued. The fact that we always enqueue bottom-up | |
317 | * reduces this to two cases and a special case for the root | |
318 | * cfs_rq. Furthermore, it also means that we will always reset | |
319 | * tmp_alone_branch either when the branch is connected | |
320 | * to a tree or when we reach the top of the tree | |
321 | */ | |
322 | if (cfs_rq->tg->parent && | |
323 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 324 | /* |
5d299eab PZ |
325 | * If parent is already on the list, we add the child |
326 | * just before. Thanks to circular linked property of | |
327 | * the list, this means to put the child at the tail | |
328 | * of the list that starts by parent. | |
67e86250 | 329 | */ |
5d299eab PZ |
330 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
331 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
332 | /* | |
333 | * The branch is now connected to its tree so we can | |
334 | * reset tmp_alone_branch to the beginning of the | |
335 | * list. | |
336 | */ | |
337 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 338 | return true; |
5d299eab | 339 | } |
3d4b47b4 | 340 | |
5d299eab PZ |
341 | if (!cfs_rq->tg->parent) { |
342 | /* | |
343 | * cfs rq without parent should be put | |
344 | * at the tail of the list. | |
345 | */ | |
346 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
347 | &rq->leaf_cfs_rq_list); | |
348 | /* | |
349 | * We have reach the top of a tree so we can reset | |
350 | * tmp_alone_branch to the beginning of the list. | |
351 | */ | |
352 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 353 | return true; |
3d4b47b4 | 354 | } |
5d299eab PZ |
355 | |
356 | /* | |
357 | * The parent has not already been added so we want to | |
358 | * make sure that it will be put after us. | |
359 | * tmp_alone_branch points to the begin of the branch | |
360 | * where we will add parent. | |
361 | */ | |
362 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
363 | /* | |
364 | * update tmp_alone_branch to points to the new begin | |
365 | * of the branch | |
366 | */ | |
367 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 368 | return false; |
3d4b47b4 PZ |
369 | } |
370 | ||
371 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
372 | { | |
373 | if (cfs_rq->on_list) { | |
31bc6aea VG |
374 | struct rq *rq = rq_of(cfs_rq); |
375 | ||
376 | /* | |
377 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
378 | * it can happen the tmp_alone_branch points the a leaf that | |
379 | * we finally want to del. In this case, tmp_alone_branch moves | |
380 | * to the prev element but it will point to rq->leaf_cfs_rq_list | |
381 | * at the end of the enqueue. | |
382 | */ | |
383 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
384 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
385 | ||
3d4b47b4 PZ |
386 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
387 | cfs_rq->on_list = 0; | |
388 | } | |
389 | } | |
390 | ||
5d299eab PZ |
391 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
392 | { | |
393 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
394 | } | |
395 | ||
039ae8bc VG |
396 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
397 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ | |
398 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
399 | leaf_cfs_rq_list) | |
b758149c PZ |
400 | |
401 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 402 | static inline struct cfs_rq * |
b758149c PZ |
403 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
404 | { | |
405 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 406 | return se->cfs_rq; |
b758149c | 407 | |
fed14d45 | 408 | return NULL; |
b758149c PZ |
409 | } |
410 | ||
411 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
412 | { | |
413 | return se->parent; | |
414 | } | |
415 | ||
464b7527 PZ |
416 | static void |
417 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
418 | { | |
419 | int se_depth, pse_depth; | |
420 | ||
421 | /* | |
422 | * preemption test can be made between sibling entities who are in the | |
423 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
424 | * both tasks until we find their ancestors who are siblings of common | |
425 | * parent. | |
426 | */ | |
427 | ||
428 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
429 | se_depth = (*se)->depth; |
430 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
431 | |
432 | while (se_depth > pse_depth) { | |
433 | se_depth--; | |
434 | *se = parent_entity(*se); | |
435 | } | |
436 | ||
437 | while (pse_depth > se_depth) { | |
438 | pse_depth--; | |
439 | *pse = parent_entity(*pse); | |
440 | } | |
441 | ||
442 | while (!is_same_group(*se, *pse)) { | |
443 | *se = parent_entity(*se); | |
444 | *pse = parent_entity(*pse); | |
445 | } | |
446 | } | |
447 | ||
8f48894f PZ |
448 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
449 | ||
450 | static inline struct task_struct *task_of(struct sched_entity *se) | |
451 | { | |
452 | return container_of(se, struct task_struct, se); | |
453 | } | |
bf0f6f24 | 454 | |
b758149c PZ |
455 | #define for_each_sched_entity(se) \ |
456 | for (; se; se = NULL) | |
bf0f6f24 | 457 | |
b758149c | 458 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 459 | { |
b758149c | 460 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
461 | } |
462 | ||
b758149c PZ |
463 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
464 | { | |
465 | struct task_struct *p = task_of(se); | |
466 | struct rq *rq = task_rq(p); | |
467 | ||
468 | return &rq->cfs; | |
469 | } | |
470 | ||
471 | /* runqueue "owned" by this group */ | |
472 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
473 | { | |
474 | return NULL; | |
475 | } | |
476 | ||
3c93a0c0 QY |
477 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
478 | { | |
479 | if (path) | |
480 | strlcpy(path, "(null)", len); | |
481 | } | |
482 | ||
f6783319 | 483 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 484 | { |
f6783319 | 485 | return true; |
3d4b47b4 PZ |
486 | } |
487 | ||
488 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
489 | { | |
490 | } | |
491 | ||
5d299eab PZ |
492 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
493 | { | |
494 | } | |
495 | ||
039ae8bc VG |
496 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
497 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 498 | |
b758149c PZ |
499 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
500 | { | |
501 | return NULL; | |
502 | } | |
503 | ||
464b7527 PZ |
504 | static inline void |
505 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
506 | { | |
507 | } | |
508 | ||
b758149c PZ |
509 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
510 | ||
6c16a6dc | 511 | static __always_inline |
9dbdb155 | 512 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
513 | |
514 | /************************************************************** | |
515 | * Scheduling class tree data structure manipulation methods: | |
516 | */ | |
517 | ||
1bf08230 | 518 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 519 | { |
1bf08230 | 520 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 521 | if (delta > 0) |
1bf08230 | 522 | max_vruntime = vruntime; |
02e0431a | 523 | |
1bf08230 | 524 | return max_vruntime; |
02e0431a PZ |
525 | } |
526 | ||
0702e3eb | 527 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
528 | { |
529 | s64 delta = (s64)(vruntime - min_vruntime); | |
530 | if (delta < 0) | |
531 | min_vruntime = vruntime; | |
532 | ||
533 | return min_vruntime; | |
534 | } | |
535 | ||
54fdc581 FC |
536 | static inline int entity_before(struct sched_entity *a, |
537 | struct sched_entity *b) | |
538 | { | |
539 | return (s64)(a->vruntime - b->vruntime) < 0; | |
540 | } | |
541 | ||
1af5f730 PZ |
542 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
543 | { | |
b60205c7 | 544 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 545 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 546 | |
1af5f730 PZ |
547 | u64 vruntime = cfs_rq->min_vruntime; |
548 | ||
b60205c7 PZ |
549 | if (curr) { |
550 | if (curr->on_rq) | |
551 | vruntime = curr->vruntime; | |
552 | else | |
553 | curr = NULL; | |
554 | } | |
1af5f730 | 555 | |
bfb06889 DB |
556 | if (leftmost) { /* non-empty tree */ |
557 | struct sched_entity *se; | |
558 | se = rb_entry(leftmost, struct sched_entity, run_node); | |
1af5f730 | 559 | |
b60205c7 | 560 | if (!curr) |
1af5f730 PZ |
561 | vruntime = se->vruntime; |
562 | else | |
563 | vruntime = min_vruntime(vruntime, se->vruntime); | |
564 | } | |
565 | ||
1bf08230 | 566 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 567 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
568 | #ifndef CONFIG_64BIT |
569 | smp_wmb(); | |
570 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
571 | #endif | |
1af5f730 PZ |
572 | } |
573 | ||
bf0f6f24 IM |
574 | /* |
575 | * Enqueue an entity into the rb-tree: | |
576 | */ | |
0702e3eb | 577 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 578 | { |
bfb06889 | 579 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node; |
bf0f6f24 IM |
580 | struct rb_node *parent = NULL; |
581 | struct sched_entity *entry; | |
bfb06889 | 582 | bool leftmost = true; |
bf0f6f24 IM |
583 | |
584 | /* | |
585 | * Find the right place in the rbtree: | |
586 | */ | |
587 | while (*link) { | |
588 | parent = *link; | |
589 | entry = rb_entry(parent, struct sched_entity, run_node); | |
590 | /* | |
591 | * We dont care about collisions. Nodes with | |
592 | * the same key stay together. | |
593 | */ | |
2bd2d6f2 | 594 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
595 | link = &parent->rb_left; |
596 | } else { | |
597 | link = &parent->rb_right; | |
bfb06889 | 598 | leftmost = false; |
bf0f6f24 IM |
599 | } |
600 | } | |
601 | ||
bf0f6f24 | 602 | rb_link_node(&se->run_node, parent, link); |
bfb06889 DB |
603 | rb_insert_color_cached(&se->run_node, |
604 | &cfs_rq->tasks_timeline, leftmost); | |
bf0f6f24 IM |
605 | } |
606 | ||
0702e3eb | 607 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 608 | { |
bfb06889 | 609 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
610 | } |
611 | ||
029632fb | 612 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 613 | { |
bfb06889 | 614 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
615 | |
616 | if (!left) | |
617 | return NULL; | |
618 | ||
619 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
620 | } |
621 | ||
ac53db59 RR |
622 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
623 | { | |
624 | struct rb_node *next = rb_next(&se->run_node); | |
625 | ||
626 | if (!next) | |
627 | return NULL; | |
628 | ||
629 | return rb_entry(next, struct sched_entity, run_node); | |
630 | } | |
631 | ||
632 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 633 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 634 | { |
bfb06889 | 635 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 636 | |
70eee74b BS |
637 | if (!last) |
638 | return NULL; | |
7eee3e67 IM |
639 | |
640 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
641 | } |
642 | ||
bf0f6f24 IM |
643 | /************************************************************** |
644 | * Scheduling class statistics methods: | |
645 | */ | |
646 | ||
acb4a848 | 647 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 648 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
649 | loff_t *ppos) |
650 | { | |
8d65af78 | 651 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 652 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
653 | |
654 | if (ret || !write) | |
655 | return ret; | |
656 | ||
657 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
658 | sysctl_sched_min_granularity); | |
659 | ||
acb4a848 CE |
660 | #define WRT_SYSCTL(name) \ |
661 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
662 | WRT_SYSCTL(sched_min_granularity); | |
663 | WRT_SYSCTL(sched_latency); | |
664 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
665 | #undef WRT_SYSCTL |
666 | ||
b2be5e96 PZ |
667 | return 0; |
668 | } | |
669 | #endif | |
647e7cac | 670 | |
a7be37ac | 671 | /* |
f9c0b095 | 672 | * delta /= w |
a7be37ac | 673 | */ |
9dbdb155 | 674 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 675 | { |
f9c0b095 | 676 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 677 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
678 | |
679 | return delta; | |
680 | } | |
681 | ||
647e7cac IM |
682 | /* |
683 | * The idea is to set a period in which each task runs once. | |
684 | * | |
532b1858 | 685 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
686 | * this period because otherwise the slices get too small. |
687 | * | |
688 | * p = (nr <= nl) ? l : l*nr/nl | |
689 | */ | |
4d78e7b6 PZ |
690 | static u64 __sched_period(unsigned long nr_running) |
691 | { | |
8e2b0bf3 BF |
692 | if (unlikely(nr_running > sched_nr_latency)) |
693 | return nr_running * sysctl_sched_min_granularity; | |
694 | else | |
695 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
696 | } |
697 | ||
647e7cac IM |
698 | /* |
699 | * We calculate the wall-time slice from the period by taking a part | |
700 | * proportional to the weight. | |
701 | * | |
f9c0b095 | 702 | * s = p*P[w/rw] |
647e7cac | 703 | */ |
6d0f0ebd | 704 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 705 | { |
0a582440 | 706 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 707 | |
0a582440 | 708 | for_each_sched_entity(se) { |
6272d68c | 709 | struct load_weight *load; |
3104bf03 | 710 | struct load_weight lw; |
6272d68c LM |
711 | |
712 | cfs_rq = cfs_rq_of(se); | |
713 | load = &cfs_rq->load; | |
f9c0b095 | 714 | |
0a582440 | 715 | if (unlikely(!se->on_rq)) { |
3104bf03 | 716 | lw = cfs_rq->load; |
0a582440 MG |
717 | |
718 | update_load_add(&lw, se->load.weight); | |
719 | load = &lw; | |
720 | } | |
9dbdb155 | 721 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
722 | } |
723 | return slice; | |
bf0f6f24 IM |
724 | } |
725 | ||
647e7cac | 726 | /* |
660cc00f | 727 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 728 | * |
f9c0b095 | 729 | * vs = s/w |
647e7cac | 730 | */ |
f9c0b095 | 731 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 732 | { |
f9c0b095 | 733 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
734 | } |
735 | ||
c0796298 | 736 | #include "pelt.h" |
23127296 | 737 | #ifdef CONFIG_SMP |
283e2ed3 | 738 | |
772bd008 | 739 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 740 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 741 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 742 | |
540247fb YD |
743 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
744 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 745 | { |
540247fb | 746 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 747 | |
f207934f PZ |
748 | memset(sa, 0, sizeof(*sa)); |
749 | ||
b5a9b340 | 750 | /* |
dfcb245e | 751 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 752 | * they get a chance to stabilize to their real load level. |
dfcb245e | 753 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
754 | * nothing has been attached to the task group yet. |
755 | */ | |
756 | if (entity_is_task(se)) | |
0dacee1b | 757 | sa->load_avg = scale_load_down(se->load.weight); |
f207934f | 758 | |
9d89c257 | 759 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 760 | } |
7ea241af | 761 | |
df217913 | 762 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 763 | |
2b8c41da YD |
764 | /* |
765 | * With new tasks being created, their initial util_avgs are extrapolated | |
766 | * based on the cfs_rq's current util_avg: | |
767 | * | |
768 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
769 | * | |
770 | * However, in many cases, the above util_avg does not give a desired | |
771 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
772 | * as when the series is a harmonic series. | |
773 | * | |
774 | * To solve this problem, we also cap the util_avg of successive tasks to | |
775 | * only 1/2 of the left utilization budget: | |
776 | * | |
8fe5c5a9 | 777 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 778 | * |
8fe5c5a9 | 779 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 780 | * |
8fe5c5a9 QP |
781 | * For example, for a CPU with 1024 of capacity, a simplest series from |
782 | * the beginning would be like: | |
2b8c41da YD |
783 | * |
784 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
785 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
786 | * | |
787 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
788 | * if util_avg > util_avg_cap. | |
789 | */ | |
d0fe0b9c | 790 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 791 | { |
d0fe0b9c | 792 | struct sched_entity *se = &p->se; |
2b8c41da YD |
793 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
794 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 795 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 796 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
797 | |
798 | if (cap > 0) { | |
799 | if (cfs_rq->avg.util_avg != 0) { | |
800 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
801 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
802 | ||
803 | if (sa->util_avg > cap) | |
804 | sa->util_avg = cap; | |
805 | } else { | |
806 | sa->util_avg = cap; | |
807 | } | |
2b8c41da | 808 | } |
7dc603c9 | 809 | |
9f683953 VG |
810 | sa->runnable_avg = cpu_scale; |
811 | ||
d0fe0b9c DE |
812 | if (p->sched_class != &fair_sched_class) { |
813 | /* | |
814 | * For !fair tasks do: | |
815 | * | |
816 | update_cfs_rq_load_avg(now, cfs_rq); | |
a4f9a0e5 | 817 | attach_entity_load_avg(cfs_rq, se); |
d0fe0b9c DE |
818 | switched_from_fair(rq, p); |
819 | * | |
820 | * such that the next switched_to_fair() has the | |
821 | * expected state. | |
822 | */ | |
823 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
824 | return; | |
7dc603c9 PZ |
825 | } |
826 | ||
df217913 | 827 | attach_entity_cfs_rq(se); |
2b8c41da YD |
828 | } |
829 | ||
7dc603c9 | 830 | #else /* !CONFIG_SMP */ |
540247fb | 831 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
832 | { |
833 | } | |
d0fe0b9c | 834 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
835 | { |
836 | } | |
3d30544f PZ |
837 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
838 | { | |
839 | } | |
7dc603c9 | 840 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 841 | |
bf0f6f24 | 842 | /* |
9dbdb155 | 843 | * Update the current task's runtime statistics. |
bf0f6f24 | 844 | */ |
b7cc0896 | 845 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 846 | { |
429d43bc | 847 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 848 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 849 | u64 delta_exec; |
bf0f6f24 IM |
850 | |
851 | if (unlikely(!curr)) | |
852 | return; | |
853 | ||
9dbdb155 PZ |
854 | delta_exec = now - curr->exec_start; |
855 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 856 | return; |
bf0f6f24 | 857 | |
8ebc91d9 | 858 | curr->exec_start = now; |
d842de87 | 859 | |
9dbdb155 PZ |
860 | schedstat_set(curr->statistics.exec_max, |
861 | max(delta_exec, curr->statistics.exec_max)); | |
862 | ||
863 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 864 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
865 | |
866 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
867 | update_min_vruntime(cfs_rq); | |
868 | ||
d842de87 SV |
869 | if (entity_is_task(curr)) { |
870 | struct task_struct *curtask = task_of(curr); | |
871 | ||
f977bb49 | 872 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 873 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 874 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 875 | } |
ec12cb7f PT |
876 | |
877 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
878 | } |
879 | ||
6e998916 SG |
880 | static void update_curr_fair(struct rq *rq) |
881 | { | |
882 | update_curr(cfs_rq_of(&rq->curr->se)); | |
883 | } | |
884 | ||
bf0f6f24 | 885 | static inline void |
5870db5b | 886 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 887 | { |
4fa8d299 JP |
888 | u64 wait_start, prev_wait_start; |
889 | ||
890 | if (!schedstat_enabled()) | |
891 | return; | |
892 | ||
893 | wait_start = rq_clock(rq_of(cfs_rq)); | |
894 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
895 | |
896 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
897 | likely(wait_start > prev_wait_start)) |
898 | wait_start -= prev_wait_start; | |
3ea94de1 | 899 | |
2ed41a55 | 900 | __schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
901 | } |
902 | ||
4fa8d299 | 903 | static inline void |
3ea94de1 JP |
904 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
905 | { | |
906 | struct task_struct *p; | |
cb251765 MG |
907 | u64 delta; |
908 | ||
4fa8d299 JP |
909 | if (!schedstat_enabled()) |
910 | return; | |
911 | ||
912 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
913 | |
914 | if (entity_is_task(se)) { | |
915 | p = task_of(se); | |
916 | if (task_on_rq_migrating(p)) { | |
917 | /* | |
918 | * Preserve migrating task's wait time so wait_start | |
919 | * time stamp can be adjusted to accumulate wait time | |
920 | * prior to migration. | |
921 | */ | |
2ed41a55 | 922 | __schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
923 | return; |
924 | } | |
925 | trace_sched_stat_wait(p, delta); | |
926 | } | |
927 | ||
2ed41a55 | 928 | __schedstat_set(se->statistics.wait_max, |
4fa8d299 | 929 | max(schedstat_val(se->statistics.wait_max), delta)); |
2ed41a55 PZ |
930 | __schedstat_inc(se->statistics.wait_count); |
931 | __schedstat_add(se->statistics.wait_sum, delta); | |
932 | __schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 933 | } |
3ea94de1 | 934 | |
4fa8d299 | 935 | static inline void |
1a3d027c JP |
936 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
937 | { | |
938 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
939 | u64 sleep_start, block_start; |
940 | ||
941 | if (!schedstat_enabled()) | |
942 | return; | |
943 | ||
944 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
945 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
946 | |
947 | if (entity_is_task(se)) | |
948 | tsk = task_of(se); | |
949 | ||
4fa8d299 JP |
950 | if (sleep_start) { |
951 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
952 | |
953 | if ((s64)delta < 0) | |
954 | delta = 0; | |
955 | ||
4fa8d299 | 956 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
2ed41a55 | 957 | __schedstat_set(se->statistics.sleep_max, delta); |
1a3d027c | 958 | |
2ed41a55 PZ |
959 | __schedstat_set(se->statistics.sleep_start, 0); |
960 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
961 | |
962 | if (tsk) { | |
963 | account_scheduler_latency(tsk, delta >> 10, 1); | |
964 | trace_sched_stat_sleep(tsk, delta); | |
965 | } | |
966 | } | |
4fa8d299 JP |
967 | if (block_start) { |
968 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
969 | |
970 | if ((s64)delta < 0) | |
971 | delta = 0; | |
972 | ||
4fa8d299 | 973 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
2ed41a55 | 974 | __schedstat_set(se->statistics.block_max, delta); |
1a3d027c | 975 | |
2ed41a55 PZ |
976 | __schedstat_set(se->statistics.block_start, 0); |
977 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
978 | |
979 | if (tsk) { | |
980 | if (tsk->in_iowait) { | |
2ed41a55 PZ |
981 | __schedstat_add(se->statistics.iowait_sum, delta); |
982 | __schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
983 | trace_sched_stat_iowait(tsk, delta); |
984 | } | |
985 | ||
986 | trace_sched_stat_blocked(tsk, delta); | |
987 | ||
988 | /* | |
989 | * Blocking time is in units of nanosecs, so shift by | |
990 | * 20 to get a milliseconds-range estimation of the | |
991 | * amount of time that the task spent sleeping: | |
992 | */ | |
993 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
994 | profile_hits(SLEEP_PROFILING, | |
995 | (void *)get_wchan(tsk), | |
996 | delta >> 20); | |
997 | } | |
998 | account_scheduler_latency(tsk, delta >> 10, 0); | |
999 | } | |
1000 | } | |
3ea94de1 | 1001 | } |
3ea94de1 | 1002 | |
bf0f6f24 IM |
1003 | /* |
1004 | * Task is being enqueued - update stats: | |
1005 | */ | |
cb251765 | 1006 | static inline void |
1a3d027c | 1007 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1008 | { |
4fa8d299 JP |
1009 | if (!schedstat_enabled()) |
1010 | return; | |
1011 | ||
bf0f6f24 IM |
1012 | /* |
1013 | * Are we enqueueing a waiting task? (for current tasks | |
1014 | * a dequeue/enqueue event is a NOP) | |
1015 | */ | |
429d43bc | 1016 | if (se != cfs_rq->curr) |
5870db5b | 1017 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
1018 | |
1019 | if (flags & ENQUEUE_WAKEUP) | |
1020 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
1021 | } |
1022 | ||
bf0f6f24 | 1023 | static inline void |
cb251765 | 1024 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1025 | { |
4fa8d299 JP |
1026 | |
1027 | if (!schedstat_enabled()) | |
1028 | return; | |
1029 | ||
bf0f6f24 IM |
1030 | /* |
1031 | * Mark the end of the wait period if dequeueing a | |
1032 | * waiting task: | |
1033 | */ | |
429d43bc | 1034 | if (se != cfs_rq->curr) |
9ef0a961 | 1035 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 1036 | |
4fa8d299 JP |
1037 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1038 | struct task_struct *tsk = task_of(se); | |
cb251765 | 1039 | |
4fa8d299 | 1040 | if (tsk->state & TASK_INTERRUPTIBLE) |
2ed41a55 | 1041 | __schedstat_set(se->statistics.sleep_start, |
4fa8d299 JP |
1042 | rq_clock(rq_of(cfs_rq))); |
1043 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
2ed41a55 | 1044 | __schedstat_set(se->statistics.block_start, |
4fa8d299 | 1045 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1046 | } |
cb251765 MG |
1047 | } |
1048 | ||
bf0f6f24 IM |
1049 | /* |
1050 | * We are picking a new current task - update its stats: | |
1051 | */ | |
1052 | static inline void | |
79303e9e | 1053 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1054 | { |
1055 | /* | |
1056 | * We are starting a new run period: | |
1057 | */ | |
78becc27 | 1058 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1059 | } |
1060 | ||
bf0f6f24 IM |
1061 | /************************************************** |
1062 | * Scheduling class queueing methods: | |
1063 | */ | |
1064 | ||
cbee9f88 PZ |
1065 | #ifdef CONFIG_NUMA_BALANCING |
1066 | /* | |
598f0ec0 MG |
1067 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1068 | * calculated based on the tasks virtual memory size and | |
1069 | * numa_balancing_scan_size. | |
cbee9f88 | 1070 | */ |
598f0ec0 MG |
1071 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1072 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1073 | |
1074 | /* Portion of address space to scan in MB */ | |
1075 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1076 | |
4b96a29b PZ |
1077 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1078 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1079 | ||
b5dd77c8 | 1080 | struct numa_group { |
c45a7795 | 1081 | refcount_t refcount; |
b5dd77c8 RR |
1082 | |
1083 | spinlock_t lock; /* nr_tasks, tasks */ | |
1084 | int nr_tasks; | |
1085 | pid_t gid; | |
1086 | int active_nodes; | |
1087 | ||
1088 | struct rcu_head rcu; | |
1089 | unsigned long total_faults; | |
1090 | unsigned long max_faults_cpu; | |
1091 | /* | |
1092 | * Faults_cpu is used to decide whether memory should move | |
1093 | * towards the CPU. As a consequence, these stats are weighted | |
1094 | * more by CPU use than by memory faults. | |
1095 | */ | |
1096 | unsigned long *faults_cpu; | |
04f5c362 | 1097 | unsigned long faults[]; |
b5dd77c8 RR |
1098 | }; |
1099 | ||
cb361d8c JH |
1100 | /* |
1101 | * For functions that can be called in multiple contexts that permit reading | |
1102 | * ->numa_group (see struct task_struct for locking rules). | |
1103 | */ | |
1104 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1105 | { | |
1106 | return rcu_dereference_check(p->numa_group, p == current || | |
1107 | (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu))); | |
1108 | } | |
1109 | ||
1110 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1111 | { | |
1112 | return rcu_dereference_protected(p->numa_group, p == current); | |
1113 | } | |
1114 | ||
b5dd77c8 RR |
1115 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1116 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1117 | ||
598f0ec0 MG |
1118 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1119 | { | |
1120 | unsigned long rss = 0; | |
1121 | unsigned long nr_scan_pages; | |
1122 | ||
1123 | /* | |
1124 | * Calculations based on RSS as non-present and empty pages are skipped | |
1125 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1126 | * on resident pages | |
1127 | */ | |
1128 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1129 | rss = get_mm_rss(p->mm); | |
1130 | if (!rss) | |
1131 | rss = nr_scan_pages; | |
1132 | ||
1133 | rss = round_up(rss, nr_scan_pages); | |
1134 | return rss / nr_scan_pages; | |
1135 | } | |
1136 | ||
1137 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
1138 | #define MAX_SCAN_WINDOW 2560 | |
1139 | ||
1140 | static unsigned int task_scan_min(struct task_struct *p) | |
1141 | { | |
316c1608 | 1142 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1143 | unsigned int scan, floor; |
1144 | unsigned int windows = 1; | |
1145 | ||
64192658 KT |
1146 | if (scan_size < MAX_SCAN_WINDOW) |
1147 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1148 | floor = 1000 / windows; |
1149 | ||
1150 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1151 | return max_t(unsigned int, floor, scan); | |
1152 | } | |
1153 | ||
b5dd77c8 RR |
1154 | static unsigned int task_scan_start(struct task_struct *p) |
1155 | { | |
1156 | unsigned long smin = task_scan_min(p); | |
1157 | unsigned long period = smin; | |
cb361d8c | 1158 | struct numa_group *ng; |
b5dd77c8 RR |
1159 | |
1160 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1161 | rcu_read_lock(); |
1162 | ng = rcu_dereference(p->numa_group); | |
1163 | if (ng) { | |
b5dd77c8 RR |
1164 | unsigned long shared = group_faults_shared(ng); |
1165 | unsigned long private = group_faults_priv(ng); | |
1166 | ||
c45a7795 | 1167 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1168 | period *= shared + 1; |
1169 | period /= private + shared + 1; | |
1170 | } | |
cb361d8c | 1171 | rcu_read_unlock(); |
b5dd77c8 RR |
1172 | |
1173 | return max(smin, period); | |
1174 | } | |
1175 | ||
598f0ec0 MG |
1176 | static unsigned int task_scan_max(struct task_struct *p) |
1177 | { | |
b5dd77c8 RR |
1178 | unsigned long smin = task_scan_min(p); |
1179 | unsigned long smax; | |
cb361d8c | 1180 | struct numa_group *ng; |
598f0ec0 MG |
1181 | |
1182 | /* Watch for min being lower than max due to floor calculations */ | |
1183 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1184 | |
1185 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1186 | ng = deref_curr_numa_group(p); |
1187 | if (ng) { | |
b5dd77c8 RR |
1188 | unsigned long shared = group_faults_shared(ng); |
1189 | unsigned long private = group_faults_priv(ng); | |
1190 | unsigned long period = smax; | |
1191 | ||
c45a7795 | 1192 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1193 | period *= shared + 1; |
1194 | period /= private + shared + 1; | |
1195 | ||
1196 | smax = max(smax, period); | |
1197 | } | |
1198 | ||
598f0ec0 MG |
1199 | return max(smin, smax); |
1200 | } | |
1201 | ||
0ec8aa00 PZ |
1202 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1203 | { | |
98fa15f3 | 1204 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1205 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1206 | } | |
1207 | ||
1208 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1209 | { | |
98fa15f3 | 1210 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1211 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1212 | } | |
1213 | ||
be1e4e76 RR |
1214 | /* Shared or private faults. */ |
1215 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1216 | ||
1217 | /* Memory and CPU locality */ | |
1218 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1219 | ||
1220 | /* Averaged statistics, and temporary buffers. */ | |
1221 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1222 | ||
e29cf08b MG |
1223 | pid_t task_numa_group_id(struct task_struct *p) |
1224 | { | |
cb361d8c JH |
1225 | struct numa_group *ng; |
1226 | pid_t gid = 0; | |
1227 | ||
1228 | rcu_read_lock(); | |
1229 | ng = rcu_dereference(p->numa_group); | |
1230 | if (ng) | |
1231 | gid = ng->gid; | |
1232 | rcu_read_unlock(); | |
1233 | ||
1234 | return gid; | |
e29cf08b MG |
1235 | } |
1236 | ||
44dba3d5 | 1237 | /* |
97fb7a0a | 1238 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1239 | * occupy the first half of the array. The second half of the |
1240 | * array is for current counters, which are averaged into the | |
1241 | * first set by task_numa_placement. | |
1242 | */ | |
1243 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1244 | { |
44dba3d5 | 1245 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1246 | } |
1247 | ||
1248 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1249 | { | |
44dba3d5 | 1250 | if (!p->numa_faults) |
ac8e895b MG |
1251 | return 0; |
1252 | ||
44dba3d5 IM |
1253 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1254 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1255 | } |
1256 | ||
83e1d2cd MG |
1257 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1258 | { | |
cb361d8c JH |
1259 | struct numa_group *ng = deref_task_numa_group(p); |
1260 | ||
1261 | if (!ng) | |
83e1d2cd MG |
1262 | return 0; |
1263 | ||
cb361d8c JH |
1264 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1265 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1266 | } |
1267 | ||
20e07dea RR |
1268 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1269 | { | |
44dba3d5 IM |
1270 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1271 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1272 | } |
1273 | ||
b5dd77c8 RR |
1274 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1275 | { | |
1276 | unsigned long faults = 0; | |
1277 | int node; | |
1278 | ||
1279 | for_each_online_node(node) { | |
1280 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1281 | } | |
1282 | ||
1283 | return faults; | |
1284 | } | |
1285 | ||
1286 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1287 | { | |
1288 | unsigned long faults = 0; | |
1289 | int node; | |
1290 | ||
1291 | for_each_online_node(node) { | |
1292 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1293 | } | |
1294 | ||
1295 | return faults; | |
1296 | } | |
1297 | ||
4142c3eb RR |
1298 | /* |
1299 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1300 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1301 | * between these nodes are slowed down, to allow things to settle down. | |
1302 | */ | |
1303 | #define ACTIVE_NODE_FRACTION 3 | |
1304 | ||
1305 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1306 | { | |
1307 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1308 | } | |
1309 | ||
6c6b1193 RR |
1310 | /* Handle placement on systems where not all nodes are directly connected. */ |
1311 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1312 | int maxdist, bool task) | |
1313 | { | |
1314 | unsigned long score = 0; | |
1315 | int node; | |
1316 | ||
1317 | /* | |
1318 | * All nodes are directly connected, and the same distance | |
1319 | * from each other. No need for fancy placement algorithms. | |
1320 | */ | |
1321 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1322 | return 0; | |
1323 | ||
1324 | /* | |
1325 | * This code is called for each node, introducing N^2 complexity, | |
1326 | * which should be ok given the number of nodes rarely exceeds 8. | |
1327 | */ | |
1328 | for_each_online_node(node) { | |
1329 | unsigned long faults; | |
1330 | int dist = node_distance(nid, node); | |
1331 | ||
1332 | /* | |
1333 | * The furthest away nodes in the system are not interesting | |
1334 | * for placement; nid was already counted. | |
1335 | */ | |
1336 | if (dist == sched_max_numa_distance || node == nid) | |
1337 | continue; | |
1338 | ||
1339 | /* | |
1340 | * On systems with a backplane NUMA topology, compare groups | |
1341 | * of nodes, and move tasks towards the group with the most | |
1342 | * memory accesses. When comparing two nodes at distance | |
1343 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1344 | * of each group. Skip other nodes. | |
1345 | */ | |
1346 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
0ee7e74d | 1347 | dist >= maxdist) |
6c6b1193 RR |
1348 | continue; |
1349 | ||
1350 | /* Add up the faults from nearby nodes. */ | |
1351 | if (task) | |
1352 | faults = task_faults(p, node); | |
1353 | else | |
1354 | faults = group_faults(p, node); | |
1355 | ||
1356 | /* | |
1357 | * On systems with a glueless mesh NUMA topology, there are | |
1358 | * no fixed "groups of nodes". Instead, nodes that are not | |
1359 | * directly connected bounce traffic through intermediate | |
1360 | * nodes; a numa_group can occupy any set of nodes. | |
1361 | * The further away a node is, the less the faults count. | |
1362 | * This seems to result in good task placement. | |
1363 | */ | |
1364 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1365 | faults *= (sched_max_numa_distance - dist); | |
1366 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1367 | } | |
1368 | ||
1369 | score += faults; | |
1370 | } | |
1371 | ||
1372 | return score; | |
1373 | } | |
1374 | ||
83e1d2cd MG |
1375 | /* |
1376 | * These return the fraction of accesses done by a particular task, or | |
1377 | * task group, on a particular numa node. The group weight is given a | |
1378 | * larger multiplier, in order to group tasks together that are almost | |
1379 | * evenly spread out between numa nodes. | |
1380 | */ | |
7bd95320 RR |
1381 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1382 | int dist) | |
83e1d2cd | 1383 | { |
7bd95320 | 1384 | unsigned long faults, total_faults; |
83e1d2cd | 1385 | |
44dba3d5 | 1386 | if (!p->numa_faults) |
83e1d2cd MG |
1387 | return 0; |
1388 | ||
1389 | total_faults = p->total_numa_faults; | |
1390 | ||
1391 | if (!total_faults) | |
1392 | return 0; | |
1393 | ||
7bd95320 | 1394 | faults = task_faults(p, nid); |
6c6b1193 RR |
1395 | faults += score_nearby_nodes(p, nid, dist, true); |
1396 | ||
7bd95320 | 1397 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1398 | } |
1399 | ||
7bd95320 RR |
1400 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1401 | int dist) | |
83e1d2cd | 1402 | { |
cb361d8c | 1403 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1404 | unsigned long faults, total_faults; |
1405 | ||
cb361d8c | 1406 | if (!ng) |
7bd95320 RR |
1407 | return 0; |
1408 | ||
cb361d8c | 1409 | total_faults = ng->total_faults; |
7bd95320 RR |
1410 | |
1411 | if (!total_faults) | |
83e1d2cd MG |
1412 | return 0; |
1413 | ||
7bd95320 | 1414 | faults = group_faults(p, nid); |
6c6b1193 RR |
1415 | faults += score_nearby_nodes(p, nid, dist, false); |
1416 | ||
7bd95320 | 1417 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1418 | } |
1419 | ||
10f39042 RR |
1420 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1421 | int src_nid, int dst_cpu) | |
1422 | { | |
cb361d8c | 1423 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1424 | int dst_nid = cpu_to_node(dst_cpu); |
1425 | int last_cpupid, this_cpupid; | |
1426 | ||
1427 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
37355bdc MG |
1428 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1429 | ||
1430 | /* | |
1431 | * Allow first faults or private faults to migrate immediately early in | |
1432 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1433 | * two full passes of the "multi-stage node selection" test that is | |
1434 | * executed below. | |
1435 | */ | |
98fa15f3 | 1436 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1437 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1438 | return true; | |
10f39042 RR |
1439 | |
1440 | /* | |
1441 | * Multi-stage node selection is used in conjunction with a periodic | |
1442 | * migration fault to build a temporal task<->page relation. By using | |
1443 | * a two-stage filter we remove short/unlikely relations. | |
1444 | * | |
1445 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1446 | * a task's usage of a particular page (n_p) per total usage of this | |
1447 | * page (n_t) (in a given time-span) to a probability. | |
1448 | * | |
1449 | * Our periodic faults will sample this probability and getting the | |
1450 | * same result twice in a row, given these samples are fully | |
1451 | * independent, is then given by P(n)^2, provided our sample period | |
1452 | * is sufficiently short compared to the usage pattern. | |
1453 | * | |
1454 | * This quadric squishes small probabilities, making it less likely we | |
1455 | * act on an unlikely task<->page relation. | |
1456 | */ | |
10f39042 RR |
1457 | if (!cpupid_pid_unset(last_cpupid) && |
1458 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1459 | return false; | |
1460 | ||
1461 | /* Always allow migrate on private faults */ | |
1462 | if (cpupid_match_pid(p, last_cpupid)) | |
1463 | return true; | |
1464 | ||
1465 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1466 | if (!ng) | |
1467 | return true; | |
1468 | ||
1469 | /* | |
4142c3eb RR |
1470 | * Destination node is much more heavily used than the source |
1471 | * node? Allow migration. | |
10f39042 | 1472 | */ |
4142c3eb RR |
1473 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1474 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1475 | return true; |
1476 | ||
1477 | /* | |
4142c3eb RR |
1478 | * Distribute memory according to CPU & memory use on each node, |
1479 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1480 | * | |
1481 | * faults_cpu(dst) 3 faults_cpu(src) | |
1482 | * --------------- * - > --------------- | |
1483 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1484 | */ |
4142c3eb RR |
1485 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1486 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1487 | } |
1488 | ||
6499b1b2 VG |
1489 | /* |
1490 | * 'numa_type' describes the node at the moment of load balancing. | |
1491 | */ | |
1492 | enum numa_type { | |
1493 | /* The node has spare capacity that can be used to run more tasks. */ | |
1494 | node_has_spare = 0, | |
1495 | /* | |
1496 | * The node is fully used and the tasks don't compete for more CPU | |
1497 | * cycles. Nevertheless, some tasks might wait before running. | |
1498 | */ | |
1499 | node_fully_busy, | |
1500 | /* | |
1501 | * The node is overloaded and can't provide expected CPU cycles to all | |
1502 | * tasks. | |
1503 | */ | |
1504 | node_overloaded | |
1505 | }; | |
58d081b5 | 1506 | |
fb13c7ee | 1507 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1508 | struct numa_stats { |
1509 | unsigned long load; | |
6499b1b2 | 1510 | unsigned long util; |
fb13c7ee | 1511 | /* Total compute capacity of CPUs on a node */ |
5ef20ca1 | 1512 | unsigned long compute_capacity; |
6499b1b2 VG |
1513 | unsigned int nr_running; |
1514 | unsigned int weight; | |
1515 | enum numa_type node_type; | |
ff7db0bf | 1516 | int idle_cpu; |
58d081b5 | 1517 | }; |
e6628d5b | 1518 | |
ff7db0bf MG |
1519 | static inline bool is_core_idle(int cpu) |
1520 | { | |
1521 | #ifdef CONFIG_SCHED_SMT | |
1522 | int sibling; | |
1523 | ||
1524 | for_each_cpu(sibling, cpu_smt_mask(cpu)) { | |
1525 | if (cpu == sibling) | |
1526 | continue; | |
1527 | ||
1528 | if (!idle_cpu(cpu)) | |
1529 | return false; | |
1530 | } | |
1531 | #endif | |
1532 | ||
1533 | return true; | |
1534 | } | |
1535 | ||
58d081b5 MG |
1536 | struct task_numa_env { |
1537 | struct task_struct *p; | |
e6628d5b | 1538 | |
58d081b5 MG |
1539 | int src_cpu, src_nid; |
1540 | int dst_cpu, dst_nid; | |
e6628d5b | 1541 | |
58d081b5 | 1542 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1543 | |
40ea2b42 | 1544 | int imbalance_pct; |
7bd95320 | 1545 | int dist; |
fb13c7ee MG |
1546 | |
1547 | struct task_struct *best_task; | |
1548 | long best_imp; | |
58d081b5 MG |
1549 | int best_cpu; |
1550 | }; | |
1551 | ||
6499b1b2 VG |
1552 | static unsigned long cpu_load(struct rq *rq); |
1553 | static unsigned long cpu_util(int cpu); | |
fb86f5b2 | 1554 | static inline long adjust_numa_imbalance(int imbalance, int src_nr_running); |
6499b1b2 VG |
1555 | |
1556 | static inline enum | |
1557 | numa_type numa_classify(unsigned int imbalance_pct, | |
1558 | struct numa_stats *ns) | |
1559 | { | |
1560 | if ((ns->nr_running > ns->weight) && | |
1561 | ((ns->compute_capacity * 100) < (ns->util * imbalance_pct))) | |
1562 | return node_overloaded; | |
1563 | ||
1564 | if ((ns->nr_running < ns->weight) || | |
1565 | ((ns->compute_capacity * 100) > (ns->util * imbalance_pct))) | |
1566 | return node_has_spare; | |
1567 | ||
1568 | return node_fully_busy; | |
1569 | } | |
1570 | ||
76c389ab VS |
1571 | #ifdef CONFIG_SCHED_SMT |
1572 | /* Forward declarations of select_idle_sibling helpers */ | |
1573 | static inline bool test_idle_cores(int cpu, bool def); | |
ff7db0bf MG |
1574 | static inline int numa_idle_core(int idle_core, int cpu) |
1575 | { | |
ff7db0bf MG |
1576 | if (!static_branch_likely(&sched_smt_present) || |
1577 | idle_core >= 0 || !test_idle_cores(cpu, false)) | |
1578 | return idle_core; | |
1579 | ||
1580 | /* | |
1581 | * Prefer cores instead of packing HT siblings | |
1582 | * and triggering future load balancing. | |
1583 | */ | |
1584 | if (is_core_idle(cpu)) | |
1585 | idle_core = cpu; | |
ff7db0bf MG |
1586 | |
1587 | return idle_core; | |
1588 | } | |
76c389ab VS |
1589 | #else |
1590 | static inline int numa_idle_core(int idle_core, int cpu) | |
1591 | { | |
1592 | return idle_core; | |
1593 | } | |
1594 | #endif | |
ff7db0bf | 1595 | |
6499b1b2 | 1596 | /* |
ff7db0bf MG |
1597 | * Gather all necessary information to make NUMA balancing placement |
1598 | * decisions that are compatible with standard load balancer. This | |
1599 | * borrows code and logic from update_sg_lb_stats but sharing a | |
1600 | * common implementation is impractical. | |
6499b1b2 VG |
1601 | */ |
1602 | static void update_numa_stats(struct task_numa_env *env, | |
ff7db0bf MG |
1603 | struct numa_stats *ns, int nid, |
1604 | bool find_idle) | |
6499b1b2 | 1605 | { |
ff7db0bf | 1606 | int cpu, idle_core = -1; |
6499b1b2 VG |
1607 | |
1608 | memset(ns, 0, sizeof(*ns)); | |
ff7db0bf MG |
1609 | ns->idle_cpu = -1; |
1610 | ||
0621df31 | 1611 | rcu_read_lock(); |
6499b1b2 VG |
1612 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
1613 | struct rq *rq = cpu_rq(cpu); | |
1614 | ||
1615 | ns->load += cpu_load(rq); | |
1616 | ns->util += cpu_util(cpu); | |
1617 | ns->nr_running += rq->cfs.h_nr_running; | |
1618 | ns->compute_capacity += capacity_of(cpu); | |
ff7db0bf MG |
1619 | |
1620 | if (find_idle && !rq->nr_running && idle_cpu(cpu)) { | |
1621 | if (READ_ONCE(rq->numa_migrate_on) || | |
1622 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) | |
1623 | continue; | |
1624 | ||
1625 | if (ns->idle_cpu == -1) | |
1626 | ns->idle_cpu = cpu; | |
1627 | ||
1628 | idle_core = numa_idle_core(idle_core, cpu); | |
1629 | } | |
6499b1b2 | 1630 | } |
0621df31 | 1631 | rcu_read_unlock(); |
6499b1b2 VG |
1632 | |
1633 | ns->weight = cpumask_weight(cpumask_of_node(nid)); | |
1634 | ||
1635 | ns->node_type = numa_classify(env->imbalance_pct, ns); | |
ff7db0bf MG |
1636 | |
1637 | if (idle_core >= 0) | |
1638 | ns->idle_cpu = idle_core; | |
6499b1b2 VG |
1639 | } |
1640 | ||
fb13c7ee MG |
1641 | static void task_numa_assign(struct task_numa_env *env, |
1642 | struct task_struct *p, long imp) | |
1643 | { | |
a4739eca SD |
1644 | struct rq *rq = cpu_rq(env->dst_cpu); |
1645 | ||
5fb52dd9 MG |
1646 | /* Check if run-queue part of active NUMA balance. */ |
1647 | if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { | |
1648 | int cpu; | |
1649 | int start = env->dst_cpu; | |
1650 | ||
1651 | /* Find alternative idle CPU. */ | |
1652 | for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { | |
1653 | if (cpu == env->best_cpu || !idle_cpu(cpu) || | |
1654 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { | |
1655 | continue; | |
1656 | } | |
1657 | ||
1658 | env->dst_cpu = cpu; | |
1659 | rq = cpu_rq(env->dst_cpu); | |
1660 | if (!xchg(&rq->numa_migrate_on, 1)) | |
1661 | goto assign; | |
1662 | } | |
1663 | ||
1664 | /* Failed to find an alternative idle CPU */ | |
a4739eca | 1665 | return; |
5fb52dd9 | 1666 | } |
a4739eca | 1667 | |
5fb52dd9 | 1668 | assign: |
a4739eca SD |
1669 | /* |
1670 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1671 | * found a better CPU to move/swap. | |
1672 | */ | |
5fb52dd9 | 1673 | if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { |
a4739eca SD |
1674 | rq = cpu_rq(env->best_cpu); |
1675 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1676 | } | |
1677 | ||
fb13c7ee MG |
1678 | if (env->best_task) |
1679 | put_task_struct(env->best_task); | |
bac78573 ON |
1680 | if (p) |
1681 | get_task_struct(p); | |
fb13c7ee MG |
1682 | |
1683 | env->best_task = p; | |
1684 | env->best_imp = imp; | |
1685 | env->best_cpu = env->dst_cpu; | |
1686 | } | |
1687 | ||
28a21745 | 1688 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1689 | struct task_numa_env *env) |
1690 | { | |
e4991b24 RR |
1691 | long imb, old_imb; |
1692 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1693 | long src_capacity, dst_capacity; |
1694 | ||
1695 | /* | |
1696 | * The load is corrected for the CPU capacity available on each node. | |
1697 | * | |
1698 | * src_load dst_load | |
1699 | * ------------ vs --------- | |
1700 | * src_capacity dst_capacity | |
1701 | */ | |
1702 | src_capacity = env->src_stats.compute_capacity; | |
1703 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1704 | |
5f95ba7a | 1705 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1706 | |
28a21745 | 1707 | orig_src_load = env->src_stats.load; |
e4991b24 | 1708 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1709 | |
5f95ba7a | 1710 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1711 | |
1712 | /* Would this change make things worse? */ | |
1713 | return (imb > old_imb); | |
e63da036 RR |
1714 | } |
1715 | ||
6fd98e77 SD |
1716 | /* |
1717 | * Maximum NUMA importance can be 1998 (2*999); | |
1718 | * SMALLIMP @ 30 would be close to 1998/64. | |
1719 | * Used to deter task migration. | |
1720 | */ | |
1721 | #define SMALLIMP 30 | |
1722 | ||
fb13c7ee MG |
1723 | /* |
1724 | * This checks if the overall compute and NUMA accesses of the system would | |
1725 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1726 | * into account that it might be best if task running on the dst_cpu should | |
1727 | * be exchanged with the source task | |
1728 | */ | |
a0f03b61 | 1729 | static bool task_numa_compare(struct task_numa_env *env, |
305c1fac | 1730 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1731 | { |
cb361d8c | 1732 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 1733 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 1734 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 1735 | struct task_struct *cur; |
28a21745 | 1736 | long src_load, dst_load; |
7bd95320 | 1737 | int dist = env->dist; |
cb361d8c JH |
1738 | long moveimp = imp; |
1739 | long load; | |
a0f03b61 | 1740 | bool stopsearch = false; |
fb13c7ee | 1741 | |
a4739eca | 1742 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
a0f03b61 | 1743 | return false; |
a4739eca | 1744 | |
fb13c7ee | 1745 | rcu_read_lock(); |
154abafc | 1746 | cur = rcu_dereference(dst_rq->curr); |
bac78573 | 1747 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) |
fb13c7ee MG |
1748 | cur = NULL; |
1749 | ||
7af68335 PZ |
1750 | /* |
1751 | * Because we have preemption enabled we can get migrated around and | |
1752 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1753 | */ | |
a0f03b61 MG |
1754 | if (cur == env->p) { |
1755 | stopsearch = true; | |
7af68335 | 1756 | goto unlock; |
a0f03b61 | 1757 | } |
7af68335 | 1758 | |
305c1fac | 1759 | if (!cur) { |
6fd98e77 | 1760 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1761 | goto assign; |
1762 | else | |
1763 | goto unlock; | |
1764 | } | |
1765 | ||
88cca72c MG |
1766 | /* Skip this swap candidate if cannot move to the source cpu. */ |
1767 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) | |
1768 | goto unlock; | |
1769 | ||
1770 | /* | |
1771 | * Skip this swap candidate if it is not moving to its preferred | |
1772 | * node and the best task is. | |
1773 | */ | |
1774 | if (env->best_task && | |
1775 | env->best_task->numa_preferred_nid == env->src_nid && | |
1776 | cur->numa_preferred_nid != env->src_nid) { | |
1777 | goto unlock; | |
1778 | } | |
1779 | ||
fb13c7ee MG |
1780 | /* |
1781 | * "imp" is the fault differential for the source task between the | |
1782 | * source and destination node. Calculate the total differential for | |
1783 | * the source task and potential destination task. The more negative | |
305c1fac | 1784 | * the value is, the more remote accesses that would be expected to |
fb13c7ee | 1785 | * be incurred if the tasks were swapped. |
88cca72c | 1786 | * |
305c1fac SD |
1787 | * If dst and source tasks are in the same NUMA group, or not |
1788 | * in any group then look only at task weights. | |
1789 | */ | |
cb361d8c JH |
1790 | cur_ng = rcu_dereference(cur->numa_group); |
1791 | if (cur_ng == p_ng) { | |
305c1fac SD |
1792 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1793 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1794 | /* |
305c1fac SD |
1795 | * Add some hysteresis to prevent swapping the |
1796 | * tasks within a group over tiny differences. | |
887c290e | 1797 | */ |
cb361d8c | 1798 | if (cur_ng) |
305c1fac SD |
1799 | imp -= imp / 16; |
1800 | } else { | |
1801 | /* | |
1802 | * Compare the group weights. If a task is all by itself | |
1803 | * (not part of a group), use the task weight instead. | |
1804 | */ | |
cb361d8c | 1805 | if (cur_ng && p_ng) |
305c1fac SD |
1806 | imp += group_weight(cur, env->src_nid, dist) - |
1807 | group_weight(cur, env->dst_nid, dist); | |
1808 | else | |
1809 | imp += task_weight(cur, env->src_nid, dist) - | |
1810 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
1811 | } |
1812 | ||
88cca72c MG |
1813 | /* Discourage picking a task already on its preferred node */ |
1814 | if (cur->numa_preferred_nid == env->dst_nid) | |
1815 | imp -= imp / 16; | |
1816 | ||
1817 | /* | |
1818 | * Encourage picking a task that moves to its preferred node. | |
1819 | * This potentially makes imp larger than it's maximum of | |
1820 | * 1998 (see SMALLIMP and task_weight for why) but in this | |
1821 | * case, it does not matter. | |
1822 | */ | |
1823 | if (cur->numa_preferred_nid == env->src_nid) | |
1824 | imp += imp / 8; | |
1825 | ||
305c1fac | 1826 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 1827 | imp = moveimp; |
305c1fac | 1828 | cur = NULL; |
fb13c7ee | 1829 | goto assign; |
305c1fac | 1830 | } |
fb13c7ee | 1831 | |
88cca72c MG |
1832 | /* |
1833 | * Prefer swapping with a task moving to its preferred node over a | |
1834 | * task that is not. | |
1835 | */ | |
1836 | if (env->best_task && cur->numa_preferred_nid == env->src_nid && | |
1837 | env->best_task->numa_preferred_nid != env->src_nid) { | |
1838 | goto assign; | |
1839 | } | |
1840 | ||
6fd98e77 SD |
1841 | /* |
1842 | * If the NUMA importance is less than SMALLIMP, | |
1843 | * task migration might only result in ping pong | |
1844 | * of tasks and also hurt performance due to cache | |
1845 | * misses. | |
1846 | */ | |
1847 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
1848 | goto unlock; | |
1849 | ||
fb13c7ee MG |
1850 | /* |
1851 | * In the overloaded case, try and keep the load balanced. | |
1852 | */ | |
305c1fac SD |
1853 | load = task_h_load(env->p) - task_h_load(cur); |
1854 | if (!load) | |
1855 | goto assign; | |
1856 | ||
e720fff6 PZ |
1857 | dst_load = env->dst_stats.load + load; |
1858 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1859 | |
28a21745 | 1860 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1861 | goto unlock; |
1862 | ||
305c1fac | 1863 | assign: |
ff7db0bf | 1864 | /* Evaluate an idle CPU for a task numa move. */ |
10e2f1ac | 1865 | if (!cur) { |
ff7db0bf MG |
1866 | int cpu = env->dst_stats.idle_cpu; |
1867 | ||
1868 | /* Nothing cached so current CPU went idle since the search. */ | |
1869 | if (cpu < 0) | |
1870 | cpu = env->dst_cpu; | |
1871 | ||
10e2f1ac | 1872 | /* |
ff7db0bf MG |
1873 | * If the CPU is no longer truly idle and the previous best CPU |
1874 | * is, keep using it. | |
10e2f1ac | 1875 | */ |
ff7db0bf MG |
1876 | if (!idle_cpu(cpu) && env->best_cpu >= 0 && |
1877 | idle_cpu(env->best_cpu)) { | |
1878 | cpu = env->best_cpu; | |
1879 | } | |
1880 | ||
ff7db0bf | 1881 | env->dst_cpu = cpu; |
10e2f1ac | 1882 | } |
ba7e5a27 | 1883 | |
fb13c7ee | 1884 | task_numa_assign(env, cur, imp); |
a0f03b61 MG |
1885 | |
1886 | /* | |
1887 | * If a move to idle is allowed because there is capacity or load | |
1888 | * balance improves then stop the search. While a better swap | |
1889 | * candidate may exist, a search is not free. | |
1890 | */ | |
1891 | if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) | |
1892 | stopsearch = true; | |
1893 | ||
1894 | /* | |
1895 | * If a swap candidate must be identified and the current best task | |
1896 | * moves its preferred node then stop the search. | |
1897 | */ | |
1898 | if (!maymove && env->best_task && | |
1899 | env->best_task->numa_preferred_nid == env->src_nid) { | |
1900 | stopsearch = true; | |
1901 | } | |
fb13c7ee MG |
1902 | unlock: |
1903 | rcu_read_unlock(); | |
a0f03b61 MG |
1904 | |
1905 | return stopsearch; | |
fb13c7ee MG |
1906 | } |
1907 | ||
887c290e RR |
1908 | static void task_numa_find_cpu(struct task_numa_env *env, |
1909 | long taskimp, long groupimp) | |
2c8a50aa | 1910 | { |
305c1fac | 1911 | bool maymove = false; |
2c8a50aa MG |
1912 | int cpu; |
1913 | ||
305c1fac | 1914 | /* |
fb86f5b2 MG |
1915 | * If dst node has spare capacity, then check if there is an |
1916 | * imbalance that would be overruled by the load balancer. | |
305c1fac | 1917 | */ |
fb86f5b2 MG |
1918 | if (env->dst_stats.node_type == node_has_spare) { |
1919 | unsigned int imbalance; | |
1920 | int src_running, dst_running; | |
1921 | ||
1922 | /* | |
1923 | * Would movement cause an imbalance? Note that if src has | |
1924 | * more running tasks that the imbalance is ignored as the | |
1925 | * move improves the imbalance from the perspective of the | |
1926 | * CPU load balancer. | |
1927 | * */ | |
1928 | src_running = env->src_stats.nr_running - 1; | |
1929 | dst_running = env->dst_stats.nr_running + 1; | |
1930 | imbalance = max(0, dst_running - src_running); | |
1931 | imbalance = adjust_numa_imbalance(imbalance, src_running); | |
1932 | ||
1933 | /* Use idle CPU if there is no imbalance */ | |
ff7db0bf | 1934 | if (!imbalance) { |
fb86f5b2 | 1935 | maymove = true; |
ff7db0bf MG |
1936 | if (env->dst_stats.idle_cpu >= 0) { |
1937 | env->dst_cpu = env->dst_stats.idle_cpu; | |
1938 | task_numa_assign(env, NULL, 0); | |
1939 | return; | |
1940 | } | |
1941 | } | |
fb86f5b2 MG |
1942 | } else { |
1943 | long src_load, dst_load, load; | |
1944 | /* | |
1945 | * If the improvement from just moving env->p direction is better | |
1946 | * than swapping tasks around, check if a move is possible. | |
1947 | */ | |
1948 | load = task_h_load(env->p); | |
1949 | dst_load = env->dst_stats.load + load; | |
1950 | src_load = env->src_stats.load - load; | |
1951 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
1952 | } | |
305c1fac | 1953 | |
2c8a50aa MG |
1954 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1955 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 1956 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
1957 | continue; |
1958 | ||
1959 | env->dst_cpu = cpu; | |
a0f03b61 MG |
1960 | if (task_numa_compare(env, taskimp, groupimp, maymove)) |
1961 | break; | |
2c8a50aa MG |
1962 | } |
1963 | } | |
1964 | ||
58d081b5 MG |
1965 | static int task_numa_migrate(struct task_struct *p) |
1966 | { | |
58d081b5 MG |
1967 | struct task_numa_env env = { |
1968 | .p = p, | |
fb13c7ee | 1969 | |
58d081b5 | 1970 | .src_cpu = task_cpu(p), |
b32e86b4 | 1971 | .src_nid = task_node(p), |
fb13c7ee MG |
1972 | |
1973 | .imbalance_pct = 112, | |
1974 | ||
1975 | .best_task = NULL, | |
1976 | .best_imp = 0, | |
4142c3eb | 1977 | .best_cpu = -1, |
58d081b5 | 1978 | }; |
cb361d8c | 1979 | unsigned long taskweight, groupweight; |
58d081b5 | 1980 | struct sched_domain *sd; |
cb361d8c JH |
1981 | long taskimp, groupimp; |
1982 | struct numa_group *ng; | |
a4739eca | 1983 | struct rq *best_rq; |
7bd95320 | 1984 | int nid, ret, dist; |
e6628d5b | 1985 | |
58d081b5 | 1986 | /* |
fb13c7ee MG |
1987 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1988 | * imbalance and would be the first to start moving tasks about. | |
1989 | * | |
1990 | * And we want to avoid any moving of tasks about, as that would create | |
1991 | * random movement of tasks -- counter the numa conditions we're trying | |
1992 | * to satisfy here. | |
58d081b5 MG |
1993 | */ |
1994 | rcu_read_lock(); | |
fb13c7ee | 1995 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1996 | if (sd) |
1997 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1998 | rcu_read_unlock(); |
1999 | ||
46a73e8a RR |
2000 | /* |
2001 | * Cpusets can break the scheduler domain tree into smaller | |
2002 | * balance domains, some of which do not cross NUMA boundaries. | |
2003 | * Tasks that are "trapped" in such domains cannot be migrated | |
2004 | * elsewhere, so there is no point in (re)trying. | |
2005 | */ | |
2006 | if (unlikely(!sd)) { | |
8cd45eee | 2007 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
2008 | return -EINVAL; |
2009 | } | |
2010 | ||
2c8a50aa | 2011 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
2012 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
2013 | taskweight = task_weight(p, env.src_nid, dist); | |
2014 | groupweight = group_weight(p, env.src_nid, dist); | |
ff7db0bf | 2015 | update_numa_stats(&env, &env.src_stats, env.src_nid, false); |
7bd95320 RR |
2016 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
2017 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
ff7db0bf | 2018 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
58d081b5 | 2019 | |
a43455a1 | 2020 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 2021 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 2022 | |
9de05d48 RR |
2023 | /* |
2024 | * Look at other nodes in these cases: | |
2025 | * - there is no space available on the preferred_nid | |
2026 | * - the task is part of a numa_group that is interleaved across | |
2027 | * multiple NUMA nodes; in order to better consolidate the group, | |
2028 | * we need to check other locations. | |
2029 | */ | |
cb361d8c JH |
2030 | ng = deref_curr_numa_group(p); |
2031 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
2c8a50aa MG |
2032 | for_each_online_node(nid) { |
2033 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
2034 | continue; | |
58d081b5 | 2035 | |
7bd95320 | 2036 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
2037 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
2038 | dist != env.dist) { | |
2039 | taskweight = task_weight(p, env.src_nid, dist); | |
2040 | groupweight = group_weight(p, env.src_nid, dist); | |
2041 | } | |
7bd95320 | 2042 | |
83e1d2cd | 2043 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
2044 | taskimp = task_weight(p, nid, dist) - taskweight; |
2045 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 2046 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
2047 | continue; |
2048 | ||
7bd95320 | 2049 | env.dist = dist; |
2c8a50aa | 2050 | env.dst_nid = nid; |
ff7db0bf | 2051 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
2d4056fa | 2052 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
2053 | } |
2054 | } | |
2055 | ||
68d1b02a RR |
2056 | /* |
2057 | * If the task is part of a workload that spans multiple NUMA nodes, | |
2058 | * and is migrating into one of the workload's active nodes, remember | |
2059 | * this node as the task's preferred numa node, so the workload can | |
2060 | * settle down. | |
2061 | * A task that migrated to a second choice node will be better off | |
2062 | * trying for a better one later. Do not set the preferred node here. | |
2063 | */ | |
cb361d8c | 2064 | if (ng) { |
db015dae RR |
2065 | if (env.best_cpu == -1) |
2066 | nid = env.src_nid; | |
2067 | else | |
8cd45eee | 2068 | nid = cpu_to_node(env.best_cpu); |
db015dae | 2069 | |
8cd45eee SD |
2070 | if (nid != p->numa_preferred_nid) |
2071 | sched_setnuma(p, nid); | |
db015dae RR |
2072 | } |
2073 | ||
2074 | /* No better CPU than the current one was found. */ | |
f22aef4a | 2075 | if (env.best_cpu == -1) { |
b2b2042b | 2076 | trace_sched_stick_numa(p, env.src_cpu, NULL, -1); |
db015dae | 2077 | return -EAGAIN; |
f22aef4a | 2078 | } |
0ec8aa00 | 2079 | |
a4739eca | 2080 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 2081 | if (env.best_task == NULL) { |
286549dc | 2082 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 2083 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc | 2084 | if (ret != 0) |
b2b2042b | 2085 | trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); |
fb13c7ee MG |
2086 | return ret; |
2087 | } | |
2088 | ||
0ad4e3df | 2089 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 2090 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 2091 | |
286549dc | 2092 | if (ret != 0) |
b2b2042b | 2093 | trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); |
fb13c7ee MG |
2094 | put_task_struct(env.best_task); |
2095 | return ret; | |
e6628d5b MG |
2096 | } |
2097 | ||
6b9a7460 MG |
2098 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
2099 | static void numa_migrate_preferred(struct task_struct *p) | |
2100 | { | |
5085e2a3 RR |
2101 | unsigned long interval = HZ; |
2102 | ||
2739d3ee | 2103 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 2104 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
2105 | return; |
2106 | ||
2739d3ee | 2107 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 2108 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 2109 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
2110 | |
2111 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 2112 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
2113 | return; |
2114 | ||
2115 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 2116 | task_numa_migrate(p); |
6b9a7460 MG |
2117 | } |
2118 | ||
20e07dea | 2119 | /* |
4142c3eb | 2120 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
2121 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
2122 | * be different from the set of nodes where the workload's memory is currently | |
2123 | * located. | |
20e07dea | 2124 | */ |
4142c3eb | 2125 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
2126 | { |
2127 | unsigned long faults, max_faults = 0; | |
4142c3eb | 2128 | int nid, active_nodes = 0; |
20e07dea RR |
2129 | |
2130 | for_each_online_node(nid) { | |
2131 | faults = group_faults_cpu(numa_group, nid); | |
2132 | if (faults > max_faults) | |
2133 | max_faults = faults; | |
2134 | } | |
2135 | ||
2136 | for_each_online_node(nid) { | |
2137 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
2138 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
2139 | active_nodes++; | |
20e07dea | 2140 | } |
4142c3eb RR |
2141 | |
2142 | numa_group->max_faults_cpu = max_faults; | |
2143 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
2144 | } |
2145 | ||
04bb2f94 RR |
2146 | /* |
2147 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
2148 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
2149 | * period will be for the next scan window. If local/(local+remote) ratio is |
2150 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
2151 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
2152 | */ |
2153 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 2154 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
2155 | |
2156 | /* | |
2157 | * Increase the scan period (slow down scanning) if the majority of | |
2158 | * our memory is already on our local node, or if the majority of | |
2159 | * the page accesses are shared with other processes. | |
2160 | * Otherwise, decrease the scan period. | |
2161 | */ | |
2162 | static void update_task_scan_period(struct task_struct *p, | |
2163 | unsigned long shared, unsigned long private) | |
2164 | { | |
2165 | unsigned int period_slot; | |
37ec97de | 2166 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
2167 | int diff; |
2168 | ||
2169 | unsigned long remote = p->numa_faults_locality[0]; | |
2170 | unsigned long local = p->numa_faults_locality[1]; | |
2171 | ||
2172 | /* | |
2173 | * If there were no record hinting faults then either the task is | |
2174 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
2175 | * to automatic numa balancing. Related to that, if there were failed |
2176 | * migration then it implies we are migrating too quickly or the local | |
2177 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 2178 | */ |
074c2381 | 2179 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
2180 | p->numa_scan_period = min(p->numa_scan_period_max, |
2181 | p->numa_scan_period << 1); | |
2182 | ||
2183 | p->mm->numa_next_scan = jiffies + | |
2184 | msecs_to_jiffies(p->numa_scan_period); | |
2185 | ||
2186 | return; | |
2187 | } | |
2188 | ||
2189 | /* | |
2190 | * Prepare to scale scan period relative to the current period. | |
2191 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
2192 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
2193 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
2194 | */ | |
2195 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
2196 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
2197 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
2198 | ||
2199 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2200 | /* | |
2201 | * Most memory accesses are local. There is no need to | |
2202 | * do fast NUMA scanning, since memory is already local. | |
2203 | */ | |
2204 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
2205 | if (!slot) | |
2206 | slot = 1; | |
2207 | diff = slot * period_slot; | |
2208 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2209 | /* | |
2210 | * Most memory accesses are shared with other tasks. | |
2211 | * There is no point in continuing fast NUMA scanning, | |
2212 | * since other tasks may just move the memory elsewhere. | |
2213 | */ | |
2214 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
2215 | if (!slot) |
2216 | slot = 1; | |
2217 | diff = slot * period_slot; | |
2218 | } else { | |
04bb2f94 | 2219 | /* |
37ec97de RR |
2220 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
2221 | * yet they are not on the local NUMA node. Speed up | |
2222 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 2223 | */ |
37ec97de RR |
2224 | int ratio = max(lr_ratio, ps_ratio); |
2225 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
2226 | } |
2227 | ||
2228 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2229 | task_scan_min(p), task_scan_max(p)); | |
2230 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2231 | } | |
2232 | ||
7e2703e6 RR |
2233 | /* |
2234 | * Get the fraction of time the task has been running since the last | |
2235 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2236 | * decays those on a 32ms period, which is orders of magnitude off | |
2237 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2238 | * stats only if the task is so new there are no NUMA statistics yet. | |
2239 | */ | |
2240 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2241 | { | |
2242 | u64 runtime, delta, now; | |
2243 | /* Use the start of this time slice to avoid calculations. */ | |
2244 | now = p->se.exec_start; | |
2245 | runtime = p->se.sum_exec_runtime; | |
2246 | ||
2247 | if (p->last_task_numa_placement) { | |
2248 | delta = runtime - p->last_sum_exec_runtime; | |
2249 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2250 | |
2251 | /* Avoid time going backwards, prevent potential divide error: */ | |
2252 | if (unlikely((s64)*period < 0)) | |
2253 | *period = 0; | |
7e2703e6 | 2254 | } else { |
c7b50216 | 2255 | delta = p->se.avg.load_sum; |
9d89c257 | 2256 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2257 | } |
2258 | ||
2259 | p->last_sum_exec_runtime = runtime; | |
2260 | p->last_task_numa_placement = now; | |
2261 | ||
2262 | return delta; | |
2263 | } | |
2264 | ||
54009416 RR |
2265 | /* |
2266 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2267 | * be done in a way that produces consistent results with group_weight, | |
2268 | * otherwise workloads might not converge. | |
2269 | */ | |
2270 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2271 | { | |
2272 | nodemask_t nodes; | |
2273 | int dist; | |
2274 | ||
2275 | /* Direct connections between all NUMA nodes. */ | |
2276 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2277 | return nid; | |
2278 | ||
2279 | /* | |
2280 | * On a system with glueless mesh NUMA topology, group_weight | |
2281 | * scores nodes according to the number of NUMA hinting faults on | |
2282 | * both the node itself, and on nearby nodes. | |
2283 | */ | |
2284 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2285 | unsigned long score, max_score = 0; | |
2286 | int node, max_node = nid; | |
2287 | ||
2288 | dist = sched_max_numa_distance; | |
2289 | ||
2290 | for_each_online_node(node) { | |
2291 | score = group_weight(p, node, dist); | |
2292 | if (score > max_score) { | |
2293 | max_score = score; | |
2294 | max_node = node; | |
2295 | } | |
2296 | } | |
2297 | return max_node; | |
2298 | } | |
2299 | ||
2300 | /* | |
2301 | * Finding the preferred nid in a system with NUMA backplane | |
2302 | * interconnect topology is more involved. The goal is to locate | |
2303 | * tasks from numa_groups near each other in the system, and | |
2304 | * untangle workloads from different sides of the system. This requires | |
2305 | * searching down the hierarchy of node groups, recursively searching | |
2306 | * inside the highest scoring group of nodes. The nodemask tricks | |
2307 | * keep the complexity of the search down. | |
2308 | */ | |
2309 | nodes = node_online_map; | |
2310 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2311 | unsigned long max_faults = 0; | |
81907478 | 2312 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2313 | int a, b; |
2314 | ||
2315 | /* Are there nodes at this distance from each other? */ | |
2316 | if (!find_numa_distance(dist)) | |
2317 | continue; | |
2318 | ||
2319 | for_each_node_mask(a, nodes) { | |
2320 | unsigned long faults = 0; | |
2321 | nodemask_t this_group; | |
2322 | nodes_clear(this_group); | |
2323 | ||
2324 | /* Sum group's NUMA faults; includes a==b case. */ | |
2325 | for_each_node_mask(b, nodes) { | |
2326 | if (node_distance(a, b) < dist) { | |
2327 | faults += group_faults(p, b); | |
2328 | node_set(b, this_group); | |
2329 | node_clear(b, nodes); | |
2330 | } | |
2331 | } | |
2332 | ||
2333 | /* Remember the top group. */ | |
2334 | if (faults > max_faults) { | |
2335 | max_faults = faults; | |
2336 | max_group = this_group; | |
2337 | /* | |
2338 | * subtle: at the smallest distance there is | |
2339 | * just one node left in each "group", the | |
2340 | * winner is the preferred nid. | |
2341 | */ | |
2342 | nid = a; | |
2343 | } | |
2344 | } | |
2345 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2346 | if (!max_faults) |
2347 | break; | |
54009416 RR |
2348 | nodes = max_group; |
2349 | } | |
2350 | return nid; | |
2351 | } | |
2352 | ||
cbee9f88 PZ |
2353 | static void task_numa_placement(struct task_struct *p) |
2354 | { | |
98fa15f3 | 2355 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2356 | unsigned long max_faults = 0; |
04bb2f94 | 2357 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2358 | unsigned long total_faults; |
2359 | u64 runtime, period; | |
7dbd13ed | 2360 | spinlock_t *group_lock = NULL; |
cb361d8c | 2361 | struct numa_group *ng; |
cbee9f88 | 2362 | |
7e5a2c17 JL |
2363 | /* |
2364 | * The p->mm->numa_scan_seq field gets updated without | |
2365 | * exclusive access. Use READ_ONCE() here to ensure | |
2366 | * that the field is read in a single access: | |
2367 | */ | |
316c1608 | 2368 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2369 | if (p->numa_scan_seq == seq) |
2370 | return; | |
2371 | p->numa_scan_seq = seq; | |
598f0ec0 | 2372 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2373 | |
7e2703e6 RR |
2374 | total_faults = p->numa_faults_locality[0] + |
2375 | p->numa_faults_locality[1]; | |
2376 | runtime = numa_get_avg_runtime(p, &period); | |
2377 | ||
7dbd13ed | 2378 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2379 | ng = deref_curr_numa_group(p); |
2380 | if (ng) { | |
2381 | group_lock = &ng->lock; | |
60e69eed | 2382 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2383 | } |
2384 | ||
688b7585 MG |
2385 | /* Find the node with the highest number of faults */ |
2386 | for_each_online_node(nid) { | |
44dba3d5 IM |
2387 | /* Keep track of the offsets in numa_faults array */ |
2388 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2389 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2390 | int priv; |
745d6147 | 2391 | |
be1e4e76 | 2392 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2393 | long diff, f_diff, f_weight; |
8c8a743c | 2394 | |
44dba3d5 IM |
2395 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2396 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2397 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2398 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2399 | |
ac8e895b | 2400 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2401 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2402 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2403 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2404 | |
7e2703e6 RR |
2405 | /* |
2406 | * Normalize the faults_from, so all tasks in a group | |
2407 | * count according to CPU use, instead of by the raw | |
2408 | * number of faults. Tasks with little runtime have | |
2409 | * little over-all impact on throughput, and thus their | |
2410 | * faults are less important. | |
2411 | */ | |
2412 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2413 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2414 | (total_faults + 1); |
44dba3d5 IM |
2415 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2416 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2417 | |
44dba3d5 IM |
2418 | p->numa_faults[mem_idx] += diff; |
2419 | p->numa_faults[cpu_idx] += f_diff; | |
2420 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2421 | p->total_numa_faults += diff; |
cb361d8c | 2422 | if (ng) { |
44dba3d5 IM |
2423 | /* |
2424 | * safe because we can only change our own group | |
2425 | * | |
2426 | * mem_idx represents the offset for a given | |
2427 | * nid and priv in a specific region because it | |
2428 | * is at the beginning of the numa_faults array. | |
2429 | */ | |
cb361d8c JH |
2430 | ng->faults[mem_idx] += diff; |
2431 | ng->faults_cpu[mem_idx] += f_diff; | |
2432 | ng->total_faults += diff; | |
2433 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2434 | } |
ac8e895b MG |
2435 | } |
2436 | ||
cb361d8c | 2437 | if (!ng) { |
f03bb676 SD |
2438 | if (faults > max_faults) { |
2439 | max_faults = faults; | |
2440 | max_nid = nid; | |
2441 | } | |
2442 | } else if (group_faults > max_faults) { | |
2443 | max_faults = group_faults; | |
688b7585 MG |
2444 | max_nid = nid; |
2445 | } | |
83e1d2cd MG |
2446 | } |
2447 | ||
cb361d8c JH |
2448 | if (ng) { |
2449 | numa_group_count_active_nodes(ng); | |
60e69eed | 2450 | spin_unlock_irq(group_lock); |
f03bb676 | 2451 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2452 | } |
2453 | ||
bb97fc31 RR |
2454 | if (max_faults) { |
2455 | /* Set the new preferred node */ | |
2456 | if (max_nid != p->numa_preferred_nid) | |
2457 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2458 | } |
30619c89 SD |
2459 | |
2460 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2461 | } |
2462 | ||
8c8a743c PZ |
2463 | static inline int get_numa_group(struct numa_group *grp) |
2464 | { | |
c45a7795 | 2465 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2466 | } |
2467 | ||
2468 | static inline void put_numa_group(struct numa_group *grp) | |
2469 | { | |
c45a7795 | 2470 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2471 | kfree_rcu(grp, rcu); |
2472 | } | |
2473 | ||
3e6a9418 MG |
2474 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2475 | int *priv) | |
8c8a743c PZ |
2476 | { |
2477 | struct numa_group *grp, *my_grp; | |
2478 | struct task_struct *tsk; | |
2479 | bool join = false; | |
2480 | int cpu = cpupid_to_cpu(cpupid); | |
2481 | int i; | |
2482 | ||
cb361d8c | 2483 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2484 | unsigned int size = sizeof(struct numa_group) + |
50ec8a40 | 2485 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2486 | |
2487 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2488 | if (!grp) | |
2489 | return; | |
2490 | ||
c45a7795 | 2491 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2492 | grp->active_nodes = 1; |
2493 | grp->max_faults_cpu = 0; | |
8c8a743c | 2494 | spin_lock_init(&grp->lock); |
e29cf08b | 2495 | grp->gid = p->pid; |
50ec8a40 | 2496 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2497 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2498 | nr_node_ids; | |
8c8a743c | 2499 | |
be1e4e76 | 2500 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2501 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2502 | |
989348b5 | 2503 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2504 | |
8c8a743c PZ |
2505 | grp->nr_tasks++; |
2506 | rcu_assign_pointer(p->numa_group, grp); | |
2507 | } | |
2508 | ||
2509 | rcu_read_lock(); | |
316c1608 | 2510 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2511 | |
2512 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2513 | goto no_join; |
8c8a743c PZ |
2514 | |
2515 | grp = rcu_dereference(tsk->numa_group); | |
2516 | if (!grp) | |
3354781a | 2517 | goto no_join; |
8c8a743c | 2518 | |
cb361d8c | 2519 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2520 | if (grp == my_grp) |
3354781a | 2521 | goto no_join; |
8c8a743c PZ |
2522 | |
2523 | /* | |
2524 | * Only join the other group if its bigger; if we're the bigger group, | |
2525 | * the other task will join us. | |
2526 | */ | |
2527 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2528 | goto no_join; |
8c8a743c PZ |
2529 | |
2530 | /* | |
2531 | * Tie-break on the grp address. | |
2532 | */ | |
2533 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2534 | goto no_join; |
8c8a743c | 2535 | |
dabe1d99 RR |
2536 | /* Always join threads in the same process. */ |
2537 | if (tsk->mm == current->mm) | |
2538 | join = true; | |
2539 | ||
2540 | /* Simple filter to avoid false positives due to PID collisions */ | |
2541 | if (flags & TNF_SHARED) | |
2542 | join = true; | |
8c8a743c | 2543 | |
3e6a9418 MG |
2544 | /* Update priv based on whether false sharing was detected */ |
2545 | *priv = !join; | |
2546 | ||
dabe1d99 | 2547 | if (join && !get_numa_group(grp)) |
3354781a | 2548 | goto no_join; |
8c8a743c | 2549 | |
8c8a743c PZ |
2550 | rcu_read_unlock(); |
2551 | ||
2552 | if (!join) | |
2553 | return; | |
2554 | ||
60e69eed MG |
2555 | BUG_ON(irqs_disabled()); |
2556 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2557 | |
be1e4e76 | 2558 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2559 | my_grp->faults[i] -= p->numa_faults[i]; |
2560 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2561 | } |
989348b5 MG |
2562 | my_grp->total_faults -= p->total_numa_faults; |
2563 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2564 | |
8c8a743c PZ |
2565 | my_grp->nr_tasks--; |
2566 | grp->nr_tasks++; | |
2567 | ||
2568 | spin_unlock(&my_grp->lock); | |
60e69eed | 2569 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2570 | |
2571 | rcu_assign_pointer(p->numa_group, grp); | |
2572 | ||
2573 | put_numa_group(my_grp); | |
3354781a PZ |
2574 | return; |
2575 | ||
2576 | no_join: | |
2577 | rcu_read_unlock(); | |
2578 | return; | |
8c8a743c PZ |
2579 | } |
2580 | ||
16d51a59 JH |
2581 | /* |
2582 | * Get rid of NUMA staticstics associated with a task (either current or dead). | |
2583 | * If @final is set, the task is dead and has reached refcount zero, so we can | |
2584 | * safely free all relevant data structures. Otherwise, there might be | |
2585 | * concurrent reads from places like load balancing and procfs, and we should | |
2586 | * reset the data back to default state without freeing ->numa_faults. | |
2587 | */ | |
2588 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 2589 | { |
cb361d8c JH |
2590 | /* safe: p either is current or is being freed by current */ |
2591 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 2592 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
2593 | unsigned long flags; |
2594 | int i; | |
8c8a743c | 2595 | |
16d51a59 JH |
2596 | if (!numa_faults) |
2597 | return; | |
2598 | ||
8c8a743c | 2599 | if (grp) { |
e9dd685c | 2600 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2601 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2602 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2603 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2604 | |
8c8a743c | 2605 | grp->nr_tasks--; |
e9dd685c | 2606 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2607 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2608 | put_numa_group(grp); |
2609 | } | |
2610 | ||
16d51a59 JH |
2611 | if (final) { |
2612 | p->numa_faults = NULL; | |
2613 | kfree(numa_faults); | |
2614 | } else { | |
2615 | p->total_numa_faults = 0; | |
2616 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
2617 | numa_faults[i] = 0; | |
2618 | } | |
8c8a743c PZ |
2619 | } |
2620 | ||
cbee9f88 PZ |
2621 | /* |
2622 | * Got a PROT_NONE fault for a page on @node. | |
2623 | */ | |
58b46da3 | 2624 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2625 | { |
2626 | struct task_struct *p = current; | |
6688cc05 | 2627 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2628 | int cpu_node = task_node(current); |
792568ec | 2629 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2630 | struct numa_group *ng; |
ac8e895b | 2631 | int priv; |
cbee9f88 | 2632 | |
2a595721 | 2633 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2634 | return; |
2635 | ||
9ff1d9ff MG |
2636 | /* for example, ksmd faulting in a user's mm */ |
2637 | if (!p->mm) | |
2638 | return; | |
2639 | ||
f809ca9a | 2640 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2641 | if (unlikely(!p->numa_faults)) { |
2642 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2643 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2644 | |
44dba3d5 IM |
2645 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2646 | if (!p->numa_faults) | |
f809ca9a | 2647 | return; |
745d6147 | 2648 | |
83e1d2cd | 2649 | p->total_numa_faults = 0; |
04bb2f94 | 2650 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2651 | } |
cbee9f88 | 2652 | |
8c8a743c PZ |
2653 | /* |
2654 | * First accesses are treated as private, otherwise consider accesses | |
2655 | * to be private if the accessing pid has not changed | |
2656 | */ | |
2657 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2658 | priv = 1; | |
2659 | } else { | |
2660 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2661 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2662 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2663 | } |
2664 | ||
792568ec RR |
2665 | /* |
2666 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2667 | * occurs wholly within the set of nodes that the workload is | |
2668 | * actively using should be counted as local. This allows the | |
2669 | * scan rate to slow down when a workload has settled down. | |
2670 | */ | |
cb361d8c | 2671 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
2672 | if (!priv && !local && ng && ng->active_nodes > 1 && |
2673 | numa_is_active_node(cpu_node, ng) && | |
2674 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2675 | local = 1; |
2676 | ||
2739d3ee | 2677 | /* |
e1ff516a YW |
2678 | * Retry to migrate task to preferred node periodically, in case it |
2679 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2680 | */ |
b6a60cf3 SD |
2681 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2682 | task_numa_placement(p); | |
6b9a7460 | 2683 | numa_migrate_preferred(p); |
b6a60cf3 | 2684 | } |
6b9a7460 | 2685 | |
b32e86b4 IM |
2686 | if (migrated) |
2687 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2688 | if (flags & TNF_MIGRATE_FAIL) |
2689 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2690 | |
44dba3d5 IM |
2691 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2692 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2693 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2694 | } |
2695 | ||
6e5fb223 PZ |
2696 | static void reset_ptenuma_scan(struct task_struct *p) |
2697 | { | |
7e5a2c17 JL |
2698 | /* |
2699 | * We only did a read acquisition of the mmap sem, so | |
2700 | * p->mm->numa_scan_seq is written to without exclusive access | |
2701 | * and the update is not guaranteed to be atomic. That's not | |
2702 | * much of an issue though, since this is just used for | |
2703 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2704 | * expensive, to avoid any form of compiler optimizations: | |
2705 | */ | |
316c1608 | 2706 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2707 | p->mm->numa_scan_offset = 0; |
2708 | } | |
2709 | ||
cbee9f88 PZ |
2710 | /* |
2711 | * The expensive part of numa migration is done from task_work context. | |
2712 | * Triggered from task_tick_numa(). | |
2713 | */ | |
9434f9f5 | 2714 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
2715 | { |
2716 | unsigned long migrate, next_scan, now = jiffies; | |
2717 | struct task_struct *p = current; | |
2718 | struct mm_struct *mm = p->mm; | |
51170840 | 2719 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2720 | struct vm_area_struct *vma; |
9f40604c | 2721 | unsigned long start, end; |
598f0ec0 | 2722 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2723 | long pages, virtpages; |
cbee9f88 | 2724 | |
9148a3a1 | 2725 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 2726 | |
b34920d4 | 2727 | work->next = work; |
cbee9f88 PZ |
2728 | /* |
2729 | * Who cares about NUMA placement when they're dying. | |
2730 | * | |
2731 | * NOTE: make sure not to dereference p->mm before this check, | |
2732 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2733 | * without p->mm even though we still had it when we enqueued this | |
2734 | * work. | |
2735 | */ | |
2736 | if (p->flags & PF_EXITING) | |
2737 | return; | |
2738 | ||
930aa174 | 2739 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2740 | mm->numa_next_scan = now + |
2741 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2742 | } |
2743 | ||
cbee9f88 PZ |
2744 | /* |
2745 | * Enforce maximal scan/migration frequency.. | |
2746 | */ | |
2747 | migrate = mm->numa_next_scan; | |
2748 | if (time_before(now, migrate)) | |
2749 | return; | |
2750 | ||
598f0ec0 MG |
2751 | if (p->numa_scan_period == 0) { |
2752 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2753 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2754 | } |
cbee9f88 | 2755 | |
fb003b80 | 2756 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2757 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2758 | return; | |
2759 | ||
19a78d11 PZ |
2760 | /* |
2761 | * Delay this task enough that another task of this mm will likely win | |
2762 | * the next time around. | |
2763 | */ | |
2764 | p->node_stamp += 2 * TICK_NSEC; | |
2765 | ||
9f40604c MG |
2766 | start = mm->numa_scan_offset; |
2767 | pages = sysctl_numa_balancing_scan_size; | |
2768 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2769 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2770 | if (!pages) |
2771 | return; | |
cbee9f88 | 2772 | |
4620f8c1 | 2773 | |
8655d549 VB |
2774 | if (!down_read_trylock(&mm->mmap_sem)) |
2775 | return; | |
9f40604c | 2776 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2777 | if (!vma) { |
2778 | reset_ptenuma_scan(p); | |
9f40604c | 2779 | start = 0; |
6e5fb223 PZ |
2780 | vma = mm->mmap; |
2781 | } | |
9f40604c | 2782 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2783 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2784 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2785 | continue; |
6b79c57b | 2786 | } |
6e5fb223 | 2787 | |
4591ce4f MG |
2788 | /* |
2789 | * Shared library pages mapped by multiple processes are not | |
2790 | * migrated as it is expected they are cache replicated. Avoid | |
2791 | * hinting faults in read-only file-backed mappings or the vdso | |
2792 | * as migrating the pages will be of marginal benefit. | |
2793 | */ | |
2794 | if (!vma->vm_mm || | |
2795 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2796 | continue; | |
2797 | ||
3c67f474 MG |
2798 | /* |
2799 | * Skip inaccessible VMAs to avoid any confusion between | |
2800 | * PROT_NONE and NUMA hinting ptes | |
2801 | */ | |
3122e80e | 2802 | if (!vma_is_accessible(vma)) |
3c67f474 | 2803 | continue; |
4591ce4f | 2804 | |
9f40604c MG |
2805 | do { |
2806 | start = max(start, vma->vm_start); | |
2807 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2808 | end = min(end, vma->vm_end); | |
4620f8c1 | 2809 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2810 | |
2811 | /* | |
4620f8c1 RR |
2812 | * Try to scan sysctl_numa_balancing_size worth of |
2813 | * hpages that have at least one present PTE that | |
2814 | * is not already pte-numa. If the VMA contains | |
2815 | * areas that are unused or already full of prot_numa | |
2816 | * PTEs, scan up to virtpages, to skip through those | |
2817 | * areas faster. | |
598f0ec0 MG |
2818 | */ |
2819 | if (nr_pte_updates) | |
2820 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2821 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2822 | |
9f40604c | 2823 | start = end; |
4620f8c1 | 2824 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2825 | goto out; |
3cf1962c RR |
2826 | |
2827 | cond_resched(); | |
9f40604c | 2828 | } while (end != vma->vm_end); |
cbee9f88 | 2829 | } |
6e5fb223 | 2830 | |
9f40604c | 2831 | out: |
6e5fb223 | 2832 | /* |
c69307d5 PZ |
2833 | * It is possible to reach the end of the VMA list but the last few |
2834 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2835 | * would find the !migratable VMA on the next scan but not reset the | |
2836 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2837 | */ |
2838 | if (vma) | |
9f40604c | 2839 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2840 | else |
2841 | reset_ptenuma_scan(p); | |
2842 | up_read(&mm->mmap_sem); | |
51170840 RR |
2843 | |
2844 | /* | |
2845 | * Make sure tasks use at least 32x as much time to run other code | |
2846 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2847 | * Usually update_task_scan_period slows down scanning enough; on an | |
2848 | * overloaded system we need to limit overhead on a per task basis. | |
2849 | */ | |
2850 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2851 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2852 | p->node_stamp += 32 * diff; | |
2853 | } | |
cbee9f88 PZ |
2854 | } |
2855 | ||
d35927a1 VS |
2856 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
2857 | { | |
2858 | int mm_users = 0; | |
2859 | struct mm_struct *mm = p->mm; | |
2860 | ||
2861 | if (mm) { | |
2862 | mm_users = atomic_read(&mm->mm_users); | |
2863 | if (mm_users == 1) { | |
2864 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
2865 | mm->numa_scan_seq = 0; | |
2866 | } | |
2867 | } | |
2868 | p->node_stamp = 0; | |
2869 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
2870 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
b34920d4 | 2871 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
2872 | p->numa_work.next = &p->numa_work; |
2873 | p->numa_faults = NULL; | |
2874 | RCU_INIT_POINTER(p->numa_group, NULL); | |
2875 | p->last_task_numa_placement = 0; | |
2876 | p->last_sum_exec_runtime = 0; | |
2877 | ||
b34920d4 VS |
2878 | init_task_work(&p->numa_work, task_numa_work); |
2879 | ||
d35927a1 VS |
2880 | /* New address space, reset the preferred nid */ |
2881 | if (!(clone_flags & CLONE_VM)) { | |
2882 | p->numa_preferred_nid = NUMA_NO_NODE; | |
2883 | return; | |
2884 | } | |
2885 | ||
2886 | /* | |
2887 | * New thread, keep existing numa_preferred_nid which should be copied | |
2888 | * already by arch_dup_task_struct but stagger when scans start. | |
2889 | */ | |
2890 | if (mm) { | |
2891 | unsigned int delay; | |
2892 | ||
2893 | delay = min_t(unsigned int, task_scan_max(current), | |
2894 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
2895 | delay += 2 * TICK_NSEC; | |
2896 | p->node_stamp = delay; | |
2897 | } | |
2898 | } | |
2899 | ||
cbee9f88 PZ |
2900 | /* |
2901 | * Drive the periodic memory faults.. | |
2902 | */ | |
b1546edc | 2903 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
2904 | { |
2905 | struct callback_head *work = &curr->numa_work; | |
2906 | u64 period, now; | |
2907 | ||
2908 | /* | |
2909 | * We don't care about NUMA placement if we don't have memory. | |
2910 | */ | |
18f855e5 | 2911 | if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) |
cbee9f88 PZ |
2912 | return; |
2913 | ||
2914 | /* | |
2915 | * Using runtime rather than walltime has the dual advantage that | |
2916 | * we (mostly) drive the selection from busy threads and that the | |
2917 | * task needs to have done some actual work before we bother with | |
2918 | * NUMA placement. | |
2919 | */ | |
2920 | now = curr->se.sum_exec_runtime; | |
2921 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2922 | ||
25b3e5a3 | 2923 | if (now > curr->node_stamp + period) { |
4b96a29b | 2924 | if (!curr->node_stamp) |
b5dd77c8 | 2925 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2926 | curr->node_stamp += period; |
cbee9f88 | 2927 | |
b34920d4 | 2928 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
cbee9f88 | 2929 | task_work_add(curr, work, true); |
cbee9f88 PZ |
2930 | } |
2931 | } | |
3fed382b | 2932 | |
3f9672ba SD |
2933 | static void update_scan_period(struct task_struct *p, int new_cpu) |
2934 | { | |
2935 | int src_nid = cpu_to_node(task_cpu(p)); | |
2936 | int dst_nid = cpu_to_node(new_cpu); | |
2937 | ||
05cbdf4f MG |
2938 | if (!static_branch_likely(&sched_numa_balancing)) |
2939 | return; | |
2940 | ||
3f9672ba SD |
2941 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
2942 | return; | |
2943 | ||
05cbdf4f MG |
2944 | if (src_nid == dst_nid) |
2945 | return; | |
2946 | ||
2947 | /* | |
2948 | * Allow resets if faults have been trapped before one scan | |
2949 | * has completed. This is most likely due to a new task that | |
2950 | * is pulled cross-node due to wakeups or load balancing. | |
2951 | */ | |
2952 | if (p->numa_scan_seq) { | |
2953 | /* | |
2954 | * Avoid scan adjustments if moving to the preferred | |
2955 | * node or if the task was not previously running on | |
2956 | * the preferred node. | |
2957 | */ | |
2958 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
2959 | (p->numa_preferred_nid != NUMA_NO_NODE && |
2960 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
2961 | return; |
2962 | } | |
2963 | ||
2964 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
2965 | } |
2966 | ||
cbee9f88 PZ |
2967 | #else |
2968 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2969 | { | |
2970 | } | |
0ec8aa00 PZ |
2971 | |
2972 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2973 | { | |
2974 | } | |
2975 | ||
2976 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2977 | { | |
2978 | } | |
3fed382b | 2979 | |
3f9672ba SD |
2980 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
2981 | { | |
2982 | } | |
2983 | ||
cbee9f88 PZ |
2984 | #endif /* CONFIG_NUMA_BALANCING */ |
2985 | ||
30cfdcfc DA |
2986 | static void |
2987 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2988 | { | |
2989 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 2990 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2991 | if (entity_is_task(se)) { |
2992 | struct rq *rq = rq_of(cfs_rq); | |
2993 | ||
2994 | account_numa_enqueue(rq, task_of(se)); | |
2995 | list_add(&se->group_node, &rq->cfs_tasks); | |
2996 | } | |
367456c7 | 2997 | #endif |
30cfdcfc | 2998 | cfs_rq->nr_running++; |
30cfdcfc DA |
2999 | } |
3000 | ||
3001 | static void | |
3002 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3003 | { | |
3004 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 3005 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3006 | if (entity_is_task(se)) { |
3007 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 3008 | list_del_init(&se->group_node); |
0ec8aa00 | 3009 | } |
bfdb198c | 3010 | #endif |
30cfdcfc | 3011 | cfs_rq->nr_running--; |
30cfdcfc DA |
3012 | } |
3013 | ||
8d5b9025 PZ |
3014 | /* |
3015 | * Signed add and clamp on underflow. | |
3016 | * | |
3017 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3018 | * memory. This allows lockless observations without ever seeing the negative | |
3019 | * values. | |
3020 | */ | |
3021 | #define add_positive(_ptr, _val) do { \ | |
3022 | typeof(_ptr) ptr = (_ptr); \ | |
3023 | typeof(_val) val = (_val); \ | |
3024 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3025 | \ | |
3026 | res = var + val; \ | |
3027 | \ | |
3028 | if (val < 0 && res > var) \ | |
3029 | res = 0; \ | |
3030 | \ | |
3031 | WRITE_ONCE(*ptr, res); \ | |
3032 | } while (0) | |
3033 | ||
3034 | /* | |
3035 | * Unsigned subtract and clamp on underflow. | |
3036 | * | |
3037 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3038 | * memory. This allows lockless observations without ever seeing the negative | |
3039 | * values. | |
3040 | */ | |
3041 | #define sub_positive(_ptr, _val) do { \ | |
3042 | typeof(_ptr) ptr = (_ptr); \ | |
3043 | typeof(*ptr) val = (_val); \ | |
3044 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3045 | res = var - val; \ | |
3046 | if (res > var) \ | |
3047 | res = 0; \ | |
3048 | WRITE_ONCE(*ptr, res); \ | |
3049 | } while (0) | |
3050 | ||
b5c0ce7b PB |
3051 | /* |
3052 | * Remove and clamp on negative, from a local variable. | |
3053 | * | |
3054 | * A variant of sub_positive(), which does not use explicit load-store | |
3055 | * and is thus optimized for local variable updates. | |
3056 | */ | |
3057 | #define lsub_positive(_ptr, _val) do { \ | |
3058 | typeof(_ptr) ptr = (_ptr); \ | |
3059 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
3060 | } while (0) | |
3061 | ||
8d5b9025 | 3062 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
3063 | static inline void |
3064 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3065 | { | |
3066 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3067 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
3068 | } | |
3069 | ||
3070 | static inline void | |
3071 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3072 | { | |
3073 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
3074 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); | |
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); | |
3090 | account_entity_dequeue(cfs_rq, se); | |
9059393e VG |
3091 | } |
3092 | dequeue_load_avg(cfs_rq, se); | |
3093 | ||
3094 | update_load_set(&se->load, weight); | |
3095 | ||
3096 | #ifdef CONFIG_SMP | |
1ea6c46a PZ |
3097 | do { |
3098 | u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib; | |
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) |
9059393e | 3106 | account_entity_enqueue(cfs_rq, se); |
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) | |
3132 | * \Sum grq->load.weight | |
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) | |
3146 | * tg->load_avg | |
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) |
cef27403 PZ |
3162 | * grp->load.weight |
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) | |
3181 | * tg_load_avg' | |
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 |
7c3edd2c PZ |
3291 | /** |
3292 | * update_tg_load_avg - update the tg's load avg | |
3293 | * @cfs_rq: the cfs_rq whose avg changed | |
3294 | * @force: update regardless of how small the difference | |
3295 | * | |
3296 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3297 | * However, because tg->load_avg is a global value there are performance | |
3298 | * considerations. | |
3299 | * | |
3300 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3301 | * differential update where we store the last value we propagated. This in | |
3302 | * turn allows skipping updates if the differential is 'small'. | |
3303 | * | |
815abf5a | 3304 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3305 | */ |
9d89c257 | 3306 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 3307 | { |
9d89c257 | 3308 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3309 | |
aa0b7ae0 WL |
3310 | /* |
3311 | * No need to update load_avg for root_task_group as it is not used. | |
3312 | */ | |
3313 | if (cfs_rq->tg == &root_task_group) | |
3314 | return; | |
3315 | ||
9d89c257 YD |
3316 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
3317 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
3318 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3319 | } |
8165e145 | 3320 | } |
f5f9739d | 3321 | |
ad936d86 | 3322 | /* |
97fb7a0a | 3323 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3324 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3325 | * including the state of rq->lock, should be made. | |
3326 | */ | |
3327 | void set_task_rq_fair(struct sched_entity *se, | |
3328 | struct cfs_rq *prev, struct cfs_rq *next) | |
3329 | { | |
0ccb977f PZ |
3330 | u64 p_last_update_time; |
3331 | u64 n_last_update_time; | |
3332 | ||
ad936d86 BP |
3333 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3334 | return; | |
3335 | ||
3336 | /* | |
3337 | * We are supposed to update the task to "current" time, then its up to | |
3338 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3339 | * getting what current time is, so simply throw away the out-of-date | |
3340 | * time. This will result in the wakee task is less decayed, but giving | |
3341 | * the wakee more load sounds not bad. | |
3342 | */ | |
0ccb977f PZ |
3343 | if (!(se->avg.last_update_time && prev)) |
3344 | return; | |
ad936d86 BP |
3345 | |
3346 | #ifndef CONFIG_64BIT | |
0ccb977f | 3347 | { |
ad936d86 BP |
3348 | u64 p_last_update_time_copy; |
3349 | u64 n_last_update_time_copy; | |
3350 | ||
3351 | do { | |
3352 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3353 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3354 | ||
3355 | smp_rmb(); | |
3356 | ||
3357 | p_last_update_time = prev->avg.last_update_time; | |
3358 | n_last_update_time = next->avg.last_update_time; | |
3359 | ||
3360 | } while (p_last_update_time != p_last_update_time_copy || | |
3361 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3362 | } |
ad936d86 | 3363 | #else |
0ccb977f PZ |
3364 | p_last_update_time = prev->avg.last_update_time; |
3365 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3366 | #endif |
23127296 | 3367 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 3368 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 3369 | } |
09a43ace | 3370 | |
0e2d2aaa PZ |
3371 | |
3372 | /* | |
3373 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3374 | * propagate its contribution. The key to this propagation is the invariant | |
3375 | * that for each group: | |
3376 | * | |
3377 | * ge->avg == grq->avg (1) | |
3378 | * | |
3379 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3380 | * represent the very same entity, just at different points in the hierarchy. | |
3381 | * | |
9f683953 VG |
3382 | * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial |
3383 | * and simply copies the running/runnable sum over (but still wrong, because | |
3384 | * the group entity and group rq do not have their PELT windows aligned). | |
0e2d2aaa | 3385 | * |
0dacee1b | 3386 | * However, update_tg_cfs_load() is more complex. So we have: |
0e2d2aaa PZ |
3387 | * |
3388 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3389 | * | |
3390 | * And since, like util, the runnable part should be directly transferable, | |
3391 | * the following would _appear_ to be the straight forward approach: | |
3392 | * | |
a4c3c049 | 3393 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3394 | * |
3395 | * And per (1) we have: | |
3396 | * | |
a4c3c049 | 3397 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3398 | * |
3399 | * Which gives: | |
3400 | * | |
3401 | * ge->load.weight * grq->avg.load_avg | |
3402 | * ge->avg.load_avg = ----------------------------------- (4) | |
3403 | * grq->load.weight | |
3404 | * | |
3405 | * Except that is wrong! | |
3406 | * | |
3407 | * Because while for entities historical weight is not important and we | |
3408 | * really only care about our future and therefore can consider a pure | |
3409 | * runnable sum, runqueues can NOT do this. | |
3410 | * | |
3411 | * We specifically want runqueues to have a load_avg that includes | |
3412 | * historical weights. Those represent the blocked load, the load we expect | |
3413 | * to (shortly) return to us. This only works by keeping the weights as | |
3414 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3415 | * | |
a4c3c049 VG |
3416 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3417 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3418 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3419 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3420 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3421 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3422 | * |
a4c3c049 | 3423 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3424 | * |
a4c3c049 | 3425 | * Given the constraint: |
0e2d2aaa | 3426 | * |
a4c3c049 | 3427 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3428 | * |
a4c3c049 VG |
3429 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3430 | * overlap. | |
0e2d2aaa | 3431 | * |
a4c3c049 | 3432 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3433 | * |
a4c3c049 | 3434 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3435 | * |
a4c3c049 | 3436 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3437 | * |
0e2d2aaa PZ |
3438 | */ |
3439 | ||
09a43ace | 3440 | static inline void |
0e2d2aaa | 3441 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3442 | { |
09a43ace | 3443 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; |
95d68593 VG |
3444 | /* |
3445 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3446 | * See ___update_load_avg() for details. | |
3447 | */ | |
3448 | u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; | |
09a43ace VG |
3449 | |
3450 | /* Nothing to update */ | |
3451 | if (!delta) | |
3452 | return; | |
3453 | ||
3454 | /* Set new sched_entity's utilization */ | |
3455 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
95d68593 | 3456 | se->avg.util_sum = se->avg.util_avg * divider; |
09a43ace VG |
3457 | |
3458 | /* Update parent cfs_rq utilization */ | |
3459 | add_positive(&cfs_rq->avg.util_avg, delta); | |
95d68593 | 3460 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; |
09a43ace VG |
3461 | } |
3462 | ||
9f683953 VG |
3463 | static inline void |
3464 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) | |
3465 | { | |
3466 | long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; | |
95d68593 VG |
3467 | /* |
3468 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3469 | * See ___update_load_avg() for details. | |
3470 | */ | |
3471 | u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; | |
9f683953 VG |
3472 | |
3473 | /* Nothing to update */ | |
3474 | if (!delta) | |
3475 | return; | |
3476 | ||
9f683953 VG |
3477 | /* Set new sched_entity's runnable */ |
3478 | se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; | |
95d68593 | 3479 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
9f683953 VG |
3480 | |
3481 | /* Update parent cfs_rq runnable */ | |
3482 | add_positive(&cfs_rq->avg.runnable_avg, delta); | |
95d68593 | 3483 | cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; |
9f683953 VG |
3484 | } |
3485 | ||
09a43ace | 3486 | static inline void |
0dacee1b | 3487 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3488 | { |
a4c3c049 | 3489 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
0dacee1b VG |
3490 | unsigned long load_avg; |
3491 | u64 load_sum = 0; | |
a4c3c049 | 3492 | s64 delta_sum; |
95d68593 | 3493 | u32 divider; |
09a43ace | 3494 | |
0e2d2aaa PZ |
3495 | if (!runnable_sum) |
3496 | return; | |
09a43ace | 3497 | |
0e2d2aaa | 3498 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3499 | |
95d68593 VG |
3500 | /* |
3501 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3502 | * See ___update_load_avg() for details. | |
3503 | */ | |
3504 | divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; | |
3505 | ||
a4c3c049 VG |
3506 | if (runnable_sum >= 0) { |
3507 | /* | |
3508 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3509 | * the CPU is saturated running == runnable. | |
3510 | */ | |
3511 | runnable_sum += se->avg.load_sum; | |
95d68593 | 3512 | runnable_sum = min_t(long, runnable_sum, divider); |
a4c3c049 VG |
3513 | } else { |
3514 | /* | |
3515 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3516 | * assuming all tasks are equally runnable. | |
3517 | */ | |
3518 | if (scale_load_down(gcfs_rq->load.weight)) { | |
3519 | load_sum = div_s64(gcfs_rq->avg.load_sum, | |
3520 | scale_load_down(gcfs_rq->load.weight)); | |
3521 | } | |
3522 | ||
3523 | /* But make sure to not inflate se's runnable */ | |
3524 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3525 | } | |
3526 | ||
3527 | /* | |
3528 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
3529 | * Rescale running sum to be in the same range as runnable sum |
3530 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
3531 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 3532 | */ |
23127296 | 3533 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
3534 | runnable_sum = max(runnable_sum, running_sum); |
3535 | ||
0e2d2aaa | 3536 | load_sum = (s64)se_weight(se) * runnable_sum; |
95d68593 | 3537 | load_avg = div_s64(load_sum, divider); |
09a43ace | 3538 | |
a4c3c049 VG |
3539 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
3540 | delta_avg = load_avg - se->avg.load_avg; | |
09a43ace | 3541 | |
a4c3c049 VG |
3542 | se->avg.load_sum = runnable_sum; |
3543 | se->avg.load_avg = load_avg; | |
3544 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
3545 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
09a43ace VG |
3546 | } |
3547 | ||
0e2d2aaa | 3548 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3549 | { |
0e2d2aaa PZ |
3550 | cfs_rq->propagate = 1; |
3551 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3552 | } |
3553 | ||
3554 | /* Update task and its cfs_rq load average */ | |
3555 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3556 | { | |
0e2d2aaa | 3557 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3558 | |
3559 | if (entity_is_task(se)) | |
3560 | return 0; | |
3561 | ||
0e2d2aaa PZ |
3562 | gcfs_rq = group_cfs_rq(se); |
3563 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3564 | return 0; |
3565 | ||
0e2d2aaa PZ |
3566 | gcfs_rq->propagate = 0; |
3567 | ||
09a43ace VG |
3568 | cfs_rq = cfs_rq_of(se); |
3569 | ||
0e2d2aaa | 3570 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3571 | |
0e2d2aaa | 3572 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
9f683953 | 3573 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); |
0dacee1b | 3574 | update_tg_cfs_load(cfs_rq, se, gcfs_rq); |
09a43ace | 3575 | |
ba19f51f | 3576 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 3577 | trace_pelt_se_tp(se); |
ba19f51f | 3578 | |
09a43ace VG |
3579 | return 1; |
3580 | } | |
3581 | ||
bc427898 VG |
3582 | /* |
3583 | * Check if we need to update the load and the utilization of a blocked | |
3584 | * group_entity: | |
3585 | */ | |
3586 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3587 | { | |
3588 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3589 | ||
3590 | /* | |
3591 | * If sched_entity still have not zero load or utilization, we have to | |
3592 | * decay it: | |
3593 | */ | |
3594 | if (se->avg.load_avg || se->avg.util_avg) | |
3595 | return false; | |
3596 | ||
3597 | /* | |
3598 | * If there is a pending propagation, we have to update the load and | |
3599 | * the utilization of the sched_entity: | |
3600 | */ | |
0e2d2aaa | 3601 | if (gcfs_rq->propagate) |
bc427898 VG |
3602 | return false; |
3603 | ||
3604 | /* | |
3605 | * Otherwise, the load and the utilization of the sched_entity is | |
3606 | * already zero and there is no pending propagation, so it will be a | |
3607 | * waste of time to try to decay it: | |
3608 | */ | |
3609 | return true; | |
3610 | } | |
3611 | ||
6e83125c | 3612 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3613 | |
9d89c257 | 3614 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
09a43ace VG |
3615 | |
3616 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3617 | { | |
3618 | return 0; | |
3619 | } | |
3620 | ||
0e2d2aaa | 3621 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3622 | |
6e83125c | 3623 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3624 | |
3d30544f PZ |
3625 | /** |
3626 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 3627 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 3628 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
3629 | * |
3630 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3631 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3632 | * post_init_entity_util_avg(). | |
3633 | * | |
3634 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3635 | * | |
7c3edd2c PZ |
3636 | * Returns true if the load decayed or we removed load. |
3637 | * | |
3638 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3639 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3640 | */ |
a2c6c91f | 3641 | static inline int |
3a123bbb | 3642 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3643 | { |
9f683953 | 3644 | unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; |
9d89c257 | 3645 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3646 | int decayed = 0; |
2dac754e | 3647 | |
2a2f5d4e PZ |
3648 | if (cfs_rq->removed.nr) { |
3649 | unsigned long r; | |
9a2dd585 | 3650 | u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib; |
2a2f5d4e PZ |
3651 | |
3652 | raw_spin_lock(&cfs_rq->removed.lock); | |
3653 | swap(cfs_rq->removed.util_avg, removed_util); | |
3654 | swap(cfs_rq->removed.load_avg, removed_load); | |
9f683953 | 3655 | swap(cfs_rq->removed.runnable_avg, removed_runnable); |
2a2f5d4e PZ |
3656 | cfs_rq->removed.nr = 0; |
3657 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3658 | ||
2a2f5d4e | 3659 | r = removed_load; |
89741892 | 3660 | sub_positive(&sa->load_avg, r); |
9a2dd585 | 3661 | sub_positive(&sa->load_sum, r * divider); |
2dac754e | 3662 | |
2a2f5d4e | 3663 | r = removed_util; |
89741892 | 3664 | sub_positive(&sa->util_avg, r); |
9a2dd585 | 3665 | sub_positive(&sa->util_sum, r * divider); |
2a2f5d4e | 3666 | |
9f683953 VG |
3667 | r = removed_runnable; |
3668 | sub_positive(&sa->runnable_avg, r); | |
3669 | sub_positive(&sa->runnable_sum, r * divider); | |
3670 | ||
3671 | /* | |
3672 | * removed_runnable is the unweighted version of removed_load so we | |
3673 | * can use it to estimate removed_load_sum. | |
3674 | */ | |
3675 | add_tg_cfs_propagate(cfs_rq, | |
3676 | -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); | |
2a2f5d4e PZ |
3677 | |
3678 | decayed = 1; | |
9d89c257 | 3679 | } |
36ee28e4 | 3680 | |
23127296 | 3681 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
36ee28e4 | 3682 | |
9d89c257 YD |
3683 | #ifndef CONFIG_64BIT |
3684 | smp_wmb(); | |
3685 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3686 | #endif | |
36ee28e4 | 3687 | |
2a2f5d4e | 3688 | return decayed; |
21e96f88 SM |
3689 | } |
3690 | ||
3d30544f PZ |
3691 | /** |
3692 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3693 | * @cfs_rq: cfs_rq to attach to | |
3694 | * @se: sched_entity to attach | |
3695 | * | |
3696 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3697 | * cfs_rq->avg.last_update_time being current. | |
3698 | */ | |
a4f9a0e5 | 3699 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
a05e8c51 | 3700 | { |
95d68593 VG |
3701 | /* |
3702 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3703 | * See ___update_load_avg() for details. | |
3704 | */ | |
f207934f PZ |
3705 | u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; |
3706 | ||
3707 | /* | |
3708 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3709 | * window because without that, really weird and wonderful things can | |
3710 | * happen. | |
3711 | * | |
3712 | * XXX illustrate | |
3713 | */ | |
a05e8c51 | 3714 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3715 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3716 | ||
3717 | /* | |
3718 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3719 | * period_contrib. This isn't strictly correct, but since we're | |
3720 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3721 | * _sum a little. | |
3722 | */ | |
3723 | se->avg.util_sum = se->avg.util_avg * divider; | |
3724 | ||
9f683953 VG |
3725 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
3726 | ||
f207934f PZ |
3727 | se->avg.load_sum = divider; |
3728 | if (se_weight(se)) { | |
3729 | se->avg.load_sum = | |
3730 | div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); | |
3731 | } | |
3732 | ||
8d5b9025 | 3733 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3734 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3735 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
9f683953 VG |
3736 | cfs_rq->avg.runnable_avg += se->avg.runnable_avg; |
3737 | cfs_rq->avg.runnable_sum += se->avg.runnable_sum; | |
0e2d2aaa PZ |
3738 | |
3739 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 3740 | |
a4f9a0e5 | 3741 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3742 | |
3743 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3744 | } |
3745 | ||
3d30544f PZ |
3746 | /** |
3747 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3748 | * @cfs_rq: cfs_rq to detach from | |
3749 | * @se: sched_entity to detach | |
3750 | * | |
3751 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3752 | * cfs_rq->avg.last_update_time being current. | |
3753 | */ | |
a05e8c51 BP |
3754 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3755 | { | |
8d5b9025 | 3756 | dequeue_load_avg(cfs_rq, se); |
89741892 PZ |
3757 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
3758 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
9f683953 VG |
3759 | sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); |
3760 | sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); | |
0e2d2aaa PZ |
3761 | |
3762 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 3763 | |
ea14b57e | 3764 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3765 | |
3766 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3767 | } |
3768 | ||
b382a531 PZ |
3769 | /* |
3770 | * Optional action to be done while updating the load average | |
3771 | */ | |
3772 | #define UPDATE_TG 0x1 | |
3773 | #define SKIP_AGE_LOAD 0x2 | |
3774 | #define DO_ATTACH 0x4 | |
3775 | ||
3776 | /* Update task and its cfs_rq load average */ | |
3777 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
3778 | { | |
23127296 | 3779 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
3780 | int decayed; |
3781 | ||
3782 | /* | |
3783 | * Track task load average for carrying it to new CPU after migrated, and | |
3784 | * track group sched_entity load average for task_h_load calc in migration | |
3785 | */ | |
3786 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 3787 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
3788 | |
3789 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
3790 | decayed |= propagate_entity_load_avg(se); | |
3791 | ||
3792 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
3793 | ||
ea14b57e PZ |
3794 | /* |
3795 | * DO_ATTACH means we're here from enqueue_entity(). | |
3796 | * !last_update_time means we've passed through | |
3797 | * migrate_task_rq_fair() indicating we migrated. | |
3798 | * | |
3799 | * IOW we're enqueueing a task on a new CPU. | |
3800 | */ | |
a4f9a0e5 | 3801 | attach_entity_load_avg(cfs_rq, se); |
b382a531 PZ |
3802 | update_tg_load_avg(cfs_rq, 0); |
3803 | ||
bef69dd8 VG |
3804 | } else if (decayed) { |
3805 | cfs_rq_util_change(cfs_rq, 0); | |
3806 | ||
3807 | if (flags & UPDATE_TG) | |
3808 | update_tg_load_avg(cfs_rq, 0); | |
3809 | } | |
b382a531 PZ |
3810 | } |
3811 | ||
9d89c257 | 3812 | #ifndef CONFIG_64BIT |
0905f04e YD |
3813 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3814 | { | |
9d89c257 | 3815 | u64 last_update_time_copy; |
0905f04e | 3816 | u64 last_update_time; |
9ee474f5 | 3817 | |
9d89c257 YD |
3818 | do { |
3819 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3820 | smp_rmb(); | |
3821 | last_update_time = cfs_rq->avg.last_update_time; | |
3822 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3823 | |
3824 | return last_update_time; | |
3825 | } | |
9d89c257 | 3826 | #else |
0905f04e YD |
3827 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3828 | { | |
3829 | return cfs_rq->avg.last_update_time; | |
3830 | } | |
9d89c257 YD |
3831 | #endif |
3832 | ||
104cb16d MR |
3833 | /* |
3834 | * Synchronize entity load avg of dequeued entity without locking | |
3835 | * the previous rq. | |
3836 | */ | |
71b47eaf | 3837 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
3838 | { |
3839 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3840 | u64 last_update_time; | |
3841 | ||
3842 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 3843 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
3844 | } |
3845 | ||
0905f04e YD |
3846 | /* |
3847 | * Task first catches up with cfs_rq, and then subtract | |
3848 | * itself from the cfs_rq (task must be off the queue now). | |
3849 | */ | |
71b47eaf | 3850 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
3851 | { |
3852 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 3853 | unsigned long flags; |
0905f04e YD |
3854 | |
3855 | /* | |
7dc603c9 PZ |
3856 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3857 | * post_init_entity_util_avg() which will have added things to the | |
3858 | * cfs_rq, so we can remove unconditionally. | |
0905f04e | 3859 | */ |
0905f04e | 3860 | |
104cb16d | 3861 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
3862 | |
3863 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
3864 | ++cfs_rq->removed.nr; | |
3865 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
3866 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
9f683953 | 3867 | cfs_rq->removed.runnable_avg += se->avg.runnable_avg; |
2a2f5d4e | 3868 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 3869 | } |
642dbc39 | 3870 | |
9f683953 VG |
3871 | static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) |
3872 | { | |
3873 | return cfs_rq->avg.runnable_avg; | |
3874 | } | |
3875 | ||
7ea241af YD |
3876 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) |
3877 | { | |
3878 | return cfs_rq->avg.load_avg; | |
3879 | } | |
3880 | ||
d91cecc1 CY |
3881 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); |
3882 | ||
7f65ea42 PB |
3883 | static inline unsigned long task_util(struct task_struct *p) |
3884 | { | |
3885 | return READ_ONCE(p->se.avg.util_avg); | |
3886 | } | |
3887 | ||
3888 | static inline unsigned long _task_util_est(struct task_struct *p) | |
3889 | { | |
3890 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
3891 | ||
92a801e5 | 3892 | return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED); |
7f65ea42 PB |
3893 | } |
3894 | ||
3895 | static inline unsigned long task_util_est(struct task_struct *p) | |
3896 | { | |
3897 | return max(task_util(p), _task_util_est(p)); | |
3898 | } | |
3899 | ||
a7008c07 VS |
3900 | #ifdef CONFIG_UCLAMP_TASK |
3901 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
3902 | { | |
3903 | return clamp(task_util_est(p), | |
3904 | uclamp_eff_value(p, UCLAMP_MIN), | |
3905 | uclamp_eff_value(p, UCLAMP_MAX)); | |
3906 | } | |
3907 | #else | |
3908 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
3909 | { | |
3910 | return task_util_est(p); | |
3911 | } | |
3912 | #endif | |
3913 | ||
7f65ea42 PB |
3914 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, |
3915 | struct task_struct *p) | |
3916 | { | |
3917 | unsigned int enqueued; | |
3918 | ||
3919 | if (!sched_feat(UTIL_EST)) | |
3920 | return; | |
3921 | ||
3922 | /* Update root cfs_rq's estimated utilization */ | |
3923 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3924 | enqueued += _task_util_est(p); |
7f65ea42 PB |
3925 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
3926 | } | |
3927 | ||
3928 | /* | |
3929 | * Check if a (signed) value is within a specified (unsigned) margin, | |
3930 | * based on the observation that: | |
3931 | * | |
3932 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
3933 | * | |
3934 | * NOTE: this only works when value + maring < INT_MAX. | |
3935 | */ | |
3936 | static inline bool within_margin(int value, int margin) | |
3937 | { | |
3938 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
3939 | } | |
3940 | ||
3941 | static void | |
3942 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep) | |
3943 | { | |
3944 | long last_ewma_diff; | |
3945 | struct util_est ue; | |
10a35e68 | 3946 | int cpu; |
7f65ea42 PB |
3947 | |
3948 | if (!sched_feat(UTIL_EST)) | |
3949 | return; | |
3950 | ||
3482d98b VG |
3951 | /* Update root cfs_rq's estimated utilization */ |
3952 | ue.enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3953 | ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p)); |
7f65ea42 PB |
3954 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued); |
3955 | ||
3956 | /* | |
3957 | * Skip update of task's estimated utilization when the task has not | |
3958 | * yet completed an activation, e.g. being migrated. | |
3959 | */ | |
3960 | if (!task_sleep) | |
3961 | return; | |
3962 | ||
d519329f PB |
3963 | /* |
3964 | * If the PELT values haven't changed since enqueue time, | |
3965 | * skip the util_est update. | |
3966 | */ | |
3967 | ue = p->se.avg.util_est; | |
3968 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
3969 | return; | |
3970 | ||
b8c96361 PB |
3971 | /* |
3972 | * Reset EWMA on utilization increases, the moving average is used only | |
3973 | * to smooth utilization decreases. | |
3974 | */ | |
3975 | ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED); | |
3976 | if (sched_feat(UTIL_EST_FASTUP)) { | |
3977 | if (ue.ewma < ue.enqueued) { | |
3978 | ue.ewma = ue.enqueued; | |
3979 | goto done; | |
3980 | } | |
3981 | } | |
3982 | ||
7f65ea42 PB |
3983 | /* |
3984 | * Skip update of task's estimated utilization when its EWMA is | |
3985 | * already ~1% close to its last activation value. | |
3986 | */ | |
7f65ea42 PB |
3987 | last_ewma_diff = ue.enqueued - ue.ewma; |
3988 | if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100))) | |
3989 | return; | |
3990 | ||
10a35e68 VG |
3991 | /* |
3992 | * To avoid overestimation of actual task utilization, skip updates if | |
3993 | * we cannot grant there is idle time in this CPU. | |
3994 | */ | |
3995 | cpu = cpu_of(rq_of(cfs_rq)); | |
3996 | if (task_util(p) > capacity_orig_of(cpu)) | |
3997 | return; | |
3998 | ||
7f65ea42 PB |
3999 | /* |
4000 | * Update Task's estimated utilization | |
4001 | * | |
4002 | * When *p completes an activation we can consolidate another sample | |
4003 | * of the task size. This is done by storing the current PELT value | |
4004 | * as ue.enqueued and by using this value to update the Exponential | |
4005 | * Weighted Moving Average (EWMA): | |
4006 | * | |
4007 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
4008 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
4009 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
4010 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
4011 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
4012 | * | |
4013 | * Where 'w' is the weight of new samples, which is configured to be | |
4014 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
4015 | */ | |
4016 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
4017 | ue.ewma += last_ewma_diff; | |
4018 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
b8c96361 | 4019 | done: |
7f65ea42 PB |
4020 | WRITE_ONCE(p->se.avg.util_est, ue); |
4021 | } | |
4022 | ||
3b1baa64 MR |
4023 | static inline int task_fits_capacity(struct task_struct *p, long capacity) |
4024 | { | |
a7008c07 | 4025 | return fits_capacity(uclamp_task_util(p), capacity); |
3b1baa64 MR |
4026 | } |
4027 | ||
4028 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
4029 | { | |
4030 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) | |
4031 | return; | |
4032 | ||
4033 | if (!p) { | |
4034 | rq->misfit_task_load = 0; | |
4035 | return; | |
4036 | } | |
4037 | ||
4038 | if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { | |
4039 | rq->misfit_task_load = 0; | |
4040 | return; | |
4041 | } | |
4042 | ||
4043 | rq->misfit_task_load = task_h_load(p); | |
4044 | } | |
4045 | ||
38033c37 PZ |
4046 | #else /* CONFIG_SMP */ |
4047 | ||
d31b1a66 VG |
4048 | #define UPDATE_TG 0x0 |
4049 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 4050 | #define DO_ATTACH 0x0 |
d31b1a66 | 4051 | |
88c0616e | 4052 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 4053 | { |
ea14b57e | 4054 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
4055 | } |
4056 | ||
9d89c257 | 4057 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 4058 | |
a05e8c51 | 4059 | static inline void |
a4f9a0e5 | 4060 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} |
a05e8c51 BP |
4061 | static inline void |
4062 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
4063 | ||
d91cecc1 | 4064 | static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
4065 | { |
4066 | return 0; | |
4067 | } | |
4068 | ||
7f65ea42 PB |
4069 | static inline void |
4070 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
4071 | ||
4072 | static inline void | |
4073 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, | |
4074 | bool task_sleep) {} | |
3b1baa64 | 4075 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 4076 | |
38033c37 | 4077 | #endif /* CONFIG_SMP */ |
9d85f21c | 4078 | |
ddc97297 PZ |
4079 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4080 | { | |
4081 | #ifdef CONFIG_SCHED_DEBUG | |
4082 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
4083 | ||
4084 | if (d < 0) | |
4085 | d = -d; | |
4086 | ||
4087 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 4088 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
4089 | #endif |
4090 | } | |
4091 | ||
aeb73b04 PZ |
4092 | static void |
4093 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
4094 | { | |
1af5f730 | 4095 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 4096 | |
2cb8600e PZ |
4097 | /* |
4098 | * The 'current' period is already promised to the current tasks, | |
4099 | * however the extra weight of the new task will slow them down a | |
4100 | * little, place the new task so that it fits in the slot that | |
4101 | * stays open at the end. | |
4102 | */ | |
94dfb5e7 | 4103 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 4104 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 4105 | |
a2e7a7eb | 4106 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 4107 | if (!initial) { |
a2e7a7eb | 4108 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 4109 | |
a2e7a7eb MG |
4110 | /* |
4111 | * Halve their sleep time's effect, to allow | |
4112 | * for a gentler effect of sleepers: | |
4113 | */ | |
4114 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
4115 | thresh >>= 1; | |
51e0304c | 4116 | |
a2e7a7eb | 4117 | vruntime -= thresh; |
aeb73b04 PZ |
4118 | } |
4119 | ||
b5d9d734 | 4120 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 4121 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
4122 | } |
4123 | ||
d3d9dc33 PT |
4124 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
4125 | ||
cb251765 MG |
4126 | static inline void check_schedstat_required(void) |
4127 | { | |
4128 | #ifdef CONFIG_SCHEDSTATS | |
4129 | if (schedstat_enabled()) | |
4130 | return; | |
4131 | ||
4132 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
4133 | if (trace_sched_stat_wait_enabled() || | |
4134 | trace_sched_stat_sleep_enabled() || | |
4135 | trace_sched_stat_iowait_enabled() || | |
4136 | trace_sched_stat_blocked_enabled() || | |
4137 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 4138 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 | 4139 | "stat_blocked and stat_runtime require the " |
f67abed5 | 4140 | "kernel parameter schedstats=enable or " |
cb251765 MG |
4141 | "kernel.sched_schedstats=1\n"); |
4142 | } | |
4143 | #endif | |
4144 | } | |
4145 | ||
fe61468b | 4146 | static inline bool cfs_bandwidth_used(void); |
b5179ac7 PZ |
4147 | |
4148 | /* | |
4149 | * MIGRATION | |
4150 | * | |
4151 | * dequeue | |
4152 | * update_curr() | |
4153 | * update_min_vruntime() | |
4154 | * vruntime -= min_vruntime | |
4155 | * | |
4156 | * enqueue | |
4157 | * update_curr() | |
4158 | * update_min_vruntime() | |
4159 | * vruntime += min_vruntime | |
4160 | * | |
4161 | * this way the vruntime transition between RQs is done when both | |
4162 | * min_vruntime are up-to-date. | |
4163 | * | |
4164 | * WAKEUP (remote) | |
4165 | * | |
59efa0ba | 4166 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
4167 | * vruntime -= min_vruntime |
4168 | * | |
4169 | * enqueue | |
4170 | * update_curr() | |
4171 | * update_min_vruntime() | |
4172 | * vruntime += min_vruntime | |
4173 | * | |
4174 | * this way we don't have the most up-to-date min_vruntime on the originating | |
4175 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
4176 | */ | |
4177 | ||
bf0f6f24 | 4178 | static void |
88ec22d3 | 4179 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4180 | { |
2f950354 PZ |
4181 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
4182 | bool curr = cfs_rq->curr == se; | |
4183 | ||
88ec22d3 | 4184 | /* |
2f950354 PZ |
4185 | * If we're the current task, we must renormalise before calling |
4186 | * update_curr(). | |
88ec22d3 | 4187 | */ |
2f950354 | 4188 | if (renorm && curr) |
88ec22d3 PZ |
4189 | se->vruntime += cfs_rq->min_vruntime; |
4190 | ||
2f950354 PZ |
4191 | update_curr(cfs_rq); |
4192 | ||
bf0f6f24 | 4193 | /* |
2f950354 PZ |
4194 | * Otherwise, renormalise after, such that we're placed at the current |
4195 | * moment in time, instead of some random moment in the past. Being | |
4196 | * placed in the past could significantly boost this task to the | |
4197 | * fairness detriment of existing tasks. | |
bf0f6f24 | 4198 | */ |
2f950354 PZ |
4199 | if (renorm && !curr) |
4200 | se->vruntime += cfs_rq->min_vruntime; | |
4201 | ||
89ee048f VG |
4202 | /* |
4203 | * When enqueuing a sched_entity, we must: | |
4204 | * - Update loads to have both entity and cfs_rq synced with now. | |
9f683953 | 4205 | * - Add its load to cfs_rq->runnable_avg |
89ee048f VG |
4206 | * - For group_entity, update its weight to reflect the new share of |
4207 | * its group cfs_rq | |
4208 | * - Add its new weight to cfs_rq->load.weight | |
4209 | */ | |
b382a531 | 4210 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
9f683953 | 4211 | se_update_runnable(se); |
1ea6c46a | 4212 | update_cfs_group(se); |
17bc14b7 | 4213 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 4214 | |
1a3d027c | 4215 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 4216 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 4217 | |
cb251765 | 4218 | check_schedstat_required(); |
4fa8d299 JP |
4219 | update_stats_enqueue(cfs_rq, se, flags); |
4220 | check_spread(cfs_rq, se); | |
2f950354 | 4221 | if (!curr) |
83b699ed | 4222 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4223 | se->on_rq = 1; |
3d4b47b4 | 4224 | |
fe61468b VG |
4225 | /* |
4226 | * When bandwidth control is enabled, cfs might have been removed | |
4227 | * because of a parent been throttled but cfs->nr_running > 1. Try to | |
4228 | * add it unconditionnally. | |
4229 | */ | |
4230 | if (cfs_rq->nr_running == 1 || cfs_bandwidth_used()) | |
3d4b47b4 | 4231 | list_add_leaf_cfs_rq(cfs_rq); |
fe61468b VG |
4232 | |
4233 | if (cfs_rq->nr_running == 1) | |
d3d9dc33 | 4234 | check_enqueue_throttle(cfs_rq); |
bf0f6f24 IM |
4235 | } |
4236 | ||
2c13c919 | 4237 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4238 | { |
2c13c919 RR |
4239 | for_each_sched_entity(se) { |
4240 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4241 | if (cfs_rq->last != se) |
2c13c919 | 4242 | break; |
f1044799 PZ |
4243 | |
4244 | cfs_rq->last = NULL; | |
2c13c919 RR |
4245 | } |
4246 | } | |
2002c695 | 4247 | |
2c13c919 RR |
4248 | static void __clear_buddies_next(struct sched_entity *se) |
4249 | { | |
4250 | for_each_sched_entity(se) { | |
4251 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4252 | if (cfs_rq->next != se) |
2c13c919 | 4253 | break; |
f1044799 PZ |
4254 | |
4255 | cfs_rq->next = NULL; | |
2c13c919 | 4256 | } |
2002c695 PZ |
4257 | } |
4258 | ||
ac53db59 RR |
4259 | static void __clear_buddies_skip(struct sched_entity *se) |
4260 | { | |
4261 | for_each_sched_entity(se) { | |
4262 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4263 | if (cfs_rq->skip != se) |
ac53db59 | 4264 | break; |
f1044799 PZ |
4265 | |
4266 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4267 | } |
4268 | } | |
4269 | ||
a571bbea PZ |
4270 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4271 | { | |
2c13c919 RR |
4272 | if (cfs_rq->last == se) |
4273 | __clear_buddies_last(se); | |
4274 | ||
4275 | if (cfs_rq->next == se) | |
4276 | __clear_buddies_next(se); | |
ac53db59 RR |
4277 | |
4278 | if (cfs_rq->skip == se) | |
4279 | __clear_buddies_skip(se); | |
a571bbea PZ |
4280 | } |
4281 | ||
6c16a6dc | 4282 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4283 | |
bf0f6f24 | 4284 | static void |
371fd7e7 | 4285 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4286 | { |
a2a2d680 DA |
4287 | /* |
4288 | * Update run-time statistics of the 'current'. | |
4289 | */ | |
4290 | update_curr(cfs_rq); | |
89ee048f VG |
4291 | |
4292 | /* | |
4293 | * When dequeuing a sched_entity, we must: | |
4294 | * - Update loads to have both entity and cfs_rq synced with now. | |
9f683953 | 4295 | * - Subtract its load from the cfs_rq->runnable_avg. |
dfcb245e | 4296 | * - Subtract its previous weight from cfs_rq->load.weight. |
89ee048f VG |
4297 | * - For group entity, update its weight to reflect the new share |
4298 | * of its group cfs_rq. | |
4299 | */ | |
88c0616e | 4300 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 4301 | se_update_runnable(se); |
a2a2d680 | 4302 | |
4fa8d299 | 4303 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 4304 | |
2002c695 | 4305 | clear_buddies(cfs_rq, se); |
4793241b | 4306 | |
83b699ed | 4307 | if (se != cfs_rq->curr) |
30cfdcfc | 4308 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4309 | se->on_rq = 0; |
30cfdcfc | 4310 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4311 | |
4312 | /* | |
b60205c7 PZ |
4313 | * Normalize after update_curr(); which will also have moved |
4314 | * min_vruntime if @se is the one holding it back. But before doing | |
4315 | * update_min_vruntime() again, which will discount @se's position and | |
4316 | * can move min_vruntime forward still more. | |
88ec22d3 | 4317 | */ |
371fd7e7 | 4318 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4319 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4320 | |
d8b4986d PT |
4321 | /* return excess runtime on last dequeue */ |
4322 | return_cfs_rq_runtime(cfs_rq); | |
4323 | ||
1ea6c46a | 4324 | update_cfs_group(se); |
b60205c7 PZ |
4325 | |
4326 | /* | |
4327 | * Now advance min_vruntime if @se was the entity holding it back, | |
4328 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4329 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4330 | * further than we started -- ie. we'll be penalized. | |
4331 | */ | |
9845c49c | 4332 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4333 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
4334 | } |
4335 | ||
4336 | /* | |
4337 | * Preempt the current task with a newly woken task if needed: | |
4338 | */ | |
7c92e54f | 4339 | static void |
2e09bf55 | 4340 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4341 | { |
11697830 | 4342 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4343 | struct sched_entity *se; |
4344 | s64 delta; | |
11697830 | 4345 | |
6d0f0ebd | 4346 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4347 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4348 | if (delta_exec > ideal_runtime) { |
8875125e | 4349 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4350 | /* |
4351 | * The current task ran long enough, ensure it doesn't get | |
4352 | * re-elected due to buddy favours. | |
4353 | */ | |
4354 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4355 | return; |
4356 | } | |
4357 | ||
4358 | /* | |
4359 | * Ensure that a task that missed wakeup preemption by a | |
4360 | * narrow margin doesn't have to wait for a full slice. | |
4361 | * This also mitigates buddy induced latencies under load. | |
4362 | */ | |
f685ceac MG |
4363 | if (delta_exec < sysctl_sched_min_granularity) |
4364 | return; | |
4365 | ||
f4cfb33e WX |
4366 | se = __pick_first_entity(cfs_rq); |
4367 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4368 | |
f4cfb33e WX |
4369 | if (delta < 0) |
4370 | return; | |
d7d82944 | 4371 | |
f4cfb33e | 4372 | if (delta > ideal_runtime) |
8875125e | 4373 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4374 | } |
4375 | ||
83b699ed | 4376 | static void |
8494f412 | 4377 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4378 | { |
83b699ed SV |
4379 | /* 'current' is not kept within the tree. */ |
4380 | if (se->on_rq) { | |
4381 | /* | |
4382 | * Any task has to be enqueued before it get to execute on | |
4383 | * a CPU. So account for the time it spent waiting on the | |
4384 | * runqueue. | |
4385 | */ | |
4fa8d299 | 4386 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 4387 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4388 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4389 | } |
4390 | ||
79303e9e | 4391 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4392 | cfs_rq->curr = se; |
4fa8d299 | 4393 | |
eba1ed4b IM |
4394 | /* |
4395 | * Track our maximum slice length, if the CPU's load is at | |
4396 | * least twice that of our own weight (i.e. dont track it | |
4397 | * when there are only lesser-weight tasks around): | |
4398 | */ | |
f2bedc47 DE |
4399 | if (schedstat_enabled() && |
4400 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
4fa8d299 JP |
4401 | schedstat_set(se->statistics.slice_max, |
4402 | max((u64)schedstat_val(se->statistics.slice_max), | |
4403 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 4404 | } |
4fa8d299 | 4405 | |
4a55b450 | 4406 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4407 | } |
4408 | ||
3f3a4904 PZ |
4409 | static int |
4410 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4411 | ||
ac53db59 RR |
4412 | /* |
4413 | * Pick the next process, keeping these things in mind, in this order: | |
4414 | * 1) keep things fair between processes/task groups | |
4415 | * 2) pick the "next" process, since someone really wants that to run | |
4416 | * 3) pick the "last" process, for cache locality | |
4417 | * 4) do not run the "skip" process, if something else is available | |
4418 | */ | |
678d5718 PZ |
4419 | static struct sched_entity * |
4420 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4421 | { |
678d5718 PZ |
4422 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4423 | struct sched_entity *se; | |
4424 | ||
4425 | /* | |
4426 | * If curr is set we have to see if its left of the leftmost entity | |
4427 | * still in the tree, provided there was anything in the tree at all. | |
4428 | */ | |
4429 | if (!left || (curr && entity_before(curr, left))) | |
4430 | left = curr; | |
4431 | ||
4432 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4433 | |
ac53db59 RR |
4434 | /* |
4435 | * Avoid running the skip buddy, if running something else can | |
4436 | * be done without getting too unfair. | |
4437 | */ | |
4438 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
4439 | struct sched_entity *second; |
4440 | ||
4441 | if (se == curr) { | |
4442 | second = __pick_first_entity(cfs_rq); | |
4443 | } else { | |
4444 | second = __pick_next_entity(se); | |
4445 | if (!second || (curr && entity_before(curr, second))) | |
4446 | second = curr; | |
4447 | } | |
4448 | ||
ac53db59 RR |
4449 | if (second && wakeup_preempt_entity(second, left) < 1) |
4450 | se = second; | |
4451 | } | |
aa2ac252 | 4452 | |
f685ceac MG |
4453 | /* |
4454 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4455 | */ | |
4456 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
4457 | se = cfs_rq->last; | |
4458 | ||
ac53db59 RR |
4459 | /* |
4460 | * Someone really wants this to run. If it's not unfair, run it. | |
4461 | */ | |
4462 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
4463 | se = cfs_rq->next; | |
4464 | ||
f685ceac | 4465 | clear_buddies(cfs_rq, se); |
4793241b PZ |
4466 | |
4467 | return se; | |
aa2ac252 PZ |
4468 | } |
4469 | ||
678d5718 | 4470 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4471 | |
ab6cde26 | 4472 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4473 | { |
4474 | /* | |
4475 | * If still on the runqueue then deactivate_task() | |
4476 | * was not called and update_curr() has to be done: | |
4477 | */ | |
4478 | if (prev->on_rq) | |
b7cc0896 | 4479 | update_curr(cfs_rq); |
bf0f6f24 | 4480 | |
d3d9dc33 PT |
4481 | /* throttle cfs_rqs exceeding runtime */ |
4482 | check_cfs_rq_runtime(cfs_rq); | |
4483 | ||
4fa8d299 | 4484 | check_spread(cfs_rq, prev); |
cb251765 | 4485 | |
30cfdcfc | 4486 | if (prev->on_rq) { |
4fa8d299 | 4487 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
4488 | /* Put 'current' back into the tree. */ |
4489 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4490 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4491 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4492 | } |
429d43bc | 4493 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4494 | } |
4495 | ||
8f4d37ec PZ |
4496 | static void |
4497 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4498 | { |
bf0f6f24 | 4499 | /* |
30cfdcfc | 4500 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4501 | */ |
30cfdcfc | 4502 | update_curr(cfs_rq); |
bf0f6f24 | 4503 | |
9d85f21c PT |
4504 | /* |
4505 | * Ensure that runnable average is periodically updated. | |
4506 | */ | |
88c0616e | 4507 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4508 | update_cfs_group(curr); |
9d85f21c | 4509 | |
8f4d37ec PZ |
4510 | #ifdef CONFIG_SCHED_HRTICK |
4511 | /* | |
4512 | * queued ticks are scheduled to match the slice, so don't bother | |
4513 | * validating it and just reschedule. | |
4514 | */ | |
983ed7a6 | 4515 | if (queued) { |
8875125e | 4516 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4517 | return; |
4518 | } | |
8f4d37ec PZ |
4519 | /* |
4520 | * don't let the period tick interfere with the hrtick preemption | |
4521 | */ | |
4522 | if (!sched_feat(DOUBLE_TICK) && | |
4523 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4524 | return; | |
4525 | #endif | |
4526 | ||
2c2efaed | 4527 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4528 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4529 | } |
4530 | ||
ab84d31e PT |
4531 | |
4532 | /************************************************** | |
4533 | * CFS bandwidth control machinery | |
4534 | */ | |
4535 | ||
4536 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 4537 | |
e9666d10 | 4538 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 4539 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4540 | |
4541 | static inline bool cfs_bandwidth_used(void) | |
4542 | { | |
c5905afb | 4543 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4544 | } |
4545 | ||
1ee14e6c | 4546 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4547 | { |
ce48c146 | 4548 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
4549 | } |
4550 | ||
4551 | void cfs_bandwidth_usage_dec(void) | |
4552 | { | |
ce48c146 | 4553 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 4554 | } |
e9666d10 | 4555 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
4556 | static bool cfs_bandwidth_used(void) |
4557 | { | |
4558 | return true; | |
4559 | } | |
4560 | ||
1ee14e6c BS |
4561 | void cfs_bandwidth_usage_inc(void) {} |
4562 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 4563 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 4564 | |
ab84d31e PT |
4565 | /* |
4566 | * default period for cfs group bandwidth. | |
4567 | * default: 0.1s, units: nanoseconds | |
4568 | */ | |
4569 | static inline u64 default_cfs_period(void) | |
4570 | { | |
4571 | return 100000000ULL; | |
4572 | } | |
ec12cb7f PT |
4573 | |
4574 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4575 | { | |
4576 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4577 | } | |
4578 | ||
a9cf55b2 | 4579 | /* |
763a9ec0 QC |
4580 | * Replenish runtime according to assigned quota. We use sched_clock_cpu |
4581 | * directly instead of rq->clock to avoid adding additional synchronization | |
4582 | * around rq->lock. | |
a9cf55b2 PT |
4583 | * |
4584 | * requires cfs_b->lock | |
4585 | */ | |
029632fb | 4586 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 | 4587 | { |
763a9ec0 QC |
4588 | if (cfs_b->quota != RUNTIME_INF) |
4589 | cfs_b->runtime = cfs_b->quota; | |
a9cf55b2 PT |
4590 | } |
4591 | ||
029632fb PZ |
4592 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4593 | { | |
4594 | return &tg->cfs_bandwidth; | |
4595 | } | |
4596 | ||
85dac906 | 4597 | /* returns 0 on failure to allocate runtime */ |
e98fa02c PT |
4598 | static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, |
4599 | struct cfs_rq *cfs_rq, u64 target_runtime) | |
ec12cb7f | 4600 | { |
e98fa02c PT |
4601 | u64 min_amount, amount = 0; |
4602 | ||
4603 | lockdep_assert_held(&cfs_b->lock); | |
ec12cb7f PT |
4604 | |
4605 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
e98fa02c | 4606 | min_amount = target_runtime - cfs_rq->runtime_remaining; |
ec12cb7f | 4607 | |
ec12cb7f PT |
4608 | if (cfs_b->quota == RUNTIME_INF) |
4609 | amount = min_amount; | |
58088ad0 | 4610 | else { |
77a4d1a1 | 4611 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4612 | |
4613 | if (cfs_b->runtime > 0) { | |
4614 | amount = min(cfs_b->runtime, min_amount); | |
4615 | cfs_b->runtime -= amount; | |
4616 | cfs_b->idle = 0; | |
4617 | } | |
ec12cb7f | 4618 | } |
ec12cb7f PT |
4619 | |
4620 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
4621 | |
4622 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4623 | } |
4624 | ||
e98fa02c PT |
4625 | /* returns 0 on failure to allocate runtime */ |
4626 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4627 | { | |
4628 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4629 | int ret; | |
4630 | ||
4631 | raw_spin_lock(&cfs_b->lock); | |
4632 | ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); | |
4633 | raw_spin_unlock(&cfs_b->lock); | |
4634 | ||
4635 | return ret; | |
4636 | } | |
4637 | ||
9dbdb155 | 4638 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4639 | { |
4640 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4641 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4642 | |
4643 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4644 | return; |
4645 | ||
5e2d2cc2 L |
4646 | if (cfs_rq->throttled) |
4647 | return; | |
85dac906 PT |
4648 | /* |
4649 | * if we're unable to extend our runtime we resched so that the active | |
4650 | * hierarchy can be throttled | |
4651 | */ | |
4652 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4653 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4654 | } |
4655 | ||
6c16a6dc | 4656 | static __always_inline |
9dbdb155 | 4657 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4658 | { |
56f570e5 | 4659 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4660 | return; |
4661 | ||
4662 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4663 | } | |
4664 | ||
85dac906 PT |
4665 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4666 | { | |
56f570e5 | 4667 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4668 | } |
4669 | ||
64660c86 PT |
4670 | /* check whether cfs_rq, or any parent, is throttled */ |
4671 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4672 | { | |
56f570e5 | 4673 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4674 | } |
4675 | ||
4676 | /* | |
4677 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4678 | * dest_cpu are members of a throttled hierarchy when performing group | |
4679 | * load-balance operations. | |
4680 | */ | |
4681 | static inline int throttled_lb_pair(struct task_group *tg, | |
4682 | int src_cpu, int dest_cpu) | |
4683 | { | |
4684 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4685 | ||
4686 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4687 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4688 | ||
4689 | return throttled_hierarchy(src_cfs_rq) || | |
4690 | throttled_hierarchy(dest_cfs_rq); | |
4691 | } | |
4692 | ||
64660c86 PT |
4693 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
4694 | { | |
4695 | struct rq *rq = data; | |
4696 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4697 | ||
4698 | cfs_rq->throttle_count--; | |
64660c86 | 4699 | if (!cfs_rq->throttle_count) { |
78becc27 | 4700 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4701 | cfs_rq->throttled_clock_task; |
31bc6aea VG |
4702 | |
4703 | /* Add cfs_rq with already running entity in the list */ | |
4704 | if (cfs_rq->nr_running >= 1) | |
4705 | list_add_leaf_cfs_rq(cfs_rq); | |
64660c86 | 4706 | } |
64660c86 PT |
4707 | |
4708 | return 0; | |
4709 | } | |
4710 | ||
4711 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4712 | { | |
4713 | struct rq *rq = data; | |
4714 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4715 | ||
82958366 | 4716 | /* group is entering throttled state, stop time */ |
31bc6aea | 4717 | if (!cfs_rq->throttle_count) { |
78becc27 | 4718 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
31bc6aea VG |
4719 | list_del_leaf_cfs_rq(cfs_rq); |
4720 | } | |
64660c86 PT |
4721 | cfs_rq->throttle_count++; |
4722 | ||
4723 | return 0; | |
4724 | } | |
4725 | ||
e98fa02c | 4726 | static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4727 | { |
4728 | struct rq *rq = rq_of(cfs_rq); | |
4729 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4730 | struct sched_entity *se; | |
43e9f7f2 | 4731 | long task_delta, idle_task_delta, dequeue = 1; |
e98fa02c PT |
4732 | |
4733 | raw_spin_lock(&cfs_b->lock); | |
4734 | /* This will start the period timer if necessary */ | |
4735 | if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { | |
4736 | /* | |
4737 | * We have raced with bandwidth becoming available, and if we | |
4738 | * actually throttled the timer might not unthrottle us for an | |
4739 | * entire period. We additionally needed to make sure that any | |
4740 | * subsequent check_cfs_rq_runtime calls agree not to throttle | |
4741 | * us, as we may commit to do cfs put_prev+pick_next, so we ask | |
4742 | * for 1ns of runtime rather than just check cfs_b. | |
4743 | */ | |
4744 | dequeue = 0; | |
4745 | } else { | |
4746 | list_add_tail_rcu(&cfs_rq->throttled_list, | |
4747 | &cfs_b->throttled_cfs_rq); | |
4748 | } | |
4749 | raw_spin_unlock(&cfs_b->lock); | |
4750 | ||
4751 | if (!dequeue) | |
4752 | return false; /* Throttle no longer required. */ | |
85dac906 PT |
4753 | |
4754 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4755 | ||
f1b17280 | 4756 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4757 | rcu_read_lock(); |
4758 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4759 | rcu_read_unlock(); | |
85dac906 PT |
4760 | |
4761 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4762 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
4763 | for_each_sched_entity(se) { |
4764 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4765 | /* throttled entity or throttle-on-deactivate */ | |
4766 | if (!se->on_rq) | |
4767 | break; | |
4768 | ||
6212437f | 4769 | if (dequeue) { |
85dac906 | 4770 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
6212437f VG |
4771 | } else { |
4772 | update_load_avg(qcfs_rq, se, 0); | |
4773 | se_update_runnable(se); | |
4774 | } | |
4775 | ||
85dac906 | 4776 | qcfs_rq->h_nr_running -= task_delta; |
43e9f7f2 | 4777 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 PT |
4778 | |
4779 | if (qcfs_rq->load.weight) | |
4780 | dequeue = 0; | |
4781 | } | |
4782 | ||
4783 | if (!se) | |
72465447 | 4784 | sub_nr_running(rq, task_delta); |
85dac906 | 4785 | |
c06f04c7 | 4786 | /* |
e98fa02c PT |
4787 | * Note: distribution will already see us throttled via the |
4788 | * throttled-list. rq->lock protects completion. | |
c06f04c7 | 4789 | */ |
e98fa02c PT |
4790 | cfs_rq->throttled = 1; |
4791 | cfs_rq->throttled_clock = rq_clock(rq); | |
4792 | return true; | |
85dac906 PT |
4793 | } |
4794 | ||
029632fb | 4795 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4796 | { |
4797 | struct rq *rq = rq_of(cfs_rq); | |
4798 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4799 | struct sched_entity *se; | |
43e9f7f2 | 4800 | long task_delta, idle_task_delta; |
671fd9da | 4801 | |
22b958d8 | 4802 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4803 | |
4804 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4805 | |
4806 | update_rq_clock(rq); | |
4807 | ||
671fd9da | 4808 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4809 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4810 | list_del_rcu(&cfs_rq->throttled_list); |
4811 | raw_spin_unlock(&cfs_b->lock); | |
4812 | ||
64660c86 PT |
4813 | /* update hierarchical throttle state */ |
4814 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4815 | ||
671fd9da PT |
4816 | if (!cfs_rq->load.weight) |
4817 | return; | |
4818 | ||
4819 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4820 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da PT |
4821 | for_each_sched_entity(se) { |
4822 | if (se->on_rq) | |
39f23ce0 VG |
4823 | break; |
4824 | cfs_rq = cfs_rq_of(se); | |
4825 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
4826 | ||
4827 | cfs_rq->h_nr_running += task_delta; | |
4828 | cfs_rq->idle_h_nr_running += idle_task_delta; | |
4829 | ||
4830 | /* end evaluation on encountering a throttled cfs_rq */ | |
4831 | if (cfs_rq_throttled(cfs_rq)) | |
4832 | goto unthrottle_throttle; | |
4833 | } | |
671fd9da | 4834 | |
39f23ce0 | 4835 | for_each_sched_entity(se) { |
671fd9da | 4836 | cfs_rq = cfs_rq_of(se); |
39f23ce0 VG |
4837 | |
4838 | update_load_avg(cfs_rq, se, UPDATE_TG); | |
4839 | se_update_runnable(se); | |
6212437f | 4840 | |
671fd9da | 4841 | cfs_rq->h_nr_running += task_delta; |
43e9f7f2 | 4842 | cfs_rq->idle_h_nr_running += idle_task_delta; |
671fd9da | 4843 | |
39f23ce0 VG |
4844 | |
4845 | /* end evaluation on encountering a throttled cfs_rq */ | |
671fd9da | 4846 | if (cfs_rq_throttled(cfs_rq)) |
39f23ce0 VG |
4847 | goto unthrottle_throttle; |
4848 | ||
4849 | /* | |
4850 | * One parent has been throttled and cfs_rq removed from the | |
4851 | * list. Add it back to not break the leaf list. | |
4852 | */ | |
4853 | if (throttled_hierarchy(cfs_rq)) | |
4854 | list_add_leaf_cfs_rq(cfs_rq); | |
671fd9da PT |
4855 | } |
4856 | ||
39f23ce0 VG |
4857 | /* At this point se is NULL and we are at root level*/ |
4858 | add_nr_running(rq, task_delta); | |
671fd9da | 4859 | |
39f23ce0 | 4860 | unthrottle_throttle: |
fe61468b VG |
4861 | /* |
4862 | * The cfs_rq_throttled() breaks in the above iteration can result in | |
4863 | * incomplete leaf list maintenance, resulting in triggering the | |
4864 | * assertion below. | |
4865 | */ | |
4866 | for_each_sched_entity(se) { | |
4867 | cfs_rq = cfs_rq_of(se); | |
4868 | ||
39f23ce0 VG |
4869 | if (list_add_leaf_cfs_rq(cfs_rq)) |
4870 | break; | |
fe61468b VG |
4871 | } |
4872 | ||
4873 | assert_list_leaf_cfs_rq(rq); | |
4874 | ||
97fb7a0a | 4875 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 4876 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 4877 | resched_curr(rq); |
671fd9da PT |
4878 | } |
4879 | ||
26a8b127 | 4880 | static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) |
671fd9da PT |
4881 | { |
4882 | struct cfs_rq *cfs_rq; | |
26a8b127 | 4883 | u64 runtime, remaining = 1; |
671fd9da PT |
4884 | |
4885 | rcu_read_lock(); | |
4886 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4887 | throttled_list) { | |
4888 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 4889 | struct rq_flags rf; |
671fd9da | 4890 | |
c0ad4aa4 | 4891 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
4892 | if (!cfs_rq_throttled(cfs_rq)) |
4893 | goto next; | |
4894 | ||
5e2d2cc2 L |
4895 | /* By the above check, this should never be true */ |
4896 | SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); | |
4897 | ||
26a8b127 | 4898 | raw_spin_lock(&cfs_b->lock); |
671fd9da | 4899 | runtime = -cfs_rq->runtime_remaining + 1; |
26a8b127 HC |
4900 | if (runtime > cfs_b->runtime) |
4901 | runtime = cfs_b->runtime; | |
4902 | cfs_b->runtime -= runtime; | |
4903 | remaining = cfs_b->runtime; | |
4904 | raw_spin_unlock(&cfs_b->lock); | |
671fd9da PT |
4905 | |
4906 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
4907 | |
4908 | /* we check whether we're throttled above */ | |
4909 | if (cfs_rq->runtime_remaining > 0) | |
4910 | unthrottle_cfs_rq(cfs_rq); | |
4911 | ||
4912 | next: | |
c0ad4aa4 | 4913 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
4914 | |
4915 | if (!remaining) | |
4916 | break; | |
4917 | } | |
4918 | rcu_read_unlock(); | |
671fd9da PT |
4919 | } |
4920 | ||
58088ad0 PT |
4921 | /* |
4922 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4923 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4924 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4925 | * used to track this state. | |
4926 | */ | |
c0ad4aa4 | 4927 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 4928 | { |
51f2176d | 4929 | int throttled; |
58088ad0 | 4930 | |
58088ad0 PT |
4931 | /* no need to continue the timer with no bandwidth constraint */ |
4932 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4933 | goto out_deactivate; |
58088ad0 | 4934 | |
671fd9da | 4935 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4936 | cfs_b->nr_periods += overrun; |
671fd9da | 4937 | |
51f2176d BS |
4938 | /* |
4939 | * idle depends on !throttled (for the case of a large deficit), and if | |
4940 | * we're going inactive then everything else can be deferred | |
4941 | */ | |
4942 | if (cfs_b->idle && !throttled) | |
4943 | goto out_deactivate; | |
a9cf55b2 PT |
4944 | |
4945 | __refill_cfs_bandwidth_runtime(cfs_b); | |
4946 | ||
671fd9da PT |
4947 | if (!throttled) { |
4948 | /* mark as potentially idle for the upcoming period */ | |
4949 | cfs_b->idle = 1; | |
51f2176d | 4950 | return 0; |
671fd9da PT |
4951 | } |
4952 | ||
e8da1b18 NR |
4953 | /* account preceding periods in which throttling occurred */ |
4954 | cfs_b->nr_throttled += overrun; | |
4955 | ||
671fd9da | 4956 | /* |
26a8b127 | 4957 | * This check is repeated as we release cfs_b->lock while we unthrottle. |
671fd9da | 4958 | */ |
ab93a4bc | 4959 | while (throttled && cfs_b->runtime > 0) { |
c0ad4aa4 | 4960 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 4961 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
26a8b127 | 4962 | distribute_cfs_runtime(cfs_b); |
c0ad4aa4 | 4963 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da PT |
4964 | |
4965 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
4966 | } | |
58088ad0 | 4967 | |
671fd9da PT |
4968 | /* |
4969 | * While we are ensured activity in the period following an | |
4970 | * unthrottle, this also covers the case in which the new bandwidth is | |
4971 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
4972 | * timer to remain active while there are any throttled entities.) | |
4973 | */ | |
4974 | cfs_b->idle = 0; | |
58088ad0 | 4975 | |
51f2176d BS |
4976 | return 0; |
4977 | ||
4978 | out_deactivate: | |
51f2176d | 4979 | return 1; |
58088ad0 | 4980 | } |
d3d9dc33 | 4981 | |
d8b4986d PT |
4982 | /* a cfs_rq won't donate quota below this amount */ |
4983 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
4984 | /* minimum remaining period time to redistribute slack quota */ | |
4985 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
4986 | /* how long we wait to gather additional slack before distributing */ | |
4987 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
4988 | ||
db06e78c BS |
4989 | /* |
4990 | * Are we near the end of the current quota period? | |
4991 | * | |
4992 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 4993 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
4994 | * migrate_hrtimers, base is never cleared, so we are fine. |
4995 | */ | |
d8b4986d PT |
4996 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
4997 | { | |
4998 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
4999 | u64 remaining; | |
5000 | ||
5001 | /* if the call-back is running a quota refresh is already occurring */ | |
5002 | if (hrtimer_callback_running(refresh_timer)) | |
5003 | return 1; | |
5004 | ||
5005 | /* is a quota refresh about to occur? */ | |
5006 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
5007 | if (remaining < min_expire) | |
5008 | return 1; | |
5009 | ||
5010 | return 0; | |
5011 | } | |
5012 | ||
5013 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
5014 | { | |
5015 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
5016 | ||
5017 | /* if there's a quota refresh soon don't bother with slack */ | |
5018 | if (runtime_refresh_within(cfs_b, min_left)) | |
5019 | return; | |
5020 | ||
66567fcb | 5021 | /* don't push forwards an existing deferred unthrottle */ |
5022 | if (cfs_b->slack_started) | |
5023 | return; | |
5024 | cfs_b->slack_started = true; | |
5025 | ||
4cfafd30 PZ |
5026 | hrtimer_start(&cfs_b->slack_timer, |
5027 | ns_to_ktime(cfs_bandwidth_slack_period), | |
5028 | HRTIMER_MODE_REL); | |
d8b4986d PT |
5029 | } |
5030 | ||
5031 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
5032 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5033 | { | |
5034 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5035 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
5036 | ||
5037 | if (slack_runtime <= 0) | |
5038 | return; | |
5039 | ||
5040 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 5041 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
5042 | cfs_b->runtime += slack_runtime; |
5043 | ||
5044 | /* we are under rq->lock, defer unthrottling using a timer */ | |
5045 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
5046 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
5047 | start_cfs_slack_bandwidth(cfs_b); | |
5048 | } | |
5049 | raw_spin_unlock(&cfs_b->lock); | |
5050 | ||
5051 | /* even if it's not valid for return we don't want to try again */ | |
5052 | cfs_rq->runtime_remaining -= slack_runtime; | |
5053 | } | |
5054 | ||
5055 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5056 | { | |
56f570e5 PT |
5057 | if (!cfs_bandwidth_used()) |
5058 | return; | |
5059 | ||
fccfdc6f | 5060 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
5061 | return; |
5062 | ||
5063 | __return_cfs_rq_runtime(cfs_rq); | |
5064 | } | |
5065 | ||
5066 | /* | |
5067 | * This is done with a timer (instead of inline with bandwidth return) since | |
5068 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
5069 | */ | |
5070 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
5071 | { | |
5072 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 5073 | unsigned long flags; |
d8b4986d PT |
5074 | |
5075 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 5076 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 5077 | cfs_b->slack_started = false; |
baa9be4f | 5078 | |
db06e78c | 5079 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 5080 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 5081 | return; |
db06e78c | 5082 | } |
d8b4986d | 5083 | |
c06f04c7 | 5084 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 5085 | runtime = cfs_b->runtime; |
c06f04c7 | 5086 | |
c0ad4aa4 | 5087 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
5088 | |
5089 | if (!runtime) | |
5090 | return; | |
5091 | ||
26a8b127 | 5092 | distribute_cfs_runtime(cfs_b); |
d8b4986d | 5093 | |
c0ad4aa4 | 5094 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
c0ad4aa4 | 5095 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
5096 | } |
5097 | ||
d3d9dc33 PT |
5098 | /* |
5099 | * When a group wakes up we want to make sure that its quota is not already | |
5100 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
5101 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
5102 | */ | |
5103 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
5104 | { | |
56f570e5 PT |
5105 | if (!cfs_bandwidth_used()) |
5106 | return; | |
5107 | ||
d3d9dc33 PT |
5108 | /* an active group must be handled by the update_curr()->put() path */ |
5109 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
5110 | return; | |
5111 | ||
5112 | /* ensure the group is not already throttled */ | |
5113 | if (cfs_rq_throttled(cfs_rq)) | |
5114 | return; | |
5115 | ||
5116 | /* update runtime allocation */ | |
5117 | account_cfs_rq_runtime(cfs_rq, 0); | |
5118 | if (cfs_rq->runtime_remaining <= 0) | |
5119 | throttle_cfs_rq(cfs_rq); | |
5120 | } | |
5121 | ||
55e16d30 PZ |
5122 | static void sync_throttle(struct task_group *tg, int cpu) |
5123 | { | |
5124 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
5125 | ||
5126 | if (!cfs_bandwidth_used()) | |
5127 | return; | |
5128 | ||
5129 | if (!tg->parent) | |
5130 | return; | |
5131 | ||
5132 | cfs_rq = tg->cfs_rq[cpu]; | |
5133 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
5134 | ||
5135 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 5136 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
5137 | } |
5138 | ||
d3d9dc33 | 5139 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 5140 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 5141 | { |
56f570e5 | 5142 | if (!cfs_bandwidth_used()) |
678d5718 | 5143 | return false; |
56f570e5 | 5144 | |
d3d9dc33 | 5145 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 5146 | return false; |
d3d9dc33 PT |
5147 | |
5148 | /* | |
5149 | * it's possible for a throttled entity to be forced into a running | |
5150 | * state (e.g. set_curr_task), in this case we're finished. | |
5151 | */ | |
5152 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 5153 | return true; |
d3d9dc33 | 5154 | |
e98fa02c | 5155 | return throttle_cfs_rq(cfs_rq); |
d3d9dc33 | 5156 | } |
029632fb | 5157 | |
029632fb PZ |
5158 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
5159 | { | |
5160 | struct cfs_bandwidth *cfs_b = | |
5161 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 5162 | |
029632fb PZ |
5163 | do_sched_cfs_slack_timer(cfs_b); |
5164 | ||
5165 | return HRTIMER_NORESTART; | |
5166 | } | |
5167 | ||
2e8e1922 PA |
5168 | extern const u64 max_cfs_quota_period; |
5169 | ||
029632fb PZ |
5170 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
5171 | { | |
5172 | struct cfs_bandwidth *cfs_b = | |
5173 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 5174 | unsigned long flags; |
029632fb PZ |
5175 | int overrun; |
5176 | int idle = 0; | |
2e8e1922 | 5177 | int count = 0; |
029632fb | 5178 | |
c0ad4aa4 | 5179 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 5180 | for (;;) { |
77a4d1a1 | 5181 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
5182 | if (!overrun) |
5183 | break; | |
5184 | ||
5a6d6a6c HC |
5185 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
5186 | ||
2e8e1922 PA |
5187 | if (++count > 3) { |
5188 | u64 new, old = ktime_to_ns(cfs_b->period); | |
5189 | ||
4929a4e6 XZ |
5190 | /* |
5191 | * Grow period by a factor of 2 to avoid losing precision. | |
5192 | * Precision loss in the quota/period ratio can cause __cfs_schedulable | |
5193 | * to fail. | |
5194 | */ | |
5195 | new = old * 2; | |
5196 | if (new < max_cfs_quota_period) { | |
5197 | cfs_b->period = ns_to_ktime(new); | |
5198 | cfs_b->quota *= 2; | |
5199 | ||
5200 | pr_warn_ratelimited( | |
5201 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5202 | smp_processor_id(), | |
5203 | div_u64(new, NSEC_PER_USEC), | |
5204 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5205 | } else { | |
5206 | pr_warn_ratelimited( | |
5207 | "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5208 | smp_processor_id(), | |
5209 | div_u64(old, NSEC_PER_USEC), | |
5210 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5211 | } | |
2e8e1922 PA |
5212 | |
5213 | /* reset count so we don't come right back in here */ | |
5214 | count = 0; | |
5215 | } | |
029632fb | 5216 | } |
4cfafd30 PZ |
5217 | if (idle) |
5218 | cfs_b->period_active = 0; | |
c0ad4aa4 | 5219 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
5220 | |
5221 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
5222 | } | |
5223 | ||
5224 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5225 | { | |
5226 | raw_spin_lock_init(&cfs_b->lock); | |
5227 | cfs_b->runtime = 0; | |
5228 | cfs_b->quota = RUNTIME_INF; | |
5229 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
5230 | ||
5231 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 5232 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5233 | cfs_b->period_timer.function = sched_cfs_period_timer; |
5234 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5235 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
66567fcb | 5236 | cfs_b->slack_started = false; |
029632fb PZ |
5237 | } |
5238 | ||
5239 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5240 | { | |
5241 | cfs_rq->runtime_enabled = 0; | |
5242 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
5243 | } | |
5244 | ||
77a4d1a1 | 5245 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 5246 | { |
4cfafd30 | 5247 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 5248 | |
f1d1be8a XP |
5249 | if (cfs_b->period_active) |
5250 | return; | |
5251 | ||
5252 | cfs_b->period_active = 1; | |
763a9ec0 | 5253 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); |
f1d1be8a | 5254 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5255 | } |
5256 | ||
5257 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5258 | { | |
7f1a169b TH |
5259 | /* init_cfs_bandwidth() was not called */ |
5260 | if (!cfs_b->throttled_cfs_rq.next) | |
5261 | return; | |
5262 | ||
029632fb PZ |
5263 | hrtimer_cancel(&cfs_b->period_timer); |
5264 | hrtimer_cancel(&cfs_b->slack_timer); | |
5265 | } | |
5266 | ||
502ce005 | 5267 | /* |
97fb7a0a | 5268 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
5269 | * |
5270 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5271 | * bits doesn't do much. | |
5272 | */ | |
5273 | ||
5274 | /* cpu online calback */ | |
0e59bdae KT |
5275 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5276 | { | |
502ce005 | 5277 | struct task_group *tg; |
0e59bdae | 5278 | |
502ce005 PZ |
5279 | lockdep_assert_held(&rq->lock); |
5280 | ||
5281 | rcu_read_lock(); | |
5282 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5283 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5284 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5285 | |
5286 | raw_spin_lock(&cfs_b->lock); | |
5287 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5288 | raw_spin_unlock(&cfs_b->lock); | |
5289 | } | |
502ce005 | 5290 | rcu_read_unlock(); |
0e59bdae KT |
5291 | } |
5292 | ||
502ce005 | 5293 | /* cpu offline callback */ |
38dc3348 | 5294 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5295 | { |
502ce005 PZ |
5296 | struct task_group *tg; |
5297 | ||
5298 | lockdep_assert_held(&rq->lock); | |
5299 | ||
5300 | rcu_read_lock(); | |
5301 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5302 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5303 | |
029632fb PZ |
5304 | if (!cfs_rq->runtime_enabled) |
5305 | continue; | |
5306 | ||
5307 | /* | |
5308 | * clock_task is not advancing so we just need to make sure | |
5309 | * there's some valid quota amount | |
5310 | */ | |
51f2176d | 5311 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 5312 | /* |
97fb7a0a | 5313 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
5314 | * in take_cpu_down(), so we prevent new cfs throttling here. |
5315 | */ | |
5316 | cfs_rq->runtime_enabled = 0; | |
5317 | ||
029632fb PZ |
5318 | if (cfs_rq_throttled(cfs_rq)) |
5319 | unthrottle_cfs_rq(cfs_rq); | |
5320 | } | |
502ce005 | 5321 | rcu_read_unlock(); |
029632fb PZ |
5322 | } |
5323 | ||
5324 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f6783319 VG |
5325 | |
5326 | static inline bool cfs_bandwidth_used(void) | |
5327 | { | |
5328 | return false; | |
5329 | } | |
5330 | ||
9dbdb155 | 5331 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5332 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5333 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5334 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5335 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5336 | |
5337 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5338 | { | |
5339 | return 0; | |
5340 | } | |
64660c86 PT |
5341 | |
5342 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5343 | { | |
5344 | return 0; | |
5345 | } | |
5346 | ||
5347 | static inline int throttled_lb_pair(struct task_group *tg, | |
5348 | int src_cpu, int dest_cpu) | |
5349 | { | |
5350 | return 0; | |
5351 | } | |
029632fb PZ |
5352 | |
5353 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5354 | ||
5355 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5356 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5357 | #endif |
5358 | ||
029632fb PZ |
5359 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5360 | { | |
5361 | return NULL; | |
5362 | } | |
5363 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5364 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5365 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5366 | |
5367 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5368 | ||
bf0f6f24 IM |
5369 | /************************************************** |
5370 | * CFS operations on tasks: | |
5371 | */ | |
5372 | ||
8f4d37ec PZ |
5373 | #ifdef CONFIG_SCHED_HRTICK |
5374 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5375 | { | |
8f4d37ec PZ |
5376 | struct sched_entity *se = &p->se; |
5377 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5378 | ||
9148a3a1 | 5379 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5380 | |
8bf46a39 | 5381 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5382 | u64 slice = sched_slice(cfs_rq, se); |
5383 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5384 | s64 delta = slice - ran; | |
5385 | ||
5386 | if (delta < 0) { | |
5387 | if (rq->curr == p) | |
8875125e | 5388 | resched_curr(rq); |
8f4d37ec PZ |
5389 | return; |
5390 | } | |
31656519 | 5391 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5392 | } |
5393 | } | |
a4c2f00f PZ |
5394 | |
5395 | /* | |
5396 | * called from enqueue/dequeue and updates the hrtick when the | |
5397 | * current task is from our class and nr_running is low enough | |
5398 | * to matter. | |
5399 | */ | |
5400 | static void hrtick_update(struct rq *rq) | |
5401 | { | |
5402 | struct task_struct *curr = rq->curr; | |
5403 | ||
b39e66ea | 5404 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5405 | return; |
5406 | ||
5407 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5408 | hrtick_start_fair(rq, curr); | |
5409 | } | |
55e12e5e | 5410 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5411 | static inline void |
5412 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5413 | { | |
5414 | } | |
a4c2f00f PZ |
5415 | |
5416 | static inline void hrtick_update(struct rq *rq) | |
5417 | { | |
5418 | } | |
8f4d37ec PZ |
5419 | #endif |
5420 | ||
2802bf3c MR |
5421 | #ifdef CONFIG_SMP |
5422 | static inline unsigned long cpu_util(int cpu); | |
2802bf3c MR |
5423 | |
5424 | static inline bool cpu_overutilized(int cpu) | |
5425 | { | |
60e17f5c | 5426 | return !fits_capacity(cpu_util(cpu), capacity_of(cpu)); |
2802bf3c MR |
5427 | } |
5428 | ||
5429 | static inline void update_overutilized_status(struct rq *rq) | |
5430 | { | |
f9f240f9 | 5431 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { |
2802bf3c | 5432 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); |
f9f240f9 QY |
5433 | trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); |
5434 | } | |
2802bf3c MR |
5435 | } |
5436 | #else | |
5437 | static inline void update_overutilized_status(struct rq *rq) { } | |
5438 | #endif | |
5439 | ||
323af6de VK |
5440 | /* Runqueue only has SCHED_IDLE tasks enqueued */ |
5441 | static int sched_idle_rq(struct rq *rq) | |
5442 | { | |
5443 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
5444 | rq->nr_running); | |
5445 | } | |
5446 | ||
afa70d94 | 5447 | #ifdef CONFIG_SMP |
323af6de VK |
5448 | static int sched_idle_cpu(int cpu) |
5449 | { | |
5450 | return sched_idle_rq(cpu_rq(cpu)); | |
5451 | } | |
afa70d94 | 5452 | #endif |
323af6de | 5453 | |
bf0f6f24 IM |
5454 | /* |
5455 | * The enqueue_task method is called before nr_running is | |
5456 | * increased. Here we update the fair scheduling stats and | |
5457 | * then put the task into the rbtree: | |
5458 | */ | |
ea87bb78 | 5459 | static void |
371fd7e7 | 5460 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5461 | { |
5462 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5463 | struct sched_entity *se = &p->se; |
43e9f7f2 | 5464 | int idle_h_nr_running = task_has_idle_policy(p); |
bf0f6f24 | 5465 | |
2539fc82 PB |
5466 | /* |
5467 | * The code below (indirectly) updates schedutil which looks at | |
5468 | * the cfs_rq utilization to select a frequency. | |
5469 | * Let's add the task's estimated utilization to the cfs_rq's | |
5470 | * estimated utilization, before we update schedutil. | |
5471 | */ | |
5472 | util_est_enqueue(&rq->cfs, p); | |
5473 | ||
8c34ab19 RW |
5474 | /* |
5475 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5476 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5477 | * passed. | |
5478 | */ | |
5479 | if (p->in_iowait) | |
674e7541 | 5480 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5481 | |
bf0f6f24 | 5482 | for_each_sched_entity(se) { |
62fb1851 | 5483 | if (se->on_rq) |
bf0f6f24 IM |
5484 | break; |
5485 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5486 | enqueue_entity(cfs_rq, se, flags); |
85dac906 | 5487 | |
953bfcd1 | 5488 | cfs_rq->h_nr_running++; |
43e9f7f2 | 5489 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 5490 | |
6d4d2246 VG |
5491 | /* end evaluation on encountering a throttled cfs_rq */ |
5492 | if (cfs_rq_throttled(cfs_rq)) | |
5493 | goto enqueue_throttle; | |
5494 | ||
88ec22d3 | 5495 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5496 | } |
8f4d37ec | 5497 | |
2069dd75 | 5498 | for_each_sched_entity(se) { |
0f317143 | 5499 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5500 | |
88c0616e | 5501 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5502 | se_update_runnable(se); |
1ea6c46a | 5503 | update_cfs_group(se); |
6d4d2246 VG |
5504 | |
5505 | cfs_rq->h_nr_running++; | |
5506 | cfs_rq->idle_h_nr_running += idle_h_nr_running; | |
5ab297ba VG |
5507 | |
5508 | /* end evaluation on encountering a throttled cfs_rq */ | |
5509 | if (cfs_rq_throttled(cfs_rq)) | |
5510 | goto enqueue_throttle; | |
b34cb07d PA |
5511 | |
5512 | /* | |
5513 | * One parent has been throttled and cfs_rq removed from the | |
5514 | * list. Add it back to not break the leaf list. | |
5515 | */ | |
5516 | if (throttled_hierarchy(cfs_rq)) | |
5517 | list_add_leaf_cfs_rq(cfs_rq); | |
2069dd75 PZ |
5518 | } |
5519 | ||
7d148be6 VG |
5520 | /* At this point se is NULL and we are at root level*/ |
5521 | add_nr_running(rq, 1); | |
2802bf3c | 5522 | |
7d148be6 VG |
5523 | /* |
5524 | * Since new tasks are assigned an initial util_avg equal to | |
5525 | * half of the spare capacity of their CPU, tiny tasks have the | |
5526 | * ability to cross the overutilized threshold, which will | |
5527 | * result in the load balancer ruining all the task placement | |
5528 | * done by EAS. As a way to mitigate that effect, do not account | |
5529 | * for the first enqueue operation of new tasks during the | |
5530 | * overutilized flag detection. | |
5531 | * | |
5532 | * A better way of solving this problem would be to wait for | |
5533 | * the PELT signals of tasks to converge before taking them | |
5534 | * into account, but that is not straightforward to implement, | |
5535 | * and the following generally works well enough in practice. | |
5536 | */ | |
5537 | if (flags & ENQUEUE_WAKEUP) | |
5538 | update_overutilized_status(rq); | |
cd126afe | 5539 | |
7d148be6 | 5540 | enqueue_throttle: |
f6783319 VG |
5541 | if (cfs_bandwidth_used()) { |
5542 | /* | |
5543 | * When bandwidth control is enabled; the cfs_rq_throttled() | |
5544 | * breaks in the above iteration can result in incomplete | |
5545 | * leaf list maintenance, resulting in triggering the assertion | |
5546 | * below. | |
5547 | */ | |
5548 | for_each_sched_entity(se) { | |
5549 | cfs_rq = cfs_rq_of(se); | |
5550 | ||
5551 | if (list_add_leaf_cfs_rq(cfs_rq)) | |
5552 | break; | |
5553 | } | |
5554 | } | |
5555 | ||
5d299eab PZ |
5556 | assert_list_leaf_cfs_rq(rq); |
5557 | ||
a4c2f00f | 5558 | hrtick_update(rq); |
bf0f6f24 IM |
5559 | } |
5560 | ||
2f36825b VP |
5561 | static void set_next_buddy(struct sched_entity *se); |
5562 | ||
bf0f6f24 IM |
5563 | /* |
5564 | * The dequeue_task method is called before nr_running is | |
5565 | * decreased. We remove the task from the rbtree and | |
5566 | * update the fair scheduling stats: | |
5567 | */ | |
371fd7e7 | 5568 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5569 | { |
5570 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5571 | struct sched_entity *se = &p->se; |
2f36825b | 5572 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 5573 | int idle_h_nr_running = task_has_idle_policy(p); |
323af6de | 5574 | bool was_sched_idle = sched_idle_rq(rq); |
bf0f6f24 IM |
5575 | |
5576 | for_each_sched_entity(se) { | |
5577 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5578 | dequeue_entity(cfs_rq, se, flags); |
85dac906 | 5579 | |
953bfcd1 | 5580 | cfs_rq->h_nr_running--; |
43e9f7f2 | 5581 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 5582 | |
6d4d2246 VG |
5583 | /* end evaluation on encountering a throttled cfs_rq */ |
5584 | if (cfs_rq_throttled(cfs_rq)) | |
5585 | goto dequeue_throttle; | |
5586 | ||
bf0f6f24 | 5587 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5588 | if (cfs_rq->load.weight) { |
754bd598 KK |
5589 | /* Avoid re-evaluating load for this entity: */ |
5590 | se = parent_entity(se); | |
2f36825b VP |
5591 | /* |
5592 | * Bias pick_next to pick a task from this cfs_rq, as | |
5593 | * p is sleeping when it is within its sched_slice. | |
5594 | */ | |
754bd598 KK |
5595 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5596 | set_next_buddy(se); | |
bf0f6f24 | 5597 | break; |
2f36825b | 5598 | } |
371fd7e7 | 5599 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5600 | } |
8f4d37ec | 5601 | |
2069dd75 | 5602 | for_each_sched_entity(se) { |
0f317143 | 5603 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5604 | |
88c0616e | 5605 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5606 | se_update_runnable(se); |
1ea6c46a | 5607 | update_cfs_group(se); |
6d4d2246 VG |
5608 | |
5609 | cfs_rq->h_nr_running--; | |
5610 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; | |
5ab297ba VG |
5611 | |
5612 | /* end evaluation on encountering a throttled cfs_rq */ | |
5613 | if (cfs_rq_throttled(cfs_rq)) | |
5614 | goto dequeue_throttle; | |
5615 | ||
2069dd75 PZ |
5616 | } |
5617 | ||
6d4d2246 | 5618 | dequeue_throttle: |
cd126afe | 5619 | if (!se) |
72465447 | 5620 | sub_nr_running(rq, 1); |
cd126afe | 5621 | |
323af6de VK |
5622 | /* balance early to pull high priority tasks */ |
5623 | if (unlikely(!was_sched_idle && sched_idle_rq(rq))) | |
5624 | rq->next_balance = jiffies; | |
5625 | ||
7f65ea42 | 5626 | util_est_dequeue(&rq->cfs, p, task_sleep); |
a4c2f00f | 5627 | hrtick_update(rq); |
bf0f6f24 IM |
5628 | } |
5629 | ||
e7693a36 | 5630 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5631 | |
5632 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
5633 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5634 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5635 | ||
9fd81dd5 | 5636 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
5637 | |
5638 | static struct { | |
5639 | cpumask_var_t idle_cpus_mask; | |
5640 | atomic_t nr_cpus; | |
f643ea22 | 5641 | int has_blocked; /* Idle CPUS has blocked load */ |
e022e0d3 | 5642 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 5643 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
5644 | } nohz ____cacheline_aligned; |
5645 | ||
9fd81dd5 | 5646 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5647 | |
b0fb1eb4 VG |
5648 | static unsigned long cpu_load(struct rq *rq) |
5649 | { | |
5650 | return cfs_rq_load_avg(&rq->cfs); | |
5651 | } | |
5652 | ||
3318544b VG |
5653 | /* |
5654 | * cpu_load_without - compute CPU load without any contributions from *p | |
5655 | * @cpu: the CPU which load is requested | |
5656 | * @p: the task which load should be discounted | |
5657 | * | |
5658 | * The load of a CPU is defined by the load of tasks currently enqueued on that | |
5659 | * CPU as well as tasks which are currently sleeping after an execution on that | |
5660 | * CPU. | |
5661 | * | |
5662 | * This method returns the load of the specified CPU by discounting the load of | |
5663 | * the specified task, whenever the task is currently contributing to the CPU | |
5664 | * load. | |
5665 | */ | |
5666 | static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) | |
5667 | { | |
5668 | struct cfs_rq *cfs_rq; | |
5669 | unsigned int load; | |
5670 | ||
5671 | /* Task has no contribution or is new */ | |
5672 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5673 | return cpu_load(rq); | |
5674 | ||
5675 | cfs_rq = &rq->cfs; | |
5676 | load = READ_ONCE(cfs_rq->avg.load_avg); | |
5677 | ||
5678 | /* Discount task's util from CPU's util */ | |
5679 | lsub_positive(&load, task_h_load(p)); | |
5680 | ||
5681 | return load; | |
5682 | } | |
5683 | ||
9f683953 VG |
5684 | static unsigned long cpu_runnable(struct rq *rq) |
5685 | { | |
5686 | return cfs_rq_runnable_avg(&rq->cfs); | |
5687 | } | |
5688 | ||
070f5e86 VG |
5689 | static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) |
5690 | { | |
5691 | struct cfs_rq *cfs_rq; | |
5692 | unsigned int runnable; | |
5693 | ||
5694 | /* Task has no contribution or is new */ | |
5695 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5696 | return cpu_runnable(rq); | |
5697 | ||
5698 | cfs_rq = &rq->cfs; | |
5699 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
5700 | ||
5701 | /* Discount task's runnable from CPU's runnable */ | |
5702 | lsub_positive(&runnable, p->se.avg.runnable_avg); | |
5703 | ||
5704 | return runnable; | |
5705 | } | |
5706 | ||
ced549fa | 5707 | static unsigned long capacity_of(int cpu) |
029632fb | 5708 | { |
ced549fa | 5709 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5710 | } |
5711 | ||
c58d25f3 PZ |
5712 | static void record_wakee(struct task_struct *p) |
5713 | { | |
5714 | /* | |
5715 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5716 | * jiffy will not have built up many flips. | |
5717 | */ | |
5718 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5719 | current->wakee_flips >>= 1; | |
5720 | current->wakee_flip_decay_ts = jiffies; | |
5721 | } | |
5722 | ||
5723 | if (current->last_wakee != p) { | |
5724 | current->last_wakee = p; | |
5725 | current->wakee_flips++; | |
5726 | } | |
5727 | } | |
5728 | ||
63b0e9ed MG |
5729 | /* |
5730 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5731 | * |
63b0e9ed | 5732 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5733 | * at a frequency roughly N times higher than one of its wakees. |
5734 | * | |
5735 | * In order to determine whether we should let the load spread vs consolidating | |
5736 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5737 | * partner, and a factor of lls_size higher frequency in the other. | |
5738 | * | |
5739 | * With both conditions met, we can be relatively sure that the relationship is | |
5740 | * non-monogamous, with partner count exceeding socket size. | |
5741 | * | |
5742 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5743 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5744 | * socket size. | |
63b0e9ed | 5745 | */ |
62470419 MW |
5746 | static int wake_wide(struct task_struct *p) |
5747 | { | |
63b0e9ed MG |
5748 | unsigned int master = current->wakee_flips; |
5749 | unsigned int slave = p->wakee_flips; | |
17c891ab | 5750 | int factor = __this_cpu_read(sd_llc_size); |
62470419 | 5751 | |
63b0e9ed MG |
5752 | if (master < slave) |
5753 | swap(master, slave); | |
5754 | if (slave < factor || master < slave * factor) | |
5755 | return 0; | |
5756 | return 1; | |
62470419 MW |
5757 | } |
5758 | ||
90001d67 | 5759 | /* |
d153b153 PZ |
5760 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
5761 | * soonest. For the purpose of speed we only consider the waking and previous | |
5762 | * CPU. | |
90001d67 | 5763 | * |
7332dec0 MG |
5764 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
5765 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
5766 | * |
5767 | * wake_affine_weight() - considers the weight to reflect the average | |
5768 | * scheduling latency of the CPUs. This seems to work | |
5769 | * for the overloaded case. | |
90001d67 | 5770 | */ |
3b76c4a3 | 5771 | static int |
89a55f56 | 5772 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 5773 | { |
7332dec0 MG |
5774 | /* |
5775 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
5776 | * context. Only allow the move if cache is shared. Otherwise an | |
5777 | * interrupt intensive workload could force all tasks onto one | |
5778 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
5779 | * |
5780 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
5781 | * There is no guarantee that the cache hot data from an interrupt | |
5782 | * is more important than cache hot data on the prev_cpu and from | |
5783 | * a cpufreq perspective, it's better to have higher utilisation | |
5784 | * on one CPU. | |
7332dec0 | 5785 | */ |
943d355d RJ |
5786 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
5787 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 5788 | |
d153b153 | 5789 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 5790 | return this_cpu; |
90001d67 | 5791 | |
3b76c4a3 | 5792 | return nr_cpumask_bits; |
90001d67 PZ |
5793 | } |
5794 | ||
3b76c4a3 | 5795 | static int |
f2cdd9cc PZ |
5796 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
5797 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5798 | { |
90001d67 PZ |
5799 | s64 this_eff_load, prev_eff_load; |
5800 | unsigned long task_load; | |
5801 | ||
11f10e54 | 5802 | this_eff_load = cpu_load(cpu_rq(this_cpu)); |
90001d67 | 5803 | |
90001d67 PZ |
5804 | if (sync) { |
5805 | unsigned long current_load = task_h_load(current); | |
5806 | ||
f2cdd9cc | 5807 | if (current_load > this_eff_load) |
3b76c4a3 | 5808 | return this_cpu; |
90001d67 | 5809 | |
f2cdd9cc | 5810 | this_eff_load -= current_load; |
90001d67 PZ |
5811 | } |
5812 | ||
90001d67 PZ |
5813 | task_load = task_h_load(p); |
5814 | ||
f2cdd9cc PZ |
5815 | this_eff_load += task_load; |
5816 | if (sched_feat(WA_BIAS)) | |
5817 | this_eff_load *= 100; | |
5818 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 5819 | |
11f10e54 | 5820 | prev_eff_load = cpu_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
5821 | prev_eff_load -= task_load; |
5822 | if (sched_feat(WA_BIAS)) | |
5823 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
5824 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 5825 | |
082f764a MG |
5826 | /* |
5827 | * If sync, adjust the weight of prev_eff_load such that if | |
5828 | * prev_eff == this_eff that select_idle_sibling() will consider | |
5829 | * stacking the wakee on top of the waker if no other CPU is | |
5830 | * idle. | |
5831 | */ | |
5832 | if (sync) | |
5833 | prev_eff_load += 1; | |
5834 | ||
5835 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
5836 | } |
5837 | ||
772bd008 | 5838 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 5839 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 5840 | { |
3b76c4a3 | 5841 | int target = nr_cpumask_bits; |
098fb9db | 5842 | |
89a55f56 | 5843 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 5844 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 5845 | |
3b76c4a3 MG |
5846 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
5847 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 5848 | |
ae92882e | 5849 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
3b76c4a3 MG |
5850 | if (target == nr_cpumask_bits) |
5851 | return prev_cpu; | |
098fb9db | 5852 | |
3b76c4a3 MG |
5853 | schedstat_inc(sd->ttwu_move_affine); |
5854 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
5855 | return target; | |
098fb9db IM |
5856 | } |
5857 | ||
aaee1203 | 5858 | static struct sched_group * |
45da2773 | 5859 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); |
aaee1203 PZ |
5860 | |
5861 | /* | |
97fb7a0a | 5862 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
5863 | */ |
5864 | static int | |
18bd1b4b | 5865 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
5866 | { |
5867 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5868 | unsigned int min_exit_latency = UINT_MAX; |
5869 | u64 latest_idle_timestamp = 0; | |
5870 | int least_loaded_cpu = this_cpu; | |
17346452 | 5871 | int shallowest_idle_cpu = -1; |
aaee1203 PZ |
5872 | int i; |
5873 | ||
eaecf41f MR |
5874 | /* Check if we have any choice: */ |
5875 | if (group->group_weight == 1) | |
ae4df9d6 | 5876 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 5877 | |
aaee1203 | 5878 | /* Traverse only the allowed CPUs */ |
3bd37062 | 5879 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
17346452 VK |
5880 | if (sched_idle_cpu(i)) |
5881 | return i; | |
5882 | ||
943d355d | 5883 | if (available_idle_cpu(i)) { |
83a0a96a NP |
5884 | struct rq *rq = cpu_rq(i); |
5885 | struct cpuidle_state *idle = idle_get_state(rq); | |
5886 | if (idle && idle->exit_latency < min_exit_latency) { | |
5887 | /* | |
5888 | * We give priority to a CPU whose idle state | |
5889 | * has the smallest exit latency irrespective | |
5890 | * of any idle timestamp. | |
5891 | */ | |
5892 | min_exit_latency = idle->exit_latency; | |
5893 | latest_idle_timestamp = rq->idle_stamp; | |
5894 | shallowest_idle_cpu = i; | |
5895 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5896 | rq->idle_stamp > latest_idle_timestamp) { | |
5897 | /* | |
5898 | * If equal or no active idle state, then | |
5899 | * the most recently idled CPU might have | |
5900 | * a warmer cache. | |
5901 | */ | |
5902 | latest_idle_timestamp = rq->idle_stamp; | |
5903 | shallowest_idle_cpu = i; | |
5904 | } | |
17346452 | 5905 | } else if (shallowest_idle_cpu == -1) { |
11f10e54 | 5906 | load = cpu_load(cpu_rq(i)); |
18cec7e0 | 5907 | if (load < min_load) { |
83a0a96a NP |
5908 | min_load = load; |
5909 | least_loaded_cpu = i; | |
5910 | } | |
e7693a36 GH |
5911 | } |
5912 | } | |
5913 | ||
17346452 | 5914 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 5915 | } |
e7693a36 | 5916 | |
18bd1b4b BJ |
5917 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
5918 | int cpu, int prev_cpu, int sd_flag) | |
5919 | { | |
93f50f90 | 5920 | int new_cpu = cpu; |
18bd1b4b | 5921 | |
3bd37062 | 5922 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
5923 | return prev_cpu; |
5924 | ||
c976a862 | 5925 | /* |
57abff06 | 5926 | * We need task's util for cpu_util_without, sync it up to |
c469933e | 5927 | * prev_cpu's last_update_time. |
c976a862 VK |
5928 | */ |
5929 | if (!(sd_flag & SD_BALANCE_FORK)) | |
5930 | sync_entity_load_avg(&p->se); | |
5931 | ||
18bd1b4b BJ |
5932 | while (sd) { |
5933 | struct sched_group *group; | |
5934 | struct sched_domain *tmp; | |
5935 | int weight; | |
5936 | ||
5937 | if (!(sd->flags & sd_flag)) { | |
5938 | sd = sd->child; | |
5939 | continue; | |
5940 | } | |
5941 | ||
45da2773 | 5942 | group = find_idlest_group(sd, p, cpu); |
18bd1b4b BJ |
5943 | if (!group) { |
5944 | sd = sd->child; | |
5945 | continue; | |
5946 | } | |
5947 | ||
5948 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 5949 | if (new_cpu == cpu) { |
97fb7a0a | 5950 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
5951 | sd = sd->child; |
5952 | continue; | |
5953 | } | |
5954 | ||
97fb7a0a | 5955 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
5956 | cpu = new_cpu; |
5957 | weight = sd->span_weight; | |
5958 | sd = NULL; | |
5959 | for_each_domain(cpu, tmp) { | |
5960 | if (weight <= tmp->span_weight) | |
5961 | break; | |
5962 | if (tmp->flags & sd_flag) | |
5963 | sd = tmp; | |
5964 | } | |
18bd1b4b BJ |
5965 | } |
5966 | ||
5967 | return new_cpu; | |
5968 | } | |
5969 | ||
10e2f1ac | 5970 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 5971 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 5972 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
5973 | |
5974 | static inline void set_idle_cores(int cpu, int val) | |
5975 | { | |
5976 | struct sched_domain_shared *sds; | |
5977 | ||
5978 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5979 | if (sds) | |
5980 | WRITE_ONCE(sds->has_idle_cores, val); | |
5981 | } | |
5982 | ||
5983 | static inline bool test_idle_cores(int cpu, bool def) | |
5984 | { | |
5985 | struct sched_domain_shared *sds; | |
5986 | ||
5987 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5988 | if (sds) | |
5989 | return READ_ONCE(sds->has_idle_cores); | |
5990 | ||
5991 | return def; | |
5992 | } | |
5993 | ||
5994 | /* | |
5995 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
5996 | * information in sd_llc_shared->has_idle_cores. | |
5997 | * | |
5998 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
5999 | * state should be fairly cheap. | |
6000 | */ | |
1b568f0a | 6001 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
6002 | { |
6003 | int core = cpu_of(rq); | |
6004 | int cpu; | |
6005 | ||
6006 | rcu_read_lock(); | |
6007 | if (test_idle_cores(core, true)) | |
6008 | goto unlock; | |
6009 | ||
6010 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6011 | if (cpu == core) | |
6012 | continue; | |
6013 | ||
943d355d | 6014 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6015 | goto unlock; |
6016 | } | |
6017 | ||
6018 | set_idle_cores(core, 1); | |
6019 | unlock: | |
6020 | rcu_read_unlock(); | |
6021 | } | |
6022 | ||
6023 | /* | |
6024 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
6025 | * there are no idle cores left in the system; tracked through | |
6026 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
6027 | */ | |
6028 | static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
6029 | { | |
6030 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); | |
c743f0a5 | 6031 | int core, cpu; |
10e2f1ac | 6032 | |
1b568f0a PZ |
6033 | if (!static_branch_likely(&sched_smt_present)) |
6034 | return -1; | |
6035 | ||
10e2f1ac PZ |
6036 | if (!test_idle_cores(target, false)) |
6037 | return -1; | |
6038 | ||
3bd37062 | 6039 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
10e2f1ac | 6040 | |
c743f0a5 | 6041 | for_each_cpu_wrap(core, cpus, target) { |
10e2f1ac PZ |
6042 | bool idle = true; |
6043 | ||
6044 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
bec2860a | 6045 | if (!available_idle_cpu(cpu)) { |
10e2f1ac | 6046 | idle = false; |
bec2860a SD |
6047 | break; |
6048 | } | |
10e2f1ac | 6049 | } |
bec2860a | 6050 | cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); |
10e2f1ac PZ |
6051 | |
6052 | if (idle) | |
6053 | return core; | |
6054 | } | |
6055 | ||
6056 | /* | |
6057 | * Failed to find an idle core; stop looking for one. | |
6058 | */ | |
6059 | set_idle_cores(target, 0); | |
6060 | ||
6061 | return -1; | |
6062 | } | |
6063 | ||
6064 | /* | |
6065 | * Scan the local SMT mask for idle CPUs. | |
6066 | */ | |
1b5500d7 | 6067 | static int select_idle_smt(struct task_struct *p, int target) |
10e2f1ac | 6068 | { |
17346452 | 6069 | int cpu; |
10e2f1ac | 6070 | |
1b568f0a PZ |
6071 | if (!static_branch_likely(&sched_smt_present)) |
6072 | return -1; | |
6073 | ||
10e2f1ac | 6074 | for_each_cpu(cpu, cpu_smt_mask(target)) { |
3bd37062 | 6075 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
10e2f1ac | 6076 | continue; |
17346452 | 6077 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) |
10e2f1ac PZ |
6078 | return cpu; |
6079 | } | |
6080 | ||
17346452 | 6081 | return -1; |
10e2f1ac PZ |
6082 | } |
6083 | ||
6084 | #else /* CONFIG_SCHED_SMT */ | |
6085 | ||
6086 | static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
6087 | { | |
6088 | return -1; | |
6089 | } | |
6090 | ||
1b5500d7 | 6091 | static inline int select_idle_smt(struct task_struct *p, int target) |
10e2f1ac PZ |
6092 | { |
6093 | return -1; | |
6094 | } | |
6095 | ||
6096 | #endif /* CONFIG_SCHED_SMT */ | |
6097 | ||
6098 | /* | |
6099 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6100 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6101 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6102 | */ |
10e2f1ac PZ |
6103 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target) |
6104 | { | |
60588bfa | 6105 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); |
9cfb38a7 | 6106 | struct sched_domain *this_sd; |
1ad3aaf3 | 6107 | u64 avg_cost, avg_idle; |
d76343c6 | 6108 | u64 time; |
8dc2d993 | 6109 | int this = smp_processor_id(); |
17346452 | 6110 | int cpu, nr = INT_MAX; |
10e2f1ac | 6111 | |
9cfb38a7 WL |
6112 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6113 | if (!this_sd) | |
6114 | return -1; | |
6115 | ||
10e2f1ac PZ |
6116 | /* |
6117 | * Due to large variance we need a large fuzz factor; hackbench in | |
6118 | * particularly is sensitive here. | |
6119 | */ | |
1ad3aaf3 PZ |
6120 | avg_idle = this_rq()->avg_idle / 512; |
6121 | avg_cost = this_sd->avg_scan_cost + 1; | |
6122 | ||
6123 | if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost) | |
10e2f1ac PZ |
6124 | return -1; |
6125 | ||
1ad3aaf3 PZ |
6126 | if (sched_feat(SIS_PROP)) { |
6127 | u64 span_avg = sd->span_weight * avg_idle; | |
6128 | if (span_avg > 4*avg_cost) | |
6129 | nr = div_u64(span_avg, avg_cost); | |
6130 | else | |
6131 | nr = 4; | |
6132 | } | |
6133 | ||
8dc2d993 | 6134 | time = cpu_clock(this); |
10e2f1ac | 6135 | |
60588bfa CJ |
6136 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6137 | ||
6138 | for_each_cpu_wrap(cpu, cpus, target) { | |
1ad3aaf3 | 6139 | if (!--nr) |
17346452 VK |
6140 | return -1; |
6141 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) | |
10e2f1ac PZ |
6142 | break; |
6143 | } | |
6144 | ||
8dc2d993 | 6145 | time = cpu_clock(this) - time; |
d76343c6 | 6146 | update_avg(&this_sd->avg_scan_cost, time); |
10e2f1ac PZ |
6147 | |
6148 | return cpu; | |
6149 | } | |
6150 | ||
b7a33161 MR |
6151 | /* |
6152 | * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which | |
6153 | * the task fits. If no CPU is big enough, but there are idle ones, try to | |
6154 | * maximize capacity. | |
6155 | */ | |
6156 | static int | |
6157 | select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) | |
6158 | { | |
6159 | unsigned long best_cap = 0; | |
6160 | int cpu, best_cpu = -1; | |
6161 | struct cpumask *cpus; | |
6162 | ||
6163 | sync_entity_load_avg(&p->se); | |
6164 | ||
6165 | cpus = this_cpu_cpumask_var_ptr(select_idle_mask); | |
6166 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); | |
6167 | ||
6168 | for_each_cpu_wrap(cpu, cpus, target) { | |
6169 | unsigned long cpu_cap = capacity_of(cpu); | |
6170 | ||
6171 | if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) | |
6172 | continue; | |
6173 | if (task_fits_capacity(p, cpu_cap)) | |
6174 | return cpu; | |
6175 | ||
6176 | if (cpu_cap > best_cap) { | |
6177 | best_cap = cpu_cap; | |
6178 | best_cpu = cpu; | |
6179 | } | |
6180 | } | |
6181 | ||
6182 | return best_cpu; | |
6183 | } | |
6184 | ||
10e2f1ac PZ |
6185 | /* |
6186 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6187 | */ |
772bd008 | 6188 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6189 | { |
99bd5e2f | 6190 | struct sched_domain *sd; |
32e839dd | 6191 | int i, recent_used_cpu; |
a50bde51 | 6192 | |
b7a33161 MR |
6193 | /* |
6194 | * For asymmetric CPU capacity systems, our domain of interest is | |
6195 | * sd_asym_cpucapacity rather than sd_llc. | |
6196 | */ | |
6197 | if (static_branch_unlikely(&sched_asym_cpucapacity)) { | |
6198 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); | |
6199 | /* | |
6200 | * On an asymmetric CPU capacity system where an exclusive | |
6201 | * cpuset defines a symmetric island (i.e. one unique | |
6202 | * capacity_orig value through the cpuset), the key will be set | |
6203 | * but the CPUs within that cpuset will not have a domain with | |
6204 | * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric | |
6205 | * capacity path. | |
6206 | */ | |
6207 | if (!sd) | |
6208 | goto symmetric; | |
6209 | ||
6210 | i = select_idle_capacity(p, sd, target); | |
6211 | return ((unsigned)i < nr_cpumask_bits) ? i : target; | |
6212 | } | |
6213 | ||
6214 | symmetric: | |
3c29e651 | 6215 | if (available_idle_cpu(target) || sched_idle_cpu(target)) |
e0a79f52 | 6216 | return target; |
99bd5e2f SS |
6217 | |
6218 | /* | |
97fb7a0a | 6219 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6220 | */ |
3c29e651 VK |
6221 | if (prev != target && cpus_share_cache(prev, target) && |
6222 | (available_idle_cpu(prev) || sched_idle_cpu(prev))) | |
772bd008 | 6223 | return prev; |
a50bde51 | 6224 | |
52262ee5 MG |
6225 | /* |
6226 | * Allow a per-cpu kthread to stack with the wakee if the | |
6227 | * kworker thread and the tasks previous CPUs are the same. | |
6228 | * The assumption is that the wakee queued work for the | |
6229 | * per-cpu kthread that is now complete and the wakeup is | |
6230 | * essentially a sync wakeup. An obvious example of this | |
6231 | * pattern is IO completions. | |
6232 | */ | |
6233 | if (is_per_cpu_kthread(current) && | |
6234 | prev == smp_processor_id() && | |
6235 | this_rq()->nr_running <= 1) { | |
6236 | return prev; | |
6237 | } | |
6238 | ||
97fb7a0a | 6239 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd MG |
6240 | recent_used_cpu = p->recent_used_cpu; |
6241 | if (recent_used_cpu != prev && | |
6242 | recent_used_cpu != target && | |
6243 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 6244 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
3bd37062 | 6245 | cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) { |
32e839dd MG |
6246 | /* |
6247 | * Replace recent_used_cpu with prev as it is a potential | |
97fb7a0a | 6248 | * candidate for the next wake: |
32e839dd MG |
6249 | */ |
6250 | p->recent_used_cpu = prev; | |
6251 | return recent_used_cpu; | |
6252 | } | |
6253 | ||
518cd623 | 6254 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6255 | if (!sd) |
6256 | return target; | |
772bd008 | 6257 | |
10e2f1ac PZ |
6258 | i = select_idle_core(p, sd, target); |
6259 | if ((unsigned)i < nr_cpumask_bits) | |
6260 | return i; | |
37407ea7 | 6261 | |
10e2f1ac PZ |
6262 | i = select_idle_cpu(p, sd, target); |
6263 | if ((unsigned)i < nr_cpumask_bits) | |
6264 | return i; | |
6265 | ||
1b5500d7 | 6266 | i = select_idle_smt(p, target); |
10e2f1ac PZ |
6267 | if ((unsigned)i < nr_cpumask_bits) |
6268 | return i; | |
970e1789 | 6269 | |
a50bde51 PZ |
6270 | return target; |
6271 | } | |
231678b7 | 6272 | |
f9be3e59 PB |
6273 | /** |
6274 | * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks | |
6275 | * @cpu: the CPU to get the utilization of | |
6276 | * | |
6277 | * The unit of the return value must be the one of capacity so we can compare | |
6278 | * the utilization with the capacity of the CPU that is available for CFS task | |
6279 | * (ie cpu_capacity). | |
231678b7 DE |
6280 | * |
6281 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
6282 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
6283 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
6284 | * capacity_orig is the cpu_capacity available at the highest frequency | |
6285 | * (arch_scale_freq_capacity()). | |
6286 | * The utilization of a CPU converges towards a sum equal to or less than the | |
6287 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
6288 | * the running time on this CPU scaled by capacity_curr. | |
6289 | * | |
f9be3e59 PB |
6290 | * The estimated utilization of a CPU is defined to be the maximum between its |
6291 | * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks | |
6292 | * currently RUNNABLE on that CPU. | |
6293 | * This allows to properly represent the expected utilization of a CPU which | |
6294 | * has just got a big task running since a long sleep period. At the same time | |
6295 | * however it preserves the benefits of the "blocked utilization" in | |
6296 | * describing the potential for other tasks waking up on the same CPU. | |
6297 | * | |
231678b7 DE |
6298 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even |
6299 | * higher than capacity_orig because of unfortunate rounding in | |
6300 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
6301 | * the average stabilizes with the new running time. We need to check that the | |
6302 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
6303 | * necessary. Without utilization capping, a group could be seen as overloaded | |
6304 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
6305 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
6306 | * capacity_orig) as it useful for predicting the capacity required after task | |
6307 | * migrations (scheduler-driven DVFS). | |
f9be3e59 PB |
6308 | * |
6309 | * Return: the (estimated) utilization for the specified CPU | |
8bb5b00c | 6310 | */ |
f9be3e59 | 6311 | static inline unsigned long cpu_util(int cpu) |
8bb5b00c | 6312 | { |
f9be3e59 PB |
6313 | struct cfs_rq *cfs_rq; |
6314 | unsigned int util; | |
6315 | ||
6316 | cfs_rq = &cpu_rq(cpu)->cfs; | |
6317 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6318 | ||
6319 | if (sched_feat(UTIL_EST)) | |
6320 | util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); | |
8bb5b00c | 6321 | |
f9be3e59 | 6322 | return min_t(unsigned long, util, capacity_orig_of(cpu)); |
8bb5b00c | 6323 | } |
a50bde51 | 6324 | |
104cb16d | 6325 | /* |
c469933e PB |
6326 | * cpu_util_without: compute cpu utilization without any contributions from *p |
6327 | * @cpu: the CPU which utilization is requested | |
6328 | * @p: the task which utilization should be discounted | |
6329 | * | |
6330 | * The utilization of a CPU is defined by the utilization of tasks currently | |
6331 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
6332 | * execution on that CPU. | |
6333 | * | |
6334 | * This method returns the utilization of the specified CPU by discounting the | |
6335 | * utilization of the specified task, whenever the task is currently | |
6336 | * contributing to the CPU utilization. | |
104cb16d | 6337 | */ |
c469933e | 6338 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) |
104cb16d | 6339 | { |
f9be3e59 PB |
6340 | struct cfs_rq *cfs_rq; |
6341 | unsigned int util; | |
104cb16d MR |
6342 | |
6343 | /* Task has no contribution or is new */ | |
f9be3e59 | 6344 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) |
104cb16d MR |
6345 | return cpu_util(cpu); |
6346 | ||
f9be3e59 PB |
6347 | cfs_rq = &cpu_rq(cpu)->cfs; |
6348 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6349 | ||
c469933e | 6350 | /* Discount task's util from CPU's util */ |
b5c0ce7b | 6351 | lsub_positive(&util, task_util(p)); |
104cb16d | 6352 | |
f9be3e59 PB |
6353 | /* |
6354 | * Covered cases: | |
6355 | * | |
6356 | * a) if *p is the only task sleeping on this CPU, then: | |
6357 | * cpu_util (== task_util) > util_est (== 0) | |
6358 | * and thus we return: | |
c469933e | 6359 | * cpu_util_without = (cpu_util - task_util) = 0 |
f9be3e59 PB |
6360 | * |
6361 | * b) if other tasks are SLEEPING on this CPU, which is now exiting | |
6362 | * IDLE, then: | |
6363 | * cpu_util >= task_util | |
6364 | * cpu_util > util_est (== 0) | |
6365 | * and thus we discount *p's blocked utilization to return: | |
c469933e | 6366 | * cpu_util_without = (cpu_util - task_util) >= 0 |
f9be3e59 PB |
6367 | * |
6368 | * c) if other tasks are RUNNABLE on that CPU and | |
6369 | * util_est > cpu_util | |
6370 | * then we use util_est since it returns a more restrictive | |
6371 | * estimation of the spare capacity on that CPU, by just | |
6372 | * considering the expected utilization of tasks already | |
6373 | * runnable on that CPU. | |
6374 | * | |
6375 | * Cases a) and b) are covered by the above code, while case c) is | |
6376 | * covered by the following code when estimated utilization is | |
6377 | * enabled. | |
6378 | */ | |
c469933e PB |
6379 | if (sched_feat(UTIL_EST)) { |
6380 | unsigned int estimated = | |
6381 | READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6382 | ||
6383 | /* | |
6384 | * Despite the following checks we still have a small window | |
6385 | * for a possible race, when an execl's select_task_rq_fair() | |
6386 | * races with LB's detach_task(): | |
6387 | * | |
6388 | * detach_task() | |
6389 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
6390 | * ---------------------------------- A | |
6391 | * deactivate_task() \ | |
6392 | * dequeue_task() + RaceTime | |
6393 | * util_est_dequeue() / | |
6394 | * ---------------------------------- B | |
6395 | * | |
6396 | * The additional check on "current == p" it's required to | |
6397 | * properly fix the execl regression and it helps in further | |
6398 | * reducing the chances for the above race. | |
6399 | */ | |
b5c0ce7b PB |
6400 | if (unlikely(task_on_rq_queued(p) || current == p)) |
6401 | lsub_positive(&estimated, _task_util_est(p)); | |
6402 | ||
c469933e PB |
6403 | util = max(util, estimated); |
6404 | } | |
f9be3e59 PB |
6405 | |
6406 | /* | |
6407 | * Utilization (estimated) can exceed the CPU capacity, thus let's | |
6408 | * clamp to the maximum CPU capacity to ensure consistency with | |
6409 | * the cpu_util call. | |
6410 | */ | |
6411 | return min_t(unsigned long, util, capacity_orig_of(cpu)); | |
104cb16d MR |
6412 | } |
6413 | ||
390031e4 QP |
6414 | /* |
6415 | * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) | |
6416 | * to @dst_cpu. | |
6417 | */ | |
6418 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
6419 | { | |
6420 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
6421 | unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); | |
6422 | ||
6423 | /* | |
6424 | * If @p migrates from @cpu to another, remove its contribution. Or, | |
6425 | * if @p migrates from another CPU to @cpu, add its contribution. In | |
6426 | * the other cases, @cpu is not impacted by the migration, so the | |
6427 | * util_avg should already be correct. | |
6428 | */ | |
6429 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
6430 | sub_positive(&util, task_util(p)); | |
6431 | else if (task_cpu(p) != cpu && dst_cpu == cpu) | |
6432 | util += task_util(p); | |
6433 | ||
6434 | if (sched_feat(UTIL_EST)) { | |
6435 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6436 | ||
6437 | /* | |
6438 | * During wake-up, the task isn't enqueued yet and doesn't | |
6439 | * appear in the cfs_rq->avg.util_est.enqueued of any rq, | |
6440 | * so just add it (if needed) to "simulate" what will be | |
6441 | * cpu_util() after the task has been enqueued. | |
6442 | */ | |
6443 | if (dst_cpu == cpu) | |
6444 | util_est += _task_util_est(p); | |
6445 | ||
6446 | util = max(util, util_est); | |
6447 | } | |
6448 | ||
6449 | return min(util, capacity_orig_of(cpu)); | |
6450 | } | |
6451 | ||
6452 | /* | |
eb92692b | 6453 | * compute_energy(): Estimates the energy that @pd would consume if @p was |
390031e4 | 6454 | * migrated to @dst_cpu. compute_energy() predicts what will be the utilization |
eb92692b | 6455 | * landscape of @pd's CPUs after the task migration, and uses the Energy Model |
390031e4 QP |
6456 | * to compute what would be the energy if we decided to actually migrate that |
6457 | * task. | |
6458 | */ | |
6459 | static long | |
6460 | compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) | |
6461 | { | |
eb92692b QP |
6462 | struct cpumask *pd_mask = perf_domain_span(pd); |
6463 | unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); | |
6464 | unsigned long max_util = 0, sum_util = 0; | |
390031e4 QP |
6465 | int cpu; |
6466 | ||
eb92692b QP |
6467 | /* |
6468 | * The capacity state of CPUs of the current rd can be driven by CPUs | |
6469 | * of another rd if they belong to the same pd. So, account for the | |
6470 | * utilization of these CPUs too by masking pd with cpu_online_mask | |
6471 | * instead of the rd span. | |
6472 | * | |
6473 | * If an entire pd is outside of the current rd, it will not appear in | |
6474 | * its pd list and will not be accounted by compute_energy(). | |
6475 | */ | |
6476 | for_each_cpu_and(cpu, pd_mask, cpu_online_mask) { | |
6477 | unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu); | |
6478 | struct task_struct *tsk = cpu == dst_cpu ? p : NULL; | |
af24bde8 PB |
6479 | |
6480 | /* | |
eb92692b QP |
6481 | * Busy time computation: utilization clamping is not |
6482 | * required since the ratio (sum_util / cpu_capacity) | |
6483 | * is already enough to scale the EM reported power | |
6484 | * consumption at the (eventually clamped) cpu_capacity. | |
af24bde8 | 6485 | */ |
eb92692b QP |
6486 | sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap, |
6487 | ENERGY_UTIL, NULL); | |
af24bde8 | 6488 | |
390031e4 | 6489 | /* |
eb92692b QP |
6490 | * Performance domain frequency: utilization clamping |
6491 | * must be considered since it affects the selection | |
6492 | * of the performance domain frequency. | |
6493 | * NOTE: in case RT tasks are running, by default the | |
6494 | * FREQUENCY_UTIL's utilization can be max OPP. | |
390031e4 | 6495 | */ |
eb92692b QP |
6496 | cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap, |
6497 | FREQUENCY_UTIL, tsk); | |
6498 | max_util = max(max_util, cpu_util); | |
390031e4 QP |
6499 | } |
6500 | ||
eb92692b | 6501 | return em_pd_energy(pd->em_pd, max_util, sum_util); |
390031e4 QP |
6502 | } |
6503 | ||
732cd75b QP |
6504 | /* |
6505 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
6506 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
6507 | * spare capacity in each performance domain and uses it as a potential | |
6508 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
6509 | * out which of the CPU candidates is the most energy-efficient. | |
6510 | * | |
6511 | * The rationale for this heuristic is as follows. In a performance domain, | |
6512 | * all the most energy efficient CPU candidates (according to the Energy | |
6513 | * Model) are those for which we'll request a low frequency. When there are | |
6514 | * several CPUs for which the frequency request will be the same, we don't | |
6515 | * have enough data to break the tie between them, because the Energy Model | |
6516 | * only includes active power costs. With this model, if we assume that | |
6517 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
6518 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
6519 | * the best candidates of the performance domain. | |
6520 | * | |
6521 | * In practice, it could be preferable from an energy standpoint to pack | |
6522 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
6523 | * but that could also hurt our chances to go cluster idle, and we have no | |
6524 | * ways to tell with the current Energy Model if this is actually a good | |
6525 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
6526 | * cluster-packing, and spreading inside a cluster. That should at least be | |
6527 | * a good thing for latency, and this is consistent with the idea that most | |
6528 | * of the energy savings of EAS come from the asymmetry of the system, and | |
6529 | * not so much from breaking the tie between identical CPUs. That's also the | |
6530 | * reason why EAS is enabled in the topology code only for systems where | |
6531 | * SD_ASYM_CPUCAPACITY is set. | |
6532 | * | |
6533 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
6534 | * they don't have any useful utilization data yet and it's not possible to | |
6535 | * forecast their impact on energy consumption. Consequently, they will be | |
6536 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
6537 | * to be energy-inefficient in some use-cases. The alternative would be to | |
6538 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
6539 | * their util_avg from the parent task, but those heuristics could hurt | |
6540 | * other use-cases too. So, until someone finds a better way to solve this, | |
6541 | * let's keep things simple by re-using the existing slow path. | |
6542 | */ | |
732cd75b QP |
6543 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) |
6544 | { | |
eb92692b | 6545 | unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; |
732cd75b | 6546 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; |
eb92692b | 6547 | unsigned long cpu_cap, util, base_energy = 0; |
732cd75b | 6548 | int cpu, best_energy_cpu = prev_cpu; |
732cd75b | 6549 | struct sched_domain *sd; |
eb92692b | 6550 | struct perf_domain *pd; |
732cd75b QP |
6551 | |
6552 | rcu_read_lock(); | |
6553 | pd = rcu_dereference(rd->pd); | |
6554 | if (!pd || READ_ONCE(rd->overutilized)) | |
6555 | goto fail; | |
732cd75b QP |
6556 | |
6557 | /* | |
6558 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
6559 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
6560 | */ | |
6561 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
6562 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
6563 | sd = sd->parent; | |
6564 | if (!sd) | |
6565 | goto fail; | |
6566 | ||
6567 | sync_entity_load_avg(&p->se); | |
6568 | if (!task_util_est(p)) | |
6569 | goto unlock; | |
6570 | ||
6571 | for (; pd; pd = pd->next) { | |
eb92692b QP |
6572 | unsigned long cur_delta, spare_cap, max_spare_cap = 0; |
6573 | unsigned long base_energy_pd; | |
732cd75b QP |
6574 | int max_spare_cap_cpu = -1; |
6575 | ||
eb92692b QP |
6576 | /* Compute the 'base' energy of the pd, without @p */ |
6577 | base_energy_pd = compute_energy(p, -1, pd); | |
6578 | base_energy += base_energy_pd; | |
6579 | ||
732cd75b | 6580 | for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { |
3bd37062 | 6581 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
6582 | continue; |
6583 | ||
732cd75b QP |
6584 | util = cpu_util_next(cpu, p, cpu); |
6585 | cpu_cap = capacity_of(cpu); | |
1d42509e VS |
6586 | spare_cap = cpu_cap - util; |
6587 | ||
6588 | /* | |
6589 | * Skip CPUs that cannot satisfy the capacity request. | |
6590 | * IOW, placing the task there would make the CPU | |
6591 | * overutilized. Take uclamp into account to see how | |
6592 | * much capacity we can get out of the CPU; this is | |
6593 | * aligned with schedutil_cpu_util(). | |
6594 | */ | |
6595 | util = uclamp_rq_util_with(cpu_rq(cpu), util, p); | |
60e17f5c | 6596 | if (!fits_capacity(util, cpu_cap)) |
732cd75b QP |
6597 | continue; |
6598 | ||
6599 | /* Always use prev_cpu as a candidate. */ | |
6600 | if (cpu == prev_cpu) { | |
eb92692b QP |
6601 | prev_delta = compute_energy(p, prev_cpu, pd); |
6602 | prev_delta -= base_energy_pd; | |
6603 | best_delta = min(best_delta, prev_delta); | |
732cd75b QP |
6604 | } |
6605 | ||
6606 | /* | |
6607 | * Find the CPU with the maximum spare capacity in | |
6608 | * the performance domain | |
6609 | */ | |
732cd75b QP |
6610 | if (spare_cap > max_spare_cap) { |
6611 | max_spare_cap = spare_cap; | |
6612 | max_spare_cap_cpu = cpu; | |
6613 | } | |
6614 | } | |
6615 | ||
6616 | /* Evaluate the energy impact of using this CPU. */ | |
4892f51a | 6617 | if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) { |
eb92692b QP |
6618 | cur_delta = compute_energy(p, max_spare_cap_cpu, pd); |
6619 | cur_delta -= base_energy_pd; | |
6620 | if (cur_delta < best_delta) { | |
6621 | best_delta = cur_delta; | |
732cd75b QP |
6622 | best_energy_cpu = max_spare_cap_cpu; |
6623 | } | |
6624 | } | |
6625 | } | |
6626 | unlock: | |
6627 | rcu_read_unlock(); | |
6628 | ||
6629 | /* | |
6630 | * Pick the best CPU if prev_cpu cannot be used, or if it saves at | |
6631 | * least 6% of the energy used by prev_cpu. | |
6632 | */ | |
eb92692b | 6633 | if (prev_delta == ULONG_MAX) |
732cd75b QP |
6634 | return best_energy_cpu; |
6635 | ||
eb92692b | 6636 | if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) |
732cd75b QP |
6637 | return best_energy_cpu; |
6638 | ||
6639 | return prev_cpu; | |
6640 | ||
6641 | fail: | |
6642 | rcu_read_unlock(); | |
6643 | ||
6644 | return -1; | |
6645 | } | |
6646 | ||
aaee1203 | 6647 | /* |
de91b9cb MR |
6648 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
6649 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
6650 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 6651 | * |
97fb7a0a IM |
6652 | * Balances load by selecting the idlest CPU in the idlest group, or under |
6653 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 6654 | * |
97fb7a0a | 6655 | * Returns the target CPU number. |
aaee1203 PZ |
6656 | * |
6657 | * preempt must be disabled. | |
6658 | */ | |
0017d735 | 6659 | static int |
ac66f547 | 6660 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 6661 | { |
f1d88b44 | 6662 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 6663 | int cpu = smp_processor_id(); |
63b0e9ed | 6664 | int new_cpu = prev_cpu; |
99bd5e2f | 6665 | int want_affine = 0; |
24d0c1d6 | 6666 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
c88d5910 | 6667 | |
c58d25f3 PZ |
6668 | if (sd_flag & SD_BALANCE_WAKE) { |
6669 | record_wakee(p); | |
732cd75b | 6670 | |
f8a696f2 | 6671 | if (sched_energy_enabled()) { |
732cd75b QP |
6672 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
6673 | if (new_cpu >= 0) | |
6674 | return new_cpu; | |
6675 | new_cpu = prev_cpu; | |
6676 | } | |
6677 | ||
00061968 | 6678 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 6679 | } |
aaee1203 | 6680 | |
dce840a0 | 6681 | rcu_read_lock(); |
aaee1203 | 6682 | for_each_domain(cpu, tmp) { |
fe3bcfe1 | 6683 | /* |
97fb7a0a | 6684 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 6685 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 6686 | */ |
99bd5e2f SS |
6687 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6688 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
6689 | if (cpu != prev_cpu) |
6690 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
6691 | ||
6692 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 6693 | break; |
f03542a7 | 6694 | } |
29cd8bae | 6695 | |
f03542a7 | 6696 | if (tmp->flags & sd_flag) |
29cd8bae | 6697 | sd = tmp; |
63b0e9ed MG |
6698 | else if (!want_affine) |
6699 | break; | |
29cd8bae PZ |
6700 | } |
6701 | ||
f1d88b44 VK |
6702 | if (unlikely(sd)) { |
6703 | /* Slow path */ | |
18bd1b4b | 6704 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
f1d88b44 VK |
6705 | } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */ |
6706 | /* Fast path */ | |
6707 | ||
6708 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); | |
6709 | ||
6710 | if (want_affine) | |
6711 | current->recent_used_cpu = cpu; | |
e7693a36 | 6712 | } |
dce840a0 | 6713 | rcu_read_unlock(); |
e7693a36 | 6714 | |
c88d5910 | 6715 | return new_cpu; |
e7693a36 | 6716 | } |
0a74bef8 | 6717 | |
144d8487 PZ |
6718 | static void detach_entity_cfs_rq(struct sched_entity *se); |
6719 | ||
0a74bef8 | 6720 | /* |
97fb7a0a | 6721 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 6722 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 6723 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6724 | */ |
3f9672ba | 6725 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 6726 | { |
59efa0ba PZ |
6727 | /* |
6728 | * As blocked tasks retain absolute vruntime the migration needs to | |
6729 | * deal with this by subtracting the old and adding the new | |
6730 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6731 | * the task on the new runqueue. | |
6732 | */ | |
6733 | if (p->state == TASK_WAKING) { | |
6734 | struct sched_entity *se = &p->se; | |
6735 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6736 | u64 min_vruntime; | |
6737 | ||
6738 | #ifndef CONFIG_64BIT | |
6739 | u64 min_vruntime_copy; | |
6740 | ||
6741 | do { | |
6742 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
6743 | smp_rmb(); | |
6744 | min_vruntime = cfs_rq->min_vruntime; | |
6745 | } while (min_vruntime != min_vruntime_copy); | |
6746 | #else | |
6747 | min_vruntime = cfs_rq->min_vruntime; | |
6748 | #endif | |
6749 | ||
6750 | se->vruntime -= min_vruntime; | |
6751 | } | |
6752 | ||
144d8487 PZ |
6753 | if (p->on_rq == TASK_ON_RQ_MIGRATING) { |
6754 | /* | |
6755 | * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' | |
6756 | * rq->lock and can modify state directly. | |
6757 | */ | |
6758 | lockdep_assert_held(&task_rq(p)->lock); | |
6759 | detach_entity_cfs_rq(&p->se); | |
6760 | ||
6761 | } else { | |
6762 | /* | |
6763 | * We are supposed to update the task to "current" time, then | |
6764 | * its up to date and ready to go to new CPU/cfs_rq. But we | |
6765 | * have difficulty in getting what current time is, so simply | |
6766 | * throw away the out-of-date time. This will result in the | |
6767 | * wakee task is less decayed, but giving the wakee more load | |
6768 | * sounds not bad. | |
6769 | */ | |
6770 | remove_entity_load_avg(&p->se); | |
6771 | } | |
9d89c257 YD |
6772 | |
6773 | /* Tell new CPU we are migrated */ | |
6774 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
6775 | |
6776 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 6777 | p->se.exec_start = 0; |
3f9672ba SD |
6778 | |
6779 | update_scan_period(p, new_cpu); | |
0a74bef8 | 6780 | } |
12695578 YD |
6781 | |
6782 | static void task_dead_fair(struct task_struct *p) | |
6783 | { | |
6784 | remove_entity_load_avg(&p->se); | |
6785 | } | |
6e2df058 PZ |
6786 | |
6787 | static int | |
6788 | balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
6789 | { | |
6790 | if (rq->nr_running) | |
6791 | return 1; | |
6792 | ||
6793 | return newidle_balance(rq, rf) != 0; | |
6794 | } | |
e7693a36 GH |
6795 | #endif /* CONFIG_SMP */ |
6796 | ||
a555e9d8 | 6797 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
6798 | { |
6799 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
6800 | ||
6801 | /* | |
e52fb7c0 PZ |
6802 | * Since its curr running now, convert the gran from real-time |
6803 | * to virtual-time in his units. | |
13814d42 MG |
6804 | * |
6805 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
6806 | * they get preempted easier. That is, if 'se' < 'curr' then | |
6807 | * the resulting gran will be larger, therefore penalizing the | |
6808 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
6809 | * be smaller, again penalizing the lighter task. | |
6810 | * | |
6811 | * This is especially important for buddies when the leftmost | |
6812 | * task is higher priority than the buddy. | |
0bbd3336 | 6813 | */ |
f4ad9bd2 | 6814 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
6815 | } |
6816 | ||
464b7527 PZ |
6817 | /* |
6818 | * Should 'se' preempt 'curr'. | |
6819 | * | |
6820 | * |s1 | |
6821 | * |s2 | |
6822 | * |s3 | |
6823 | * g | |
6824 | * |<--->|c | |
6825 | * | |
6826 | * w(c, s1) = -1 | |
6827 | * w(c, s2) = 0 | |
6828 | * w(c, s3) = 1 | |
6829 | * | |
6830 | */ | |
6831 | static int | |
6832 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
6833 | { | |
6834 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
6835 | ||
6836 | if (vdiff <= 0) | |
6837 | return -1; | |
6838 | ||
a555e9d8 | 6839 | gran = wakeup_gran(se); |
464b7527 PZ |
6840 | if (vdiff > gran) |
6841 | return 1; | |
6842 | ||
6843 | return 0; | |
6844 | } | |
6845 | ||
02479099 PZ |
6846 | static void set_last_buddy(struct sched_entity *se) |
6847 | { | |
1da1843f | 6848 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6849 | return; |
6850 | ||
c5ae366e DA |
6851 | for_each_sched_entity(se) { |
6852 | if (SCHED_WARN_ON(!se->on_rq)) | |
6853 | return; | |
69c80f3e | 6854 | cfs_rq_of(se)->last = se; |
c5ae366e | 6855 | } |
02479099 PZ |
6856 | } |
6857 | ||
6858 | static void set_next_buddy(struct sched_entity *se) | |
6859 | { | |
1da1843f | 6860 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6861 | return; |
6862 | ||
c5ae366e DA |
6863 | for_each_sched_entity(se) { |
6864 | if (SCHED_WARN_ON(!se->on_rq)) | |
6865 | return; | |
69c80f3e | 6866 | cfs_rq_of(se)->next = se; |
c5ae366e | 6867 | } |
02479099 PZ |
6868 | } |
6869 | ||
ac53db59 RR |
6870 | static void set_skip_buddy(struct sched_entity *se) |
6871 | { | |
69c80f3e VP |
6872 | for_each_sched_entity(se) |
6873 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
6874 | } |
6875 | ||
bf0f6f24 IM |
6876 | /* |
6877 | * Preempt the current task with a newly woken task if needed: | |
6878 | */ | |
5a9b86f6 | 6879 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
6880 | { |
6881 | struct task_struct *curr = rq->curr; | |
8651a86c | 6882 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 6883 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 6884 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 6885 | int next_buddy_marked = 0; |
bf0f6f24 | 6886 | |
4ae7d5ce IM |
6887 | if (unlikely(se == pse)) |
6888 | return; | |
6889 | ||
5238cdd3 | 6890 | /* |
163122b7 | 6891 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
6892 | * unconditionally check_prempt_curr() after an enqueue (which may have |
6893 | * lead to a throttle). This both saves work and prevents false | |
6894 | * next-buddy nomination below. | |
6895 | */ | |
6896 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
6897 | return; | |
6898 | ||
2f36825b | 6899 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 6900 | set_next_buddy(pse); |
2f36825b VP |
6901 | next_buddy_marked = 1; |
6902 | } | |
57fdc26d | 6903 | |
aec0a514 BR |
6904 | /* |
6905 | * We can come here with TIF_NEED_RESCHED already set from new task | |
6906 | * wake up path. | |
5238cdd3 PT |
6907 | * |
6908 | * Note: this also catches the edge-case of curr being in a throttled | |
6909 | * group (e.g. via set_curr_task), since update_curr() (in the | |
6910 | * enqueue of curr) will have resulted in resched being set. This | |
6911 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
6912 | * below. | |
aec0a514 BR |
6913 | */ |
6914 | if (test_tsk_need_resched(curr)) | |
6915 | return; | |
6916 | ||
a2f5c9ab | 6917 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
6918 | if (unlikely(task_has_idle_policy(curr)) && |
6919 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
6920 | goto preempt; |
6921 | ||
91c234b4 | 6922 | /* |
a2f5c9ab DH |
6923 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
6924 | * is driven by the tick): | |
91c234b4 | 6925 | */ |
8ed92e51 | 6926 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 6927 | return; |
bf0f6f24 | 6928 | |
464b7527 | 6929 | find_matching_se(&se, &pse); |
9bbd7374 | 6930 | update_curr(cfs_rq_of(se)); |
002f128b | 6931 | BUG_ON(!pse); |
2f36825b VP |
6932 | if (wakeup_preempt_entity(se, pse) == 1) { |
6933 | /* | |
6934 | * Bias pick_next to pick the sched entity that is | |
6935 | * triggering this preemption. | |
6936 | */ | |
6937 | if (!next_buddy_marked) | |
6938 | set_next_buddy(pse); | |
3a7e73a2 | 6939 | goto preempt; |
2f36825b | 6940 | } |
464b7527 | 6941 | |
3a7e73a2 | 6942 | return; |
a65ac745 | 6943 | |
3a7e73a2 | 6944 | preempt: |
8875125e | 6945 | resched_curr(rq); |
3a7e73a2 PZ |
6946 | /* |
6947 | * Only set the backward buddy when the current task is still | |
6948 | * on the rq. This can happen when a wakeup gets interleaved | |
6949 | * with schedule on the ->pre_schedule() or idle_balance() | |
6950 | * point, either of which can * drop the rq lock. | |
6951 | * | |
6952 | * Also, during early boot the idle thread is in the fair class, | |
6953 | * for obvious reasons its a bad idea to schedule back to it. | |
6954 | */ | |
6955 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
6956 | return; | |
6957 | ||
6958 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
6959 | set_last_buddy(se); | |
bf0f6f24 IM |
6960 | } |
6961 | ||
5d7d6056 | 6962 | struct task_struct * |
d8ac8971 | 6963 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
6964 | { |
6965 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6966 | struct sched_entity *se; | |
678d5718 | 6967 | struct task_struct *p; |
37e117c0 | 6968 | int new_tasks; |
678d5718 | 6969 | |
6e83125c | 6970 | again: |
6e2df058 | 6971 | if (!sched_fair_runnable(rq)) |
38033c37 | 6972 | goto idle; |
678d5718 | 6973 | |
9674f5ca | 6974 | #ifdef CONFIG_FAIR_GROUP_SCHED |
67692435 | 6975 | if (!prev || prev->sched_class != &fair_sched_class) |
678d5718 PZ |
6976 | goto simple; |
6977 | ||
6978 | /* | |
6979 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
6980 | * likely that a next task is from the same cgroup as the current. | |
6981 | * | |
6982 | * Therefore attempt to avoid putting and setting the entire cgroup | |
6983 | * hierarchy, only change the part that actually changes. | |
6984 | */ | |
6985 | ||
6986 | do { | |
6987 | struct sched_entity *curr = cfs_rq->curr; | |
6988 | ||
6989 | /* | |
6990 | * Since we got here without doing put_prev_entity() we also | |
6991 | * have to consider cfs_rq->curr. If it is still a runnable | |
6992 | * entity, update_curr() will update its vruntime, otherwise | |
6993 | * forget we've ever seen it. | |
6994 | */ | |
54d27365 BS |
6995 | if (curr) { |
6996 | if (curr->on_rq) | |
6997 | update_curr(cfs_rq); | |
6998 | else | |
6999 | curr = NULL; | |
678d5718 | 7000 | |
54d27365 BS |
7001 | /* |
7002 | * This call to check_cfs_rq_runtime() will do the | |
7003 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 7004 | * Therefore the nr_running test will indeed |
54d27365 BS |
7005 | * be correct. |
7006 | */ | |
9674f5ca VK |
7007 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
7008 | cfs_rq = &rq->cfs; | |
7009 | ||
7010 | if (!cfs_rq->nr_running) | |
7011 | goto idle; | |
7012 | ||
54d27365 | 7013 | goto simple; |
9674f5ca | 7014 | } |
54d27365 | 7015 | } |
678d5718 PZ |
7016 | |
7017 | se = pick_next_entity(cfs_rq, curr); | |
7018 | cfs_rq = group_cfs_rq(se); | |
7019 | } while (cfs_rq); | |
7020 | ||
7021 | p = task_of(se); | |
7022 | ||
7023 | /* | |
7024 | * Since we haven't yet done put_prev_entity and if the selected task | |
7025 | * is a different task than we started out with, try and touch the | |
7026 | * least amount of cfs_rqs. | |
7027 | */ | |
7028 | if (prev != p) { | |
7029 | struct sched_entity *pse = &prev->se; | |
7030 | ||
7031 | while (!(cfs_rq = is_same_group(se, pse))) { | |
7032 | int se_depth = se->depth; | |
7033 | int pse_depth = pse->depth; | |
7034 | ||
7035 | if (se_depth <= pse_depth) { | |
7036 | put_prev_entity(cfs_rq_of(pse), pse); | |
7037 | pse = parent_entity(pse); | |
7038 | } | |
7039 | if (se_depth >= pse_depth) { | |
7040 | set_next_entity(cfs_rq_of(se), se); | |
7041 | se = parent_entity(se); | |
7042 | } | |
7043 | } | |
7044 | ||
7045 | put_prev_entity(cfs_rq, pse); | |
7046 | set_next_entity(cfs_rq, se); | |
7047 | } | |
7048 | ||
93824900 | 7049 | goto done; |
678d5718 | 7050 | simple: |
678d5718 | 7051 | #endif |
67692435 PZ |
7052 | if (prev) |
7053 | put_prev_task(rq, prev); | |
606dba2e | 7054 | |
bf0f6f24 | 7055 | do { |
678d5718 | 7056 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 7057 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
7058 | cfs_rq = group_cfs_rq(se); |
7059 | } while (cfs_rq); | |
7060 | ||
8f4d37ec | 7061 | p = task_of(se); |
678d5718 | 7062 | |
13a453c2 | 7063 | done: __maybe_unused; |
93824900 UR |
7064 | #ifdef CONFIG_SMP |
7065 | /* | |
7066 | * Move the next running task to the front of | |
7067 | * the list, so our cfs_tasks list becomes MRU | |
7068 | * one. | |
7069 | */ | |
7070 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
7071 | #endif | |
7072 | ||
b39e66ea MG |
7073 | if (hrtick_enabled(rq)) |
7074 | hrtick_start_fair(rq, p); | |
8f4d37ec | 7075 | |
3b1baa64 MR |
7076 | update_misfit_status(p, rq); |
7077 | ||
8f4d37ec | 7078 | return p; |
38033c37 PZ |
7079 | |
7080 | idle: | |
67692435 PZ |
7081 | if (!rf) |
7082 | return NULL; | |
7083 | ||
5ba553ef | 7084 | new_tasks = newidle_balance(rq, rf); |
46f69fa3 | 7085 | |
37e117c0 | 7086 | /* |
5ba553ef | 7087 | * Because newidle_balance() releases (and re-acquires) rq->lock, it is |
37e117c0 PZ |
7088 | * possible for any higher priority task to appear. In that case we |
7089 | * must re-start the pick_next_entity() loop. | |
7090 | */ | |
e4aa358b | 7091 | if (new_tasks < 0) |
37e117c0 PZ |
7092 | return RETRY_TASK; |
7093 | ||
e4aa358b | 7094 | if (new_tasks > 0) |
38033c37 | 7095 | goto again; |
38033c37 | 7096 | |
23127296 VG |
7097 | /* |
7098 | * rq is about to be idle, check if we need to update the | |
7099 | * lost_idle_time of clock_pelt | |
7100 | */ | |
7101 | update_idle_rq_clock_pelt(rq); | |
7102 | ||
38033c37 | 7103 | return NULL; |
bf0f6f24 IM |
7104 | } |
7105 | ||
98c2f700 PZ |
7106 | static struct task_struct *__pick_next_task_fair(struct rq *rq) |
7107 | { | |
7108 | return pick_next_task_fair(rq, NULL, NULL); | |
7109 | } | |
7110 | ||
bf0f6f24 IM |
7111 | /* |
7112 | * Account for a descheduled task: | |
7113 | */ | |
6e2df058 | 7114 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
7115 | { |
7116 | struct sched_entity *se = &prev->se; | |
7117 | struct cfs_rq *cfs_rq; | |
7118 | ||
7119 | for_each_sched_entity(se) { | |
7120 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 7121 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
7122 | } |
7123 | } | |
7124 | ||
ac53db59 RR |
7125 | /* |
7126 | * sched_yield() is very simple | |
7127 | * | |
7128 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
7129 | */ | |
7130 | static void yield_task_fair(struct rq *rq) | |
7131 | { | |
7132 | struct task_struct *curr = rq->curr; | |
7133 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
7134 | struct sched_entity *se = &curr->se; | |
7135 | ||
7136 | /* | |
7137 | * Are we the only task in the tree? | |
7138 | */ | |
7139 | if (unlikely(rq->nr_running == 1)) | |
7140 | return; | |
7141 | ||
7142 | clear_buddies(cfs_rq, se); | |
7143 | ||
7144 | if (curr->policy != SCHED_BATCH) { | |
7145 | update_rq_clock(rq); | |
7146 | /* | |
7147 | * Update run-time statistics of the 'current'. | |
7148 | */ | |
7149 | update_curr(cfs_rq); | |
916671c0 MG |
7150 | /* |
7151 | * Tell update_rq_clock() that we've just updated, | |
7152 | * so we don't do microscopic update in schedule() | |
7153 | * and double the fastpath cost. | |
7154 | */ | |
adcc8da8 | 7155 | rq_clock_skip_update(rq); |
ac53db59 RR |
7156 | } |
7157 | ||
7158 | set_skip_buddy(se); | |
7159 | } | |
7160 | ||
d95f4122 MG |
7161 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
7162 | { | |
7163 | struct sched_entity *se = &p->se; | |
7164 | ||
5238cdd3 PT |
7165 | /* throttled hierarchies are not runnable */ |
7166 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
7167 | return false; |
7168 | ||
7169 | /* Tell the scheduler that we'd really like pse to run next. */ | |
7170 | set_next_buddy(se); | |
7171 | ||
d95f4122 MG |
7172 | yield_task_fair(rq); |
7173 | ||
7174 | return true; | |
7175 | } | |
7176 | ||
681f3e68 | 7177 | #ifdef CONFIG_SMP |
bf0f6f24 | 7178 | /************************************************** |
e9c84cb8 PZ |
7179 | * Fair scheduling class load-balancing methods. |
7180 | * | |
7181 | * BASICS | |
7182 | * | |
7183 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 7184 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
7185 | * time to each task. This is expressed in the following equation: |
7186 | * | |
7187 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
7188 | * | |
97fb7a0a | 7189 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
7190 | * W_i,0 is defined as: |
7191 | * | |
7192 | * W_i,0 = \Sum_j w_i,j (2) | |
7193 | * | |
97fb7a0a | 7194 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 7195 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
7196 | * |
7197 | * The weight average is an exponential decay average of the instantaneous | |
7198 | * weight: | |
7199 | * | |
7200 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
7201 | * | |
97fb7a0a | 7202 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
7203 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
7204 | * can also include other factors [XXX]. | |
7205 | * | |
7206 | * To achieve this balance we define a measure of imbalance which follows | |
7207 | * directly from (1): | |
7208 | * | |
ced549fa | 7209 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
7210 | * |
7211 | * We them move tasks around to minimize the imbalance. In the continuous | |
7212 | * function space it is obvious this converges, in the discrete case we get | |
7213 | * a few fun cases generally called infeasible weight scenarios. | |
7214 | * | |
7215 | * [XXX expand on: | |
7216 | * - infeasible weights; | |
7217 | * - local vs global optima in the discrete case. ] | |
7218 | * | |
7219 | * | |
7220 | * SCHED DOMAINS | |
7221 | * | |
7222 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 7223 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 7224 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 7225 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 7226 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 7227 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
7228 | * the groups. |
7229 | * | |
7230 | * This yields: | |
7231 | * | |
7232 | * log_2 n 1 n | |
7233 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
7234 | * i = 0 2^i 2^i | |
7235 | * `- size of each group | |
97fb7a0a | 7236 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
7237 | * | `- freq |
7238 | * `- sum over all levels | |
7239 | * | |
7240 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
7241 | * this makes (5) the runtime complexity of the balancer. | |
7242 | * | |
7243 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 7244 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
7245 | * |
7246 | * The adjacency matrix of the resulting graph is given by: | |
7247 | * | |
97a7142f | 7248 | * log_2 n |
e9c84cb8 PZ |
7249 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
7250 | * k = 0 | |
7251 | * | |
7252 | * And you'll find that: | |
7253 | * | |
7254 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
7255 | * | |
97fb7a0a | 7256 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
7257 | * The task movement gives a factor of O(m), giving a convergence complexity |
7258 | * of: | |
7259 | * | |
7260 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
7261 | * | |
7262 | * | |
7263 | * WORK CONSERVING | |
7264 | * | |
7265 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 7266 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
7267 | * tree itself instead of relying on other CPUs to bring it work. |
7268 | * | |
7269 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
7270 | * time. | |
7271 | * | |
7272 | * [XXX more?] | |
7273 | * | |
7274 | * | |
7275 | * CGROUPS | |
7276 | * | |
7277 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
7278 | * | |
7279 | * s_k,i | |
7280 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
7281 | * S_k | |
7282 | * | |
7283 | * Where | |
7284 | * | |
7285 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
7286 | * | |
97fb7a0a | 7287 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
7288 | * |
7289 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
7290 | * property. | |
7291 | * | |
7292 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
7293 | * rewrite all of this once again.] | |
97a7142f | 7294 | */ |
bf0f6f24 | 7295 | |
ed387b78 HS |
7296 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
7297 | ||
0ec8aa00 PZ |
7298 | enum fbq_type { regular, remote, all }; |
7299 | ||
0b0695f2 | 7300 | /* |
a9723389 VG |
7301 | * 'group_type' describes the group of CPUs at the moment of load balancing. |
7302 | * | |
0b0695f2 | 7303 | * The enum is ordered by pulling priority, with the group with lowest priority |
a9723389 VG |
7304 | * first so the group_type can simply be compared when selecting the busiest |
7305 | * group. See update_sd_pick_busiest(). | |
0b0695f2 | 7306 | */ |
3b1baa64 | 7307 | enum group_type { |
a9723389 | 7308 | /* The group has spare capacity that can be used to run more tasks. */ |
0b0695f2 | 7309 | group_has_spare = 0, |
a9723389 VG |
7310 | /* |
7311 | * The group is fully used and the tasks don't compete for more CPU | |
7312 | * cycles. Nevertheless, some tasks might wait before running. | |
7313 | */ | |
0b0695f2 | 7314 | group_fully_busy, |
a9723389 VG |
7315 | /* |
7316 | * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity | |
7317 | * and must be migrated to a more powerful CPU. | |
7318 | */ | |
3b1baa64 | 7319 | group_misfit_task, |
a9723389 VG |
7320 | /* |
7321 | * SD_ASYM_PACKING only: One local CPU with higher capacity is available, | |
7322 | * and the task should be migrated to it instead of running on the | |
7323 | * current CPU. | |
7324 | */ | |
0b0695f2 | 7325 | group_asym_packing, |
a9723389 VG |
7326 | /* |
7327 | * The tasks' affinity constraints previously prevented the scheduler | |
7328 | * from balancing the load across the system. | |
7329 | */ | |
3b1baa64 | 7330 | group_imbalanced, |
a9723389 VG |
7331 | /* |
7332 | * The CPU is overloaded and can't provide expected CPU cycles to all | |
7333 | * tasks. | |
7334 | */ | |
0b0695f2 VG |
7335 | group_overloaded |
7336 | }; | |
7337 | ||
7338 | enum migration_type { | |
7339 | migrate_load = 0, | |
7340 | migrate_util, | |
7341 | migrate_task, | |
7342 | migrate_misfit | |
3b1baa64 MR |
7343 | }; |
7344 | ||
ddcdf6e7 | 7345 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 7346 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
7347 | #define LBF_DST_PINNED 0x04 |
7348 | #define LBF_SOME_PINNED 0x08 | |
e022e0d3 | 7349 | #define LBF_NOHZ_STATS 0x10 |
f643ea22 | 7350 | #define LBF_NOHZ_AGAIN 0x20 |
ddcdf6e7 PZ |
7351 | |
7352 | struct lb_env { | |
7353 | struct sched_domain *sd; | |
7354 | ||
ddcdf6e7 | 7355 | struct rq *src_rq; |
85c1e7da | 7356 | int src_cpu; |
ddcdf6e7 PZ |
7357 | |
7358 | int dst_cpu; | |
7359 | struct rq *dst_rq; | |
7360 | ||
88b8dac0 SV |
7361 | struct cpumask *dst_grpmask; |
7362 | int new_dst_cpu; | |
ddcdf6e7 | 7363 | enum cpu_idle_type idle; |
bd939f45 | 7364 | long imbalance; |
b9403130 MW |
7365 | /* The set of CPUs under consideration for load-balancing */ |
7366 | struct cpumask *cpus; | |
7367 | ||
ddcdf6e7 | 7368 | unsigned int flags; |
367456c7 PZ |
7369 | |
7370 | unsigned int loop; | |
7371 | unsigned int loop_break; | |
7372 | unsigned int loop_max; | |
0ec8aa00 PZ |
7373 | |
7374 | enum fbq_type fbq_type; | |
0b0695f2 | 7375 | enum migration_type migration_type; |
163122b7 | 7376 | struct list_head tasks; |
ddcdf6e7 PZ |
7377 | }; |
7378 | ||
029632fb PZ |
7379 | /* |
7380 | * Is this task likely cache-hot: | |
7381 | */ | |
5d5e2b1b | 7382 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
7383 | { |
7384 | s64 delta; | |
7385 | ||
e5673f28 KT |
7386 | lockdep_assert_held(&env->src_rq->lock); |
7387 | ||
029632fb PZ |
7388 | if (p->sched_class != &fair_sched_class) |
7389 | return 0; | |
7390 | ||
1da1843f | 7391 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
7392 | return 0; |
7393 | ||
7394 | /* | |
7395 | * Buddy candidates are cache hot: | |
7396 | */ | |
5d5e2b1b | 7397 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
7398 | (&p->se == cfs_rq_of(&p->se)->next || |
7399 | &p->se == cfs_rq_of(&p->se)->last)) | |
7400 | return 1; | |
7401 | ||
7402 | if (sysctl_sched_migration_cost == -1) | |
7403 | return 1; | |
7404 | if (sysctl_sched_migration_cost == 0) | |
7405 | return 0; | |
7406 | ||
5d5e2b1b | 7407 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
7408 | |
7409 | return delta < (s64)sysctl_sched_migration_cost; | |
7410 | } | |
7411 | ||
3a7053b3 | 7412 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 7413 | /* |
2a1ed24c SD |
7414 | * Returns 1, if task migration degrades locality |
7415 | * Returns 0, if task migration improves locality i.e migration preferred. | |
7416 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 7417 | */ |
2a1ed24c | 7418 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 7419 | { |
b1ad065e | 7420 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
7421 | unsigned long src_weight, dst_weight; |
7422 | int src_nid, dst_nid, dist; | |
3a7053b3 | 7423 | |
2a595721 | 7424 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
7425 | return -1; |
7426 | ||
c3b9bc5b | 7427 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 7428 | return -1; |
7a0f3083 MG |
7429 | |
7430 | src_nid = cpu_to_node(env->src_cpu); | |
7431 | dst_nid = cpu_to_node(env->dst_cpu); | |
7432 | ||
83e1d2cd | 7433 | if (src_nid == dst_nid) |
2a1ed24c | 7434 | return -1; |
7a0f3083 | 7435 | |
2a1ed24c SD |
7436 | /* Migrating away from the preferred node is always bad. */ |
7437 | if (src_nid == p->numa_preferred_nid) { | |
7438 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
7439 | return 1; | |
7440 | else | |
7441 | return -1; | |
7442 | } | |
b1ad065e | 7443 | |
c1ceac62 RR |
7444 | /* Encourage migration to the preferred node. */ |
7445 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 7446 | return 0; |
b1ad065e | 7447 | |
739294fb | 7448 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 7449 | if (env->idle == CPU_IDLE) |
739294fb RR |
7450 | return -1; |
7451 | ||
f35678b6 | 7452 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 7453 | if (numa_group) { |
f35678b6 SD |
7454 | src_weight = group_weight(p, src_nid, dist); |
7455 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 7456 | } else { |
f35678b6 SD |
7457 | src_weight = task_weight(p, src_nid, dist); |
7458 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
7459 | } |
7460 | ||
f35678b6 | 7461 | return dst_weight < src_weight; |
7a0f3083 MG |
7462 | } |
7463 | ||
3a7053b3 | 7464 | #else |
2a1ed24c | 7465 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
7466 | struct lb_env *env) |
7467 | { | |
2a1ed24c | 7468 | return -1; |
7a0f3083 | 7469 | } |
3a7053b3 MG |
7470 | #endif |
7471 | ||
1e3c88bd PZ |
7472 | /* |
7473 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
7474 | */ | |
7475 | static | |
8e45cb54 | 7476 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 7477 | { |
2a1ed24c | 7478 | int tsk_cache_hot; |
e5673f28 KT |
7479 | |
7480 | lockdep_assert_held(&env->src_rq->lock); | |
7481 | ||
1e3c88bd PZ |
7482 | /* |
7483 | * We do not migrate tasks that are: | |
d3198084 | 7484 | * 1) throttled_lb_pair, or |
3bd37062 | 7485 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
7486 | * 3) running (obviously), or |
7487 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7488 | */ |
d3198084 JK |
7489 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7490 | return 0; | |
7491 | ||
3bd37062 | 7492 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 7493 | int cpu; |
88b8dac0 | 7494 | |
ae92882e | 7495 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 7496 | |
6263322c PZ |
7497 | env->flags |= LBF_SOME_PINNED; |
7498 | ||
88b8dac0 | 7499 | /* |
97fb7a0a | 7500 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
7501 | * our sched_group. We may want to revisit it if we couldn't |
7502 | * meet load balance goals by pulling other tasks on src_cpu. | |
7503 | * | |
65a4433a JH |
7504 | * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have |
7505 | * already computed one in current iteration. | |
88b8dac0 | 7506 | */ |
65a4433a | 7507 | if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
7508 | return 0; |
7509 | ||
97fb7a0a | 7510 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 7511 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 7512 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 7513 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7514 | env->new_dst_cpu = cpu; |
7515 | break; | |
7516 | } | |
88b8dac0 | 7517 | } |
e02e60c1 | 7518 | |
1e3c88bd PZ |
7519 | return 0; |
7520 | } | |
88b8dac0 SV |
7521 | |
7522 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 7523 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7524 | |
ddcdf6e7 | 7525 | if (task_running(env->src_rq, p)) { |
ae92882e | 7526 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
7527 | return 0; |
7528 | } | |
7529 | ||
7530 | /* | |
7531 | * Aggressive migration if: | |
3a7053b3 MG |
7532 | * 1) destination numa is preferred |
7533 | * 2) task is cache cold, or | |
7534 | * 3) too many balance attempts have failed. | |
1e3c88bd | 7535 | */ |
2a1ed24c SD |
7536 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7537 | if (tsk_cache_hot == -1) | |
7538 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7539 | |
2a1ed24c | 7540 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7541 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7542 | if (tsk_cache_hot == 1) { |
ae92882e JP |
7543 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
7544 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 7545 | } |
1e3c88bd PZ |
7546 | return 1; |
7547 | } | |
7548 | ||
ae92882e | 7549 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 7550 | return 0; |
1e3c88bd PZ |
7551 | } |
7552 | ||
897c395f | 7553 | /* |
163122b7 KT |
7554 | * detach_task() -- detach the task for the migration specified in env |
7555 | */ | |
7556 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7557 | { | |
7558 | lockdep_assert_held(&env->src_rq->lock); | |
7559 | ||
5704ac0a | 7560 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7561 | set_task_cpu(p, env->dst_cpu); |
7562 | } | |
7563 | ||
897c395f | 7564 | /* |
e5673f28 | 7565 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7566 | * part of active balancing operations within "domain". |
897c395f | 7567 | * |
e5673f28 | 7568 | * Returns a task if successful and NULL otherwise. |
897c395f | 7569 | */ |
e5673f28 | 7570 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7571 | { |
93824900 | 7572 | struct task_struct *p; |
897c395f | 7573 | |
e5673f28 KT |
7574 | lockdep_assert_held(&env->src_rq->lock); |
7575 | ||
93824900 UR |
7576 | list_for_each_entry_reverse(p, |
7577 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7578 | if (!can_migrate_task(p, env)) |
7579 | continue; | |
897c395f | 7580 | |
163122b7 | 7581 | detach_task(p, env); |
e5673f28 | 7582 | |
367456c7 | 7583 | /* |
e5673f28 | 7584 | * Right now, this is only the second place where |
163122b7 | 7585 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 7586 | * so we can safely collect stats here rather than |
163122b7 | 7587 | * inside detach_tasks(). |
367456c7 | 7588 | */ |
ae92882e | 7589 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 7590 | return p; |
897c395f | 7591 | } |
e5673f28 | 7592 | return NULL; |
897c395f PZ |
7593 | } |
7594 | ||
eb95308e PZ |
7595 | static const unsigned int sched_nr_migrate_break = 32; |
7596 | ||
5d6523eb | 7597 | /* |
0b0695f2 | 7598 | * detach_tasks() -- tries to detach up to imbalance load/util/tasks from |
163122b7 | 7599 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 7600 | * |
163122b7 | 7601 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 7602 | */ |
163122b7 | 7603 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 7604 | { |
5d6523eb | 7605 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
0b0695f2 | 7606 | unsigned long util, load; |
5d6523eb | 7607 | struct task_struct *p; |
163122b7 KT |
7608 | int detached = 0; |
7609 | ||
7610 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 7611 | |
bd939f45 | 7612 | if (env->imbalance <= 0) |
5d6523eb | 7613 | return 0; |
1e3c88bd | 7614 | |
5d6523eb | 7615 | while (!list_empty(tasks)) { |
985d3a4c YD |
7616 | /* |
7617 | * We don't want to steal all, otherwise we may be treated likewise, | |
7618 | * which could at worst lead to a livelock crash. | |
7619 | */ | |
7620 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
7621 | break; | |
7622 | ||
93824900 | 7623 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 7624 | |
367456c7 PZ |
7625 | env->loop++; |
7626 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 7627 | if (env->loop > env->loop_max) |
367456c7 | 7628 | break; |
5d6523eb PZ |
7629 | |
7630 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 7631 | if (env->loop > env->loop_break) { |
eb95308e | 7632 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 7633 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 7634 | break; |
a195f004 | 7635 | } |
1e3c88bd | 7636 | |
d3198084 | 7637 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
7638 | goto next; |
7639 | ||
0b0695f2 VG |
7640 | switch (env->migration_type) { |
7641 | case migrate_load: | |
7642 | load = task_h_load(p); | |
5d6523eb | 7643 | |
0b0695f2 VG |
7644 | if (sched_feat(LB_MIN) && |
7645 | load < 16 && !env->sd->nr_balance_failed) | |
7646 | goto next; | |
367456c7 | 7647 | |
6cf82d55 VG |
7648 | /* |
7649 | * Make sure that we don't migrate too much load. | |
7650 | * Nevertheless, let relax the constraint if | |
7651 | * scheduler fails to find a good waiting task to | |
7652 | * migrate. | |
7653 | */ | |
7654 | if (load/2 > env->imbalance && | |
7655 | env->sd->nr_balance_failed <= env->sd->cache_nice_tries) | |
0b0695f2 VG |
7656 | goto next; |
7657 | ||
7658 | env->imbalance -= load; | |
7659 | break; | |
7660 | ||
7661 | case migrate_util: | |
7662 | util = task_util_est(p); | |
7663 | ||
7664 | if (util > env->imbalance) | |
7665 | goto next; | |
7666 | ||
7667 | env->imbalance -= util; | |
7668 | break; | |
7669 | ||
7670 | case migrate_task: | |
7671 | env->imbalance--; | |
7672 | break; | |
7673 | ||
7674 | case migrate_misfit: | |
c63be7be VG |
7675 | /* This is not a misfit task */ |
7676 | if (task_fits_capacity(p, capacity_of(env->src_cpu))) | |
0b0695f2 VG |
7677 | goto next; |
7678 | ||
7679 | env->imbalance = 0; | |
7680 | break; | |
7681 | } | |
1e3c88bd | 7682 | |
163122b7 KT |
7683 | detach_task(p, env); |
7684 | list_add(&p->se.group_node, &env->tasks); | |
7685 | ||
7686 | detached++; | |
1e3c88bd | 7687 | |
c1a280b6 | 7688 | #ifdef CONFIG_PREEMPTION |
ee00e66f PZ |
7689 | /* |
7690 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 7691 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
7692 | * the critical section. |
7693 | */ | |
5d6523eb | 7694 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 7695 | break; |
1e3c88bd PZ |
7696 | #endif |
7697 | ||
ee00e66f PZ |
7698 | /* |
7699 | * We only want to steal up to the prescribed amount of | |
0b0695f2 | 7700 | * load/util/tasks. |
ee00e66f | 7701 | */ |
bd939f45 | 7702 | if (env->imbalance <= 0) |
ee00e66f | 7703 | break; |
367456c7 PZ |
7704 | |
7705 | continue; | |
7706 | next: | |
93824900 | 7707 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 7708 | } |
5d6523eb | 7709 | |
1e3c88bd | 7710 | /* |
163122b7 KT |
7711 | * Right now, this is one of only two places we collect this stat |
7712 | * so we can safely collect detach_one_task() stats here rather | |
7713 | * than inside detach_one_task(). | |
1e3c88bd | 7714 | */ |
ae92882e | 7715 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 7716 | |
163122b7 KT |
7717 | return detached; |
7718 | } | |
7719 | ||
7720 | /* | |
7721 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
7722 | */ | |
7723 | static void attach_task(struct rq *rq, struct task_struct *p) | |
7724 | { | |
7725 | lockdep_assert_held(&rq->lock); | |
7726 | ||
7727 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 7728 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
163122b7 KT |
7729 | check_preempt_curr(rq, p, 0); |
7730 | } | |
7731 | ||
7732 | /* | |
7733 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
7734 | * its new rq. | |
7735 | */ | |
7736 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
7737 | { | |
8a8c69c3 PZ |
7738 | struct rq_flags rf; |
7739 | ||
7740 | rq_lock(rq, &rf); | |
5704ac0a | 7741 | update_rq_clock(rq); |
163122b7 | 7742 | attach_task(rq, p); |
8a8c69c3 | 7743 | rq_unlock(rq, &rf); |
163122b7 KT |
7744 | } |
7745 | ||
7746 | /* | |
7747 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
7748 | * new rq. | |
7749 | */ | |
7750 | static void attach_tasks(struct lb_env *env) | |
7751 | { | |
7752 | struct list_head *tasks = &env->tasks; | |
7753 | struct task_struct *p; | |
8a8c69c3 | 7754 | struct rq_flags rf; |
163122b7 | 7755 | |
8a8c69c3 | 7756 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 7757 | update_rq_clock(env->dst_rq); |
163122b7 KT |
7758 | |
7759 | while (!list_empty(tasks)) { | |
7760 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
7761 | list_del_init(&p->se.group_node); | |
1e3c88bd | 7762 | |
163122b7 KT |
7763 | attach_task(env->dst_rq, p); |
7764 | } | |
7765 | ||
8a8c69c3 | 7766 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
7767 | } |
7768 | ||
b0c79224 | 7769 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
7770 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
7771 | { | |
7772 | if (cfs_rq->avg.load_avg) | |
7773 | return true; | |
7774 | ||
7775 | if (cfs_rq->avg.util_avg) | |
7776 | return true; | |
7777 | ||
7778 | return false; | |
7779 | } | |
7780 | ||
91c27493 | 7781 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
7782 | { |
7783 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
7784 | return true; | |
7785 | ||
3727e0e1 VG |
7786 | if (READ_ONCE(rq->avg_dl.util_avg)) |
7787 | return true; | |
7788 | ||
b4eccf5f TG |
7789 | if (thermal_load_avg(rq)) |
7790 | return true; | |
7791 | ||
11d4afd4 | 7792 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
7793 | if (READ_ONCE(rq->avg_irq.util_avg)) |
7794 | return true; | |
7795 | #endif | |
7796 | ||
371bf427 VG |
7797 | return false; |
7798 | } | |
7799 | ||
b0c79224 VS |
7800 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
7801 | { | |
7802 | rq->last_blocked_load_update_tick = jiffies; | |
7803 | ||
7804 | if (!has_blocked) | |
7805 | rq->has_blocked_load = 0; | |
7806 | } | |
7807 | #else | |
7808 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
7809 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
7810 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} | |
7811 | #endif | |
7812 | ||
bef69dd8 VG |
7813 | static bool __update_blocked_others(struct rq *rq, bool *done) |
7814 | { | |
7815 | const struct sched_class *curr_class; | |
7816 | u64 now = rq_clock_pelt(rq); | |
b4eccf5f | 7817 | unsigned long thermal_pressure; |
bef69dd8 VG |
7818 | bool decayed; |
7819 | ||
7820 | /* | |
7821 | * update_load_avg() can call cpufreq_update_util(). Make sure that RT, | |
7822 | * DL and IRQ signals have been updated before updating CFS. | |
7823 | */ | |
7824 | curr_class = rq->curr->sched_class; | |
7825 | ||
b4eccf5f TG |
7826 | thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); |
7827 | ||
bef69dd8 VG |
7828 | decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | |
7829 | update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | | |
05289b90 | 7830 | update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | |
bef69dd8 VG |
7831 | update_irq_load_avg(rq, 0); |
7832 | ||
7833 | if (others_have_blocked(rq)) | |
7834 | *done = false; | |
7835 | ||
7836 | return decayed; | |
7837 | } | |
7838 | ||
1936c53c VG |
7839 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7840 | ||
039ae8bc VG |
7841 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
7842 | { | |
7843 | if (cfs_rq->load.weight) | |
7844 | return false; | |
7845 | ||
7846 | if (cfs_rq->avg.load_sum) | |
7847 | return false; | |
7848 | ||
7849 | if (cfs_rq->avg.util_sum) | |
7850 | return false; | |
7851 | ||
9f683953 VG |
7852 | if (cfs_rq->avg.runnable_sum) |
7853 | return false; | |
7854 | ||
039ae8bc VG |
7855 | return true; |
7856 | } | |
7857 | ||
bef69dd8 | 7858 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 7859 | { |
039ae8bc | 7860 | struct cfs_rq *cfs_rq, *pos; |
bef69dd8 VG |
7861 | bool decayed = false; |
7862 | int cpu = cpu_of(rq); | |
b90f7c9d | 7863 | |
9763b67f PZ |
7864 | /* |
7865 | * Iterates the task_group tree in a bottom up fashion, see | |
7866 | * list_add_leaf_cfs_rq() for details. | |
7867 | */ | |
039ae8bc | 7868 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
7869 | struct sched_entity *se; |
7870 | ||
bef69dd8 | 7871 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { |
9d89c257 | 7872 | update_tg_load_avg(cfs_rq, 0); |
4e516076 | 7873 | |
bef69dd8 VG |
7874 | if (cfs_rq == &rq->cfs) |
7875 | decayed = true; | |
7876 | } | |
7877 | ||
bc427898 VG |
7878 | /* Propagate pending load changes to the parent, if any: */ |
7879 | se = cfs_rq->tg->se[cpu]; | |
7880 | if (se && !skip_blocked_update(se)) | |
88c0616e | 7881 | update_load_avg(cfs_rq_of(se), se, 0); |
a9e7f654 | 7882 | |
039ae8bc VG |
7883 | /* |
7884 | * There can be a lot of idle CPU cgroups. Don't let fully | |
7885 | * decayed cfs_rqs linger on the list. | |
7886 | */ | |
7887 | if (cfs_rq_is_decayed(cfs_rq)) | |
7888 | list_del_leaf_cfs_rq(cfs_rq); | |
7889 | ||
1936c53c VG |
7890 | /* Don't need periodic decay once load/util_avg are null */ |
7891 | if (cfs_rq_has_blocked(cfs_rq)) | |
bef69dd8 | 7892 | *done = false; |
9d89c257 | 7893 | } |
12b04875 | 7894 | |
bef69dd8 | 7895 | return decayed; |
9e3081ca PZ |
7896 | } |
7897 | ||
9763b67f | 7898 | /* |
68520796 | 7899 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
7900 | * This needs to be done in a top-down fashion because the load of a child |
7901 | * group is a fraction of its parents load. | |
7902 | */ | |
68520796 | 7903 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 7904 | { |
68520796 VD |
7905 | struct rq *rq = rq_of(cfs_rq); |
7906 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 7907 | unsigned long now = jiffies; |
68520796 | 7908 | unsigned long load; |
a35b6466 | 7909 | |
68520796 | 7910 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
7911 | return; |
7912 | ||
0e9f0245 | 7913 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
7914 | for_each_sched_entity(se) { |
7915 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 7916 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
7917 | if (cfs_rq->last_h_load_update == now) |
7918 | break; | |
7919 | } | |
a35b6466 | 7920 | |
68520796 | 7921 | if (!se) { |
7ea241af | 7922 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
7923 | cfs_rq->last_h_load_update = now; |
7924 | } | |
7925 | ||
0e9f0245 | 7926 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 7927 | load = cfs_rq->h_load; |
7ea241af YD |
7928 | load = div64_ul(load * se->avg.load_avg, |
7929 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
7930 | cfs_rq = group_cfs_rq(se); |
7931 | cfs_rq->h_load = load; | |
7932 | cfs_rq->last_h_load_update = now; | |
7933 | } | |
9763b67f PZ |
7934 | } |
7935 | ||
367456c7 | 7936 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 7937 | { |
367456c7 | 7938 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 7939 | |
68520796 | 7940 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 7941 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 7942 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
7943 | } |
7944 | #else | |
bef69dd8 | 7945 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 7946 | { |
6c1d47c0 | 7947 | struct cfs_rq *cfs_rq = &rq->cfs; |
bef69dd8 | 7948 | bool decayed; |
b90f7c9d | 7949 | |
bef69dd8 VG |
7950 | decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
7951 | if (cfs_rq_has_blocked(cfs_rq)) | |
7952 | *done = false; | |
b90f7c9d | 7953 | |
bef69dd8 | 7954 | return decayed; |
9e3081ca PZ |
7955 | } |
7956 | ||
367456c7 | 7957 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 7958 | { |
9d89c257 | 7959 | return p->se.avg.load_avg; |
1e3c88bd | 7960 | } |
230059de | 7961 | #endif |
1e3c88bd | 7962 | |
bef69dd8 VG |
7963 | static void update_blocked_averages(int cpu) |
7964 | { | |
7965 | bool decayed = false, done = true; | |
7966 | struct rq *rq = cpu_rq(cpu); | |
7967 | struct rq_flags rf; | |
7968 | ||
7969 | rq_lock_irqsave(rq, &rf); | |
7970 | update_rq_clock(rq); | |
7971 | ||
7972 | decayed |= __update_blocked_others(rq, &done); | |
7973 | decayed |= __update_blocked_fair(rq, &done); | |
7974 | ||
7975 | update_blocked_load_status(rq, !done); | |
7976 | if (decayed) | |
7977 | cpufreq_update_util(rq, 0); | |
7978 | rq_unlock_irqrestore(rq, &rf); | |
7979 | } | |
7980 | ||
1e3c88bd | 7981 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 7982 | |
1e3c88bd PZ |
7983 | /* |
7984 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
7985 | */ | |
7986 | struct sg_lb_stats { | |
7987 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
7988 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
63b2ca30 | 7989 | unsigned long group_capacity; |
070f5e86 VG |
7990 | unsigned long group_util; /* Total utilization over the CPUs of the group */ |
7991 | unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ | |
5e23e474 | 7992 | unsigned int sum_nr_running; /* Nr of tasks running in the group */ |
a3498347 | 7993 | unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ |
147c5fc2 PZ |
7994 | unsigned int idle_cpus; |
7995 | unsigned int group_weight; | |
caeb178c | 7996 | enum group_type group_type; |
490ba971 | 7997 | unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ |
3b1baa64 | 7998 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
7999 | #ifdef CONFIG_NUMA_BALANCING |
8000 | unsigned int nr_numa_running; | |
8001 | unsigned int nr_preferred_running; | |
8002 | #endif | |
1e3c88bd PZ |
8003 | }; |
8004 | ||
56cf515b JK |
8005 | /* |
8006 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
8007 | * during load balancing. | |
8008 | */ | |
8009 | struct sd_lb_stats { | |
8010 | struct sched_group *busiest; /* Busiest group in this sd */ | |
8011 | struct sched_group *local; /* Local group in this sd */ | |
8012 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 8013 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b | 8014 | unsigned long avg_load; /* Average load across all groups in sd */ |
0b0695f2 | 8015 | unsigned int prefer_sibling; /* tasks should go to sibling first */ |
56cf515b | 8016 | |
56cf515b | 8017 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 8018 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
8019 | }; |
8020 | ||
147c5fc2 PZ |
8021 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
8022 | { | |
8023 | /* | |
8024 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
8025 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
0b0695f2 VG |
8026 | * We must however set busiest_stat::group_type and |
8027 | * busiest_stat::idle_cpus to the worst busiest group because | |
8028 | * update_sd_pick_busiest() reads these before assignment. | |
147c5fc2 PZ |
8029 | */ |
8030 | *sds = (struct sd_lb_stats){ | |
8031 | .busiest = NULL, | |
8032 | .local = NULL, | |
8033 | .total_load = 0UL, | |
63b2ca30 | 8034 | .total_capacity = 0UL, |
147c5fc2 | 8035 | .busiest_stat = { |
0b0695f2 VG |
8036 | .idle_cpus = UINT_MAX, |
8037 | .group_type = group_has_spare, | |
147c5fc2 PZ |
8038 | }, |
8039 | }; | |
8040 | } | |
8041 | ||
287cdaac | 8042 | static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
8043 | { |
8044 | struct rq *rq = cpu_rq(cpu); | |
8ec59c0f | 8045 | unsigned long max = arch_scale_cpu_capacity(cpu); |
523e979d | 8046 | unsigned long used, free; |
523e979d | 8047 | unsigned long irq; |
b654f7de | 8048 | |
2e62c474 | 8049 | irq = cpu_util_irq(rq); |
cadefd3d | 8050 | |
523e979d VG |
8051 | if (unlikely(irq >= max)) |
8052 | return 1; | |
aa483808 | 8053 | |
467b7d01 TG |
8054 | /* |
8055 | * avg_rt.util_avg and avg_dl.util_avg track binary signals | |
8056 | * (running and not running) with weights 0 and 1024 respectively. | |
8057 | * avg_thermal.load_avg tracks thermal pressure and the weighted | |
8058 | * average uses the actual delta max capacity(load). | |
8059 | */ | |
523e979d VG |
8060 | used = READ_ONCE(rq->avg_rt.util_avg); |
8061 | used += READ_ONCE(rq->avg_dl.util_avg); | |
467b7d01 | 8062 | used += thermal_load_avg(rq); |
1e3c88bd | 8063 | |
523e979d VG |
8064 | if (unlikely(used >= max)) |
8065 | return 1; | |
1e3c88bd | 8066 | |
523e979d | 8067 | free = max - used; |
2e62c474 VG |
8068 | |
8069 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
8070 | } |
8071 | ||
ced549fa | 8072 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 8073 | { |
287cdaac | 8074 | unsigned long capacity = scale_rt_capacity(sd, cpu); |
1e3c88bd PZ |
8075 | struct sched_group *sdg = sd->groups; |
8076 | ||
8ec59c0f | 8077 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); |
1e3c88bd | 8078 | |
ced549fa NP |
8079 | if (!capacity) |
8080 | capacity = 1; | |
1e3c88bd | 8081 | |
ced549fa NP |
8082 | cpu_rq(cpu)->cpu_capacity = capacity; |
8083 | sdg->sgc->capacity = capacity; | |
bf475ce0 | 8084 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 8085 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
8086 | } |
8087 | ||
63b2ca30 | 8088 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
8089 | { |
8090 | struct sched_domain *child = sd->child; | |
8091 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 8092 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
8093 | unsigned long interval; |
8094 | ||
8095 | interval = msecs_to_jiffies(sd->balance_interval); | |
8096 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 8097 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
8098 | |
8099 | if (!child) { | |
ced549fa | 8100 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
8101 | return; |
8102 | } | |
8103 | ||
dc7ff76e | 8104 | capacity = 0; |
bf475ce0 | 8105 | min_capacity = ULONG_MAX; |
e3d6d0cb | 8106 | max_capacity = 0; |
1e3c88bd | 8107 | |
74a5ce20 PZ |
8108 | if (child->flags & SD_OVERLAP) { |
8109 | /* | |
8110 | * SD_OVERLAP domains cannot assume that child groups | |
8111 | * span the current group. | |
8112 | */ | |
8113 | ||
ae4df9d6 | 8114 | for_each_cpu(cpu, sched_group_span(sdg)) { |
4c58f57f | 8115 | unsigned long cpu_cap = capacity_of(cpu); |
863bffc8 | 8116 | |
4c58f57f PL |
8117 | capacity += cpu_cap; |
8118 | min_capacity = min(cpu_cap, min_capacity); | |
8119 | max_capacity = max(cpu_cap, max_capacity); | |
863bffc8 | 8120 | } |
74a5ce20 PZ |
8121 | } else { |
8122 | /* | |
8123 | * !SD_OVERLAP domains can assume that child groups | |
8124 | * span the current group. | |
97a7142f | 8125 | */ |
74a5ce20 PZ |
8126 | |
8127 | group = child->groups; | |
8128 | do { | |
bf475ce0 MR |
8129 | struct sched_group_capacity *sgc = group->sgc; |
8130 | ||
8131 | capacity += sgc->capacity; | |
8132 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 8133 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
8134 | group = group->next; |
8135 | } while (group != child->groups); | |
8136 | } | |
1e3c88bd | 8137 | |
63b2ca30 | 8138 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8139 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 8140 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
8141 | } |
8142 | ||
9d5efe05 | 8143 | /* |
ea67821b VG |
8144 | * Check whether the capacity of the rq has been noticeably reduced by side |
8145 | * activity. The imbalance_pct is used for the threshold. | |
8146 | * Return true is the capacity is reduced | |
9d5efe05 SV |
8147 | */ |
8148 | static inline int | |
ea67821b | 8149 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 8150 | { |
ea67821b VG |
8151 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
8152 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
8153 | } |
8154 | ||
a0fe2cf0 VS |
8155 | /* |
8156 | * Check whether a rq has a misfit task and if it looks like we can actually | |
8157 | * help that task: we can migrate the task to a CPU of higher capacity, or | |
8158 | * the task's current CPU is heavily pressured. | |
8159 | */ | |
8160 | static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) | |
8161 | { | |
8162 | return rq->misfit_task_load && | |
8163 | (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || | |
8164 | check_cpu_capacity(rq, sd)); | |
8165 | } | |
8166 | ||
30ce5dab PZ |
8167 | /* |
8168 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 8169 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 8170 | * |
97fb7a0a IM |
8171 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
8172 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
8173 | * Something like: |
8174 | * | |
2b4d5b25 IM |
8175 | * { 0 1 2 3 } { 4 5 6 7 } |
8176 | * * * * * | |
30ce5dab PZ |
8177 | * |
8178 | * If we were to balance group-wise we'd place two tasks in the first group and | |
8179 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 8180 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
8181 | * |
8182 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
8183 | * by noticing the lower domain failed to reach balance and had difficulty |
8184 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
8185 | * |
8186 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 8187 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 8188 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
8189 | * to create an effective group imbalance. |
8190 | * | |
8191 | * This is a somewhat tricky proposition since the next run might not find the | |
8192 | * group imbalance and decide the groups need to be balanced again. A most | |
8193 | * subtle and fragile situation. | |
8194 | */ | |
8195 | ||
6263322c | 8196 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 8197 | { |
63b2ca30 | 8198 | return group->sgc->imbalance; |
30ce5dab PZ |
8199 | } |
8200 | ||
b37d9316 | 8201 | /* |
ea67821b VG |
8202 | * group_has_capacity returns true if the group has spare capacity that could |
8203 | * be used by some tasks. | |
8204 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
8205 | * smaller than the number of CPUs or if the utilization is lower than the |
8206 | * available capacity for CFS tasks. | |
ea67821b VG |
8207 | * For the latter, we use a threshold to stabilize the state, to take into |
8208 | * account the variance of the tasks' load and to return true if the available | |
8209 | * capacity in meaningful for the load balancer. | |
8210 | * As an example, an available capacity of 1% can appear but it doesn't make | |
8211 | * any benefit for the load balance. | |
b37d9316 | 8212 | */ |
ea67821b | 8213 | static inline bool |
57abff06 | 8214 | group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
b37d9316 | 8215 | { |
5e23e474 | 8216 | if (sgs->sum_nr_running < sgs->group_weight) |
ea67821b | 8217 | return true; |
c61037e9 | 8218 | |
070f5e86 VG |
8219 | if ((sgs->group_capacity * imbalance_pct) < |
8220 | (sgs->group_runnable * 100)) | |
8221 | return false; | |
8222 | ||
ea67821b | 8223 | if ((sgs->group_capacity * 100) > |
57abff06 | 8224 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8225 | return true; |
b37d9316 | 8226 | |
ea67821b VG |
8227 | return false; |
8228 | } | |
8229 | ||
8230 | /* | |
8231 | * group_is_overloaded returns true if the group has more tasks than it can | |
8232 | * handle. | |
8233 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
8234 | * with the exact right number of tasks, has no more spare capacity but is not | |
8235 | * overloaded so both group_has_capacity and group_is_overloaded return | |
8236 | * false. | |
8237 | */ | |
8238 | static inline bool | |
57abff06 | 8239 | group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
ea67821b | 8240 | { |
5e23e474 | 8241 | if (sgs->sum_nr_running <= sgs->group_weight) |
ea67821b | 8242 | return false; |
b37d9316 | 8243 | |
ea67821b | 8244 | if ((sgs->group_capacity * 100) < |
57abff06 | 8245 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8246 | return true; |
b37d9316 | 8247 | |
070f5e86 VG |
8248 | if ((sgs->group_capacity * imbalance_pct) < |
8249 | (sgs->group_runnable * 100)) | |
8250 | return true; | |
8251 | ||
ea67821b | 8252 | return false; |
b37d9316 PZ |
8253 | } |
8254 | ||
9e0994c0 | 8255 | /* |
e3d6d0cb | 8256 | * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller |
9e0994c0 MR |
8257 | * per-CPU capacity than sched_group ref. |
8258 | */ | |
8259 | static inline bool | |
e3d6d0cb | 8260 | group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref) |
9e0994c0 | 8261 | { |
60e17f5c | 8262 | return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity); |
9e0994c0 MR |
8263 | } |
8264 | ||
e3d6d0cb MR |
8265 | /* |
8266 | * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller | |
8267 | * per-CPU capacity_orig than sched_group ref. | |
8268 | */ | |
8269 | static inline bool | |
8270 | group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref) | |
8271 | { | |
60e17f5c | 8272 | return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity); |
e3d6d0cb MR |
8273 | } |
8274 | ||
79a89f92 | 8275 | static inline enum |
57abff06 | 8276 | group_type group_classify(unsigned int imbalance_pct, |
0b0695f2 | 8277 | struct sched_group *group, |
79a89f92 | 8278 | struct sg_lb_stats *sgs) |
caeb178c | 8279 | { |
57abff06 | 8280 | if (group_is_overloaded(imbalance_pct, sgs)) |
caeb178c RR |
8281 | return group_overloaded; |
8282 | ||
8283 | if (sg_imbalanced(group)) | |
8284 | return group_imbalanced; | |
8285 | ||
0b0695f2 VG |
8286 | if (sgs->group_asym_packing) |
8287 | return group_asym_packing; | |
8288 | ||
3b1baa64 MR |
8289 | if (sgs->group_misfit_task_load) |
8290 | return group_misfit_task; | |
8291 | ||
57abff06 | 8292 | if (!group_has_capacity(imbalance_pct, sgs)) |
0b0695f2 VG |
8293 | return group_fully_busy; |
8294 | ||
8295 | return group_has_spare; | |
caeb178c RR |
8296 | } |
8297 | ||
63928384 | 8298 | static bool update_nohz_stats(struct rq *rq, bool force) |
e022e0d3 PZ |
8299 | { |
8300 | #ifdef CONFIG_NO_HZ_COMMON | |
8301 | unsigned int cpu = rq->cpu; | |
8302 | ||
f643ea22 VG |
8303 | if (!rq->has_blocked_load) |
8304 | return false; | |
8305 | ||
e022e0d3 | 8306 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
f643ea22 | 8307 | return false; |
e022e0d3 | 8308 | |
63928384 | 8309 | if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick)) |
f643ea22 | 8310 | return true; |
e022e0d3 PZ |
8311 | |
8312 | update_blocked_averages(cpu); | |
f643ea22 VG |
8313 | |
8314 | return rq->has_blocked_load; | |
8315 | #else | |
8316 | return false; | |
e022e0d3 PZ |
8317 | #endif |
8318 | } | |
8319 | ||
1e3c88bd PZ |
8320 | /** |
8321 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 8322 | * @env: The load balancing environment. |
1e3c88bd | 8323 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 8324 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 8325 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 8326 | */ |
bd939f45 | 8327 | static inline void update_sg_lb_stats(struct lb_env *env, |
630246a0 QP |
8328 | struct sched_group *group, |
8329 | struct sg_lb_stats *sgs, | |
8330 | int *sg_status) | |
1e3c88bd | 8331 | { |
0b0695f2 | 8332 | int i, nr_running, local_group; |
1e3c88bd | 8333 | |
b72ff13c PZ |
8334 | memset(sgs, 0, sizeof(*sgs)); |
8335 | ||
0b0695f2 VG |
8336 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group)); |
8337 | ||
ae4df9d6 | 8338 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd PZ |
8339 | struct rq *rq = cpu_rq(i); |
8340 | ||
63928384 | 8341 | if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false)) |
f643ea22 | 8342 | env->flags |= LBF_NOHZ_AGAIN; |
e022e0d3 | 8343 | |
b0fb1eb4 | 8344 | sgs->group_load += cpu_load(rq); |
9e91d61d | 8345 | sgs->group_util += cpu_util(i); |
070f5e86 | 8346 | sgs->group_runnable += cpu_runnable(rq); |
a3498347 | 8347 | sgs->sum_h_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 8348 | |
a426f99c | 8349 | nr_running = rq->nr_running; |
5e23e474 VG |
8350 | sgs->sum_nr_running += nr_running; |
8351 | ||
a426f99c | 8352 | if (nr_running > 1) |
630246a0 | 8353 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 8354 | |
2802bf3c MR |
8355 | if (cpu_overutilized(i)) |
8356 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 8357 | |
0ec8aa00 PZ |
8358 | #ifdef CONFIG_NUMA_BALANCING |
8359 | sgs->nr_numa_running += rq->nr_numa_running; | |
8360 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
8361 | #endif | |
a426f99c WL |
8362 | /* |
8363 | * No need to call idle_cpu() if nr_running is not 0 | |
8364 | */ | |
0b0695f2 | 8365 | if (!nr_running && idle_cpu(i)) { |
aae6d3dd | 8366 | sgs->idle_cpus++; |
0b0695f2 VG |
8367 | /* Idle cpu can't have misfit task */ |
8368 | continue; | |
8369 | } | |
8370 | ||
8371 | if (local_group) | |
8372 | continue; | |
3b1baa64 | 8373 | |
0b0695f2 | 8374 | /* Check for a misfit task on the cpu */ |
3b1baa64 | 8375 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && |
757ffdd7 | 8376 | sgs->group_misfit_task_load < rq->misfit_task_load) { |
3b1baa64 | 8377 | sgs->group_misfit_task_load = rq->misfit_task_load; |
630246a0 | 8378 | *sg_status |= SG_OVERLOAD; |
757ffdd7 | 8379 | } |
1e3c88bd PZ |
8380 | } |
8381 | ||
0b0695f2 VG |
8382 | /* Check if dst CPU is idle and preferred to this group */ |
8383 | if (env->sd->flags & SD_ASYM_PACKING && | |
8384 | env->idle != CPU_NOT_IDLE && | |
8385 | sgs->sum_h_nr_running && | |
8386 | sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) { | |
8387 | sgs->group_asym_packing = 1; | |
8388 | } | |
8389 | ||
63b2ca30 | 8390 | sgs->group_capacity = group->sgc->capacity; |
1e3c88bd | 8391 | |
aae6d3dd | 8392 | sgs->group_weight = group->group_weight; |
b37d9316 | 8393 | |
57abff06 | 8394 | sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); |
0b0695f2 VG |
8395 | |
8396 | /* Computing avg_load makes sense only when group is overloaded */ | |
8397 | if (sgs->group_type == group_overloaded) | |
8398 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / | |
8399 | sgs->group_capacity; | |
1e3c88bd PZ |
8400 | } |
8401 | ||
532cb4c4 MN |
8402 | /** |
8403 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 8404 | * @env: The load balancing environment. |
532cb4c4 MN |
8405 | * @sds: sched_domain statistics |
8406 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 8407 | * @sgs: sched_group statistics |
532cb4c4 MN |
8408 | * |
8409 | * Determine if @sg is a busier group than the previously selected | |
8410 | * busiest group. | |
e69f6186 YB |
8411 | * |
8412 | * Return: %true if @sg is a busier group than the previously selected | |
8413 | * busiest group. %false otherwise. | |
532cb4c4 | 8414 | */ |
bd939f45 | 8415 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
8416 | struct sd_lb_stats *sds, |
8417 | struct sched_group *sg, | |
bd939f45 | 8418 | struct sg_lb_stats *sgs) |
532cb4c4 | 8419 | { |
caeb178c | 8420 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 8421 | |
0b0695f2 VG |
8422 | /* Make sure that there is at least one task to pull */ |
8423 | if (!sgs->sum_h_nr_running) | |
8424 | return false; | |
8425 | ||
cad68e55 MR |
8426 | /* |
8427 | * Don't try to pull misfit tasks we can't help. | |
8428 | * We can use max_capacity here as reduction in capacity on some | |
8429 | * CPUs in the group should either be possible to resolve | |
8430 | * internally or be covered by avg_load imbalance (eventually). | |
8431 | */ | |
8432 | if (sgs->group_type == group_misfit_task && | |
8433 | (!group_smaller_max_cpu_capacity(sg, sds->local) || | |
0b0695f2 | 8434 | sds->local_stat.group_type != group_has_spare)) |
cad68e55 MR |
8435 | return false; |
8436 | ||
caeb178c | 8437 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
8438 | return true; |
8439 | ||
caeb178c RR |
8440 | if (sgs->group_type < busiest->group_type) |
8441 | return false; | |
8442 | ||
9e0994c0 | 8443 | /* |
0b0695f2 VG |
8444 | * The candidate and the current busiest group are the same type of |
8445 | * group. Let check which one is the busiest according to the type. | |
9e0994c0 | 8446 | */ |
9e0994c0 | 8447 | |
0b0695f2 VG |
8448 | switch (sgs->group_type) { |
8449 | case group_overloaded: | |
8450 | /* Select the overloaded group with highest avg_load. */ | |
8451 | if (sgs->avg_load <= busiest->avg_load) | |
8452 | return false; | |
8453 | break; | |
8454 | ||
8455 | case group_imbalanced: | |
8456 | /* | |
8457 | * Select the 1st imbalanced group as we don't have any way to | |
8458 | * choose one more than another. | |
8459 | */ | |
9e0994c0 MR |
8460 | return false; |
8461 | ||
0b0695f2 VG |
8462 | case group_asym_packing: |
8463 | /* Prefer to move from lowest priority CPU's work */ | |
8464 | if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) | |
8465 | return false; | |
8466 | break; | |
532cb4c4 | 8467 | |
0b0695f2 VG |
8468 | case group_misfit_task: |
8469 | /* | |
8470 | * If we have more than one misfit sg go with the biggest | |
8471 | * misfit. | |
8472 | */ | |
8473 | if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
8474 | return false; | |
8475 | break; | |
532cb4c4 | 8476 | |
0b0695f2 VG |
8477 | case group_fully_busy: |
8478 | /* | |
8479 | * Select the fully busy group with highest avg_load. In | |
8480 | * theory, there is no need to pull task from such kind of | |
8481 | * group because tasks have all compute capacity that they need | |
8482 | * but we can still improve the overall throughput by reducing | |
8483 | * contention when accessing shared HW resources. | |
8484 | * | |
8485 | * XXX for now avg_load is not computed and always 0 so we | |
8486 | * select the 1st one. | |
8487 | */ | |
8488 | if (sgs->avg_load <= busiest->avg_load) | |
8489 | return false; | |
8490 | break; | |
8491 | ||
8492 | case group_has_spare: | |
8493 | /* | |
5f68eb19 VG |
8494 | * Select not overloaded group with lowest number of idle cpus |
8495 | * and highest number of running tasks. We could also compare | |
8496 | * the spare capacity which is more stable but it can end up | |
8497 | * that the group has less spare capacity but finally more idle | |
0b0695f2 VG |
8498 | * CPUs which means less opportunity to pull tasks. |
8499 | */ | |
5f68eb19 | 8500 | if (sgs->idle_cpus > busiest->idle_cpus) |
0b0695f2 | 8501 | return false; |
5f68eb19 VG |
8502 | else if ((sgs->idle_cpus == busiest->idle_cpus) && |
8503 | (sgs->sum_nr_running <= busiest->sum_nr_running)) | |
8504 | return false; | |
8505 | ||
0b0695f2 | 8506 | break; |
532cb4c4 MN |
8507 | } |
8508 | ||
0b0695f2 VG |
8509 | /* |
8510 | * Candidate sg has no more than one task per CPU and has higher | |
8511 | * per-CPU capacity. Migrating tasks to less capable CPUs may harm | |
8512 | * throughput. Maximize throughput, power/energy consequences are not | |
8513 | * considered. | |
8514 | */ | |
8515 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && | |
8516 | (sgs->group_type <= group_fully_busy) && | |
8517 | (group_smaller_min_cpu_capacity(sds->local, sg))) | |
8518 | return false; | |
8519 | ||
8520 | return true; | |
532cb4c4 MN |
8521 | } |
8522 | ||
0ec8aa00 PZ |
8523 | #ifdef CONFIG_NUMA_BALANCING |
8524 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8525 | { | |
a3498347 | 8526 | if (sgs->sum_h_nr_running > sgs->nr_numa_running) |
0ec8aa00 | 8527 | return regular; |
a3498347 | 8528 | if (sgs->sum_h_nr_running > sgs->nr_preferred_running) |
0ec8aa00 PZ |
8529 | return remote; |
8530 | return all; | |
8531 | } | |
8532 | ||
8533 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8534 | { | |
8535 | if (rq->nr_running > rq->nr_numa_running) | |
8536 | return regular; | |
8537 | if (rq->nr_running > rq->nr_preferred_running) | |
8538 | return remote; | |
8539 | return all; | |
8540 | } | |
8541 | #else | |
8542 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8543 | { | |
8544 | return all; | |
8545 | } | |
8546 | ||
8547 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8548 | { | |
8549 | return regular; | |
8550 | } | |
8551 | #endif /* CONFIG_NUMA_BALANCING */ | |
8552 | ||
57abff06 VG |
8553 | |
8554 | struct sg_lb_stats; | |
8555 | ||
3318544b VG |
8556 | /* |
8557 | * task_running_on_cpu - return 1 if @p is running on @cpu. | |
8558 | */ | |
8559 | ||
8560 | static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) | |
8561 | { | |
8562 | /* Task has no contribution or is new */ | |
8563 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
8564 | return 0; | |
8565 | ||
8566 | if (task_on_rq_queued(p)) | |
8567 | return 1; | |
8568 | ||
8569 | return 0; | |
8570 | } | |
8571 | ||
8572 | /** | |
8573 | * idle_cpu_without - would a given CPU be idle without p ? | |
8574 | * @cpu: the processor on which idleness is tested. | |
8575 | * @p: task which should be ignored. | |
8576 | * | |
8577 | * Return: 1 if the CPU would be idle. 0 otherwise. | |
8578 | */ | |
8579 | static int idle_cpu_without(int cpu, struct task_struct *p) | |
8580 | { | |
8581 | struct rq *rq = cpu_rq(cpu); | |
8582 | ||
8583 | if (rq->curr != rq->idle && rq->curr != p) | |
8584 | return 0; | |
8585 | ||
8586 | /* | |
8587 | * rq->nr_running can't be used but an updated version without the | |
8588 | * impact of p on cpu must be used instead. The updated nr_running | |
8589 | * be computed and tested before calling idle_cpu_without(). | |
8590 | */ | |
8591 | ||
8592 | #ifdef CONFIG_SMP | |
8593 | if (!llist_empty(&rq->wake_list)) | |
8594 | return 0; | |
8595 | #endif | |
8596 | ||
8597 | return 1; | |
8598 | } | |
8599 | ||
57abff06 VG |
8600 | /* |
8601 | * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. | |
3318544b | 8602 | * @sd: The sched_domain level to look for idlest group. |
57abff06 VG |
8603 | * @group: sched_group whose statistics are to be updated. |
8604 | * @sgs: variable to hold the statistics for this group. | |
3318544b | 8605 | * @p: The task for which we look for the idlest group/CPU. |
57abff06 VG |
8606 | */ |
8607 | static inline void update_sg_wakeup_stats(struct sched_domain *sd, | |
8608 | struct sched_group *group, | |
8609 | struct sg_lb_stats *sgs, | |
8610 | struct task_struct *p) | |
8611 | { | |
8612 | int i, nr_running; | |
8613 | ||
8614 | memset(sgs, 0, sizeof(*sgs)); | |
8615 | ||
8616 | for_each_cpu(i, sched_group_span(group)) { | |
8617 | struct rq *rq = cpu_rq(i); | |
3318544b | 8618 | unsigned int local; |
57abff06 | 8619 | |
3318544b | 8620 | sgs->group_load += cpu_load_without(rq, p); |
57abff06 | 8621 | sgs->group_util += cpu_util_without(i, p); |
070f5e86 | 8622 | sgs->group_runnable += cpu_runnable_without(rq, p); |
3318544b VG |
8623 | local = task_running_on_cpu(i, p); |
8624 | sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; | |
57abff06 | 8625 | |
3318544b | 8626 | nr_running = rq->nr_running - local; |
57abff06 VG |
8627 | sgs->sum_nr_running += nr_running; |
8628 | ||
8629 | /* | |
3318544b | 8630 | * No need to call idle_cpu_without() if nr_running is not 0 |
57abff06 | 8631 | */ |
3318544b | 8632 | if (!nr_running && idle_cpu_without(i, p)) |
57abff06 VG |
8633 | sgs->idle_cpus++; |
8634 | ||
57abff06 VG |
8635 | } |
8636 | ||
8637 | /* Check if task fits in the group */ | |
8638 | if (sd->flags & SD_ASYM_CPUCAPACITY && | |
8639 | !task_fits_capacity(p, group->sgc->max_capacity)) { | |
8640 | sgs->group_misfit_task_load = 1; | |
8641 | } | |
8642 | ||
8643 | sgs->group_capacity = group->sgc->capacity; | |
8644 | ||
289de359 VG |
8645 | sgs->group_weight = group->group_weight; |
8646 | ||
57abff06 VG |
8647 | sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); |
8648 | ||
8649 | /* | |
8650 | * Computing avg_load makes sense only when group is fully busy or | |
8651 | * overloaded | |
8652 | */ | |
6c8116c9 TZ |
8653 | if (sgs->group_type == group_fully_busy || |
8654 | sgs->group_type == group_overloaded) | |
57abff06 VG |
8655 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / |
8656 | sgs->group_capacity; | |
8657 | } | |
8658 | ||
8659 | static bool update_pick_idlest(struct sched_group *idlest, | |
8660 | struct sg_lb_stats *idlest_sgs, | |
8661 | struct sched_group *group, | |
8662 | struct sg_lb_stats *sgs) | |
8663 | { | |
8664 | if (sgs->group_type < idlest_sgs->group_type) | |
8665 | return true; | |
8666 | ||
8667 | if (sgs->group_type > idlest_sgs->group_type) | |
8668 | return false; | |
8669 | ||
8670 | /* | |
8671 | * The candidate and the current idlest group are the same type of | |
8672 | * group. Let check which one is the idlest according to the type. | |
8673 | */ | |
8674 | ||
8675 | switch (sgs->group_type) { | |
8676 | case group_overloaded: | |
8677 | case group_fully_busy: | |
8678 | /* Select the group with lowest avg_load. */ | |
8679 | if (idlest_sgs->avg_load <= sgs->avg_load) | |
8680 | return false; | |
8681 | break; | |
8682 | ||
8683 | case group_imbalanced: | |
8684 | case group_asym_packing: | |
8685 | /* Those types are not used in the slow wakeup path */ | |
8686 | return false; | |
8687 | ||
8688 | case group_misfit_task: | |
8689 | /* Select group with the highest max capacity */ | |
8690 | if (idlest->sgc->max_capacity >= group->sgc->max_capacity) | |
8691 | return false; | |
8692 | break; | |
8693 | ||
8694 | case group_has_spare: | |
8695 | /* Select group with most idle CPUs */ | |
8696 | if (idlest_sgs->idle_cpus >= sgs->idle_cpus) | |
8697 | return false; | |
8698 | break; | |
8699 | } | |
8700 | ||
8701 | return true; | |
8702 | } | |
8703 | ||
8704 | /* | |
8705 | * find_idlest_group() finds and returns the least busy CPU group within the | |
8706 | * domain. | |
8707 | * | |
8708 | * Assumes p is allowed on at least one CPU in sd. | |
8709 | */ | |
8710 | static struct sched_group * | |
45da2773 | 8711 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
57abff06 VG |
8712 | { |
8713 | struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; | |
8714 | struct sg_lb_stats local_sgs, tmp_sgs; | |
8715 | struct sg_lb_stats *sgs; | |
8716 | unsigned long imbalance; | |
8717 | struct sg_lb_stats idlest_sgs = { | |
8718 | .avg_load = UINT_MAX, | |
8719 | .group_type = group_overloaded, | |
8720 | }; | |
8721 | ||
8722 | imbalance = scale_load_down(NICE_0_LOAD) * | |
8723 | (sd->imbalance_pct-100) / 100; | |
8724 | ||
8725 | do { | |
8726 | int local_group; | |
8727 | ||
8728 | /* Skip over this group if it has no CPUs allowed */ | |
8729 | if (!cpumask_intersects(sched_group_span(group), | |
8730 | p->cpus_ptr)) | |
8731 | continue; | |
8732 | ||
8733 | local_group = cpumask_test_cpu(this_cpu, | |
8734 | sched_group_span(group)); | |
8735 | ||
8736 | if (local_group) { | |
8737 | sgs = &local_sgs; | |
8738 | local = group; | |
8739 | } else { | |
8740 | sgs = &tmp_sgs; | |
8741 | } | |
8742 | ||
8743 | update_sg_wakeup_stats(sd, group, sgs, p); | |
8744 | ||
8745 | if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { | |
8746 | idlest = group; | |
8747 | idlest_sgs = *sgs; | |
8748 | } | |
8749 | ||
8750 | } while (group = group->next, group != sd->groups); | |
8751 | ||
8752 | ||
8753 | /* There is no idlest group to push tasks to */ | |
8754 | if (!idlest) | |
8755 | return NULL; | |
8756 | ||
7ed735c3 VG |
8757 | /* The local group has been skipped because of CPU affinity */ |
8758 | if (!local) | |
8759 | return idlest; | |
8760 | ||
57abff06 VG |
8761 | /* |
8762 | * If the local group is idler than the selected idlest group | |
8763 | * don't try and push the task. | |
8764 | */ | |
8765 | if (local_sgs.group_type < idlest_sgs.group_type) | |
8766 | return NULL; | |
8767 | ||
8768 | /* | |
8769 | * If the local group is busier than the selected idlest group | |
8770 | * try and push the task. | |
8771 | */ | |
8772 | if (local_sgs.group_type > idlest_sgs.group_type) | |
8773 | return idlest; | |
8774 | ||
8775 | switch (local_sgs.group_type) { | |
8776 | case group_overloaded: | |
8777 | case group_fully_busy: | |
8778 | /* | |
8779 | * When comparing groups across NUMA domains, it's possible for | |
8780 | * the local domain to be very lightly loaded relative to the | |
8781 | * remote domains but "imbalance" skews the comparison making | |
8782 | * remote CPUs look much more favourable. When considering | |
8783 | * cross-domain, add imbalance to the load on the remote node | |
8784 | * and consider staying local. | |
8785 | */ | |
8786 | ||
8787 | if ((sd->flags & SD_NUMA) && | |
8788 | ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) | |
8789 | return NULL; | |
8790 | ||
8791 | /* | |
8792 | * If the local group is less loaded than the selected | |
8793 | * idlest group don't try and push any tasks. | |
8794 | */ | |
8795 | if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) | |
8796 | return NULL; | |
8797 | ||
8798 | if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) | |
8799 | return NULL; | |
8800 | break; | |
8801 | ||
8802 | case group_imbalanced: | |
8803 | case group_asym_packing: | |
8804 | /* Those type are not used in the slow wakeup path */ | |
8805 | return NULL; | |
8806 | ||
8807 | case group_misfit_task: | |
8808 | /* Select group with the highest max capacity */ | |
8809 | if (local->sgc->max_capacity >= idlest->sgc->max_capacity) | |
8810 | return NULL; | |
8811 | break; | |
8812 | ||
8813 | case group_has_spare: | |
8814 | if (sd->flags & SD_NUMA) { | |
8815 | #ifdef CONFIG_NUMA_BALANCING | |
8816 | int idlest_cpu; | |
8817 | /* | |
8818 | * If there is spare capacity at NUMA, try to select | |
8819 | * the preferred node | |
8820 | */ | |
8821 | if (cpu_to_node(this_cpu) == p->numa_preferred_nid) | |
8822 | return NULL; | |
8823 | ||
8824 | idlest_cpu = cpumask_first(sched_group_span(idlest)); | |
8825 | if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) | |
8826 | return idlest; | |
8827 | #endif | |
8828 | /* | |
8829 | * Otherwise, keep the task on this node to stay close | |
8830 | * its wakeup source and improve locality. If there is | |
8831 | * a real need of migration, periodic load balance will | |
8832 | * take care of it. | |
8833 | */ | |
8834 | if (local_sgs.idle_cpus) | |
8835 | return NULL; | |
8836 | } | |
8837 | ||
8838 | /* | |
8839 | * Select group with highest number of idle CPUs. We could also | |
8840 | * compare the utilization which is more stable but it can end | |
8841 | * up that the group has less spare capacity but finally more | |
8842 | * idle CPUs which means more opportunity to run task. | |
8843 | */ | |
8844 | if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) | |
8845 | return NULL; | |
8846 | break; | |
8847 | } | |
8848 | ||
8849 | return idlest; | |
8850 | } | |
8851 | ||
1e3c88bd | 8852 | /** |
461819ac | 8853 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 8854 | * @env: The load balancing environment. |
1e3c88bd PZ |
8855 | * @sds: variable to hold the statistics for this sched_domain. |
8856 | */ | |
0b0695f2 | 8857 | |
0ec8aa00 | 8858 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8859 | { |
bd939f45 PZ |
8860 | struct sched_domain *child = env->sd->child; |
8861 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 8862 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 8863 | struct sg_lb_stats tmp_sgs; |
630246a0 | 8864 | int sg_status = 0; |
1e3c88bd | 8865 | |
e022e0d3 | 8866 | #ifdef CONFIG_NO_HZ_COMMON |
f643ea22 | 8867 | if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked)) |
e022e0d3 | 8868 | env->flags |= LBF_NOHZ_STATS; |
e022e0d3 PZ |
8869 | #endif |
8870 | ||
1e3c88bd | 8871 | do { |
56cf515b | 8872 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
8873 | int local_group; |
8874 | ||
ae4df9d6 | 8875 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
8876 | if (local_group) { |
8877 | sds->local = sg; | |
05b40e05 | 8878 | sgs = local; |
b72ff13c PZ |
8879 | |
8880 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
8881 | time_after_eq(jiffies, sg->sgc->next_update)) |
8882 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 8883 | } |
1e3c88bd | 8884 | |
630246a0 | 8885 | update_sg_lb_stats(env, sg, sgs, &sg_status); |
1e3c88bd | 8886 | |
b72ff13c PZ |
8887 | if (local_group) |
8888 | goto next_group; | |
8889 | ||
1e3c88bd | 8890 | |
b72ff13c | 8891 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 8892 | sds->busiest = sg; |
56cf515b | 8893 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
8894 | } |
8895 | ||
b72ff13c PZ |
8896 | next_group: |
8897 | /* Now, start updating sd_lb_stats */ | |
8898 | sds->total_load += sgs->group_load; | |
63b2ca30 | 8899 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 8900 | |
532cb4c4 | 8901 | sg = sg->next; |
bd939f45 | 8902 | } while (sg != env->sd->groups); |
0ec8aa00 | 8903 | |
0b0695f2 VG |
8904 | /* Tag domain that child domain prefers tasks go to siblings first */ |
8905 | sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; | |
8906 | ||
f643ea22 VG |
8907 | #ifdef CONFIG_NO_HZ_COMMON |
8908 | if ((env->flags & LBF_NOHZ_AGAIN) && | |
8909 | cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) { | |
8910 | ||
8911 | WRITE_ONCE(nohz.next_blocked, | |
8912 | jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
8913 | } | |
8914 | #endif | |
8915 | ||
0ec8aa00 PZ |
8916 | if (env->sd->flags & SD_NUMA) |
8917 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
8918 | |
8919 | if (!env->sd->parent) { | |
2802bf3c MR |
8920 | struct root_domain *rd = env->dst_rq->rd; |
8921 | ||
4486edd1 | 8922 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
8923 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
8924 | ||
8925 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
8926 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
f9f240f9 | 8927 | trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); |
2802bf3c | 8928 | } else if (sg_status & SG_OVERUTILIZED) { |
f9f240f9 QY |
8929 | struct root_domain *rd = env->dst_rq->rd; |
8930 | ||
8931 | WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); | |
8932 | trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); | |
4486edd1 | 8933 | } |
532cb4c4 MN |
8934 | } |
8935 | ||
fb86f5b2 MG |
8936 | static inline long adjust_numa_imbalance(int imbalance, int src_nr_running) |
8937 | { | |
8938 | unsigned int imbalance_min; | |
8939 | ||
8940 | /* | |
8941 | * Allow a small imbalance based on a simple pair of communicating | |
8942 | * tasks that remain local when the source domain is almost idle. | |
8943 | */ | |
8944 | imbalance_min = 2; | |
8945 | if (src_nr_running <= imbalance_min) | |
8946 | return 0; | |
8947 | ||
8948 | return imbalance; | |
8949 | } | |
8950 | ||
1e3c88bd PZ |
8951 | /** |
8952 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
8953 | * groups of a given sched_domain during load balance. | |
bd939f45 | 8954 | * @env: load balance environment |
1e3c88bd | 8955 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 8956 | */ |
bd939f45 | 8957 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8958 | { |
56cf515b JK |
8959 | struct sg_lb_stats *local, *busiest; |
8960 | ||
8961 | local = &sds->local_stat; | |
56cf515b | 8962 | busiest = &sds->busiest_stat; |
dd5feea1 | 8963 | |
0b0695f2 VG |
8964 | if (busiest->group_type == group_misfit_task) { |
8965 | /* Set imbalance to allow misfit tasks to be balanced. */ | |
8966 | env->migration_type = migrate_misfit; | |
c63be7be | 8967 | env->imbalance = 1; |
0b0695f2 VG |
8968 | return; |
8969 | } | |
8970 | ||
8971 | if (busiest->group_type == group_asym_packing) { | |
8972 | /* | |
8973 | * In case of asym capacity, we will try to migrate all load to | |
8974 | * the preferred CPU. | |
8975 | */ | |
8976 | env->migration_type = migrate_task; | |
8977 | env->imbalance = busiest->sum_h_nr_running; | |
8978 | return; | |
8979 | } | |
8980 | ||
8981 | if (busiest->group_type == group_imbalanced) { | |
8982 | /* | |
8983 | * In the group_imb case we cannot rely on group-wide averages | |
8984 | * to ensure CPU-load equilibrium, try to move any task to fix | |
8985 | * the imbalance. The next load balance will take care of | |
8986 | * balancing back the system. | |
8987 | */ | |
8988 | env->migration_type = migrate_task; | |
8989 | env->imbalance = 1; | |
490ba971 VG |
8990 | return; |
8991 | } | |
8992 | ||
1e3c88bd | 8993 | /* |
0b0695f2 | 8994 | * Try to use spare capacity of local group without overloading it or |
a9723389 | 8995 | * emptying busiest. |
1e3c88bd | 8996 | */ |
0b0695f2 VG |
8997 | if (local->group_type == group_has_spare) { |
8998 | if (busiest->group_type > group_fully_busy) { | |
8999 | /* | |
9000 | * If busiest is overloaded, try to fill spare | |
9001 | * capacity. This might end up creating spare capacity | |
9002 | * in busiest or busiest still being overloaded but | |
9003 | * there is no simple way to directly compute the | |
9004 | * amount of load to migrate in order to balance the | |
9005 | * system. | |
9006 | */ | |
9007 | env->migration_type = migrate_util; | |
9008 | env->imbalance = max(local->group_capacity, local->group_util) - | |
9009 | local->group_util; | |
9010 | ||
9011 | /* | |
9012 | * In some cases, the group's utilization is max or even | |
9013 | * higher than capacity because of migrations but the | |
9014 | * local CPU is (newly) idle. There is at least one | |
9015 | * waiting task in this overloaded busiest group. Let's | |
9016 | * try to pull it. | |
9017 | */ | |
9018 | if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { | |
9019 | env->migration_type = migrate_task; | |
9020 | env->imbalance = 1; | |
9021 | } | |
9022 | ||
9023 | return; | |
9024 | } | |
9025 | ||
9026 | if (busiest->group_weight == 1 || sds->prefer_sibling) { | |
5e23e474 | 9027 | unsigned int nr_diff = busiest->sum_nr_running; |
0b0695f2 VG |
9028 | /* |
9029 | * When prefer sibling, evenly spread running tasks on | |
9030 | * groups. | |
9031 | */ | |
9032 | env->migration_type = migrate_task; | |
5e23e474 | 9033 | lsub_positive(&nr_diff, local->sum_nr_running); |
0b0695f2 | 9034 | env->imbalance = nr_diff >> 1; |
b396f523 | 9035 | } else { |
0b0695f2 | 9036 | |
b396f523 MG |
9037 | /* |
9038 | * If there is no overload, we just want to even the number of | |
9039 | * idle cpus. | |
9040 | */ | |
9041 | env->migration_type = migrate_task; | |
9042 | env->imbalance = max_t(long, 0, (local->idle_cpus - | |
0b0695f2 | 9043 | busiest->idle_cpus) >> 1); |
b396f523 MG |
9044 | } |
9045 | ||
9046 | /* Consider allowing a small imbalance between NUMA groups */ | |
fb86f5b2 MG |
9047 | if (env->sd->flags & SD_NUMA) |
9048 | env->imbalance = adjust_numa_imbalance(env->imbalance, | |
9049 | busiest->sum_nr_running); | |
b396f523 | 9050 | |
fcf0553d | 9051 | return; |
1e3c88bd PZ |
9052 | } |
9053 | ||
9a5d9ba6 | 9054 | /* |
0b0695f2 VG |
9055 | * Local is fully busy but has to take more load to relieve the |
9056 | * busiest group | |
9a5d9ba6 | 9057 | */ |
0b0695f2 VG |
9058 | if (local->group_type < group_overloaded) { |
9059 | /* | |
9060 | * Local will become overloaded so the avg_load metrics are | |
9061 | * finally needed. | |
9062 | */ | |
9063 | ||
9064 | local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / | |
9065 | local->group_capacity; | |
9066 | ||
9067 | sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / | |
9068 | sds->total_capacity; | |
111688ca AL |
9069 | /* |
9070 | * If the local group is more loaded than the selected | |
9071 | * busiest group don't try to pull any tasks. | |
9072 | */ | |
9073 | if (local->avg_load >= busiest->avg_load) { | |
9074 | env->imbalance = 0; | |
9075 | return; | |
9076 | } | |
dd5feea1 SS |
9077 | } |
9078 | ||
9079 | /* | |
0b0695f2 VG |
9080 | * Both group are or will become overloaded and we're trying to get all |
9081 | * the CPUs to the average_load, so we don't want to push ourselves | |
9082 | * above the average load, nor do we wish to reduce the max loaded CPU | |
9083 | * below the average load. At the same time, we also don't want to | |
9084 | * reduce the group load below the group capacity. Thus we look for | |
9085 | * the minimum possible imbalance. | |
dd5feea1 | 9086 | */ |
0b0695f2 | 9087 | env->migration_type = migrate_load; |
56cf515b | 9088 | env->imbalance = min( |
0b0695f2 | 9089 | (busiest->avg_load - sds->avg_load) * busiest->group_capacity, |
63b2ca30 | 9090 | (sds->avg_load - local->avg_load) * local->group_capacity |
ca8ce3d0 | 9091 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 9092 | } |
fab47622 | 9093 | |
1e3c88bd PZ |
9094 | /******* find_busiest_group() helpers end here *********************/ |
9095 | ||
0b0695f2 VG |
9096 | /* |
9097 | * Decision matrix according to the local and busiest group type: | |
9098 | * | |
9099 | * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded | |
9100 | * has_spare nr_idle balanced N/A N/A balanced balanced | |
9101 | * fully_busy nr_idle nr_idle N/A N/A balanced balanced | |
9102 | * misfit_task force N/A N/A N/A force force | |
9103 | * asym_packing force force N/A N/A force force | |
9104 | * imbalanced force force N/A N/A force force | |
9105 | * overloaded force force N/A N/A force avg_load | |
9106 | * | |
9107 | * N/A : Not Applicable because already filtered while updating | |
9108 | * statistics. | |
9109 | * balanced : The system is balanced for these 2 groups. | |
9110 | * force : Calculate the imbalance as load migration is probably needed. | |
9111 | * avg_load : Only if imbalance is significant enough. | |
9112 | * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite | |
9113 | * different in groups. | |
9114 | */ | |
9115 | ||
1e3c88bd PZ |
9116 | /** |
9117 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 9118 | * if there is an imbalance. |
1e3c88bd | 9119 | * |
a3df0679 | 9120 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
9121 | * to restore balance. |
9122 | * | |
cd96891d | 9123 | * @env: The load balancing environment. |
1e3c88bd | 9124 | * |
e69f6186 | 9125 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 9126 | */ |
56cf515b | 9127 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 9128 | { |
56cf515b | 9129 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
9130 | struct sd_lb_stats sds; |
9131 | ||
147c5fc2 | 9132 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
9133 | |
9134 | /* | |
b0fb1eb4 | 9135 | * Compute the various statistics relevant for load balancing at |
1e3c88bd PZ |
9136 | * this level. |
9137 | */ | |
23f0d209 | 9138 | update_sd_lb_stats(env, &sds); |
2802bf3c | 9139 | |
f8a696f2 | 9140 | if (sched_energy_enabled()) { |
2802bf3c MR |
9141 | struct root_domain *rd = env->dst_rq->rd; |
9142 | ||
9143 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
9144 | goto out_balanced; | |
9145 | } | |
9146 | ||
56cf515b JK |
9147 | local = &sds.local_stat; |
9148 | busiest = &sds.busiest_stat; | |
1e3c88bd | 9149 | |
cc57aa8f | 9150 | /* There is no busy sibling group to pull tasks from */ |
0b0695f2 | 9151 | if (!sds.busiest) |
1e3c88bd PZ |
9152 | goto out_balanced; |
9153 | ||
0b0695f2 VG |
9154 | /* Misfit tasks should be dealt with regardless of the avg load */ |
9155 | if (busiest->group_type == group_misfit_task) | |
9156 | goto force_balance; | |
9157 | ||
9158 | /* ASYM feature bypasses nice load balance check */ | |
9159 | if (busiest->group_type == group_asym_packing) | |
9160 | goto force_balance; | |
b0432d8f | 9161 | |
866ab43e PZ |
9162 | /* |
9163 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 9164 | * work because they assume all things are equal, which typically |
3bd37062 | 9165 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 9166 | */ |
caeb178c | 9167 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
9168 | goto force_balance; |
9169 | ||
cc57aa8f | 9170 | /* |
9c58c79a | 9171 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
9172 | * don't try and pull any tasks. |
9173 | */ | |
0b0695f2 | 9174 | if (local->group_type > busiest->group_type) |
1e3c88bd PZ |
9175 | goto out_balanced; |
9176 | ||
cc57aa8f | 9177 | /* |
0b0695f2 VG |
9178 | * When groups are overloaded, use the avg_load to ensure fairness |
9179 | * between tasks. | |
cc57aa8f | 9180 | */ |
0b0695f2 VG |
9181 | if (local->group_type == group_overloaded) { |
9182 | /* | |
9183 | * If the local group is more loaded than the selected | |
9184 | * busiest group don't try to pull any tasks. | |
9185 | */ | |
9186 | if (local->avg_load >= busiest->avg_load) | |
9187 | goto out_balanced; | |
9188 | ||
9189 | /* XXX broken for overlapping NUMA groups */ | |
9190 | sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / | |
9191 | sds.total_capacity; | |
1e3c88bd | 9192 | |
aae6d3dd | 9193 | /* |
0b0695f2 VG |
9194 | * Don't pull any tasks if this group is already above the |
9195 | * domain average load. | |
aae6d3dd | 9196 | */ |
0b0695f2 | 9197 | if (local->avg_load >= sds.avg_load) |
aae6d3dd | 9198 | goto out_balanced; |
0b0695f2 | 9199 | |
c186fafe | 9200 | /* |
0b0695f2 VG |
9201 | * If the busiest group is more loaded, use imbalance_pct to be |
9202 | * conservative. | |
c186fafe | 9203 | */ |
56cf515b JK |
9204 | if (100 * busiest->avg_load <= |
9205 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 9206 | goto out_balanced; |
aae6d3dd | 9207 | } |
1e3c88bd | 9208 | |
0b0695f2 VG |
9209 | /* Try to move all excess tasks to child's sibling domain */ |
9210 | if (sds.prefer_sibling && local->group_type == group_has_spare && | |
5e23e474 | 9211 | busiest->sum_nr_running > local->sum_nr_running + 1) |
0b0695f2 VG |
9212 | goto force_balance; |
9213 | ||
2ab4092f VG |
9214 | if (busiest->group_type != group_overloaded) { |
9215 | if (env->idle == CPU_NOT_IDLE) | |
9216 | /* | |
9217 | * If the busiest group is not overloaded (and as a | |
9218 | * result the local one too) but this CPU is already | |
9219 | * busy, let another idle CPU try to pull task. | |
9220 | */ | |
9221 | goto out_balanced; | |
9222 | ||
9223 | if (busiest->group_weight > 1 && | |
9224 | local->idle_cpus <= (busiest->idle_cpus + 1)) | |
9225 | /* | |
9226 | * If the busiest group is not overloaded | |
9227 | * and there is no imbalance between this and busiest | |
9228 | * group wrt idle CPUs, it is balanced. The imbalance | |
9229 | * becomes significant if the diff is greater than 1 | |
9230 | * otherwise we might end up to just move the imbalance | |
9231 | * on another group. Of course this applies only if | |
9232 | * there is more than 1 CPU per group. | |
9233 | */ | |
9234 | goto out_balanced; | |
9235 | ||
9236 | if (busiest->sum_h_nr_running == 1) | |
9237 | /* | |
9238 | * busiest doesn't have any tasks waiting to run | |
9239 | */ | |
9240 | goto out_balanced; | |
9241 | } | |
0b0695f2 | 9242 | |
fab47622 | 9243 | force_balance: |
1e3c88bd | 9244 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 9245 | calculate_imbalance(env, &sds); |
bb3485c8 | 9246 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
9247 | |
9248 | out_balanced: | |
bd939f45 | 9249 | env->imbalance = 0; |
1e3c88bd PZ |
9250 | return NULL; |
9251 | } | |
9252 | ||
9253 | /* | |
97fb7a0a | 9254 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 9255 | */ |
bd939f45 | 9256 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 9257 | struct sched_group *group) |
1e3c88bd PZ |
9258 | { |
9259 | struct rq *busiest = NULL, *rq; | |
0b0695f2 VG |
9260 | unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; |
9261 | unsigned int busiest_nr = 0; | |
1e3c88bd PZ |
9262 | int i; |
9263 | ||
ae4df9d6 | 9264 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
0b0695f2 VG |
9265 | unsigned long capacity, load, util; |
9266 | unsigned int nr_running; | |
0ec8aa00 PZ |
9267 | enum fbq_type rt; |
9268 | ||
9269 | rq = cpu_rq(i); | |
9270 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 9271 | |
0ec8aa00 PZ |
9272 | /* |
9273 | * We classify groups/runqueues into three groups: | |
9274 | * - regular: there are !numa tasks | |
9275 | * - remote: there are numa tasks that run on the 'wrong' node | |
9276 | * - all: there is no distinction | |
9277 | * | |
9278 | * In order to avoid migrating ideally placed numa tasks, | |
9279 | * ignore those when there's better options. | |
9280 | * | |
9281 | * If we ignore the actual busiest queue to migrate another | |
9282 | * task, the next balance pass can still reduce the busiest | |
9283 | * queue by moving tasks around inside the node. | |
9284 | * | |
9285 | * If we cannot move enough load due to this classification | |
9286 | * the next pass will adjust the group classification and | |
9287 | * allow migration of more tasks. | |
9288 | * | |
9289 | * Both cases only affect the total convergence complexity. | |
9290 | */ | |
9291 | if (rt > env->fbq_type) | |
9292 | continue; | |
9293 | ||
ced549fa | 9294 | capacity = capacity_of(i); |
0b0695f2 | 9295 | nr_running = rq->cfs.h_nr_running; |
9d5efe05 | 9296 | |
4ad3831a CR |
9297 | /* |
9298 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
9299 | * eventually lead to active_balancing high->low capacity. | |
9300 | * Higher per-CPU capacity is considered better than balancing | |
9301 | * average load. | |
9302 | */ | |
9303 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
9304 | capacity_of(env->dst_cpu) < capacity && | |
0b0695f2 | 9305 | nr_running == 1) |
4ad3831a CR |
9306 | continue; |
9307 | ||
0b0695f2 VG |
9308 | switch (env->migration_type) { |
9309 | case migrate_load: | |
9310 | /* | |
b0fb1eb4 VG |
9311 | * When comparing with load imbalance, use cpu_load() |
9312 | * which is not scaled with the CPU capacity. | |
0b0695f2 | 9313 | */ |
b0fb1eb4 | 9314 | load = cpu_load(rq); |
1e3c88bd | 9315 | |
0b0695f2 VG |
9316 | if (nr_running == 1 && load > env->imbalance && |
9317 | !check_cpu_capacity(rq, env->sd)) | |
9318 | break; | |
ea67821b | 9319 | |
0b0695f2 VG |
9320 | /* |
9321 | * For the load comparisons with the other CPUs, | |
b0fb1eb4 VG |
9322 | * consider the cpu_load() scaled with the CPU |
9323 | * capacity, so that the load can be moved away | |
9324 | * from the CPU that is potentially running at a | |
9325 | * lower capacity. | |
0b0695f2 VG |
9326 | * |
9327 | * Thus we're looking for max(load_i / capacity_i), | |
9328 | * crosswise multiplication to rid ourselves of the | |
9329 | * division works out to: | |
9330 | * load_i * capacity_j > load_j * capacity_i; | |
9331 | * where j is our previous maximum. | |
9332 | */ | |
9333 | if (load * busiest_capacity > busiest_load * capacity) { | |
9334 | busiest_load = load; | |
9335 | busiest_capacity = capacity; | |
9336 | busiest = rq; | |
9337 | } | |
9338 | break; | |
9339 | ||
9340 | case migrate_util: | |
9341 | util = cpu_util(cpu_of(rq)); | |
9342 | ||
c32b4308 VG |
9343 | /* |
9344 | * Don't try to pull utilization from a CPU with one | |
9345 | * running task. Whatever its utilization, we will fail | |
9346 | * detach the task. | |
9347 | */ | |
9348 | if (nr_running <= 1) | |
9349 | continue; | |
9350 | ||
0b0695f2 VG |
9351 | if (busiest_util < util) { |
9352 | busiest_util = util; | |
9353 | busiest = rq; | |
9354 | } | |
9355 | break; | |
9356 | ||
9357 | case migrate_task: | |
9358 | if (busiest_nr < nr_running) { | |
9359 | busiest_nr = nr_running; | |
9360 | busiest = rq; | |
9361 | } | |
9362 | break; | |
9363 | ||
9364 | case migrate_misfit: | |
9365 | /* | |
9366 | * For ASYM_CPUCAPACITY domains with misfit tasks we | |
9367 | * simply seek the "biggest" misfit task. | |
9368 | */ | |
9369 | if (rq->misfit_task_load > busiest_load) { | |
9370 | busiest_load = rq->misfit_task_load; | |
9371 | busiest = rq; | |
9372 | } | |
9373 | ||
9374 | break; | |
1e3c88bd | 9375 | |
1e3c88bd PZ |
9376 | } |
9377 | } | |
9378 | ||
9379 | return busiest; | |
9380 | } | |
9381 | ||
9382 | /* | |
9383 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
9384 | * so long as it is large enough. | |
9385 | */ | |
9386 | #define MAX_PINNED_INTERVAL 512 | |
9387 | ||
46a745d9 VG |
9388 | static inline bool |
9389 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 9390 | { |
46a745d9 VG |
9391 | /* |
9392 | * ASYM_PACKING needs to force migrate tasks from busy but | |
9393 | * lower priority CPUs in order to pack all tasks in the | |
9394 | * highest priority CPUs. | |
9395 | */ | |
9396 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
9397 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
9398 | } | |
bd939f45 | 9399 | |
46a745d9 VG |
9400 | static inline bool |
9401 | voluntary_active_balance(struct lb_env *env) | |
9402 | { | |
9403 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 9404 | |
46a745d9 VG |
9405 | if (asym_active_balance(env)) |
9406 | return 1; | |
1af3ed3d | 9407 | |
1aaf90a4 VG |
9408 | /* |
9409 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
9410 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
9411 | * because of other sched_class or IRQs if more capacity stays | |
9412 | * available on dst_cpu. | |
9413 | */ | |
9414 | if ((env->idle != CPU_NOT_IDLE) && | |
9415 | (env->src_rq->cfs.h_nr_running == 1)) { | |
9416 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
9417 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
9418 | return 1; | |
9419 | } | |
9420 | ||
0b0695f2 | 9421 | if (env->migration_type == migrate_misfit) |
cad68e55 MR |
9422 | return 1; |
9423 | ||
46a745d9 VG |
9424 | return 0; |
9425 | } | |
9426 | ||
9427 | static int need_active_balance(struct lb_env *env) | |
9428 | { | |
9429 | struct sched_domain *sd = env->sd; | |
9430 | ||
9431 | if (voluntary_active_balance(env)) | |
9432 | return 1; | |
9433 | ||
1af3ed3d PZ |
9434 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
9435 | } | |
9436 | ||
969c7921 TH |
9437 | static int active_load_balance_cpu_stop(void *data); |
9438 | ||
23f0d209 JK |
9439 | static int should_we_balance(struct lb_env *env) |
9440 | { | |
9441 | struct sched_group *sg = env->sd->groups; | |
64297f2b | 9442 | int cpu; |
23f0d209 | 9443 | |
024c9d2f PZ |
9444 | /* |
9445 | * Ensure the balancing environment is consistent; can happen | |
9446 | * when the softirq triggers 'during' hotplug. | |
9447 | */ | |
9448 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
9449 | return 0; | |
9450 | ||
23f0d209 | 9451 | /* |
97fb7a0a | 9452 | * In the newly idle case, we will allow all the CPUs |
23f0d209 JK |
9453 | * to do the newly idle load balance. |
9454 | */ | |
9455 | if (env->idle == CPU_NEWLY_IDLE) | |
9456 | return 1; | |
9457 | ||
97fb7a0a | 9458 | /* Try to find first idle CPU */ |
e5c14b1f | 9459 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 9460 | if (!idle_cpu(cpu)) |
23f0d209 JK |
9461 | continue; |
9462 | ||
64297f2b PW |
9463 | /* Are we the first idle CPU? */ |
9464 | return cpu == env->dst_cpu; | |
23f0d209 JK |
9465 | } |
9466 | ||
64297f2b PW |
9467 | /* Are we the first CPU of this group ? */ |
9468 | return group_balance_cpu(sg) == env->dst_cpu; | |
23f0d209 JK |
9469 | } |
9470 | ||
1e3c88bd PZ |
9471 | /* |
9472 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
9473 | * tasks if there is an imbalance. | |
9474 | */ | |
9475 | static int load_balance(int this_cpu, struct rq *this_rq, | |
9476 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 9477 | int *continue_balancing) |
1e3c88bd | 9478 | { |
88b8dac0 | 9479 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 9480 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 9481 | struct sched_group *group; |
1e3c88bd | 9482 | struct rq *busiest; |
8a8c69c3 | 9483 | struct rq_flags rf; |
4ba29684 | 9484 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 9485 | |
8e45cb54 PZ |
9486 | struct lb_env env = { |
9487 | .sd = sd, | |
ddcdf6e7 PZ |
9488 | .dst_cpu = this_cpu, |
9489 | .dst_rq = this_rq, | |
ae4df9d6 | 9490 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 9491 | .idle = idle, |
eb95308e | 9492 | .loop_break = sched_nr_migrate_break, |
b9403130 | 9493 | .cpus = cpus, |
0ec8aa00 | 9494 | .fbq_type = all, |
163122b7 | 9495 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
9496 | }; |
9497 | ||
65a4433a | 9498 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 9499 | |
ae92882e | 9500 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
9501 | |
9502 | redo: | |
23f0d209 JK |
9503 | if (!should_we_balance(&env)) { |
9504 | *continue_balancing = 0; | |
1e3c88bd | 9505 | goto out_balanced; |
23f0d209 | 9506 | } |
1e3c88bd | 9507 | |
23f0d209 | 9508 | group = find_busiest_group(&env); |
1e3c88bd | 9509 | if (!group) { |
ae92882e | 9510 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
9511 | goto out_balanced; |
9512 | } | |
9513 | ||
b9403130 | 9514 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 9515 | if (!busiest) { |
ae92882e | 9516 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
9517 | goto out_balanced; |
9518 | } | |
9519 | ||
78feefc5 | 9520 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 9521 | |
ae92882e | 9522 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 9523 | |
1aaf90a4 VG |
9524 | env.src_cpu = busiest->cpu; |
9525 | env.src_rq = busiest; | |
9526 | ||
1e3c88bd PZ |
9527 | ld_moved = 0; |
9528 | if (busiest->nr_running > 1) { | |
9529 | /* | |
9530 | * Attempt to move tasks. If find_busiest_group has found | |
9531 | * an imbalance but busiest->nr_running <= 1, the group is | |
9532 | * still unbalanced. ld_moved simply stays zero, so it is | |
9533 | * correctly treated as an imbalance. | |
9534 | */ | |
8e45cb54 | 9535 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 9536 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 9537 | |
5d6523eb | 9538 | more_balance: |
8a8c69c3 | 9539 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 9540 | update_rq_clock(busiest); |
88b8dac0 SV |
9541 | |
9542 | /* | |
9543 | * cur_ld_moved - load moved in current iteration | |
9544 | * ld_moved - cumulative load moved across iterations | |
9545 | */ | |
163122b7 | 9546 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
9547 | |
9548 | /* | |
163122b7 KT |
9549 | * We've detached some tasks from busiest_rq. Every |
9550 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
9551 | * unlock busiest->lock, and we are able to be sure | |
9552 | * that nobody can manipulate the tasks in parallel. | |
9553 | * See task_rq_lock() family for the details. | |
1e3c88bd | 9554 | */ |
163122b7 | 9555 | |
8a8c69c3 | 9556 | rq_unlock(busiest, &rf); |
163122b7 KT |
9557 | |
9558 | if (cur_ld_moved) { | |
9559 | attach_tasks(&env); | |
9560 | ld_moved += cur_ld_moved; | |
9561 | } | |
9562 | ||
8a8c69c3 | 9563 | local_irq_restore(rf.flags); |
88b8dac0 | 9564 | |
f1cd0858 JK |
9565 | if (env.flags & LBF_NEED_BREAK) { |
9566 | env.flags &= ~LBF_NEED_BREAK; | |
9567 | goto more_balance; | |
9568 | } | |
9569 | ||
88b8dac0 SV |
9570 | /* |
9571 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
9572 | * us and move them to an alternate dst_cpu in our sched_group | |
9573 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 9574 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
9575 | * sched_group. |
9576 | * | |
9577 | * This changes load balance semantics a bit on who can move | |
9578 | * load to a given_cpu. In addition to the given_cpu itself | |
9579 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
9580 | * nohz-idle), we now have balance_cpu in a position to move | |
9581 | * load to given_cpu. In rare situations, this may cause | |
9582 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
9583 | * _independently_ and at _same_ time to move some load to | |
9584 | * given_cpu) causing exceess load to be moved to given_cpu. | |
9585 | * This however should not happen so much in practice and | |
9586 | * moreover subsequent load balance cycles should correct the | |
9587 | * excess load moved. | |
9588 | */ | |
6263322c | 9589 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 9590 | |
97fb7a0a | 9591 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 9592 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 9593 | |
78feefc5 | 9594 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 9595 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 9596 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
9597 | env.loop = 0; |
9598 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 9599 | |
88b8dac0 SV |
9600 | /* |
9601 | * Go back to "more_balance" rather than "redo" since we | |
9602 | * need to continue with same src_cpu. | |
9603 | */ | |
9604 | goto more_balance; | |
9605 | } | |
1e3c88bd | 9606 | |
6263322c PZ |
9607 | /* |
9608 | * We failed to reach balance because of affinity. | |
9609 | */ | |
9610 | if (sd_parent) { | |
63b2ca30 | 9611 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 9612 | |
afdeee05 | 9613 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 9614 | *group_imbalance = 1; |
6263322c PZ |
9615 | } |
9616 | ||
1e3c88bd | 9617 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 9618 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 9619 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
9620 | /* |
9621 | * Attempting to continue load balancing at the current | |
9622 | * sched_domain level only makes sense if there are | |
9623 | * active CPUs remaining as possible busiest CPUs to | |
9624 | * pull load from which are not contained within the | |
9625 | * destination group that is receiving any migrated | |
9626 | * load. | |
9627 | */ | |
9628 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
9629 | env.loop = 0; |
9630 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 9631 | goto redo; |
bbf18b19 | 9632 | } |
afdeee05 | 9633 | goto out_all_pinned; |
1e3c88bd PZ |
9634 | } |
9635 | } | |
9636 | ||
9637 | if (!ld_moved) { | |
ae92882e | 9638 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
9639 | /* |
9640 | * Increment the failure counter only on periodic balance. | |
9641 | * We do not want newidle balance, which can be very | |
9642 | * frequent, pollute the failure counter causing | |
9643 | * excessive cache_hot migrations and active balances. | |
9644 | */ | |
9645 | if (idle != CPU_NEWLY_IDLE) | |
9646 | sd->nr_balance_failed++; | |
1e3c88bd | 9647 | |
bd939f45 | 9648 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
9649 | unsigned long flags; |
9650 | ||
1e3c88bd PZ |
9651 | raw_spin_lock_irqsave(&busiest->lock, flags); |
9652 | ||
97fb7a0a IM |
9653 | /* |
9654 | * Don't kick the active_load_balance_cpu_stop, | |
9655 | * if the curr task on busiest CPU can't be | |
9656 | * moved to this_cpu: | |
1e3c88bd | 9657 | */ |
3bd37062 | 9658 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
1e3c88bd PZ |
9659 | raw_spin_unlock_irqrestore(&busiest->lock, |
9660 | flags); | |
8e45cb54 | 9661 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
9662 | goto out_one_pinned; |
9663 | } | |
9664 | ||
969c7921 TH |
9665 | /* |
9666 | * ->active_balance synchronizes accesses to | |
9667 | * ->active_balance_work. Once set, it's cleared | |
9668 | * only after active load balance is finished. | |
9669 | */ | |
1e3c88bd PZ |
9670 | if (!busiest->active_balance) { |
9671 | busiest->active_balance = 1; | |
9672 | busiest->push_cpu = this_cpu; | |
9673 | active_balance = 1; | |
9674 | } | |
9675 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 9676 | |
bd939f45 | 9677 | if (active_balance) { |
969c7921 TH |
9678 | stop_one_cpu_nowait(cpu_of(busiest), |
9679 | active_load_balance_cpu_stop, busiest, | |
9680 | &busiest->active_balance_work); | |
bd939f45 | 9681 | } |
1e3c88bd | 9682 | |
d02c0711 | 9683 | /* We've kicked active balancing, force task migration. */ |
1e3c88bd PZ |
9684 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
9685 | } | |
9686 | } else | |
9687 | sd->nr_balance_failed = 0; | |
9688 | ||
46a745d9 | 9689 | if (likely(!active_balance) || voluntary_active_balance(&env)) { |
1e3c88bd PZ |
9690 | /* We were unbalanced, so reset the balancing interval */ |
9691 | sd->balance_interval = sd->min_interval; | |
9692 | } else { | |
9693 | /* | |
9694 | * If we've begun active balancing, start to back off. This | |
9695 | * case may not be covered by the all_pinned logic if there | |
9696 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 9697 | * detach_tasks). |
1e3c88bd PZ |
9698 | */ |
9699 | if (sd->balance_interval < sd->max_interval) | |
9700 | sd->balance_interval *= 2; | |
9701 | } | |
9702 | ||
1e3c88bd PZ |
9703 | goto out; |
9704 | ||
9705 | out_balanced: | |
afdeee05 VG |
9706 | /* |
9707 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
9708 | * constraints. Clear the imbalance flag only if other tasks got |
9709 | * a chance to move and fix the imbalance. | |
afdeee05 | 9710 | */ |
f6cad8df | 9711 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
9712 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
9713 | ||
9714 | if (*group_imbalance) | |
9715 | *group_imbalance = 0; | |
9716 | } | |
9717 | ||
9718 | out_all_pinned: | |
9719 | /* | |
9720 | * We reach balance because all tasks are pinned at this level so | |
9721 | * we can't migrate them. Let the imbalance flag set so parent level | |
9722 | * can try to migrate them. | |
9723 | */ | |
ae92882e | 9724 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
9725 | |
9726 | sd->nr_balance_failed = 0; | |
9727 | ||
9728 | out_one_pinned: | |
3f130a37 VS |
9729 | ld_moved = 0; |
9730 | ||
9731 | /* | |
5ba553ef PZ |
9732 | * newidle_balance() disregards balance intervals, so we could |
9733 | * repeatedly reach this code, which would lead to balance_interval | |
9734 | * skyrocketting in a short amount of time. Skip the balance_interval | |
9735 | * increase logic to avoid that. | |
3f130a37 VS |
9736 | */ |
9737 | if (env.idle == CPU_NEWLY_IDLE) | |
9738 | goto out; | |
9739 | ||
1e3c88bd | 9740 | /* tune up the balancing interval */ |
47b7aee1 VS |
9741 | if ((env.flags & LBF_ALL_PINNED && |
9742 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
9743 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 9744 | sd->balance_interval *= 2; |
1e3c88bd | 9745 | out: |
1e3c88bd PZ |
9746 | return ld_moved; |
9747 | } | |
9748 | ||
52a08ef1 JL |
9749 | static inline unsigned long |
9750 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
9751 | { | |
9752 | unsigned long interval = sd->balance_interval; | |
9753 | ||
9754 | if (cpu_busy) | |
9755 | interval *= sd->busy_factor; | |
9756 | ||
9757 | /* scale ms to jiffies */ | |
9758 | interval = msecs_to_jiffies(interval); | |
9759 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
9760 | ||
9761 | return interval; | |
9762 | } | |
9763 | ||
9764 | static inline void | |
31851a98 | 9765 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
9766 | { |
9767 | unsigned long interval, next; | |
9768 | ||
31851a98 LY |
9769 | /* used by idle balance, so cpu_busy = 0 */ |
9770 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
9771 | next = sd->last_balance + interval; |
9772 | ||
9773 | if (time_after(*next_balance, next)) | |
9774 | *next_balance = next; | |
9775 | } | |
9776 | ||
1e3c88bd | 9777 | /* |
97fb7a0a | 9778 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
9779 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
9780 | * least 1 task to be running on each physical CPU where possible, and | |
9781 | * avoids physical / logical imbalances. | |
1e3c88bd | 9782 | */ |
969c7921 | 9783 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 9784 | { |
969c7921 TH |
9785 | struct rq *busiest_rq = data; |
9786 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 9787 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 9788 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 9789 | struct sched_domain *sd; |
e5673f28 | 9790 | struct task_struct *p = NULL; |
8a8c69c3 | 9791 | struct rq_flags rf; |
969c7921 | 9792 | |
8a8c69c3 | 9793 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
9794 | /* |
9795 | * Between queueing the stop-work and running it is a hole in which | |
9796 | * CPUs can become inactive. We should not move tasks from or to | |
9797 | * inactive CPUs. | |
9798 | */ | |
9799 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
9800 | goto out_unlock; | |
969c7921 | 9801 | |
97fb7a0a | 9802 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
9803 | if (unlikely(busiest_cpu != smp_processor_id() || |
9804 | !busiest_rq->active_balance)) | |
9805 | goto out_unlock; | |
1e3c88bd PZ |
9806 | |
9807 | /* Is there any task to move? */ | |
9808 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 9809 | goto out_unlock; |
1e3c88bd PZ |
9810 | |
9811 | /* | |
9812 | * This condition is "impossible", if it occurs | |
9813 | * we need to fix it. Originally reported by | |
97fb7a0a | 9814 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd PZ |
9815 | */ |
9816 | BUG_ON(busiest_rq == target_rq); | |
9817 | ||
1e3c88bd | 9818 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 9819 | rcu_read_lock(); |
1e3c88bd | 9820 | for_each_domain(target_cpu, sd) { |
e669ac8a VS |
9821 | if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
9822 | break; | |
1e3c88bd PZ |
9823 | } |
9824 | ||
9825 | if (likely(sd)) { | |
8e45cb54 PZ |
9826 | struct lb_env env = { |
9827 | .sd = sd, | |
ddcdf6e7 PZ |
9828 | .dst_cpu = target_cpu, |
9829 | .dst_rq = target_rq, | |
9830 | .src_cpu = busiest_rq->cpu, | |
9831 | .src_rq = busiest_rq, | |
8e45cb54 | 9832 | .idle = CPU_IDLE, |
65a4433a JH |
9833 | /* |
9834 | * can_migrate_task() doesn't need to compute new_dst_cpu | |
9835 | * for active balancing. Since we have CPU_IDLE, but no | |
9836 | * @dst_grpmask we need to make that test go away with lying | |
9837 | * about DST_PINNED. | |
9838 | */ | |
9839 | .flags = LBF_DST_PINNED, | |
8e45cb54 PZ |
9840 | }; |
9841 | ||
ae92882e | 9842 | schedstat_inc(sd->alb_count); |
3bed5e21 | 9843 | update_rq_clock(busiest_rq); |
1e3c88bd | 9844 | |
e5673f28 | 9845 | p = detach_one_task(&env); |
d02c0711 | 9846 | if (p) { |
ae92882e | 9847 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
9848 | /* Active balancing done, reset the failure counter. */ |
9849 | sd->nr_balance_failed = 0; | |
9850 | } else { | |
ae92882e | 9851 | schedstat_inc(sd->alb_failed); |
d02c0711 | 9852 | } |
1e3c88bd | 9853 | } |
dce840a0 | 9854 | rcu_read_unlock(); |
969c7921 TH |
9855 | out_unlock: |
9856 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 9857 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
9858 | |
9859 | if (p) | |
9860 | attach_one_task(target_rq, p); | |
9861 | ||
9862 | local_irq_enable(); | |
9863 | ||
969c7921 | 9864 | return 0; |
1e3c88bd PZ |
9865 | } |
9866 | ||
af3fe03c PZ |
9867 | static DEFINE_SPINLOCK(balancing); |
9868 | ||
9869 | /* | |
9870 | * Scale the max load_balance interval with the number of CPUs in the system. | |
9871 | * This trades load-balance latency on larger machines for less cross talk. | |
9872 | */ | |
9873 | void update_max_interval(void) | |
9874 | { | |
9875 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
9876 | } | |
9877 | ||
9878 | /* | |
9879 | * It checks each scheduling domain to see if it is due to be balanced, | |
9880 | * and initiates a balancing operation if so. | |
9881 | * | |
9882 | * Balancing parameters are set up in init_sched_domains. | |
9883 | */ | |
9884 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
9885 | { | |
9886 | int continue_balancing = 1; | |
9887 | int cpu = rq->cpu; | |
323af6de | 9888 | int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
9889 | unsigned long interval; |
9890 | struct sched_domain *sd; | |
9891 | /* Earliest time when we have to do rebalance again */ | |
9892 | unsigned long next_balance = jiffies + 60*HZ; | |
9893 | int update_next_balance = 0; | |
9894 | int need_serialize, need_decay = 0; | |
9895 | u64 max_cost = 0; | |
9896 | ||
9897 | rcu_read_lock(); | |
9898 | for_each_domain(cpu, sd) { | |
9899 | /* | |
9900 | * Decay the newidle max times here because this is a regular | |
9901 | * visit to all the domains. Decay ~1% per second. | |
9902 | */ | |
9903 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
9904 | sd->max_newidle_lb_cost = | |
9905 | (sd->max_newidle_lb_cost * 253) / 256; | |
9906 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
9907 | need_decay = 1; | |
9908 | } | |
9909 | max_cost += sd->max_newidle_lb_cost; | |
9910 | ||
af3fe03c PZ |
9911 | /* |
9912 | * Stop the load balance at this level. There is another | |
9913 | * CPU in our sched group which is doing load balancing more | |
9914 | * actively. | |
9915 | */ | |
9916 | if (!continue_balancing) { | |
9917 | if (need_decay) | |
9918 | continue; | |
9919 | break; | |
9920 | } | |
9921 | ||
323af6de | 9922 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
9923 | |
9924 | need_serialize = sd->flags & SD_SERIALIZE; | |
9925 | if (need_serialize) { | |
9926 | if (!spin_trylock(&balancing)) | |
9927 | goto out; | |
9928 | } | |
9929 | ||
9930 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
9931 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
9932 | /* | |
9933 | * The LBF_DST_PINNED logic could have changed | |
9934 | * env->dst_cpu, so we can't know our idle | |
9935 | * state even if we migrated tasks. Update it. | |
9936 | */ | |
9937 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
323af6de | 9938 | busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
9939 | } |
9940 | sd->last_balance = jiffies; | |
323af6de | 9941 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
9942 | } |
9943 | if (need_serialize) | |
9944 | spin_unlock(&balancing); | |
9945 | out: | |
9946 | if (time_after(next_balance, sd->last_balance + interval)) { | |
9947 | next_balance = sd->last_balance + interval; | |
9948 | update_next_balance = 1; | |
9949 | } | |
9950 | } | |
9951 | if (need_decay) { | |
9952 | /* | |
9953 | * Ensure the rq-wide value also decays but keep it at a | |
9954 | * reasonable floor to avoid funnies with rq->avg_idle. | |
9955 | */ | |
9956 | rq->max_idle_balance_cost = | |
9957 | max((u64)sysctl_sched_migration_cost, max_cost); | |
9958 | } | |
9959 | rcu_read_unlock(); | |
9960 | ||
9961 | /* | |
9962 | * next_balance will be updated only when there is a need. | |
9963 | * When the cpu is attached to null domain for ex, it will not be | |
9964 | * updated. | |
9965 | */ | |
9966 | if (likely(update_next_balance)) { | |
9967 | rq->next_balance = next_balance; | |
9968 | ||
9969 | #ifdef CONFIG_NO_HZ_COMMON | |
9970 | /* | |
9971 | * If this CPU has been elected to perform the nohz idle | |
9972 | * balance. Other idle CPUs have already rebalanced with | |
9973 | * nohz_idle_balance() and nohz.next_balance has been | |
9974 | * updated accordingly. This CPU is now running the idle load | |
9975 | * balance for itself and we need to update the | |
9976 | * nohz.next_balance accordingly. | |
9977 | */ | |
9978 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
9979 | nohz.next_balance = rq->next_balance; | |
9980 | #endif | |
9981 | } | |
9982 | } | |
9983 | ||
d987fc7f MG |
9984 | static inline int on_null_domain(struct rq *rq) |
9985 | { | |
9986 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
9987 | } | |
9988 | ||
3451d024 | 9989 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
9990 | /* |
9991 | * idle load balancing details | |
83cd4fe2 VP |
9992 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
9993 | * needed, they will kick the idle load balancer, which then does idle | |
9994 | * load balancing for all the idle CPUs. | |
9b019acb NP |
9995 | * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set |
9996 | * anywhere yet. | |
83cd4fe2 | 9997 | */ |
1e3c88bd | 9998 | |
3dd0337d | 9999 | static inline int find_new_ilb(void) |
1e3c88bd | 10000 | { |
9b019acb | 10001 | int ilb; |
1e3c88bd | 10002 | |
9b019acb NP |
10003 | for_each_cpu_and(ilb, nohz.idle_cpus_mask, |
10004 | housekeeping_cpumask(HK_FLAG_MISC)) { | |
10005 | if (idle_cpu(ilb)) | |
10006 | return ilb; | |
10007 | } | |
786d6dc7 SS |
10008 | |
10009 | return nr_cpu_ids; | |
1e3c88bd | 10010 | } |
1e3c88bd | 10011 | |
83cd4fe2 | 10012 | /* |
9b019acb NP |
10013 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick any |
10014 | * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one). | |
83cd4fe2 | 10015 | */ |
a4064fb6 | 10016 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
10017 | { |
10018 | int ilb_cpu; | |
10019 | ||
10020 | nohz.next_balance++; | |
10021 | ||
3dd0337d | 10022 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 10023 | |
0b005cf5 SS |
10024 | if (ilb_cpu >= nr_cpu_ids) |
10025 | return; | |
83cd4fe2 | 10026 | |
a4064fb6 | 10027 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 10028 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 10029 | return; |
4550487a | 10030 | |
1c792db7 | 10031 | /* |
90b5363a | 10032 | * This way we generate an IPI on the target CPU which |
1c792db7 SS |
10033 | * is idle. And the softirq performing nohz idle load balance |
10034 | * will be run before returning from the IPI. | |
10035 | */ | |
90b5363a | 10036 | smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); |
4550487a PZ |
10037 | } |
10038 | ||
10039 | /* | |
9f132742 VS |
10040 | * Current decision point for kicking the idle load balancer in the presence |
10041 | * of idle CPUs in the system. | |
4550487a PZ |
10042 | */ |
10043 | static void nohz_balancer_kick(struct rq *rq) | |
10044 | { | |
10045 | unsigned long now = jiffies; | |
10046 | struct sched_domain_shared *sds; | |
10047 | struct sched_domain *sd; | |
10048 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 10049 | unsigned int flags = 0; |
4550487a PZ |
10050 | |
10051 | if (unlikely(rq->idle_balance)) | |
10052 | return; | |
10053 | ||
10054 | /* | |
10055 | * We may be recently in ticked or tickless idle mode. At the first | |
10056 | * busy tick after returning from idle, we will update the busy stats. | |
10057 | */ | |
00357f5e | 10058 | nohz_balance_exit_idle(rq); |
4550487a PZ |
10059 | |
10060 | /* | |
10061 | * None are in tickless mode and hence no need for NOHZ idle load | |
10062 | * balancing. | |
10063 | */ | |
10064 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
10065 | return; | |
10066 | ||
f643ea22 VG |
10067 | if (READ_ONCE(nohz.has_blocked) && |
10068 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
10069 | flags = NOHZ_STATS_KICK; |
10070 | ||
4550487a | 10071 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 10072 | goto out; |
4550487a | 10073 | |
a0fe2cf0 | 10074 | if (rq->nr_running >= 2) { |
a4064fb6 | 10075 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10076 | goto out; |
10077 | } | |
10078 | ||
10079 | rcu_read_lock(); | |
4550487a PZ |
10080 | |
10081 | sd = rcu_dereference(rq->sd); | |
10082 | if (sd) { | |
e25a7a94 VS |
10083 | /* |
10084 | * If there's a CFS task and the current CPU has reduced | |
10085 | * capacity; kick the ILB to see if there's a better CPU to run | |
10086 | * on. | |
10087 | */ | |
10088 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
a4064fb6 | 10089 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10090 | goto unlock; |
10091 | } | |
10092 | } | |
10093 | ||
011b27bb | 10094 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 10095 | if (sd) { |
b9a7b883 VS |
10096 | /* |
10097 | * When ASYM_PACKING; see if there's a more preferred CPU | |
10098 | * currently idle; in which case, kick the ILB to move tasks | |
10099 | * around. | |
10100 | */ | |
7edab78d | 10101 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
4550487a | 10102 | if (sched_asym_prefer(i, cpu)) { |
a4064fb6 | 10103 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
10104 | goto unlock; |
10105 | } | |
10106 | } | |
10107 | } | |
b9a7b883 | 10108 | |
a0fe2cf0 VS |
10109 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
10110 | if (sd) { | |
10111 | /* | |
10112 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
10113 | * to run the misfit task on. | |
10114 | */ | |
10115 | if (check_misfit_status(rq, sd)) { | |
10116 | flags = NOHZ_KICK_MASK; | |
10117 | goto unlock; | |
10118 | } | |
b9a7b883 VS |
10119 | |
10120 | /* | |
10121 | * For asymmetric systems, we do not want to nicely balance | |
10122 | * cache use, instead we want to embrace asymmetry and only | |
10123 | * ensure tasks have enough CPU capacity. | |
10124 | * | |
10125 | * Skip the LLC logic because it's not relevant in that case. | |
10126 | */ | |
10127 | goto unlock; | |
a0fe2cf0 VS |
10128 | } |
10129 | ||
b9a7b883 VS |
10130 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
10131 | if (sds) { | |
e25a7a94 | 10132 | /* |
b9a7b883 VS |
10133 | * If there is an imbalance between LLC domains (IOW we could |
10134 | * increase the overall cache use), we need some less-loaded LLC | |
10135 | * domain to pull some load. Likewise, we may need to spread | |
10136 | * load within the current LLC domain (e.g. packed SMT cores but | |
10137 | * other CPUs are idle). We can't really know from here how busy | |
10138 | * the others are - so just get a nohz balance going if it looks | |
10139 | * like this LLC domain has tasks we could move. | |
e25a7a94 | 10140 | */ |
b9a7b883 VS |
10141 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
10142 | if (nr_busy > 1) { | |
10143 | flags = NOHZ_KICK_MASK; | |
10144 | goto unlock; | |
4550487a PZ |
10145 | } |
10146 | } | |
10147 | unlock: | |
10148 | rcu_read_unlock(); | |
10149 | out: | |
a4064fb6 PZ |
10150 | if (flags) |
10151 | kick_ilb(flags); | |
83cd4fe2 VP |
10152 | } |
10153 | ||
00357f5e | 10154 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 10155 | { |
00357f5e | 10156 | struct sched_domain *sd; |
a22e47a4 | 10157 | |
00357f5e PZ |
10158 | rcu_read_lock(); |
10159 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 10160 | |
00357f5e PZ |
10161 | if (!sd || !sd->nohz_idle) |
10162 | goto unlock; | |
10163 | sd->nohz_idle = 0; | |
10164 | ||
10165 | atomic_inc(&sd->shared->nr_busy_cpus); | |
10166 | unlock: | |
10167 | rcu_read_unlock(); | |
71325960 SS |
10168 | } |
10169 | ||
00357f5e PZ |
10170 | void nohz_balance_exit_idle(struct rq *rq) |
10171 | { | |
10172 | SCHED_WARN_ON(rq != this_rq()); | |
10173 | ||
10174 | if (likely(!rq->nohz_tick_stopped)) | |
10175 | return; | |
10176 | ||
10177 | rq->nohz_tick_stopped = 0; | |
10178 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
10179 | atomic_dec(&nohz.nr_cpus); | |
10180 | ||
10181 | set_cpu_sd_state_busy(rq->cpu); | |
10182 | } | |
10183 | ||
10184 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
10185 | { |
10186 | struct sched_domain *sd; | |
69e1e811 | 10187 | |
69e1e811 | 10188 | rcu_read_lock(); |
0e369d75 | 10189 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
10190 | |
10191 | if (!sd || sd->nohz_idle) | |
10192 | goto unlock; | |
10193 | sd->nohz_idle = 1; | |
10194 | ||
0e369d75 | 10195 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 10196 | unlock: |
69e1e811 SS |
10197 | rcu_read_unlock(); |
10198 | } | |
10199 | ||
1e3c88bd | 10200 | /* |
97fb7a0a | 10201 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 10202 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 10203 | */ |
c1cc017c | 10204 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 10205 | { |
00357f5e PZ |
10206 | struct rq *rq = cpu_rq(cpu); |
10207 | ||
10208 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
10209 | ||
97fb7a0a | 10210 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
10211 | if (!cpu_active(cpu)) |
10212 | return; | |
10213 | ||
387bc8b5 | 10214 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
de201559 | 10215 | if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) |
387bc8b5 FW |
10216 | return; |
10217 | ||
f643ea22 VG |
10218 | /* |
10219 | * Can be set safely without rq->lock held | |
10220 | * If a clear happens, it will have evaluated last additions because | |
10221 | * rq->lock is held during the check and the clear | |
10222 | */ | |
10223 | rq->has_blocked_load = 1; | |
10224 | ||
10225 | /* | |
10226 | * The tick is still stopped but load could have been added in the | |
10227 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
10228 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
10229 | * of nohz.has_blocked can only happen after checking the new load | |
10230 | */ | |
00357f5e | 10231 | if (rq->nohz_tick_stopped) |
f643ea22 | 10232 | goto out; |
1e3c88bd | 10233 | |
97fb7a0a | 10234 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 10235 | if (on_null_domain(rq)) |
d987fc7f MG |
10236 | return; |
10237 | ||
00357f5e PZ |
10238 | rq->nohz_tick_stopped = 1; |
10239 | ||
c1cc017c AS |
10240 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
10241 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 10242 | |
f643ea22 VG |
10243 | /* |
10244 | * Ensures that if nohz_idle_balance() fails to observe our | |
10245 | * @idle_cpus_mask store, it must observe the @has_blocked | |
10246 | * store. | |
10247 | */ | |
10248 | smp_mb__after_atomic(); | |
10249 | ||
00357f5e | 10250 | set_cpu_sd_state_idle(cpu); |
f643ea22 VG |
10251 | |
10252 | out: | |
10253 | /* | |
10254 | * Each time a cpu enter idle, we assume that it has blocked load and | |
10255 | * enable the periodic update of the load of idle cpus | |
10256 | */ | |
10257 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 10258 | } |
1e3c88bd | 10259 | |
1e3c88bd | 10260 | /* |
31e77c93 VG |
10261 | * Internal function that runs load balance for all idle cpus. The load balance |
10262 | * can be a simple update of blocked load or a complete load balance with | |
10263 | * tasks movement depending of flags. | |
10264 | * The function returns false if the loop has stopped before running | |
10265 | * through all idle CPUs. | |
1e3c88bd | 10266 | */ |
31e77c93 VG |
10267 | static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags, |
10268 | enum cpu_idle_type idle) | |
83cd4fe2 | 10269 | { |
c5afb6a8 | 10270 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
10271 | unsigned long now = jiffies; |
10272 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 10273 | bool has_blocked_load = false; |
c5afb6a8 | 10274 | int update_next_balance = 0; |
b7031a02 | 10275 | int this_cpu = this_rq->cpu; |
b7031a02 | 10276 | int balance_cpu; |
31e77c93 | 10277 | int ret = false; |
b7031a02 | 10278 | struct rq *rq; |
83cd4fe2 | 10279 | |
b7031a02 | 10280 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 10281 | |
f643ea22 VG |
10282 | /* |
10283 | * We assume there will be no idle load after this update and clear | |
10284 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
10285 | * set the has_blocked flag and trig another update of idle load. | |
10286 | * Because a cpu that becomes idle, is added to idle_cpus_mask before | |
10287 | * setting the flag, we are sure to not clear the state and not | |
10288 | * check the load of an idle cpu. | |
10289 | */ | |
10290 | WRITE_ONCE(nohz.has_blocked, 0); | |
10291 | ||
10292 | /* | |
10293 | * Ensures that if we miss the CPU, we must see the has_blocked | |
10294 | * store from nohz_balance_enter_idle(). | |
10295 | */ | |
10296 | smp_mb(); | |
10297 | ||
83cd4fe2 | 10298 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { |
8a6d42d1 | 10299 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
10300 | continue; |
10301 | ||
10302 | /* | |
97fb7a0a IM |
10303 | * If this CPU gets work to do, stop the load balancing |
10304 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
10305 | * balancing owner will pick it up. |
10306 | */ | |
f643ea22 VG |
10307 | if (need_resched()) { |
10308 | has_blocked_load = true; | |
10309 | goto abort; | |
10310 | } | |
83cd4fe2 | 10311 | |
5ed4f1d9 VG |
10312 | rq = cpu_rq(balance_cpu); |
10313 | ||
63928384 | 10314 | has_blocked_load |= update_nohz_stats(rq, true); |
f643ea22 | 10315 | |
ed61bbc6 TC |
10316 | /* |
10317 | * If time for next balance is due, | |
10318 | * do the balance. | |
10319 | */ | |
10320 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
10321 | struct rq_flags rf; |
10322 | ||
31e77c93 | 10323 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 10324 | update_rq_clock(rq); |
31e77c93 | 10325 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 10326 | |
b7031a02 PZ |
10327 | if (flags & NOHZ_BALANCE_KICK) |
10328 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 10329 | } |
83cd4fe2 | 10330 | |
c5afb6a8 VG |
10331 | if (time_after(next_balance, rq->next_balance)) { |
10332 | next_balance = rq->next_balance; | |
10333 | update_next_balance = 1; | |
10334 | } | |
83cd4fe2 | 10335 | } |
c5afb6a8 | 10336 | |
31e77c93 VG |
10337 | /* Newly idle CPU doesn't need an update */ |
10338 | if (idle != CPU_NEWLY_IDLE) { | |
10339 | update_blocked_averages(this_cpu); | |
10340 | has_blocked_load |= this_rq->has_blocked_load; | |
10341 | } | |
10342 | ||
b7031a02 PZ |
10343 | if (flags & NOHZ_BALANCE_KICK) |
10344 | rebalance_domains(this_rq, CPU_IDLE); | |
10345 | ||
f643ea22 VG |
10346 | WRITE_ONCE(nohz.next_blocked, |
10347 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
10348 | ||
31e77c93 VG |
10349 | /* The full idle balance loop has been done */ |
10350 | ret = true; | |
10351 | ||
f643ea22 VG |
10352 | abort: |
10353 | /* There is still blocked load, enable periodic update */ | |
10354 | if (has_blocked_load) | |
10355 | WRITE_ONCE(nohz.has_blocked, 1); | |
a4064fb6 | 10356 | |
c5afb6a8 VG |
10357 | /* |
10358 | * next_balance will be updated only when there is a need. | |
10359 | * When the CPU is attached to null domain for ex, it will not be | |
10360 | * updated. | |
10361 | */ | |
10362 | if (likely(update_next_balance)) | |
10363 | nohz.next_balance = next_balance; | |
b7031a02 | 10364 | |
31e77c93 VG |
10365 | return ret; |
10366 | } | |
10367 | ||
10368 | /* | |
10369 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
10370 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
10371 | */ | |
10372 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
10373 | { | |
10374 | int this_cpu = this_rq->cpu; | |
10375 | unsigned int flags; | |
10376 | ||
10377 | if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK)) | |
10378 | return false; | |
10379 | ||
10380 | if (idle != CPU_IDLE) { | |
10381 | atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu)); | |
10382 | return false; | |
10383 | } | |
10384 | ||
80eb8657 | 10385 | /* could be _relaxed() */ |
31e77c93 VG |
10386 | flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu)); |
10387 | if (!(flags & NOHZ_KICK_MASK)) | |
10388 | return false; | |
10389 | ||
10390 | _nohz_idle_balance(this_rq, flags, idle); | |
10391 | ||
b7031a02 | 10392 | return true; |
83cd4fe2 | 10393 | } |
31e77c93 VG |
10394 | |
10395 | static void nohz_newidle_balance(struct rq *this_rq) | |
10396 | { | |
10397 | int this_cpu = this_rq->cpu; | |
10398 | ||
10399 | /* | |
10400 | * This CPU doesn't want to be disturbed by scheduler | |
10401 | * housekeeping | |
10402 | */ | |
10403 | if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) | |
10404 | return; | |
10405 | ||
10406 | /* Will wake up very soon. No time for doing anything else*/ | |
10407 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
10408 | return; | |
10409 | ||
10410 | /* Don't need to update blocked load of idle CPUs*/ | |
10411 | if (!READ_ONCE(nohz.has_blocked) || | |
10412 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
10413 | return; | |
10414 | ||
10415 | raw_spin_unlock(&this_rq->lock); | |
10416 | /* | |
10417 | * This CPU is going to be idle and blocked load of idle CPUs | |
10418 | * need to be updated. Run the ilb locally as it is a good | |
10419 | * candidate for ilb instead of waking up another idle CPU. | |
10420 | * Kick an normal ilb if we failed to do the update. | |
10421 | */ | |
10422 | if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE)) | |
10423 | kick_ilb(NOHZ_STATS_KICK); | |
10424 | raw_spin_lock(&this_rq->lock); | |
10425 | } | |
10426 | ||
dd707247 PZ |
10427 | #else /* !CONFIG_NO_HZ_COMMON */ |
10428 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
10429 | ||
31e77c93 | 10430 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
10431 | { |
10432 | return false; | |
10433 | } | |
31e77c93 VG |
10434 | |
10435 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 10436 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 10437 | |
47ea5412 PZ |
10438 | /* |
10439 | * idle_balance is called by schedule() if this_cpu is about to become | |
10440 | * idle. Attempts to pull tasks from other CPUs. | |
7277a34c PZ |
10441 | * |
10442 | * Returns: | |
10443 | * < 0 - we released the lock and there are !fair tasks present | |
10444 | * 0 - failed, no new tasks | |
10445 | * > 0 - success, new (fair) tasks present | |
47ea5412 | 10446 | */ |
d91cecc1 | 10447 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) |
47ea5412 PZ |
10448 | { |
10449 | unsigned long next_balance = jiffies + HZ; | |
10450 | int this_cpu = this_rq->cpu; | |
10451 | struct sched_domain *sd; | |
10452 | int pulled_task = 0; | |
10453 | u64 curr_cost = 0; | |
10454 | ||
5ba553ef | 10455 | update_misfit_status(NULL, this_rq); |
47ea5412 PZ |
10456 | /* |
10457 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
10458 | * measure the duration of idle_balance() as idle time. | |
10459 | */ | |
10460 | this_rq->idle_stamp = rq_clock(this_rq); | |
10461 | ||
10462 | /* | |
10463 | * Do not pull tasks towards !active CPUs... | |
10464 | */ | |
10465 | if (!cpu_active(this_cpu)) | |
10466 | return 0; | |
10467 | ||
10468 | /* | |
10469 | * This is OK, because current is on_cpu, which avoids it being picked | |
10470 | * for load-balance and preemption/IRQs are still disabled avoiding | |
10471 | * further scheduler activity on it and we're being very careful to | |
10472 | * re-start the picking loop. | |
10473 | */ | |
10474 | rq_unpin_lock(this_rq, rf); | |
10475 | ||
10476 | if (this_rq->avg_idle < sysctl_sched_migration_cost || | |
e90c8fe1 | 10477 | !READ_ONCE(this_rq->rd->overload)) { |
31e77c93 | 10478 | |
47ea5412 PZ |
10479 | rcu_read_lock(); |
10480 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
10481 | if (sd) | |
10482 | update_next_balance(sd, &next_balance); | |
10483 | rcu_read_unlock(); | |
10484 | ||
31e77c93 VG |
10485 | nohz_newidle_balance(this_rq); |
10486 | ||
47ea5412 PZ |
10487 | goto out; |
10488 | } | |
10489 | ||
10490 | raw_spin_unlock(&this_rq->lock); | |
10491 | ||
10492 | update_blocked_averages(this_cpu); | |
10493 | rcu_read_lock(); | |
10494 | for_each_domain(this_cpu, sd) { | |
10495 | int continue_balancing = 1; | |
10496 | u64 t0, domain_cost; | |
10497 | ||
47ea5412 PZ |
10498 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
10499 | update_next_balance(sd, &next_balance); | |
10500 | break; | |
10501 | } | |
10502 | ||
10503 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
10504 | t0 = sched_clock_cpu(this_cpu); | |
10505 | ||
10506 | pulled_task = load_balance(this_cpu, this_rq, | |
10507 | sd, CPU_NEWLY_IDLE, | |
10508 | &continue_balancing); | |
10509 | ||
10510 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
10511 | if (domain_cost > sd->max_newidle_lb_cost) | |
10512 | sd->max_newidle_lb_cost = domain_cost; | |
10513 | ||
10514 | curr_cost += domain_cost; | |
10515 | } | |
10516 | ||
10517 | update_next_balance(sd, &next_balance); | |
10518 | ||
10519 | /* | |
10520 | * Stop searching for tasks to pull if there are | |
10521 | * now runnable tasks on this rq. | |
10522 | */ | |
10523 | if (pulled_task || this_rq->nr_running > 0) | |
10524 | break; | |
10525 | } | |
10526 | rcu_read_unlock(); | |
10527 | ||
10528 | raw_spin_lock(&this_rq->lock); | |
10529 | ||
10530 | if (curr_cost > this_rq->max_idle_balance_cost) | |
10531 | this_rq->max_idle_balance_cost = curr_cost; | |
10532 | ||
457be908 | 10533 | out: |
47ea5412 PZ |
10534 | /* |
10535 | * While browsing the domains, we released the rq lock, a task could | |
10536 | * have been enqueued in the meantime. Since we're not going idle, | |
10537 | * pretend we pulled a task. | |
10538 | */ | |
10539 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
10540 | pulled_task = 1; | |
10541 | ||
47ea5412 PZ |
10542 | /* Move the next balance forward */ |
10543 | if (time_after(this_rq->next_balance, next_balance)) | |
10544 | this_rq->next_balance = next_balance; | |
10545 | ||
10546 | /* Is there a task of a high priority class? */ | |
10547 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
10548 | pulled_task = -1; | |
10549 | ||
10550 | if (pulled_task) | |
10551 | this_rq->idle_stamp = 0; | |
10552 | ||
10553 | rq_repin_lock(this_rq, rf); | |
10554 | ||
10555 | return pulled_task; | |
10556 | } | |
10557 | ||
83cd4fe2 VP |
10558 | /* |
10559 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
10560 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
10561 | */ | |
0766f788 | 10562 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 10563 | { |
208cb16b | 10564 | struct rq *this_rq = this_rq(); |
6eb57e0d | 10565 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
10566 | CPU_IDLE : CPU_NOT_IDLE; |
10567 | ||
1e3c88bd | 10568 | /* |
97fb7a0a IM |
10569 | * If this CPU has a pending nohz_balance_kick, then do the |
10570 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 10571 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 10572 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
10573 | * load balance only within the local sched_domain hierarchy |
10574 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 10575 | */ |
b7031a02 PZ |
10576 | if (nohz_idle_balance(this_rq, idle)) |
10577 | return; | |
10578 | ||
10579 | /* normal load balance */ | |
10580 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 10581 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
10582 | } |
10583 | ||
1e3c88bd PZ |
10584 | /* |
10585 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 10586 | */ |
7caff66f | 10587 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 10588 | { |
1e3c88bd | 10589 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
10590 | if (unlikely(on_null_domain(rq))) |
10591 | return; | |
10592 | ||
10593 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 10594 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
10595 | |
10596 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
10597 | } |
10598 | ||
0bcdcf28 CE |
10599 | static void rq_online_fair(struct rq *rq) |
10600 | { | |
10601 | update_sysctl(); | |
0e59bdae KT |
10602 | |
10603 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
10604 | } |
10605 | ||
10606 | static void rq_offline_fair(struct rq *rq) | |
10607 | { | |
10608 | update_sysctl(); | |
a4c96ae3 PB |
10609 | |
10610 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
10611 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
10612 | } |
10613 | ||
55e12e5e | 10614 | #endif /* CONFIG_SMP */ |
e1d1484f | 10615 | |
bf0f6f24 | 10616 | /* |
d84b3131 FW |
10617 | * scheduler tick hitting a task of our scheduling class. |
10618 | * | |
10619 | * NOTE: This function can be called remotely by the tick offload that | |
10620 | * goes along full dynticks. Therefore no local assumption can be made | |
10621 | * and everything must be accessed through the @rq and @curr passed in | |
10622 | * parameters. | |
bf0f6f24 | 10623 | */ |
8f4d37ec | 10624 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
10625 | { |
10626 | struct cfs_rq *cfs_rq; | |
10627 | struct sched_entity *se = &curr->se; | |
10628 | ||
10629 | for_each_sched_entity(se) { | |
10630 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 10631 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 10632 | } |
18bf2805 | 10633 | |
b52da86e | 10634 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 10635 | task_tick_numa(rq, curr); |
3b1baa64 MR |
10636 | |
10637 | update_misfit_status(curr, rq); | |
2802bf3c | 10638 | update_overutilized_status(task_rq(curr)); |
bf0f6f24 IM |
10639 | } |
10640 | ||
10641 | /* | |
cd29fe6f PZ |
10642 | * called on fork with the child task as argument from the parent's context |
10643 | * - child not yet on the tasklist | |
10644 | * - preemption disabled | |
bf0f6f24 | 10645 | */ |
cd29fe6f | 10646 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 10647 | { |
4fc420c9 DN |
10648 | struct cfs_rq *cfs_rq; |
10649 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 10650 | struct rq *rq = this_rq(); |
8a8c69c3 | 10651 | struct rq_flags rf; |
bf0f6f24 | 10652 | |
8a8c69c3 | 10653 | rq_lock(rq, &rf); |
861d034e PZ |
10654 | update_rq_clock(rq); |
10655 | ||
4fc420c9 DN |
10656 | cfs_rq = task_cfs_rq(current); |
10657 | curr = cfs_rq->curr; | |
e210bffd PZ |
10658 | if (curr) { |
10659 | update_curr(cfs_rq); | |
b5d9d734 | 10660 | se->vruntime = curr->vruntime; |
e210bffd | 10661 | } |
aeb73b04 | 10662 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 10663 | |
cd29fe6f | 10664 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 10665 | /* |
edcb60a3 IM |
10666 | * Upon rescheduling, sched_class::put_prev_task() will place |
10667 | * 'current' within the tree based on its new key value. | |
10668 | */ | |
4d78e7b6 | 10669 | swap(curr->vruntime, se->vruntime); |
8875125e | 10670 | resched_curr(rq); |
4d78e7b6 | 10671 | } |
bf0f6f24 | 10672 | |
88ec22d3 | 10673 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 10674 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
10675 | } |
10676 | ||
cb469845 SR |
10677 | /* |
10678 | * Priority of the task has changed. Check to see if we preempt | |
10679 | * the current task. | |
10680 | */ | |
da7a735e PZ |
10681 | static void |
10682 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 10683 | { |
da0c1e65 | 10684 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
10685 | return; |
10686 | ||
7c2e8bbd FW |
10687 | if (rq->cfs.nr_running == 1) |
10688 | return; | |
10689 | ||
cb469845 SR |
10690 | /* |
10691 | * Reschedule if we are currently running on this runqueue and | |
10692 | * our priority decreased, or if we are not currently running on | |
10693 | * this runqueue and our priority is higher than the current's | |
10694 | */ | |
da7a735e | 10695 | if (rq->curr == p) { |
cb469845 | 10696 | if (p->prio > oldprio) |
8875125e | 10697 | resched_curr(rq); |
cb469845 | 10698 | } else |
15afe09b | 10699 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
10700 | } |
10701 | ||
daa59407 | 10702 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
10703 | { |
10704 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
10705 | |
10706 | /* | |
daa59407 BP |
10707 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
10708 | * the dequeue_entity(.flags=0) will already have normalized the | |
10709 | * vruntime. | |
10710 | */ | |
10711 | if (p->on_rq) | |
10712 | return true; | |
10713 | ||
10714 | /* | |
10715 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
10716 | * But there are some cases where it has already been normalized: | |
da7a735e | 10717 | * |
daa59407 BP |
10718 | * - A forked child which is waiting for being woken up by |
10719 | * wake_up_new_task(). | |
10720 | * - A task which has been woken up by try_to_wake_up() and | |
10721 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 10722 | */ |
d0cdb3ce SM |
10723 | if (!se->sum_exec_runtime || |
10724 | (p->state == TASK_WAKING && p->sched_remote_wakeup)) | |
daa59407 BP |
10725 | return true; |
10726 | ||
10727 | return false; | |
10728 | } | |
10729 | ||
09a43ace VG |
10730 | #ifdef CONFIG_FAIR_GROUP_SCHED |
10731 | /* | |
10732 | * Propagate the changes of the sched_entity across the tg tree to make it | |
10733 | * visible to the root | |
10734 | */ | |
10735 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
10736 | { | |
10737 | struct cfs_rq *cfs_rq; | |
10738 | ||
10739 | /* Start to propagate at parent */ | |
10740 | se = se->parent; | |
10741 | ||
10742 | for_each_sched_entity(se) { | |
10743 | cfs_rq = cfs_rq_of(se); | |
10744 | ||
10745 | if (cfs_rq_throttled(cfs_rq)) | |
10746 | break; | |
10747 | ||
88c0616e | 10748 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace VG |
10749 | } |
10750 | } | |
10751 | #else | |
10752 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
10753 | #endif | |
10754 | ||
df217913 | 10755 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 10756 | { |
daa59407 BP |
10757 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
10758 | ||
9d89c257 | 10759 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 10760 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 10761 | detach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 10762 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 10763 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
10764 | } |
10765 | ||
df217913 | 10766 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 10767 | { |
daa59407 | 10768 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
10769 | |
10770 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
10771 | /* |
10772 | * Since the real-depth could have been changed (only FAIR | |
10773 | * class maintain depth value), reset depth properly. | |
10774 | */ | |
10775 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10776 | #endif | |
7855a35a | 10777 | |
df217913 | 10778 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 10779 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
a4f9a0e5 | 10780 | attach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 10781 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 10782 | propagate_entity_cfs_rq(se); |
df217913 VG |
10783 | } |
10784 | ||
10785 | static void detach_task_cfs_rq(struct task_struct *p) | |
10786 | { | |
10787 | struct sched_entity *se = &p->se; | |
10788 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10789 | ||
10790 | if (!vruntime_normalized(p)) { | |
10791 | /* | |
10792 | * Fix up our vruntime so that the current sleep doesn't | |
10793 | * cause 'unlimited' sleep bonus. | |
10794 | */ | |
10795 | place_entity(cfs_rq, se, 0); | |
10796 | se->vruntime -= cfs_rq->min_vruntime; | |
10797 | } | |
10798 | ||
10799 | detach_entity_cfs_rq(se); | |
10800 | } | |
10801 | ||
10802 | static void attach_task_cfs_rq(struct task_struct *p) | |
10803 | { | |
10804 | struct sched_entity *se = &p->se; | |
10805 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10806 | ||
10807 | attach_entity_cfs_rq(se); | |
daa59407 BP |
10808 | |
10809 | if (!vruntime_normalized(p)) | |
10810 | se->vruntime += cfs_rq->min_vruntime; | |
10811 | } | |
6efdb105 | 10812 | |
daa59407 BP |
10813 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
10814 | { | |
10815 | detach_task_cfs_rq(p); | |
10816 | } | |
10817 | ||
10818 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
10819 | { | |
10820 | attach_task_cfs_rq(p); | |
7855a35a | 10821 | |
daa59407 | 10822 | if (task_on_rq_queued(p)) { |
7855a35a | 10823 | /* |
daa59407 BP |
10824 | * We were most likely switched from sched_rt, so |
10825 | * kick off the schedule if running, otherwise just see | |
10826 | * if we can still preempt the current task. | |
7855a35a | 10827 | */ |
daa59407 BP |
10828 | if (rq->curr == p) |
10829 | resched_curr(rq); | |
10830 | else | |
10831 | check_preempt_curr(rq, p, 0); | |
7855a35a | 10832 | } |
cb469845 SR |
10833 | } |
10834 | ||
83b699ed SV |
10835 | /* Account for a task changing its policy or group. |
10836 | * | |
10837 | * This routine is mostly called to set cfs_rq->curr field when a task | |
10838 | * migrates between groups/classes. | |
10839 | */ | |
a0e813f2 | 10840 | static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) |
83b699ed | 10841 | { |
03b7fad1 PZ |
10842 | struct sched_entity *se = &p->se; |
10843 | ||
10844 | #ifdef CONFIG_SMP | |
10845 | if (task_on_rq_queued(p)) { | |
10846 | /* | |
10847 | * Move the next running task to the front of the list, so our | |
10848 | * cfs_tasks list becomes MRU one. | |
10849 | */ | |
10850 | list_move(&se->group_node, &rq->cfs_tasks); | |
10851 | } | |
10852 | #endif | |
83b699ed | 10853 | |
ec12cb7f PT |
10854 | for_each_sched_entity(se) { |
10855 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10856 | ||
10857 | set_next_entity(cfs_rq, se); | |
10858 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
10859 | account_cfs_rq_runtime(cfs_rq, 0); | |
10860 | } | |
83b699ed SV |
10861 | } |
10862 | ||
029632fb PZ |
10863 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
10864 | { | |
bfb06889 | 10865 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
029632fb PZ |
10866 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
10867 | #ifndef CONFIG_64BIT | |
10868 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
10869 | #endif | |
141965c7 | 10870 | #ifdef CONFIG_SMP |
2a2f5d4e | 10871 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 10872 | #endif |
029632fb PZ |
10873 | } |
10874 | ||
810b3817 | 10875 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
10876 | static void task_set_group_fair(struct task_struct *p) |
10877 | { | |
10878 | struct sched_entity *se = &p->se; | |
10879 | ||
10880 | set_task_rq(p, task_cpu(p)); | |
10881 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10882 | } | |
10883 | ||
bc54da21 | 10884 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 10885 | { |
daa59407 | 10886 | detach_task_cfs_rq(p); |
b2b5ce02 | 10887 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
10888 | |
10889 | #ifdef CONFIG_SMP | |
10890 | /* Tell se's cfs_rq has been changed -- migrated */ | |
10891 | p->se.avg.last_update_time = 0; | |
10892 | #endif | |
daa59407 | 10893 | attach_task_cfs_rq(p); |
810b3817 | 10894 | } |
029632fb | 10895 | |
ea86cb4b VG |
10896 | static void task_change_group_fair(struct task_struct *p, int type) |
10897 | { | |
10898 | switch (type) { | |
10899 | case TASK_SET_GROUP: | |
10900 | task_set_group_fair(p); | |
10901 | break; | |
10902 | ||
10903 | case TASK_MOVE_GROUP: | |
10904 | task_move_group_fair(p); | |
10905 | break; | |
10906 | } | |
10907 | } | |
10908 | ||
029632fb PZ |
10909 | void free_fair_sched_group(struct task_group *tg) |
10910 | { | |
10911 | int i; | |
10912 | ||
10913 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
10914 | ||
10915 | for_each_possible_cpu(i) { | |
10916 | if (tg->cfs_rq) | |
10917 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 10918 | if (tg->se) |
029632fb PZ |
10919 | kfree(tg->se[i]); |
10920 | } | |
10921 | ||
10922 | kfree(tg->cfs_rq); | |
10923 | kfree(tg->se); | |
10924 | } | |
10925 | ||
10926 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
10927 | { | |
029632fb | 10928 | struct sched_entity *se; |
b7fa30c9 | 10929 | struct cfs_rq *cfs_rq; |
029632fb PZ |
10930 | int i; |
10931 | ||
6396bb22 | 10932 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
10933 | if (!tg->cfs_rq) |
10934 | goto err; | |
6396bb22 | 10935 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
10936 | if (!tg->se) |
10937 | goto err; | |
10938 | ||
10939 | tg->shares = NICE_0_LOAD; | |
10940 | ||
10941 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
10942 | ||
10943 | for_each_possible_cpu(i) { | |
10944 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
10945 | GFP_KERNEL, cpu_to_node(i)); | |
10946 | if (!cfs_rq) | |
10947 | goto err; | |
10948 | ||
10949 | se = kzalloc_node(sizeof(struct sched_entity), | |
10950 | GFP_KERNEL, cpu_to_node(i)); | |
10951 | if (!se) | |
10952 | goto err_free_rq; | |
10953 | ||
10954 | init_cfs_rq(cfs_rq); | |
10955 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 10956 | init_entity_runnable_average(se); |
029632fb PZ |
10957 | } |
10958 | ||
10959 | return 1; | |
10960 | ||
10961 | err_free_rq: | |
10962 | kfree(cfs_rq); | |
10963 | err: | |
10964 | return 0; | |
10965 | } | |
10966 | ||
8663e24d PZ |
10967 | void online_fair_sched_group(struct task_group *tg) |
10968 | { | |
10969 | struct sched_entity *se; | |
a46d14ec | 10970 | struct rq_flags rf; |
8663e24d PZ |
10971 | struct rq *rq; |
10972 | int i; | |
10973 | ||
10974 | for_each_possible_cpu(i) { | |
10975 | rq = cpu_rq(i); | |
10976 | se = tg->se[i]; | |
a46d14ec | 10977 | rq_lock_irq(rq, &rf); |
4126bad6 | 10978 | update_rq_clock(rq); |
d0326691 | 10979 | attach_entity_cfs_rq(se); |
55e16d30 | 10980 | sync_throttle(tg, i); |
a46d14ec | 10981 | rq_unlock_irq(rq, &rf); |
8663e24d PZ |
10982 | } |
10983 | } | |
10984 | ||
6fe1f348 | 10985 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 10986 | { |
029632fb | 10987 | unsigned long flags; |
6fe1f348 PZ |
10988 | struct rq *rq; |
10989 | int cpu; | |
029632fb | 10990 | |
6fe1f348 PZ |
10991 | for_each_possible_cpu(cpu) { |
10992 | if (tg->se[cpu]) | |
10993 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 10994 | |
6fe1f348 PZ |
10995 | /* |
10996 | * Only empty task groups can be destroyed; so we can speculatively | |
10997 | * check on_list without danger of it being re-added. | |
10998 | */ | |
10999 | if (!tg->cfs_rq[cpu]->on_list) | |
11000 | continue; | |
11001 | ||
11002 | rq = cpu_rq(cpu); | |
11003 | ||
11004 | raw_spin_lock_irqsave(&rq->lock, flags); | |
11005 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
11006 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
11007 | } | |
029632fb PZ |
11008 | } |
11009 | ||
11010 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
11011 | struct sched_entity *se, int cpu, | |
11012 | struct sched_entity *parent) | |
11013 | { | |
11014 | struct rq *rq = cpu_rq(cpu); | |
11015 | ||
11016 | cfs_rq->tg = tg; | |
11017 | cfs_rq->rq = rq; | |
029632fb PZ |
11018 | init_cfs_rq_runtime(cfs_rq); |
11019 | ||
11020 | tg->cfs_rq[cpu] = cfs_rq; | |
11021 | tg->se[cpu] = se; | |
11022 | ||
11023 | /* se could be NULL for root_task_group */ | |
11024 | if (!se) | |
11025 | return; | |
11026 | ||
fed14d45 | 11027 | if (!parent) { |
029632fb | 11028 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
11029 | se->depth = 0; |
11030 | } else { | |
029632fb | 11031 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
11032 | se->depth = parent->depth + 1; |
11033 | } | |
029632fb PZ |
11034 | |
11035 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
11036 | /* guarantee group entities always have weight */ |
11037 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
11038 | se->parent = parent; |
11039 | } | |
11040 | ||
11041 | static DEFINE_MUTEX(shares_mutex); | |
11042 | ||
11043 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
11044 | { | |
11045 | int i; | |
029632fb PZ |
11046 | |
11047 | /* | |
11048 | * We can't change the weight of the root cgroup. | |
11049 | */ | |
11050 | if (!tg->se[0]) | |
11051 | return -EINVAL; | |
11052 | ||
11053 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
11054 | ||
11055 | mutex_lock(&shares_mutex); | |
11056 | if (tg->shares == shares) | |
11057 | goto done; | |
11058 | ||
11059 | tg->shares = shares; | |
11060 | for_each_possible_cpu(i) { | |
11061 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
11062 | struct sched_entity *se = tg->se[i]; |
11063 | struct rq_flags rf; | |
029632fb | 11064 | |
029632fb | 11065 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 11066 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 11067 | update_rq_clock(rq); |
89ee048f | 11068 | for_each_sched_entity(se) { |
88c0616e | 11069 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 11070 | update_cfs_group(se); |
89ee048f | 11071 | } |
8a8c69c3 | 11072 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
11073 | } |
11074 | ||
11075 | done: | |
11076 | mutex_unlock(&shares_mutex); | |
11077 | return 0; | |
11078 | } | |
11079 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
11080 | ||
11081 | void free_fair_sched_group(struct task_group *tg) { } | |
11082 | ||
11083 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11084 | { | |
11085 | return 1; | |
11086 | } | |
11087 | ||
8663e24d PZ |
11088 | void online_fair_sched_group(struct task_group *tg) { } |
11089 | ||
6fe1f348 | 11090 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
11091 | |
11092 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
11093 | ||
810b3817 | 11094 | |
6d686f45 | 11095 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
11096 | { |
11097 | struct sched_entity *se = &task->se; | |
0d721cea PW |
11098 | unsigned int rr_interval = 0; |
11099 | ||
11100 | /* | |
11101 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
11102 | * idle runqueue: | |
11103 | */ | |
0d721cea | 11104 | if (rq->cfs.load.weight) |
a59f4e07 | 11105 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
11106 | |
11107 | return rr_interval; | |
11108 | } | |
11109 | ||
bf0f6f24 IM |
11110 | /* |
11111 | * All the scheduling class methods: | |
11112 | */ | |
029632fb | 11113 | const struct sched_class fair_sched_class = { |
5522d5d5 | 11114 | .next = &idle_sched_class, |
bf0f6f24 IM |
11115 | .enqueue_task = enqueue_task_fair, |
11116 | .dequeue_task = dequeue_task_fair, | |
11117 | .yield_task = yield_task_fair, | |
d95f4122 | 11118 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 11119 | |
2e09bf55 | 11120 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 | 11121 | |
98c2f700 | 11122 | .pick_next_task = __pick_next_task_fair, |
bf0f6f24 | 11123 | .put_prev_task = put_prev_task_fair, |
03b7fad1 | 11124 | .set_next_task = set_next_task_fair, |
bf0f6f24 | 11125 | |
681f3e68 | 11126 | #ifdef CONFIG_SMP |
6e2df058 | 11127 | .balance = balance_fair, |
4ce72a2c | 11128 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 11129 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 11130 | |
0bcdcf28 CE |
11131 | .rq_online = rq_online_fair, |
11132 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 11133 | |
12695578 | 11134 | .task_dead = task_dead_fair, |
c5b28038 | 11135 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 11136 | #endif |
bf0f6f24 | 11137 | |
bf0f6f24 | 11138 | .task_tick = task_tick_fair, |
cd29fe6f | 11139 | .task_fork = task_fork_fair, |
cb469845 SR |
11140 | |
11141 | .prio_changed = prio_changed_fair, | |
da7a735e | 11142 | .switched_from = switched_from_fair, |
cb469845 | 11143 | .switched_to = switched_to_fair, |
810b3817 | 11144 | |
0d721cea PW |
11145 | .get_rr_interval = get_rr_interval_fair, |
11146 | ||
6e998916 SG |
11147 | .update_curr = update_curr_fair, |
11148 | ||
810b3817 | 11149 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 11150 | .task_change_group = task_change_group_fair, |
810b3817 | 11151 | #endif |
982d9cdc PB |
11152 | |
11153 | #ifdef CONFIG_UCLAMP_TASK | |
11154 | .uclamp_enabled = 1, | |
11155 | #endif | |
bf0f6f24 IM |
11156 | }; |
11157 | ||
11158 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 11159 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 11160 | { |
039ae8bc | 11161 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 11162 | |
5973e5b9 | 11163 | rcu_read_lock(); |
039ae8bc | 11164 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 11165 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 11166 | rcu_read_unlock(); |
bf0f6f24 | 11167 | } |
397f2378 SD |
11168 | |
11169 | #ifdef CONFIG_NUMA_BALANCING | |
11170 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
11171 | { | |
11172 | int node; | |
11173 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 11174 | struct numa_group *ng; |
397f2378 | 11175 | |
cb361d8c JH |
11176 | rcu_read_lock(); |
11177 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
11178 | for_each_online_node(node) { |
11179 | if (p->numa_faults) { | |
11180 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
11181 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
11182 | } | |
cb361d8c JH |
11183 | if (ng) { |
11184 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
11185 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
11186 | } |
11187 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
11188 | } | |
cb361d8c | 11189 | rcu_read_unlock(); |
397f2378 SD |
11190 | } |
11191 | #endif /* CONFIG_NUMA_BALANCING */ | |
11192 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
11193 | |
11194 | __init void init_sched_fair_class(void) | |
11195 | { | |
11196 | #ifdef CONFIG_SMP | |
11197 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
11198 | ||
3451d024 | 11199 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 11200 | nohz.next_balance = jiffies; |
f643ea22 | 11201 | nohz.next_blocked = jiffies; |
029632fb | 11202 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
11203 | #endif |
11204 | #endif /* SMP */ | |
11205 | ||
11206 | } | |
3c93a0c0 QY |
11207 | |
11208 | /* | |
11209 | * Helper functions to facilitate extracting info from tracepoints. | |
11210 | */ | |
11211 | ||
11212 | const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq) | |
11213 | { | |
11214 | #ifdef CONFIG_SMP | |
11215 | return cfs_rq ? &cfs_rq->avg : NULL; | |
11216 | #else | |
11217 | return NULL; | |
11218 | #endif | |
11219 | } | |
11220 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg); | |
11221 | ||
11222 | char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len) | |
11223 | { | |
11224 | if (!cfs_rq) { | |
11225 | if (str) | |
11226 | strlcpy(str, "(null)", len); | |
11227 | else | |
11228 | return NULL; | |
11229 | } | |
11230 | ||
11231 | cfs_rq_tg_path(cfs_rq, str, len); | |
11232 | return str; | |
11233 | } | |
11234 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path); | |
11235 | ||
11236 | int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq) | |
11237 | { | |
11238 | return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; | |
11239 | } | |
11240 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu); | |
11241 | ||
11242 | const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq) | |
11243 | { | |
11244 | #ifdef CONFIG_SMP | |
11245 | return rq ? &rq->avg_rt : NULL; | |
11246 | #else | |
11247 | return NULL; | |
11248 | #endif | |
11249 | } | |
11250 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt); | |
11251 | ||
11252 | const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq) | |
11253 | { | |
11254 | #ifdef CONFIG_SMP | |
11255 | return rq ? &rq->avg_dl : NULL; | |
11256 | #else | |
11257 | return NULL; | |
11258 | #endif | |
11259 | } | |
11260 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl); | |
11261 | ||
11262 | const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq) | |
11263 | { | |
11264 | #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ) | |
11265 | return rq ? &rq->avg_irq : NULL; | |
11266 | #else | |
11267 | return NULL; | |
11268 | #endif | |
11269 | } | |
11270 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq); | |
11271 | ||
11272 | int sched_trace_rq_cpu(struct rq *rq) | |
11273 | { | |
11274 | return rq ? cpu_of(rq) : -1; | |
11275 | } | |
11276 | EXPORT_SYMBOL_GPL(sched_trace_rq_cpu); | |
11277 | ||
11278 | const struct cpumask *sched_trace_rd_span(struct root_domain *rd) | |
11279 | { | |
11280 | #ifdef CONFIG_SMP | |
11281 | return rd ? rd->span : NULL; | |
11282 | #else | |
11283 | return NULL; | |
11284 | #endif | |
11285 | } | |
11286 | EXPORT_SYMBOL_GPL(sched_trace_rd_span); |