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 | |
afe06efd TC |
89 | #ifdef CONFIG_SMP |
90 | /* | |
97fb7a0a | 91 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
92 | */ |
93 | int __weak arch_asym_cpu_priority(int cpu) | |
94 | { | |
95 | return -cpu; | |
96 | } | |
6d101ba6 OJ |
97 | |
98 | /* | |
60e17f5c | 99 | * The margin used when comparing utilization with CPU capacity. |
6d101ba6 OJ |
100 | * |
101 | * (default: ~20%) | |
102 | */ | |
60e17f5c VK |
103 | #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) |
104 | ||
afe06efd TC |
105 | #endif |
106 | ||
ec12cb7f PT |
107 | #ifdef CONFIG_CFS_BANDWIDTH |
108 | /* | |
109 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
110 | * each time a cfs_rq requests quota. | |
111 | * | |
112 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
113 | * to consumption or the quota being specified to be smaller than the slice) | |
114 | * we will always only issue the remaining available time. | |
115 | * | |
2b4d5b25 IM |
116 | * (default: 5 msec, units: microseconds) |
117 | */ | |
118 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
ec12cb7f PT |
119 | #endif |
120 | ||
8527632d PG |
121 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
122 | { | |
123 | lw->weight += inc; | |
124 | lw->inv_weight = 0; | |
125 | } | |
126 | ||
127 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
128 | { | |
129 | lw->weight -= dec; | |
130 | lw->inv_weight = 0; | |
131 | } | |
132 | ||
133 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
134 | { | |
135 | lw->weight = w; | |
136 | lw->inv_weight = 0; | |
137 | } | |
138 | ||
029632fb PZ |
139 | /* |
140 | * Increase the granularity value when there are more CPUs, | |
141 | * because with more CPUs the 'effective latency' as visible | |
142 | * to users decreases. But the relationship is not linear, | |
143 | * so pick a second-best guess by going with the log2 of the | |
144 | * number of CPUs. | |
145 | * | |
146 | * This idea comes from the SD scheduler of Con Kolivas: | |
147 | */ | |
58ac93e4 | 148 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 149 | { |
58ac93e4 | 150 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
151 | unsigned int factor; |
152 | ||
153 | switch (sysctl_sched_tunable_scaling) { | |
154 | case SCHED_TUNABLESCALING_NONE: | |
155 | factor = 1; | |
156 | break; | |
157 | case SCHED_TUNABLESCALING_LINEAR: | |
158 | factor = cpus; | |
159 | break; | |
160 | case SCHED_TUNABLESCALING_LOG: | |
161 | default: | |
162 | factor = 1 + ilog2(cpus); | |
163 | break; | |
164 | } | |
165 | ||
166 | return factor; | |
167 | } | |
168 | ||
169 | static void update_sysctl(void) | |
170 | { | |
171 | unsigned int factor = get_update_sysctl_factor(); | |
172 | ||
173 | #define SET_SYSCTL(name) \ | |
174 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
175 | SET_SYSCTL(sched_min_granularity); | |
176 | SET_SYSCTL(sched_latency); | |
177 | SET_SYSCTL(sched_wakeup_granularity); | |
178 | #undef SET_SYSCTL | |
179 | } | |
180 | ||
181 | void sched_init_granularity(void) | |
182 | { | |
183 | update_sysctl(); | |
184 | } | |
185 | ||
9dbdb155 | 186 | #define WMULT_CONST (~0U) |
029632fb PZ |
187 | #define WMULT_SHIFT 32 |
188 | ||
9dbdb155 PZ |
189 | static void __update_inv_weight(struct load_weight *lw) |
190 | { | |
191 | unsigned long w; | |
192 | ||
193 | if (likely(lw->inv_weight)) | |
194 | return; | |
195 | ||
196 | w = scale_load_down(lw->weight); | |
197 | ||
198 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
199 | lw->inv_weight = 1; | |
200 | else if (unlikely(!w)) | |
201 | lw->inv_weight = WMULT_CONST; | |
202 | else | |
203 | lw->inv_weight = WMULT_CONST / w; | |
204 | } | |
029632fb PZ |
205 | |
206 | /* | |
9dbdb155 PZ |
207 | * delta_exec * weight / lw.weight |
208 | * OR | |
209 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
210 | * | |
1c3de5e1 | 211 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
212 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
213 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
214 | * | |
215 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
216 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 217 | */ |
9dbdb155 | 218 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 219 | { |
9dbdb155 PZ |
220 | u64 fact = scale_load_down(weight); |
221 | int shift = WMULT_SHIFT; | |
029632fb | 222 | |
9dbdb155 | 223 | __update_inv_weight(lw); |
029632fb | 224 | |
9dbdb155 PZ |
225 | if (unlikely(fact >> 32)) { |
226 | while (fact >> 32) { | |
227 | fact >>= 1; | |
228 | shift--; | |
229 | } | |
029632fb PZ |
230 | } |
231 | ||
9dbdb155 PZ |
232 | /* hint to use a 32x32->64 mul */ |
233 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 234 | |
9dbdb155 PZ |
235 | while (fact >> 32) { |
236 | fact >>= 1; | |
237 | shift--; | |
238 | } | |
029632fb | 239 | |
9dbdb155 | 240 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
241 | } |
242 | ||
243 | ||
244 | const struct sched_class fair_sched_class; | |
a4c2f00f | 245 | |
bf0f6f24 IM |
246 | /************************************************************** |
247 | * CFS operations on generic schedulable entities: | |
248 | */ | |
249 | ||
62160e3f | 250 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8f48894f PZ |
251 | static inline struct task_struct *task_of(struct sched_entity *se) |
252 | { | |
9148a3a1 | 253 | SCHED_WARN_ON(!entity_is_task(se)); |
8f48894f PZ |
254 | return container_of(se, struct task_struct, se); |
255 | } | |
256 | ||
b758149c PZ |
257 | /* Walk up scheduling entities hierarchy */ |
258 | #define for_each_sched_entity(se) \ | |
259 | for (; se; se = se->parent) | |
260 | ||
261 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
262 | { | |
263 | return p->se.cfs_rq; | |
264 | } | |
265 | ||
266 | /* runqueue on which this entity is (to be) queued */ | |
267 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
268 | { | |
269 | return se->cfs_rq; | |
270 | } | |
271 | ||
272 | /* runqueue "owned" by this group */ | |
273 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
274 | { | |
275 | return grp->my_q; | |
276 | } | |
277 | ||
3c93a0c0 QY |
278 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
279 | { | |
280 | if (!path) | |
281 | return; | |
282 | ||
283 | if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) | |
284 | autogroup_path(cfs_rq->tg, path, len); | |
285 | else if (cfs_rq && cfs_rq->tg->css.cgroup) | |
286 | cgroup_path(cfs_rq->tg->css.cgroup, path, len); | |
287 | else | |
288 | strlcpy(path, "(null)", len); | |
289 | } | |
290 | ||
f6783319 | 291 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 292 | { |
5d299eab PZ |
293 | struct rq *rq = rq_of(cfs_rq); |
294 | int cpu = cpu_of(rq); | |
295 | ||
296 | if (cfs_rq->on_list) | |
f6783319 | 297 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
298 | |
299 | cfs_rq->on_list = 1; | |
300 | ||
301 | /* | |
302 | * Ensure we either appear before our parent (if already | |
303 | * enqueued) or force our parent to appear after us when it is | |
304 | * enqueued. The fact that we always enqueue bottom-up | |
305 | * reduces this to two cases and a special case for the root | |
306 | * cfs_rq. Furthermore, it also means that we will always reset | |
307 | * tmp_alone_branch either when the branch is connected | |
308 | * to a tree or when we reach the top of the tree | |
309 | */ | |
310 | if (cfs_rq->tg->parent && | |
311 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 312 | /* |
5d299eab PZ |
313 | * If parent is already on the list, we add the child |
314 | * just before. Thanks to circular linked property of | |
315 | * the list, this means to put the child at the tail | |
316 | * of the list that starts by parent. | |
67e86250 | 317 | */ |
5d299eab PZ |
318 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
319 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
320 | /* | |
321 | * The branch is now connected to its tree so we can | |
322 | * reset tmp_alone_branch to the beginning of the | |
323 | * list. | |
324 | */ | |
325 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 326 | return true; |
5d299eab | 327 | } |
3d4b47b4 | 328 | |
5d299eab PZ |
329 | if (!cfs_rq->tg->parent) { |
330 | /* | |
331 | * cfs rq without parent should be put | |
332 | * at the tail of the list. | |
333 | */ | |
334 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
335 | &rq->leaf_cfs_rq_list); | |
336 | /* | |
337 | * We have reach the top of a tree so we can reset | |
338 | * tmp_alone_branch to the beginning of the list. | |
339 | */ | |
340 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 341 | return true; |
3d4b47b4 | 342 | } |
5d299eab PZ |
343 | |
344 | /* | |
345 | * The parent has not already been added so we want to | |
346 | * make sure that it will be put after us. | |
347 | * tmp_alone_branch points to the begin of the branch | |
348 | * where we will add parent. | |
349 | */ | |
350 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
351 | /* | |
352 | * update tmp_alone_branch to points to the new begin | |
353 | * of the branch | |
354 | */ | |
355 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 356 | return false; |
3d4b47b4 PZ |
357 | } |
358 | ||
359 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
360 | { | |
361 | if (cfs_rq->on_list) { | |
31bc6aea VG |
362 | struct rq *rq = rq_of(cfs_rq); |
363 | ||
364 | /* | |
365 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
366 | * it can happen the tmp_alone_branch points the a leaf that | |
367 | * we finally want to del. In this case, tmp_alone_branch moves | |
368 | * to the prev element but it will point to rq->leaf_cfs_rq_list | |
369 | * at the end of the enqueue. | |
370 | */ | |
371 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
372 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
373 | ||
3d4b47b4 PZ |
374 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
375 | cfs_rq->on_list = 0; | |
376 | } | |
377 | } | |
378 | ||
5d299eab PZ |
379 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
380 | { | |
381 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
382 | } | |
383 | ||
039ae8bc VG |
384 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
385 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ | |
386 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
387 | leaf_cfs_rq_list) | |
b758149c PZ |
388 | |
389 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 390 | static inline struct cfs_rq * |
b758149c PZ |
391 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
392 | { | |
393 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 394 | return se->cfs_rq; |
b758149c | 395 | |
fed14d45 | 396 | return NULL; |
b758149c PZ |
397 | } |
398 | ||
399 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
400 | { | |
401 | return se->parent; | |
402 | } | |
403 | ||
464b7527 PZ |
404 | static void |
405 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
406 | { | |
407 | int se_depth, pse_depth; | |
408 | ||
409 | /* | |
410 | * preemption test can be made between sibling entities who are in the | |
411 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
412 | * both tasks until we find their ancestors who are siblings of common | |
413 | * parent. | |
414 | */ | |
415 | ||
416 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
417 | se_depth = (*se)->depth; |
418 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
419 | |
420 | while (se_depth > pse_depth) { | |
421 | se_depth--; | |
422 | *se = parent_entity(*se); | |
423 | } | |
424 | ||
425 | while (pse_depth > se_depth) { | |
426 | pse_depth--; | |
427 | *pse = parent_entity(*pse); | |
428 | } | |
429 | ||
430 | while (!is_same_group(*se, *pse)) { | |
431 | *se = parent_entity(*se); | |
432 | *pse = parent_entity(*pse); | |
433 | } | |
434 | } | |
435 | ||
8f48894f PZ |
436 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
437 | ||
438 | static inline struct task_struct *task_of(struct sched_entity *se) | |
439 | { | |
440 | return container_of(se, struct task_struct, se); | |
441 | } | |
bf0f6f24 | 442 | |
b758149c PZ |
443 | #define for_each_sched_entity(se) \ |
444 | for (; se; se = NULL) | |
bf0f6f24 | 445 | |
b758149c | 446 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 447 | { |
b758149c | 448 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
449 | } |
450 | ||
b758149c PZ |
451 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
452 | { | |
453 | struct task_struct *p = task_of(se); | |
454 | struct rq *rq = task_rq(p); | |
455 | ||
456 | return &rq->cfs; | |
457 | } | |
458 | ||
459 | /* runqueue "owned" by this group */ | |
460 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
461 | { | |
462 | return NULL; | |
463 | } | |
464 | ||
3c93a0c0 QY |
465 | static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) |
466 | { | |
467 | if (path) | |
468 | strlcpy(path, "(null)", len); | |
469 | } | |
470 | ||
f6783319 | 471 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 472 | { |
f6783319 | 473 | return true; |
3d4b47b4 PZ |
474 | } |
475 | ||
476 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
477 | { | |
478 | } | |
479 | ||
5d299eab PZ |
480 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
481 | { | |
482 | } | |
483 | ||
039ae8bc VG |
484 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
485 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 486 | |
b758149c PZ |
487 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
488 | { | |
489 | return NULL; | |
490 | } | |
491 | ||
464b7527 PZ |
492 | static inline void |
493 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
494 | { | |
495 | } | |
496 | ||
b758149c PZ |
497 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
498 | ||
6c16a6dc | 499 | static __always_inline |
9dbdb155 | 500 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
501 | |
502 | /************************************************************** | |
503 | * Scheduling class tree data structure manipulation methods: | |
504 | */ | |
505 | ||
1bf08230 | 506 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 507 | { |
1bf08230 | 508 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 509 | if (delta > 0) |
1bf08230 | 510 | max_vruntime = vruntime; |
02e0431a | 511 | |
1bf08230 | 512 | return max_vruntime; |
02e0431a PZ |
513 | } |
514 | ||
0702e3eb | 515 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
516 | { |
517 | s64 delta = (s64)(vruntime - min_vruntime); | |
518 | if (delta < 0) | |
519 | min_vruntime = vruntime; | |
520 | ||
521 | return min_vruntime; | |
522 | } | |
523 | ||
54fdc581 FC |
524 | static inline int entity_before(struct sched_entity *a, |
525 | struct sched_entity *b) | |
526 | { | |
527 | return (s64)(a->vruntime - b->vruntime) < 0; | |
528 | } | |
529 | ||
1af5f730 PZ |
530 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
531 | { | |
b60205c7 | 532 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 533 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 534 | |
1af5f730 PZ |
535 | u64 vruntime = cfs_rq->min_vruntime; |
536 | ||
b60205c7 PZ |
537 | if (curr) { |
538 | if (curr->on_rq) | |
539 | vruntime = curr->vruntime; | |
540 | else | |
541 | curr = NULL; | |
542 | } | |
1af5f730 | 543 | |
bfb06889 DB |
544 | if (leftmost) { /* non-empty tree */ |
545 | struct sched_entity *se; | |
546 | se = rb_entry(leftmost, struct sched_entity, run_node); | |
1af5f730 | 547 | |
b60205c7 | 548 | if (!curr) |
1af5f730 PZ |
549 | vruntime = se->vruntime; |
550 | else | |
551 | vruntime = min_vruntime(vruntime, se->vruntime); | |
552 | } | |
553 | ||
1bf08230 | 554 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 555 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
556 | #ifndef CONFIG_64BIT |
557 | smp_wmb(); | |
558 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
559 | #endif | |
1af5f730 PZ |
560 | } |
561 | ||
bf0f6f24 IM |
562 | /* |
563 | * Enqueue an entity into the rb-tree: | |
564 | */ | |
0702e3eb | 565 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 566 | { |
bfb06889 | 567 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node; |
bf0f6f24 IM |
568 | struct rb_node *parent = NULL; |
569 | struct sched_entity *entry; | |
bfb06889 | 570 | bool leftmost = true; |
bf0f6f24 IM |
571 | |
572 | /* | |
573 | * Find the right place in the rbtree: | |
574 | */ | |
575 | while (*link) { | |
576 | parent = *link; | |
577 | entry = rb_entry(parent, struct sched_entity, run_node); | |
578 | /* | |
579 | * We dont care about collisions. Nodes with | |
580 | * the same key stay together. | |
581 | */ | |
2bd2d6f2 | 582 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
583 | link = &parent->rb_left; |
584 | } else { | |
585 | link = &parent->rb_right; | |
bfb06889 | 586 | leftmost = false; |
bf0f6f24 IM |
587 | } |
588 | } | |
589 | ||
bf0f6f24 | 590 | rb_link_node(&se->run_node, parent, link); |
bfb06889 DB |
591 | rb_insert_color_cached(&se->run_node, |
592 | &cfs_rq->tasks_timeline, leftmost); | |
bf0f6f24 IM |
593 | } |
594 | ||
0702e3eb | 595 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 596 | { |
bfb06889 | 597 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
598 | } |
599 | ||
029632fb | 600 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 601 | { |
bfb06889 | 602 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
603 | |
604 | if (!left) | |
605 | return NULL; | |
606 | ||
607 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
608 | } |
609 | ||
ac53db59 RR |
610 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
611 | { | |
612 | struct rb_node *next = rb_next(&se->run_node); | |
613 | ||
614 | if (!next) | |
615 | return NULL; | |
616 | ||
617 | return rb_entry(next, struct sched_entity, run_node); | |
618 | } | |
619 | ||
620 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 621 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 622 | { |
bfb06889 | 623 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 624 | |
70eee74b BS |
625 | if (!last) |
626 | return NULL; | |
7eee3e67 IM |
627 | |
628 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
629 | } |
630 | ||
bf0f6f24 IM |
631 | /************************************************************** |
632 | * Scheduling class statistics methods: | |
633 | */ | |
634 | ||
acb4a848 | 635 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 636 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
637 | loff_t *ppos) |
638 | { | |
8d65af78 | 639 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 640 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
641 | |
642 | if (ret || !write) | |
643 | return ret; | |
644 | ||
645 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
646 | sysctl_sched_min_granularity); | |
647 | ||
acb4a848 CE |
648 | #define WRT_SYSCTL(name) \ |
649 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
650 | WRT_SYSCTL(sched_min_granularity); | |
651 | WRT_SYSCTL(sched_latency); | |
652 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
653 | #undef WRT_SYSCTL |
654 | ||
b2be5e96 PZ |
655 | return 0; |
656 | } | |
657 | #endif | |
647e7cac | 658 | |
a7be37ac | 659 | /* |
f9c0b095 | 660 | * delta /= w |
a7be37ac | 661 | */ |
9dbdb155 | 662 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 663 | { |
f9c0b095 | 664 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 665 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
666 | |
667 | return delta; | |
668 | } | |
669 | ||
647e7cac IM |
670 | /* |
671 | * The idea is to set a period in which each task runs once. | |
672 | * | |
532b1858 | 673 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
674 | * this period because otherwise the slices get too small. |
675 | * | |
676 | * p = (nr <= nl) ? l : l*nr/nl | |
677 | */ | |
4d78e7b6 PZ |
678 | static u64 __sched_period(unsigned long nr_running) |
679 | { | |
8e2b0bf3 BF |
680 | if (unlikely(nr_running > sched_nr_latency)) |
681 | return nr_running * sysctl_sched_min_granularity; | |
682 | else | |
683 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
684 | } |
685 | ||
647e7cac IM |
686 | /* |
687 | * We calculate the wall-time slice from the period by taking a part | |
688 | * proportional to the weight. | |
689 | * | |
f9c0b095 | 690 | * s = p*P[w/rw] |
647e7cac | 691 | */ |
6d0f0ebd | 692 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 693 | { |
0a582440 | 694 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 695 | |
0a582440 | 696 | for_each_sched_entity(se) { |
6272d68c | 697 | struct load_weight *load; |
3104bf03 | 698 | struct load_weight lw; |
6272d68c LM |
699 | |
700 | cfs_rq = cfs_rq_of(se); | |
701 | load = &cfs_rq->load; | |
f9c0b095 | 702 | |
0a582440 | 703 | if (unlikely(!se->on_rq)) { |
3104bf03 | 704 | lw = cfs_rq->load; |
0a582440 MG |
705 | |
706 | update_load_add(&lw, se->load.weight); | |
707 | load = &lw; | |
708 | } | |
9dbdb155 | 709 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
710 | } |
711 | return slice; | |
bf0f6f24 IM |
712 | } |
713 | ||
647e7cac | 714 | /* |
660cc00f | 715 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 716 | * |
f9c0b095 | 717 | * vs = s/w |
647e7cac | 718 | */ |
f9c0b095 | 719 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 720 | { |
f9c0b095 | 721 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
722 | } |
723 | ||
c0796298 | 724 | #include "pelt.h" |
23127296 | 725 | #ifdef CONFIG_SMP |
283e2ed3 | 726 | |
772bd008 | 727 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 728 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 729 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 730 | |
540247fb YD |
731 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
732 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 733 | { |
540247fb | 734 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 735 | |
f207934f PZ |
736 | memset(sa, 0, sizeof(*sa)); |
737 | ||
b5a9b340 | 738 | /* |
dfcb245e | 739 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 740 | * they get a chance to stabilize to their real load level. |
dfcb245e | 741 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
742 | * nothing has been attached to the task group yet. |
743 | */ | |
744 | if (entity_is_task(se)) | |
1ea6c46a | 745 | sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight); |
1ea6c46a | 746 | |
f207934f PZ |
747 | se->runnable_weight = se->load.weight; |
748 | ||
9d89c257 | 749 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 750 | } |
7ea241af | 751 | |
7dc603c9 | 752 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
df217913 | 753 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 754 | |
2b8c41da YD |
755 | /* |
756 | * With new tasks being created, their initial util_avgs are extrapolated | |
757 | * based on the cfs_rq's current util_avg: | |
758 | * | |
759 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
760 | * | |
761 | * However, in many cases, the above util_avg does not give a desired | |
762 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
763 | * as when the series is a harmonic series. | |
764 | * | |
765 | * To solve this problem, we also cap the util_avg of successive tasks to | |
766 | * only 1/2 of the left utilization budget: | |
767 | * | |
8fe5c5a9 | 768 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 769 | * |
8fe5c5a9 | 770 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 771 | * |
8fe5c5a9 QP |
772 | * For example, for a CPU with 1024 of capacity, a simplest series from |
773 | * the beginning would be like: | |
2b8c41da YD |
774 | * |
775 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
776 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
777 | * | |
778 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
779 | * if util_avg > util_avg_cap. | |
780 | */ | |
d0fe0b9c | 781 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 782 | { |
d0fe0b9c | 783 | struct sched_entity *se = &p->se; |
2b8c41da YD |
784 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
785 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 786 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 787 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
788 | |
789 | if (cap > 0) { | |
790 | if (cfs_rq->avg.util_avg != 0) { | |
791 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
792 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
793 | ||
794 | if (sa->util_avg > cap) | |
795 | sa->util_avg = cap; | |
796 | } else { | |
797 | sa->util_avg = cap; | |
798 | } | |
2b8c41da | 799 | } |
7dc603c9 | 800 | |
d0fe0b9c DE |
801 | if (p->sched_class != &fair_sched_class) { |
802 | /* | |
803 | * For !fair tasks do: | |
804 | * | |
805 | update_cfs_rq_load_avg(now, cfs_rq); | |
806 | attach_entity_load_avg(cfs_rq, se, 0); | |
807 | switched_from_fair(rq, p); | |
808 | * | |
809 | * such that the next switched_to_fair() has the | |
810 | * expected state. | |
811 | */ | |
812 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
813 | return; | |
7dc603c9 PZ |
814 | } |
815 | ||
df217913 | 816 | attach_entity_cfs_rq(se); |
2b8c41da YD |
817 | } |
818 | ||
7dc603c9 | 819 | #else /* !CONFIG_SMP */ |
540247fb | 820 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
821 | { |
822 | } | |
d0fe0b9c | 823 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
824 | { |
825 | } | |
3d30544f PZ |
826 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
827 | { | |
828 | } | |
7dc603c9 | 829 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 830 | |
bf0f6f24 | 831 | /* |
9dbdb155 | 832 | * Update the current task's runtime statistics. |
bf0f6f24 | 833 | */ |
b7cc0896 | 834 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 835 | { |
429d43bc | 836 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 837 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 838 | u64 delta_exec; |
bf0f6f24 IM |
839 | |
840 | if (unlikely(!curr)) | |
841 | return; | |
842 | ||
9dbdb155 PZ |
843 | delta_exec = now - curr->exec_start; |
844 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 845 | return; |
bf0f6f24 | 846 | |
8ebc91d9 | 847 | curr->exec_start = now; |
d842de87 | 848 | |
9dbdb155 PZ |
849 | schedstat_set(curr->statistics.exec_max, |
850 | max(delta_exec, curr->statistics.exec_max)); | |
851 | ||
852 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 853 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
854 | |
855 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
856 | update_min_vruntime(cfs_rq); | |
857 | ||
d842de87 SV |
858 | if (entity_is_task(curr)) { |
859 | struct task_struct *curtask = task_of(curr); | |
860 | ||
f977bb49 | 861 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 862 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 863 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 864 | } |
ec12cb7f PT |
865 | |
866 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
867 | } |
868 | ||
6e998916 SG |
869 | static void update_curr_fair(struct rq *rq) |
870 | { | |
871 | update_curr(cfs_rq_of(&rq->curr->se)); | |
872 | } | |
873 | ||
bf0f6f24 | 874 | static inline void |
5870db5b | 875 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 876 | { |
4fa8d299 JP |
877 | u64 wait_start, prev_wait_start; |
878 | ||
879 | if (!schedstat_enabled()) | |
880 | return; | |
881 | ||
882 | wait_start = rq_clock(rq_of(cfs_rq)); | |
883 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
884 | |
885 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
886 | likely(wait_start > prev_wait_start)) |
887 | wait_start -= prev_wait_start; | |
3ea94de1 | 888 | |
2ed41a55 | 889 | __schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
890 | } |
891 | ||
4fa8d299 | 892 | static inline void |
3ea94de1 JP |
893 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
894 | { | |
895 | struct task_struct *p; | |
cb251765 MG |
896 | u64 delta; |
897 | ||
4fa8d299 JP |
898 | if (!schedstat_enabled()) |
899 | return; | |
900 | ||
901 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
902 | |
903 | if (entity_is_task(se)) { | |
904 | p = task_of(se); | |
905 | if (task_on_rq_migrating(p)) { | |
906 | /* | |
907 | * Preserve migrating task's wait time so wait_start | |
908 | * time stamp can be adjusted to accumulate wait time | |
909 | * prior to migration. | |
910 | */ | |
2ed41a55 | 911 | __schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
912 | return; |
913 | } | |
914 | trace_sched_stat_wait(p, delta); | |
915 | } | |
916 | ||
2ed41a55 | 917 | __schedstat_set(se->statistics.wait_max, |
4fa8d299 | 918 | max(schedstat_val(se->statistics.wait_max), delta)); |
2ed41a55 PZ |
919 | __schedstat_inc(se->statistics.wait_count); |
920 | __schedstat_add(se->statistics.wait_sum, delta); | |
921 | __schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 922 | } |
3ea94de1 | 923 | |
4fa8d299 | 924 | static inline void |
1a3d027c JP |
925 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
926 | { | |
927 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
928 | u64 sleep_start, block_start; |
929 | ||
930 | if (!schedstat_enabled()) | |
931 | return; | |
932 | ||
933 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
934 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
935 | |
936 | if (entity_is_task(se)) | |
937 | tsk = task_of(se); | |
938 | ||
4fa8d299 JP |
939 | if (sleep_start) { |
940 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
941 | |
942 | if ((s64)delta < 0) | |
943 | delta = 0; | |
944 | ||
4fa8d299 | 945 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
2ed41a55 | 946 | __schedstat_set(se->statistics.sleep_max, delta); |
1a3d027c | 947 | |
2ed41a55 PZ |
948 | __schedstat_set(se->statistics.sleep_start, 0); |
949 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
950 | |
951 | if (tsk) { | |
952 | account_scheduler_latency(tsk, delta >> 10, 1); | |
953 | trace_sched_stat_sleep(tsk, delta); | |
954 | } | |
955 | } | |
4fa8d299 JP |
956 | if (block_start) { |
957 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
958 | |
959 | if ((s64)delta < 0) | |
960 | delta = 0; | |
961 | ||
4fa8d299 | 962 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
2ed41a55 | 963 | __schedstat_set(se->statistics.block_max, delta); |
1a3d027c | 964 | |
2ed41a55 PZ |
965 | __schedstat_set(se->statistics.block_start, 0); |
966 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
967 | |
968 | if (tsk) { | |
969 | if (tsk->in_iowait) { | |
2ed41a55 PZ |
970 | __schedstat_add(se->statistics.iowait_sum, delta); |
971 | __schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
972 | trace_sched_stat_iowait(tsk, delta); |
973 | } | |
974 | ||
975 | trace_sched_stat_blocked(tsk, delta); | |
976 | ||
977 | /* | |
978 | * Blocking time is in units of nanosecs, so shift by | |
979 | * 20 to get a milliseconds-range estimation of the | |
980 | * amount of time that the task spent sleeping: | |
981 | */ | |
982 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
983 | profile_hits(SLEEP_PROFILING, | |
984 | (void *)get_wchan(tsk), | |
985 | delta >> 20); | |
986 | } | |
987 | account_scheduler_latency(tsk, delta >> 10, 0); | |
988 | } | |
989 | } | |
3ea94de1 | 990 | } |
3ea94de1 | 991 | |
bf0f6f24 IM |
992 | /* |
993 | * Task is being enqueued - update stats: | |
994 | */ | |
cb251765 | 995 | static inline void |
1a3d027c | 996 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 997 | { |
4fa8d299 JP |
998 | if (!schedstat_enabled()) |
999 | return; | |
1000 | ||
bf0f6f24 IM |
1001 | /* |
1002 | * Are we enqueueing a waiting task? (for current tasks | |
1003 | * a dequeue/enqueue event is a NOP) | |
1004 | */ | |
429d43bc | 1005 | if (se != cfs_rq->curr) |
5870db5b | 1006 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
1007 | |
1008 | if (flags & ENQUEUE_WAKEUP) | |
1009 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
1010 | } |
1011 | ||
bf0f6f24 | 1012 | static inline void |
cb251765 | 1013 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1014 | { |
4fa8d299 JP |
1015 | |
1016 | if (!schedstat_enabled()) | |
1017 | return; | |
1018 | ||
bf0f6f24 IM |
1019 | /* |
1020 | * Mark the end of the wait period if dequeueing a | |
1021 | * waiting task: | |
1022 | */ | |
429d43bc | 1023 | if (se != cfs_rq->curr) |
9ef0a961 | 1024 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 1025 | |
4fa8d299 JP |
1026 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1027 | struct task_struct *tsk = task_of(se); | |
cb251765 | 1028 | |
4fa8d299 | 1029 | if (tsk->state & TASK_INTERRUPTIBLE) |
2ed41a55 | 1030 | __schedstat_set(se->statistics.sleep_start, |
4fa8d299 JP |
1031 | rq_clock(rq_of(cfs_rq))); |
1032 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
2ed41a55 | 1033 | __schedstat_set(se->statistics.block_start, |
4fa8d299 | 1034 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1035 | } |
cb251765 MG |
1036 | } |
1037 | ||
bf0f6f24 IM |
1038 | /* |
1039 | * We are picking a new current task - update its stats: | |
1040 | */ | |
1041 | static inline void | |
79303e9e | 1042 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1043 | { |
1044 | /* | |
1045 | * We are starting a new run period: | |
1046 | */ | |
78becc27 | 1047 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1048 | } |
1049 | ||
bf0f6f24 IM |
1050 | /************************************************** |
1051 | * Scheduling class queueing methods: | |
1052 | */ | |
1053 | ||
cbee9f88 PZ |
1054 | #ifdef CONFIG_NUMA_BALANCING |
1055 | /* | |
598f0ec0 MG |
1056 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1057 | * calculated based on the tasks virtual memory size and | |
1058 | * numa_balancing_scan_size. | |
cbee9f88 | 1059 | */ |
598f0ec0 MG |
1060 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1061 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1062 | |
1063 | /* Portion of address space to scan in MB */ | |
1064 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1065 | |
4b96a29b PZ |
1066 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1067 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1068 | ||
b5dd77c8 | 1069 | struct numa_group { |
c45a7795 | 1070 | refcount_t refcount; |
b5dd77c8 RR |
1071 | |
1072 | spinlock_t lock; /* nr_tasks, tasks */ | |
1073 | int nr_tasks; | |
1074 | pid_t gid; | |
1075 | int active_nodes; | |
1076 | ||
1077 | struct rcu_head rcu; | |
1078 | unsigned long total_faults; | |
1079 | unsigned long max_faults_cpu; | |
1080 | /* | |
1081 | * Faults_cpu is used to decide whether memory should move | |
1082 | * towards the CPU. As a consequence, these stats are weighted | |
1083 | * more by CPU use than by memory faults. | |
1084 | */ | |
1085 | unsigned long *faults_cpu; | |
1086 | unsigned long faults[0]; | |
1087 | }; | |
1088 | ||
cb361d8c JH |
1089 | /* |
1090 | * For functions that can be called in multiple contexts that permit reading | |
1091 | * ->numa_group (see struct task_struct for locking rules). | |
1092 | */ | |
1093 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1094 | { | |
1095 | return rcu_dereference_check(p->numa_group, p == current || | |
1096 | (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu))); | |
1097 | } | |
1098 | ||
1099 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1100 | { | |
1101 | return rcu_dereference_protected(p->numa_group, p == current); | |
1102 | } | |
1103 | ||
b5dd77c8 RR |
1104 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1105 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1106 | ||
598f0ec0 MG |
1107 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1108 | { | |
1109 | unsigned long rss = 0; | |
1110 | unsigned long nr_scan_pages; | |
1111 | ||
1112 | /* | |
1113 | * Calculations based on RSS as non-present and empty pages are skipped | |
1114 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1115 | * on resident pages | |
1116 | */ | |
1117 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1118 | rss = get_mm_rss(p->mm); | |
1119 | if (!rss) | |
1120 | rss = nr_scan_pages; | |
1121 | ||
1122 | rss = round_up(rss, nr_scan_pages); | |
1123 | return rss / nr_scan_pages; | |
1124 | } | |
1125 | ||
1126 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
1127 | #define MAX_SCAN_WINDOW 2560 | |
1128 | ||
1129 | static unsigned int task_scan_min(struct task_struct *p) | |
1130 | { | |
316c1608 | 1131 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1132 | unsigned int scan, floor; |
1133 | unsigned int windows = 1; | |
1134 | ||
64192658 KT |
1135 | if (scan_size < MAX_SCAN_WINDOW) |
1136 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1137 | floor = 1000 / windows; |
1138 | ||
1139 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1140 | return max_t(unsigned int, floor, scan); | |
1141 | } | |
1142 | ||
b5dd77c8 RR |
1143 | static unsigned int task_scan_start(struct task_struct *p) |
1144 | { | |
1145 | unsigned long smin = task_scan_min(p); | |
1146 | unsigned long period = smin; | |
cb361d8c | 1147 | struct numa_group *ng; |
b5dd77c8 RR |
1148 | |
1149 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1150 | rcu_read_lock(); |
1151 | ng = rcu_dereference(p->numa_group); | |
1152 | if (ng) { | |
b5dd77c8 RR |
1153 | unsigned long shared = group_faults_shared(ng); |
1154 | unsigned long private = group_faults_priv(ng); | |
1155 | ||
c45a7795 | 1156 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1157 | period *= shared + 1; |
1158 | period /= private + shared + 1; | |
1159 | } | |
cb361d8c | 1160 | rcu_read_unlock(); |
b5dd77c8 RR |
1161 | |
1162 | return max(smin, period); | |
1163 | } | |
1164 | ||
598f0ec0 MG |
1165 | static unsigned int task_scan_max(struct task_struct *p) |
1166 | { | |
b5dd77c8 RR |
1167 | unsigned long smin = task_scan_min(p); |
1168 | unsigned long smax; | |
cb361d8c | 1169 | struct numa_group *ng; |
598f0ec0 MG |
1170 | |
1171 | /* Watch for min being lower than max due to floor calculations */ | |
1172 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1173 | |
1174 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1175 | ng = deref_curr_numa_group(p); |
1176 | if (ng) { | |
b5dd77c8 RR |
1177 | unsigned long shared = group_faults_shared(ng); |
1178 | unsigned long private = group_faults_priv(ng); | |
1179 | unsigned long period = smax; | |
1180 | ||
c45a7795 | 1181 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1182 | period *= shared + 1; |
1183 | period /= private + shared + 1; | |
1184 | ||
1185 | smax = max(smax, period); | |
1186 | } | |
1187 | ||
598f0ec0 MG |
1188 | return max(smin, smax); |
1189 | } | |
1190 | ||
0ec8aa00 PZ |
1191 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1192 | { | |
98fa15f3 | 1193 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1194 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1195 | } | |
1196 | ||
1197 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1198 | { | |
98fa15f3 | 1199 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1200 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1201 | } | |
1202 | ||
be1e4e76 RR |
1203 | /* Shared or private faults. */ |
1204 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1205 | ||
1206 | /* Memory and CPU locality */ | |
1207 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1208 | ||
1209 | /* Averaged statistics, and temporary buffers. */ | |
1210 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1211 | ||
e29cf08b MG |
1212 | pid_t task_numa_group_id(struct task_struct *p) |
1213 | { | |
cb361d8c JH |
1214 | struct numa_group *ng; |
1215 | pid_t gid = 0; | |
1216 | ||
1217 | rcu_read_lock(); | |
1218 | ng = rcu_dereference(p->numa_group); | |
1219 | if (ng) | |
1220 | gid = ng->gid; | |
1221 | rcu_read_unlock(); | |
1222 | ||
1223 | return gid; | |
e29cf08b MG |
1224 | } |
1225 | ||
44dba3d5 | 1226 | /* |
97fb7a0a | 1227 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1228 | * occupy the first half of the array. The second half of the |
1229 | * array is for current counters, which are averaged into the | |
1230 | * first set by task_numa_placement. | |
1231 | */ | |
1232 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1233 | { |
44dba3d5 | 1234 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1235 | } |
1236 | ||
1237 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1238 | { | |
44dba3d5 | 1239 | if (!p->numa_faults) |
ac8e895b MG |
1240 | return 0; |
1241 | ||
44dba3d5 IM |
1242 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1243 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1244 | } |
1245 | ||
83e1d2cd MG |
1246 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1247 | { | |
cb361d8c JH |
1248 | struct numa_group *ng = deref_task_numa_group(p); |
1249 | ||
1250 | if (!ng) | |
83e1d2cd MG |
1251 | return 0; |
1252 | ||
cb361d8c JH |
1253 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1254 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1255 | } |
1256 | ||
20e07dea RR |
1257 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1258 | { | |
44dba3d5 IM |
1259 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1260 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1261 | } |
1262 | ||
b5dd77c8 RR |
1263 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1264 | { | |
1265 | unsigned long faults = 0; | |
1266 | int node; | |
1267 | ||
1268 | for_each_online_node(node) { | |
1269 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1270 | } | |
1271 | ||
1272 | return faults; | |
1273 | } | |
1274 | ||
1275 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1276 | { | |
1277 | unsigned long faults = 0; | |
1278 | int node; | |
1279 | ||
1280 | for_each_online_node(node) { | |
1281 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1282 | } | |
1283 | ||
1284 | return faults; | |
1285 | } | |
1286 | ||
4142c3eb RR |
1287 | /* |
1288 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1289 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1290 | * between these nodes are slowed down, to allow things to settle down. | |
1291 | */ | |
1292 | #define ACTIVE_NODE_FRACTION 3 | |
1293 | ||
1294 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1295 | { | |
1296 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1297 | } | |
1298 | ||
6c6b1193 RR |
1299 | /* Handle placement on systems where not all nodes are directly connected. */ |
1300 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1301 | int maxdist, bool task) | |
1302 | { | |
1303 | unsigned long score = 0; | |
1304 | int node; | |
1305 | ||
1306 | /* | |
1307 | * All nodes are directly connected, and the same distance | |
1308 | * from each other. No need for fancy placement algorithms. | |
1309 | */ | |
1310 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1311 | return 0; | |
1312 | ||
1313 | /* | |
1314 | * This code is called for each node, introducing N^2 complexity, | |
1315 | * which should be ok given the number of nodes rarely exceeds 8. | |
1316 | */ | |
1317 | for_each_online_node(node) { | |
1318 | unsigned long faults; | |
1319 | int dist = node_distance(nid, node); | |
1320 | ||
1321 | /* | |
1322 | * The furthest away nodes in the system are not interesting | |
1323 | * for placement; nid was already counted. | |
1324 | */ | |
1325 | if (dist == sched_max_numa_distance || node == nid) | |
1326 | continue; | |
1327 | ||
1328 | /* | |
1329 | * On systems with a backplane NUMA topology, compare groups | |
1330 | * of nodes, and move tasks towards the group with the most | |
1331 | * memory accesses. When comparing two nodes at distance | |
1332 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1333 | * of each group. Skip other nodes. | |
1334 | */ | |
1335 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
0ee7e74d | 1336 | dist >= maxdist) |
6c6b1193 RR |
1337 | continue; |
1338 | ||
1339 | /* Add up the faults from nearby nodes. */ | |
1340 | if (task) | |
1341 | faults = task_faults(p, node); | |
1342 | else | |
1343 | faults = group_faults(p, node); | |
1344 | ||
1345 | /* | |
1346 | * On systems with a glueless mesh NUMA topology, there are | |
1347 | * no fixed "groups of nodes". Instead, nodes that are not | |
1348 | * directly connected bounce traffic through intermediate | |
1349 | * nodes; a numa_group can occupy any set of nodes. | |
1350 | * The further away a node is, the less the faults count. | |
1351 | * This seems to result in good task placement. | |
1352 | */ | |
1353 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1354 | faults *= (sched_max_numa_distance - dist); | |
1355 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1356 | } | |
1357 | ||
1358 | score += faults; | |
1359 | } | |
1360 | ||
1361 | return score; | |
1362 | } | |
1363 | ||
83e1d2cd MG |
1364 | /* |
1365 | * These return the fraction of accesses done by a particular task, or | |
1366 | * task group, on a particular numa node. The group weight is given a | |
1367 | * larger multiplier, in order to group tasks together that are almost | |
1368 | * evenly spread out between numa nodes. | |
1369 | */ | |
7bd95320 RR |
1370 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1371 | int dist) | |
83e1d2cd | 1372 | { |
7bd95320 | 1373 | unsigned long faults, total_faults; |
83e1d2cd | 1374 | |
44dba3d5 | 1375 | if (!p->numa_faults) |
83e1d2cd MG |
1376 | return 0; |
1377 | ||
1378 | total_faults = p->total_numa_faults; | |
1379 | ||
1380 | if (!total_faults) | |
1381 | return 0; | |
1382 | ||
7bd95320 | 1383 | faults = task_faults(p, nid); |
6c6b1193 RR |
1384 | faults += score_nearby_nodes(p, nid, dist, true); |
1385 | ||
7bd95320 | 1386 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1387 | } |
1388 | ||
7bd95320 RR |
1389 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1390 | int dist) | |
83e1d2cd | 1391 | { |
cb361d8c | 1392 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1393 | unsigned long faults, total_faults; |
1394 | ||
cb361d8c | 1395 | if (!ng) |
7bd95320 RR |
1396 | return 0; |
1397 | ||
cb361d8c | 1398 | total_faults = ng->total_faults; |
7bd95320 RR |
1399 | |
1400 | if (!total_faults) | |
83e1d2cd MG |
1401 | return 0; |
1402 | ||
7bd95320 | 1403 | faults = group_faults(p, nid); |
6c6b1193 RR |
1404 | faults += score_nearby_nodes(p, nid, dist, false); |
1405 | ||
7bd95320 | 1406 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1407 | } |
1408 | ||
10f39042 RR |
1409 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1410 | int src_nid, int dst_cpu) | |
1411 | { | |
cb361d8c | 1412 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1413 | int dst_nid = cpu_to_node(dst_cpu); |
1414 | int last_cpupid, this_cpupid; | |
1415 | ||
1416 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
37355bdc MG |
1417 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1418 | ||
1419 | /* | |
1420 | * Allow first faults or private faults to migrate immediately early in | |
1421 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1422 | * two full passes of the "multi-stage node selection" test that is | |
1423 | * executed below. | |
1424 | */ | |
98fa15f3 | 1425 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1426 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1427 | return true; | |
10f39042 RR |
1428 | |
1429 | /* | |
1430 | * Multi-stage node selection is used in conjunction with a periodic | |
1431 | * migration fault to build a temporal task<->page relation. By using | |
1432 | * a two-stage filter we remove short/unlikely relations. | |
1433 | * | |
1434 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1435 | * a task's usage of a particular page (n_p) per total usage of this | |
1436 | * page (n_t) (in a given time-span) to a probability. | |
1437 | * | |
1438 | * Our periodic faults will sample this probability and getting the | |
1439 | * same result twice in a row, given these samples are fully | |
1440 | * independent, is then given by P(n)^2, provided our sample period | |
1441 | * is sufficiently short compared to the usage pattern. | |
1442 | * | |
1443 | * This quadric squishes small probabilities, making it less likely we | |
1444 | * act on an unlikely task<->page relation. | |
1445 | */ | |
10f39042 RR |
1446 | if (!cpupid_pid_unset(last_cpupid) && |
1447 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1448 | return false; | |
1449 | ||
1450 | /* Always allow migrate on private faults */ | |
1451 | if (cpupid_match_pid(p, last_cpupid)) | |
1452 | return true; | |
1453 | ||
1454 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1455 | if (!ng) | |
1456 | return true; | |
1457 | ||
1458 | /* | |
4142c3eb RR |
1459 | * Destination node is much more heavily used than the source |
1460 | * node? Allow migration. | |
10f39042 | 1461 | */ |
4142c3eb RR |
1462 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1463 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1464 | return true; |
1465 | ||
1466 | /* | |
4142c3eb RR |
1467 | * Distribute memory according to CPU & memory use on each node, |
1468 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1469 | * | |
1470 | * faults_cpu(dst) 3 faults_cpu(src) | |
1471 | * --------------- * - > --------------- | |
1472 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1473 | */ |
4142c3eb RR |
1474 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1475 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1476 | } |
1477 | ||
a3df0679 | 1478 | static unsigned long cpu_runnable_load(struct rq *rq); |
58d081b5 | 1479 | |
fb13c7ee | 1480 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1481 | struct numa_stats { |
1482 | unsigned long load; | |
fb13c7ee MG |
1483 | |
1484 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1485 | unsigned long compute_capacity; |
58d081b5 | 1486 | }; |
e6628d5b | 1487 | |
fb13c7ee MG |
1488 | /* |
1489 | * XXX borrowed from update_sg_lb_stats | |
1490 | */ | |
1491 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1492 | { | |
d90707eb | 1493 | int cpu; |
fb13c7ee MG |
1494 | |
1495 | memset(ns, 0, sizeof(*ns)); | |
1496 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1497 | struct rq *rq = cpu_rq(cpu); | |
1498 | ||
a3df0679 | 1499 | ns->load += cpu_runnable_load(rq); |
ced549fa | 1500 | ns->compute_capacity += capacity_of(cpu); |
fb13c7ee MG |
1501 | } |
1502 | ||
fb13c7ee MG |
1503 | } |
1504 | ||
58d081b5 MG |
1505 | struct task_numa_env { |
1506 | struct task_struct *p; | |
e6628d5b | 1507 | |
58d081b5 MG |
1508 | int src_cpu, src_nid; |
1509 | int dst_cpu, dst_nid; | |
e6628d5b | 1510 | |
58d081b5 | 1511 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1512 | |
40ea2b42 | 1513 | int imbalance_pct; |
7bd95320 | 1514 | int dist; |
fb13c7ee MG |
1515 | |
1516 | struct task_struct *best_task; | |
1517 | long best_imp; | |
58d081b5 MG |
1518 | int best_cpu; |
1519 | }; | |
1520 | ||
fb13c7ee MG |
1521 | static void task_numa_assign(struct task_numa_env *env, |
1522 | struct task_struct *p, long imp) | |
1523 | { | |
a4739eca SD |
1524 | struct rq *rq = cpu_rq(env->dst_cpu); |
1525 | ||
1526 | /* Bail out if run-queue part of active NUMA balance. */ | |
1527 | if (xchg(&rq->numa_migrate_on, 1)) | |
1528 | return; | |
1529 | ||
1530 | /* | |
1531 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1532 | * found a better CPU to move/swap. | |
1533 | */ | |
1534 | if (env->best_cpu != -1) { | |
1535 | rq = cpu_rq(env->best_cpu); | |
1536 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1537 | } | |
1538 | ||
fb13c7ee MG |
1539 | if (env->best_task) |
1540 | put_task_struct(env->best_task); | |
bac78573 ON |
1541 | if (p) |
1542 | get_task_struct(p); | |
fb13c7ee MG |
1543 | |
1544 | env->best_task = p; | |
1545 | env->best_imp = imp; | |
1546 | env->best_cpu = env->dst_cpu; | |
1547 | } | |
1548 | ||
28a21745 | 1549 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1550 | struct task_numa_env *env) |
1551 | { | |
e4991b24 RR |
1552 | long imb, old_imb; |
1553 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1554 | long src_capacity, dst_capacity; |
1555 | ||
1556 | /* | |
1557 | * The load is corrected for the CPU capacity available on each node. | |
1558 | * | |
1559 | * src_load dst_load | |
1560 | * ------------ vs --------- | |
1561 | * src_capacity dst_capacity | |
1562 | */ | |
1563 | src_capacity = env->src_stats.compute_capacity; | |
1564 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1565 | |
5f95ba7a | 1566 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1567 | |
28a21745 | 1568 | orig_src_load = env->src_stats.load; |
e4991b24 | 1569 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1570 | |
5f95ba7a | 1571 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1572 | |
1573 | /* Would this change make things worse? */ | |
1574 | return (imb > old_imb); | |
e63da036 RR |
1575 | } |
1576 | ||
6fd98e77 SD |
1577 | /* |
1578 | * Maximum NUMA importance can be 1998 (2*999); | |
1579 | * SMALLIMP @ 30 would be close to 1998/64. | |
1580 | * Used to deter task migration. | |
1581 | */ | |
1582 | #define SMALLIMP 30 | |
1583 | ||
fb13c7ee MG |
1584 | /* |
1585 | * This checks if the overall compute and NUMA accesses of the system would | |
1586 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1587 | * into account that it might be best if task running on the dst_cpu should | |
1588 | * be exchanged with the source task | |
1589 | */ | |
887c290e | 1590 | static void task_numa_compare(struct task_numa_env *env, |
305c1fac | 1591 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1592 | { |
cb361d8c | 1593 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 1594 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 1595 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 1596 | struct task_struct *cur; |
28a21745 | 1597 | long src_load, dst_load; |
7bd95320 | 1598 | int dist = env->dist; |
cb361d8c JH |
1599 | long moveimp = imp; |
1600 | long load; | |
fb13c7ee | 1601 | |
a4739eca SD |
1602 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
1603 | return; | |
1604 | ||
fb13c7ee | 1605 | rcu_read_lock(); |
bac78573 ON |
1606 | cur = task_rcu_dereference(&dst_rq->curr); |
1607 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) | |
fb13c7ee MG |
1608 | cur = NULL; |
1609 | ||
7af68335 PZ |
1610 | /* |
1611 | * Because we have preemption enabled we can get migrated around and | |
1612 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1613 | */ | |
1614 | if (cur == env->p) | |
1615 | goto unlock; | |
1616 | ||
305c1fac | 1617 | if (!cur) { |
6fd98e77 | 1618 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1619 | goto assign; |
1620 | else | |
1621 | goto unlock; | |
1622 | } | |
1623 | ||
fb13c7ee MG |
1624 | /* |
1625 | * "imp" is the fault differential for the source task between the | |
1626 | * source and destination node. Calculate the total differential for | |
1627 | * the source task and potential destination task. The more negative | |
305c1fac | 1628 | * the value is, the more remote accesses that would be expected to |
fb13c7ee MG |
1629 | * be incurred if the tasks were swapped. |
1630 | */ | |
305c1fac | 1631 | /* Skip this swap candidate if cannot move to the source cpu */ |
3bd37062 | 1632 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) |
305c1fac | 1633 | goto unlock; |
fb13c7ee | 1634 | |
305c1fac SD |
1635 | /* |
1636 | * If dst and source tasks are in the same NUMA group, or not | |
1637 | * in any group then look only at task weights. | |
1638 | */ | |
cb361d8c JH |
1639 | cur_ng = rcu_dereference(cur->numa_group); |
1640 | if (cur_ng == p_ng) { | |
305c1fac SD |
1641 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1642 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1643 | /* |
305c1fac SD |
1644 | * Add some hysteresis to prevent swapping the |
1645 | * tasks within a group over tiny differences. | |
887c290e | 1646 | */ |
cb361d8c | 1647 | if (cur_ng) |
305c1fac SD |
1648 | imp -= imp / 16; |
1649 | } else { | |
1650 | /* | |
1651 | * Compare the group weights. If a task is all by itself | |
1652 | * (not part of a group), use the task weight instead. | |
1653 | */ | |
cb361d8c | 1654 | if (cur_ng && p_ng) |
305c1fac SD |
1655 | imp += group_weight(cur, env->src_nid, dist) - |
1656 | group_weight(cur, env->dst_nid, dist); | |
1657 | else | |
1658 | imp += task_weight(cur, env->src_nid, dist) - | |
1659 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
1660 | } |
1661 | ||
305c1fac | 1662 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 1663 | imp = moveimp; |
305c1fac | 1664 | cur = NULL; |
fb13c7ee | 1665 | goto assign; |
305c1fac | 1666 | } |
fb13c7ee | 1667 | |
6fd98e77 SD |
1668 | /* |
1669 | * If the NUMA importance is less than SMALLIMP, | |
1670 | * task migration might only result in ping pong | |
1671 | * of tasks and also hurt performance due to cache | |
1672 | * misses. | |
1673 | */ | |
1674 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
1675 | goto unlock; | |
1676 | ||
fb13c7ee MG |
1677 | /* |
1678 | * In the overloaded case, try and keep the load balanced. | |
1679 | */ | |
305c1fac SD |
1680 | load = task_h_load(env->p) - task_h_load(cur); |
1681 | if (!load) | |
1682 | goto assign; | |
1683 | ||
e720fff6 PZ |
1684 | dst_load = env->dst_stats.load + load; |
1685 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1686 | |
28a21745 | 1687 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1688 | goto unlock; |
1689 | ||
305c1fac | 1690 | assign: |
ba7e5a27 RR |
1691 | /* |
1692 | * One idle CPU per node is evaluated for a task numa move. | |
1693 | * Call select_idle_sibling to maybe find a better one. | |
1694 | */ | |
10e2f1ac PZ |
1695 | if (!cur) { |
1696 | /* | |
97fb7a0a | 1697 | * select_idle_siblings() uses an per-CPU cpumask that |
10e2f1ac PZ |
1698 | * can be used from IRQ context. |
1699 | */ | |
1700 | local_irq_disable(); | |
772bd008 MR |
1701 | env->dst_cpu = select_idle_sibling(env->p, env->src_cpu, |
1702 | env->dst_cpu); | |
10e2f1ac PZ |
1703 | local_irq_enable(); |
1704 | } | |
ba7e5a27 | 1705 | |
fb13c7ee MG |
1706 | task_numa_assign(env, cur, imp); |
1707 | unlock: | |
1708 | rcu_read_unlock(); | |
1709 | } | |
1710 | ||
887c290e RR |
1711 | static void task_numa_find_cpu(struct task_numa_env *env, |
1712 | long taskimp, long groupimp) | |
2c8a50aa | 1713 | { |
305c1fac SD |
1714 | long src_load, dst_load, load; |
1715 | bool maymove = false; | |
2c8a50aa MG |
1716 | int cpu; |
1717 | ||
305c1fac SD |
1718 | load = task_h_load(env->p); |
1719 | dst_load = env->dst_stats.load + load; | |
1720 | src_load = env->src_stats.load - load; | |
1721 | ||
1722 | /* | |
1723 | * If the improvement from just moving env->p direction is better | |
1724 | * than swapping tasks around, check if a move is possible. | |
1725 | */ | |
1726 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
1727 | ||
2c8a50aa MG |
1728 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1729 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 1730 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
1731 | continue; |
1732 | ||
1733 | env->dst_cpu = cpu; | |
305c1fac | 1734 | task_numa_compare(env, taskimp, groupimp, maymove); |
2c8a50aa MG |
1735 | } |
1736 | } | |
1737 | ||
58d081b5 MG |
1738 | static int task_numa_migrate(struct task_struct *p) |
1739 | { | |
58d081b5 MG |
1740 | struct task_numa_env env = { |
1741 | .p = p, | |
fb13c7ee | 1742 | |
58d081b5 | 1743 | .src_cpu = task_cpu(p), |
b32e86b4 | 1744 | .src_nid = task_node(p), |
fb13c7ee MG |
1745 | |
1746 | .imbalance_pct = 112, | |
1747 | ||
1748 | .best_task = NULL, | |
1749 | .best_imp = 0, | |
4142c3eb | 1750 | .best_cpu = -1, |
58d081b5 | 1751 | }; |
cb361d8c | 1752 | unsigned long taskweight, groupweight; |
58d081b5 | 1753 | struct sched_domain *sd; |
cb361d8c JH |
1754 | long taskimp, groupimp; |
1755 | struct numa_group *ng; | |
a4739eca | 1756 | struct rq *best_rq; |
7bd95320 | 1757 | int nid, ret, dist; |
e6628d5b | 1758 | |
58d081b5 | 1759 | /* |
fb13c7ee MG |
1760 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1761 | * imbalance and would be the first to start moving tasks about. | |
1762 | * | |
1763 | * And we want to avoid any moving of tasks about, as that would create | |
1764 | * random movement of tasks -- counter the numa conditions we're trying | |
1765 | * to satisfy here. | |
58d081b5 MG |
1766 | */ |
1767 | rcu_read_lock(); | |
fb13c7ee | 1768 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1769 | if (sd) |
1770 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1771 | rcu_read_unlock(); |
1772 | ||
46a73e8a RR |
1773 | /* |
1774 | * Cpusets can break the scheduler domain tree into smaller | |
1775 | * balance domains, some of which do not cross NUMA boundaries. | |
1776 | * Tasks that are "trapped" in such domains cannot be migrated | |
1777 | * elsewhere, so there is no point in (re)trying. | |
1778 | */ | |
1779 | if (unlikely(!sd)) { | |
8cd45eee | 1780 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
1781 | return -EINVAL; |
1782 | } | |
1783 | ||
2c8a50aa | 1784 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1785 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1786 | taskweight = task_weight(p, env.src_nid, dist); | |
1787 | groupweight = group_weight(p, env.src_nid, dist); | |
1788 | update_numa_stats(&env.src_stats, env.src_nid); | |
1789 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1790 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1791 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1792 | |
a43455a1 | 1793 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 1794 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 1795 | |
9de05d48 RR |
1796 | /* |
1797 | * Look at other nodes in these cases: | |
1798 | * - there is no space available on the preferred_nid | |
1799 | * - the task is part of a numa_group that is interleaved across | |
1800 | * multiple NUMA nodes; in order to better consolidate the group, | |
1801 | * we need to check other locations. | |
1802 | */ | |
cb361d8c JH |
1803 | ng = deref_curr_numa_group(p); |
1804 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
2c8a50aa MG |
1805 | for_each_online_node(nid) { |
1806 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1807 | continue; | |
58d081b5 | 1808 | |
7bd95320 | 1809 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1810 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1811 | dist != env.dist) { | |
1812 | taskweight = task_weight(p, env.src_nid, dist); | |
1813 | groupweight = group_weight(p, env.src_nid, dist); | |
1814 | } | |
7bd95320 | 1815 | |
83e1d2cd | 1816 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1817 | taskimp = task_weight(p, nid, dist) - taskweight; |
1818 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1819 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1820 | continue; |
1821 | ||
7bd95320 | 1822 | env.dist = dist; |
2c8a50aa MG |
1823 | env.dst_nid = nid; |
1824 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
2d4056fa | 1825 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1826 | } |
1827 | } | |
1828 | ||
68d1b02a RR |
1829 | /* |
1830 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1831 | * and is migrating into one of the workload's active nodes, remember | |
1832 | * this node as the task's preferred numa node, so the workload can | |
1833 | * settle down. | |
1834 | * A task that migrated to a second choice node will be better off | |
1835 | * trying for a better one later. Do not set the preferred node here. | |
1836 | */ | |
cb361d8c | 1837 | if (ng) { |
db015dae RR |
1838 | if (env.best_cpu == -1) |
1839 | nid = env.src_nid; | |
1840 | else | |
8cd45eee | 1841 | nid = cpu_to_node(env.best_cpu); |
db015dae | 1842 | |
8cd45eee SD |
1843 | if (nid != p->numa_preferred_nid) |
1844 | sched_setnuma(p, nid); | |
db015dae RR |
1845 | } |
1846 | ||
1847 | /* No better CPU than the current one was found. */ | |
1848 | if (env.best_cpu == -1) | |
1849 | return -EAGAIN; | |
0ec8aa00 | 1850 | |
a4739eca | 1851 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 1852 | if (env.best_task == NULL) { |
286549dc | 1853 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 1854 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc MG |
1855 | if (ret != 0) |
1856 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1857 | return ret; |
1858 | } | |
1859 | ||
0ad4e3df | 1860 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 1861 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 1862 | |
286549dc MG |
1863 | if (ret != 0) |
1864 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1865 | put_task_struct(env.best_task); |
1866 | return ret; | |
e6628d5b MG |
1867 | } |
1868 | ||
6b9a7460 MG |
1869 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1870 | static void numa_migrate_preferred(struct task_struct *p) | |
1871 | { | |
5085e2a3 RR |
1872 | unsigned long interval = HZ; |
1873 | ||
2739d3ee | 1874 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 1875 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
1876 | return; |
1877 | ||
2739d3ee | 1878 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 1879 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 1880 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
1881 | |
1882 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1883 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1884 | return; |
1885 | ||
1886 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1887 | task_numa_migrate(p); |
6b9a7460 MG |
1888 | } |
1889 | ||
20e07dea | 1890 | /* |
4142c3eb | 1891 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
1892 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
1893 | * be different from the set of nodes where the workload's memory is currently | |
1894 | * located. | |
20e07dea | 1895 | */ |
4142c3eb | 1896 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
1897 | { |
1898 | unsigned long faults, max_faults = 0; | |
4142c3eb | 1899 | int nid, active_nodes = 0; |
20e07dea RR |
1900 | |
1901 | for_each_online_node(nid) { | |
1902 | faults = group_faults_cpu(numa_group, nid); | |
1903 | if (faults > max_faults) | |
1904 | max_faults = faults; | |
1905 | } | |
1906 | ||
1907 | for_each_online_node(nid) { | |
1908 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
1909 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
1910 | active_nodes++; | |
20e07dea | 1911 | } |
4142c3eb RR |
1912 | |
1913 | numa_group->max_faults_cpu = max_faults; | |
1914 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
1915 | } |
1916 | ||
04bb2f94 RR |
1917 | /* |
1918 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1919 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1920 | * period will be for the next scan window. If local/(local+remote) ratio is |
1921 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1922 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1923 | */ |
1924 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1925 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1926 | |
1927 | /* | |
1928 | * Increase the scan period (slow down scanning) if the majority of | |
1929 | * our memory is already on our local node, or if the majority of | |
1930 | * the page accesses are shared with other processes. | |
1931 | * Otherwise, decrease the scan period. | |
1932 | */ | |
1933 | static void update_task_scan_period(struct task_struct *p, | |
1934 | unsigned long shared, unsigned long private) | |
1935 | { | |
1936 | unsigned int period_slot; | |
37ec97de | 1937 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
1938 | int diff; |
1939 | ||
1940 | unsigned long remote = p->numa_faults_locality[0]; | |
1941 | unsigned long local = p->numa_faults_locality[1]; | |
1942 | ||
1943 | /* | |
1944 | * If there were no record hinting faults then either the task is | |
1945 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
1946 | * to automatic numa balancing. Related to that, if there were failed |
1947 | * migration then it implies we are migrating too quickly or the local | |
1948 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 1949 | */ |
074c2381 | 1950 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
1951 | p->numa_scan_period = min(p->numa_scan_period_max, |
1952 | p->numa_scan_period << 1); | |
1953 | ||
1954 | p->mm->numa_next_scan = jiffies + | |
1955 | msecs_to_jiffies(p->numa_scan_period); | |
1956 | ||
1957 | return; | |
1958 | } | |
1959 | ||
1960 | /* | |
1961 | * Prepare to scale scan period relative to the current period. | |
1962 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1963 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1964 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1965 | */ | |
1966 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
1967 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
1968 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
1969 | ||
1970 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
1971 | /* | |
1972 | * Most memory accesses are local. There is no need to | |
1973 | * do fast NUMA scanning, since memory is already local. | |
1974 | */ | |
1975 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
1976 | if (!slot) | |
1977 | slot = 1; | |
1978 | diff = slot * period_slot; | |
1979 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
1980 | /* | |
1981 | * Most memory accesses are shared with other tasks. | |
1982 | * There is no point in continuing fast NUMA scanning, | |
1983 | * since other tasks may just move the memory elsewhere. | |
1984 | */ | |
1985 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
1986 | if (!slot) |
1987 | slot = 1; | |
1988 | diff = slot * period_slot; | |
1989 | } else { | |
04bb2f94 | 1990 | /* |
37ec97de RR |
1991 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
1992 | * yet they are not on the local NUMA node. Speed up | |
1993 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 1994 | */ |
37ec97de RR |
1995 | int ratio = max(lr_ratio, ps_ratio); |
1996 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
1997 | } |
1998 | ||
1999 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2000 | task_scan_min(p), task_scan_max(p)); | |
2001 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2002 | } | |
2003 | ||
7e2703e6 RR |
2004 | /* |
2005 | * Get the fraction of time the task has been running since the last | |
2006 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2007 | * decays those on a 32ms period, which is orders of magnitude off | |
2008 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2009 | * stats only if the task is so new there are no NUMA statistics yet. | |
2010 | */ | |
2011 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2012 | { | |
2013 | u64 runtime, delta, now; | |
2014 | /* Use the start of this time slice to avoid calculations. */ | |
2015 | now = p->se.exec_start; | |
2016 | runtime = p->se.sum_exec_runtime; | |
2017 | ||
2018 | if (p->last_task_numa_placement) { | |
2019 | delta = runtime - p->last_sum_exec_runtime; | |
2020 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2021 | |
2022 | /* Avoid time going backwards, prevent potential divide error: */ | |
2023 | if (unlikely((s64)*period < 0)) | |
2024 | *period = 0; | |
7e2703e6 | 2025 | } else { |
c7b50216 | 2026 | delta = p->se.avg.load_sum; |
9d89c257 | 2027 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2028 | } |
2029 | ||
2030 | p->last_sum_exec_runtime = runtime; | |
2031 | p->last_task_numa_placement = now; | |
2032 | ||
2033 | return delta; | |
2034 | } | |
2035 | ||
54009416 RR |
2036 | /* |
2037 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2038 | * be done in a way that produces consistent results with group_weight, | |
2039 | * otherwise workloads might not converge. | |
2040 | */ | |
2041 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2042 | { | |
2043 | nodemask_t nodes; | |
2044 | int dist; | |
2045 | ||
2046 | /* Direct connections between all NUMA nodes. */ | |
2047 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2048 | return nid; | |
2049 | ||
2050 | /* | |
2051 | * On a system with glueless mesh NUMA topology, group_weight | |
2052 | * scores nodes according to the number of NUMA hinting faults on | |
2053 | * both the node itself, and on nearby nodes. | |
2054 | */ | |
2055 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2056 | unsigned long score, max_score = 0; | |
2057 | int node, max_node = nid; | |
2058 | ||
2059 | dist = sched_max_numa_distance; | |
2060 | ||
2061 | for_each_online_node(node) { | |
2062 | score = group_weight(p, node, dist); | |
2063 | if (score > max_score) { | |
2064 | max_score = score; | |
2065 | max_node = node; | |
2066 | } | |
2067 | } | |
2068 | return max_node; | |
2069 | } | |
2070 | ||
2071 | /* | |
2072 | * Finding the preferred nid in a system with NUMA backplane | |
2073 | * interconnect topology is more involved. The goal is to locate | |
2074 | * tasks from numa_groups near each other in the system, and | |
2075 | * untangle workloads from different sides of the system. This requires | |
2076 | * searching down the hierarchy of node groups, recursively searching | |
2077 | * inside the highest scoring group of nodes. The nodemask tricks | |
2078 | * keep the complexity of the search down. | |
2079 | */ | |
2080 | nodes = node_online_map; | |
2081 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2082 | unsigned long max_faults = 0; | |
81907478 | 2083 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2084 | int a, b; |
2085 | ||
2086 | /* Are there nodes at this distance from each other? */ | |
2087 | if (!find_numa_distance(dist)) | |
2088 | continue; | |
2089 | ||
2090 | for_each_node_mask(a, nodes) { | |
2091 | unsigned long faults = 0; | |
2092 | nodemask_t this_group; | |
2093 | nodes_clear(this_group); | |
2094 | ||
2095 | /* Sum group's NUMA faults; includes a==b case. */ | |
2096 | for_each_node_mask(b, nodes) { | |
2097 | if (node_distance(a, b) < dist) { | |
2098 | faults += group_faults(p, b); | |
2099 | node_set(b, this_group); | |
2100 | node_clear(b, nodes); | |
2101 | } | |
2102 | } | |
2103 | ||
2104 | /* Remember the top group. */ | |
2105 | if (faults > max_faults) { | |
2106 | max_faults = faults; | |
2107 | max_group = this_group; | |
2108 | /* | |
2109 | * subtle: at the smallest distance there is | |
2110 | * just one node left in each "group", the | |
2111 | * winner is the preferred nid. | |
2112 | */ | |
2113 | nid = a; | |
2114 | } | |
2115 | } | |
2116 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2117 | if (!max_faults) |
2118 | break; | |
54009416 RR |
2119 | nodes = max_group; |
2120 | } | |
2121 | return nid; | |
2122 | } | |
2123 | ||
cbee9f88 PZ |
2124 | static void task_numa_placement(struct task_struct *p) |
2125 | { | |
98fa15f3 | 2126 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2127 | unsigned long max_faults = 0; |
04bb2f94 | 2128 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2129 | unsigned long total_faults; |
2130 | u64 runtime, period; | |
7dbd13ed | 2131 | spinlock_t *group_lock = NULL; |
cb361d8c | 2132 | struct numa_group *ng; |
cbee9f88 | 2133 | |
7e5a2c17 JL |
2134 | /* |
2135 | * The p->mm->numa_scan_seq field gets updated without | |
2136 | * exclusive access. Use READ_ONCE() here to ensure | |
2137 | * that the field is read in a single access: | |
2138 | */ | |
316c1608 | 2139 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2140 | if (p->numa_scan_seq == seq) |
2141 | return; | |
2142 | p->numa_scan_seq = seq; | |
598f0ec0 | 2143 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2144 | |
7e2703e6 RR |
2145 | total_faults = p->numa_faults_locality[0] + |
2146 | p->numa_faults_locality[1]; | |
2147 | runtime = numa_get_avg_runtime(p, &period); | |
2148 | ||
7dbd13ed | 2149 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2150 | ng = deref_curr_numa_group(p); |
2151 | if (ng) { | |
2152 | group_lock = &ng->lock; | |
60e69eed | 2153 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2154 | } |
2155 | ||
688b7585 MG |
2156 | /* Find the node with the highest number of faults */ |
2157 | for_each_online_node(nid) { | |
44dba3d5 IM |
2158 | /* Keep track of the offsets in numa_faults array */ |
2159 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2160 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2161 | int priv; |
745d6147 | 2162 | |
be1e4e76 | 2163 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2164 | long diff, f_diff, f_weight; |
8c8a743c | 2165 | |
44dba3d5 IM |
2166 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2167 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2168 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2169 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2170 | |
ac8e895b | 2171 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2172 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2173 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2174 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2175 | |
7e2703e6 RR |
2176 | /* |
2177 | * Normalize the faults_from, so all tasks in a group | |
2178 | * count according to CPU use, instead of by the raw | |
2179 | * number of faults. Tasks with little runtime have | |
2180 | * little over-all impact on throughput, and thus their | |
2181 | * faults are less important. | |
2182 | */ | |
2183 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2184 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2185 | (total_faults + 1); |
44dba3d5 IM |
2186 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2187 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2188 | |
44dba3d5 IM |
2189 | p->numa_faults[mem_idx] += diff; |
2190 | p->numa_faults[cpu_idx] += f_diff; | |
2191 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2192 | p->total_numa_faults += diff; |
cb361d8c | 2193 | if (ng) { |
44dba3d5 IM |
2194 | /* |
2195 | * safe because we can only change our own group | |
2196 | * | |
2197 | * mem_idx represents the offset for a given | |
2198 | * nid and priv in a specific region because it | |
2199 | * is at the beginning of the numa_faults array. | |
2200 | */ | |
cb361d8c JH |
2201 | ng->faults[mem_idx] += diff; |
2202 | ng->faults_cpu[mem_idx] += f_diff; | |
2203 | ng->total_faults += diff; | |
2204 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2205 | } |
ac8e895b MG |
2206 | } |
2207 | ||
cb361d8c | 2208 | if (!ng) { |
f03bb676 SD |
2209 | if (faults > max_faults) { |
2210 | max_faults = faults; | |
2211 | max_nid = nid; | |
2212 | } | |
2213 | } else if (group_faults > max_faults) { | |
2214 | max_faults = group_faults; | |
688b7585 MG |
2215 | max_nid = nid; |
2216 | } | |
83e1d2cd MG |
2217 | } |
2218 | ||
cb361d8c JH |
2219 | if (ng) { |
2220 | numa_group_count_active_nodes(ng); | |
60e69eed | 2221 | spin_unlock_irq(group_lock); |
f03bb676 | 2222 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2223 | } |
2224 | ||
bb97fc31 RR |
2225 | if (max_faults) { |
2226 | /* Set the new preferred node */ | |
2227 | if (max_nid != p->numa_preferred_nid) | |
2228 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2229 | } |
30619c89 SD |
2230 | |
2231 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2232 | } |
2233 | ||
8c8a743c PZ |
2234 | static inline int get_numa_group(struct numa_group *grp) |
2235 | { | |
c45a7795 | 2236 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2237 | } |
2238 | ||
2239 | static inline void put_numa_group(struct numa_group *grp) | |
2240 | { | |
c45a7795 | 2241 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2242 | kfree_rcu(grp, rcu); |
2243 | } | |
2244 | ||
3e6a9418 MG |
2245 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2246 | int *priv) | |
8c8a743c PZ |
2247 | { |
2248 | struct numa_group *grp, *my_grp; | |
2249 | struct task_struct *tsk; | |
2250 | bool join = false; | |
2251 | int cpu = cpupid_to_cpu(cpupid); | |
2252 | int i; | |
2253 | ||
cb361d8c | 2254 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2255 | unsigned int size = sizeof(struct numa_group) + |
50ec8a40 | 2256 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2257 | |
2258 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2259 | if (!grp) | |
2260 | return; | |
2261 | ||
c45a7795 | 2262 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2263 | grp->active_nodes = 1; |
2264 | grp->max_faults_cpu = 0; | |
8c8a743c | 2265 | spin_lock_init(&grp->lock); |
e29cf08b | 2266 | grp->gid = p->pid; |
50ec8a40 | 2267 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2268 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2269 | nr_node_ids; | |
8c8a743c | 2270 | |
be1e4e76 | 2271 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2272 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2273 | |
989348b5 | 2274 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2275 | |
8c8a743c PZ |
2276 | grp->nr_tasks++; |
2277 | rcu_assign_pointer(p->numa_group, grp); | |
2278 | } | |
2279 | ||
2280 | rcu_read_lock(); | |
316c1608 | 2281 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2282 | |
2283 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2284 | goto no_join; |
8c8a743c PZ |
2285 | |
2286 | grp = rcu_dereference(tsk->numa_group); | |
2287 | if (!grp) | |
3354781a | 2288 | goto no_join; |
8c8a743c | 2289 | |
cb361d8c | 2290 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2291 | if (grp == my_grp) |
3354781a | 2292 | goto no_join; |
8c8a743c PZ |
2293 | |
2294 | /* | |
2295 | * Only join the other group if its bigger; if we're the bigger group, | |
2296 | * the other task will join us. | |
2297 | */ | |
2298 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2299 | goto no_join; |
8c8a743c PZ |
2300 | |
2301 | /* | |
2302 | * Tie-break on the grp address. | |
2303 | */ | |
2304 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2305 | goto no_join; |
8c8a743c | 2306 | |
dabe1d99 RR |
2307 | /* Always join threads in the same process. */ |
2308 | if (tsk->mm == current->mm) | |
2309 | join = true; | |
2310 | ||
2311 | /* Simple filter to avoid false positives due to PID collisions */ | |
2312 | if (flags & TNF_SHARED) | |
2313 | join = true; | |
8c8a743c | 2314 | |
3e6a9418 MG |
2315 | /* Update priv based on whether false sharing was detected */ |
2316 | *priv = !join; | |
2317 | ||
dabe1d99 | 2318 | if (join && !get_numa_group(grp)) |
3354781a | 2319 | goto no_join; |
8c8a743c | 2320 | |
8c8a743c PZ |
2321 | rcu_read_unlock(); |
2322 | ||
2323 | if (!join) | |
2324 | return; | |
2325 | ||
60e69eed MG |
2326 | BUG_ON(irqs_disabled()); |
2327 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2328 | |
be1e4e76 | 2329 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2330 | my_grp->faults[i] -= p->numa_faults[i]; |
2331 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2332 | } |
989348b5 MG |
2333 | my_grp->total_faults -= p->total_numa_faults; |
2334 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2335 | |
8c8a743c PZ |
2336 | my_grp->nr_tasks--; |
2337 | grp->nr_tasks++; | |
2338 | ||
2339 | spin_unlock(&my_grp->lock); | |
60e69eed | 2340 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2341 | |
2342 | rcu_assign_pointer(p->numa_group, grp); | |
2343 | ||
2344 | put_numa_group(my_grp); | |
3354781a PZ |
2345 | return; |
2346 | ||
2347 | no_join: | |
2348 | rcu_read_unlock(); | |
2349 | return; | |
8c8a743c PZ |
2350 | } |
2351 | ||
16d51a59 JH |
2352 | /* |
2353 | * Get rid of NUMA staticstics associated with a task (either current or dead). | |
2354 | * If @final is set, the task is dead and has reached refcount zero, so we can | |
2355 | * safely free all relevant data structures. Otherwise, there might be | |
2356 | * concurrent reads from places like load balancing and procfs, and we should | |
2357 | * reset the data back to default state without freeing ->numa_faults. | |
2358 | */ | |
2359 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 2360 | { |
cb361d8c JH |
2361 | /* safe: p either is current or is being freed by current */ |
2362 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 2363 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
2364 | unsigned long flags; |
2365 | int i; | |
8c8a743c | 2366 | |
16d51a59 JH |
2367 | if (!numa_faults) |
2368 | return; | |
2369 | ||
8c8a743c | 2370 | if (grp) { |
e9dd685c | 2371 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2372 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2373 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2374 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2375 | |
8c8a743c | 2376 | grp->nr_tasks--; |
e9dd685c | 2377 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2378 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2379 | put_numa_group(grp); |
2380 | } | |
2381 | ||
16d51a59 JH |
2382 | if (final) { |
2383 | p->numa_faults = NULL; | |
2384 | kfree(numa_faults); | |
2385 | } else { | |
2386 | p->total_numa_faults = 0; | |
2387 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
2388 | numa_faults[i] = 0; | |
2389 | } | |
8c8a743c PZ |
2390 | } |
2391 | ||
cbee9f88 PZ |
2392 | /* |
2393 | * Got a PROT_NONE fault for a page on @node. | |
2394 | */ | |
58b46da3 | 2395 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2396 | { |
2397 | struct task_struct *p = current; | |
6688cc05 | 2398 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2399 | int cpu_node = task_node(current); |
792568ec | 2400 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2401 | struct numa_group *ng; |
ac8e895b | 2402 | int priv; |
cbee9f88 | 2403 | |
2a595721 | 2404 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2405 | return; |
2406 | ||
9ff1d9ff MG |
2407 | /* for example, ksmd faulting in a user's mm */ |
2408 | if (!p->mm) | |
2409 | return; | |
2410 | ||
f809ca9a | 2411 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2412 | if (unlikely(!p->numa_faults)) { |
2413 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2414 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2415 | |
44dba3d5 IM |
2416 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2417 | if (!p->numa_faults) | |
f809ca9a | 2418 | return; |
745d6147 | 2419 | |
83e1d2cd | 2420 | p->total_numa_faults = 0; |
04bb2f94 | 2421 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2422 | } |
cbee9f88 | 2423 | |
8c8a743c PZ |
2424 | /* |
2425 | * First accesses are treated as private, otherwise consider accesses | |
2426 | * to be private if the accessing pid has not changed | |
2427 | */ | |
2428 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2429 | priv = 1; | |
2430 | } else { | |
2431 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2432 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2433 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2434 | } |
2435 | ||
792568ec RR |
2436 | /* |
2437 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2438 | * occurs wholly within the set of nodes that the workload is | |
2439 | * actively using should be counted as local. This allows the | |
2440 | * scan rate to slow down when a workload has settled down. | |
2441 | */ | |
cb361d8c | 2442 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
2443 | if (!priv && !local && ng && ng->active_nodes > 1 && |
2444 | numa_is_active_node(cpu_node, ng) && | |
2445 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2446 | local = 1; |
2447 | ||
2739d3ee | 2448 | /* |
e1ff516a YW |
2449 | * Retry to migrate task to preferred node periodically, in case it |
2450 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2451 | */ |
b6a60cf3 SD |
2452 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2453 | task_numa_placement(p); | |
6b9a7460 | 2454 | numa_migrate_preferred(p); |
b6a60cf3 | 2455 | } |
6b9a7460 | 2456 | |
b32e86b4 IM |
2457 | if (migrated) |
2458 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2459 | if (flags & TNF_MIGRATE_FAIL) |
2460 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2461 | |
44dba3d5 IM |
2462 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2463 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2464 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2465 | } |
2466 | ||
6e5fb223 PZ |
2467 | static void reset_ptenuma_scan(struct task_struct *p) |
2468 | { | |
7e5a2c17 JL |
2469 | /* |
2470 | * We only did a read acquisition of the mmap sem, so | |
2471 | * p->mm->numa_scan_seq is written to without exclusive access | |
2472 | * and the update is not guaranteed to be atomic. That's not | |
2473 | * much of an issue though, since this is just used for | |
2474 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2475 | * expensive, to avoid any form of compiler optimizations: | |
2476 | */ | |
316c1608 | 2477 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2478 | p->mm->numa_scan_offset = 0; |
2479 | } | |
2480 | ||
cbee9f88 PZ |
2481 | /* |
2482 | * The expensive part of numa migration is done from task_work context. | |
2483 | * Triggered from task_tick_numa(). | |
2484 | */ | |
9434f9f5 | 2485 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
2486 | { |
2487 | unsigned long migrate, next_scan, now = jiffies; | |
2488 | struct task_struct *p = current; | |
2489 | struct mm_struct *mm = p->mm; | |
51170840 | 2490 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2491 | struct vm_area_struct *vma; |
9f40604c | 2492 | unsigned long start, end; |
598f0ec0 | 2493 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2494 | long pages, virtpages; |
cbee9f88 | 2495 | |
9148a3a1 | 2496 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 2497 | |
b34920d4 | 2498 | work->next = work; |
cbee9f88 PZ |
2499 | /* |
2500 | * Who cares about NUMA placement when they're dying. | |
2501 | * | |
2502 | * NOTE: make sure not to dereference p->mm before this check, | |
2503 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2504 | * without p->mm even though we still had it when we enqueued this | |
2505 | * work. | |
2506 | */ | |
2507 | if (p->flags & PF_EXITING) | |
2508 | return; | |
2509 | ||
930aa174 | 2510 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2511 | mm->numa_next_scan = now + |
2512 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2513 | } |
2514 | ||
cbee9f88 PZ |
2515 | /* |
2516 | * Enforce maximal scan/migration frequency.. | |
2517 | */ | |
2518 | migrate = mm->numa_next_scan; | |
2519 | if (time_before(now, migrate)) | |
2520 | return; | |
2521 | ||
598f0ec0 MG |
2522 | if (p->numa_scan_period == 0) { |
2523 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2524 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2525 | } |
cbee9f88 | 2526 | |
fb003b80 | 2527 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2528 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2529 | return; | |
2530 | ||
19a78d11 PZ |
2531 | /* |
2532 | * Delay this task enough that another task of this mm will likely win | |
2533 | * the next time around. | |
2534 | */ | |
2535 | p->node_stamp += 2 * TICK_NSEC; | |
2536 | ||
9f40604c MG |
2537 | start = mm->numa_scan_offset; |
2538 | pages = sysctl_numa_balancing_scan_size; | |
2539 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2540 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2541 | if (!pages) |
2542 | return; | |
cbee9f88 | 2543 | |
4620f8c1 | 2544 | |
8655d549 VB |
2545 | if (!down_read_trylock(&mm->mmap_sem)) |
2546 | return; | |
9f40604c | 2547 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2548 | if (!vma) { |
2549 | reset_ptenuma_scan(p); | |
9f40604c | 2550 | start = 0; |
6e5fb223 PZ |
2551 | vma = mm->mmap; |
2552 | } | |
9f40604c | 2553 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2554 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2555 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2556 | continue; |
6b79c57b | 2557 | } |
6e5fb223 | 2558 | |
4591ce4f MG |
2559 | /* |
2560 | * Shared library pages mapped by multiple processes are not | |
2561 | * migrated as it is expected they are cache replicated. Avoid | |
2562 | * hinting faults in read-only file-backed mappings or the vdso | |
2563 | * as migrating the pages will be of marginal benefit. | |
2564 | */ | |
2565 | if (!vma->vm_mm || | |
2566 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2567 | continue; | |
2568 | ||
3c67f474 MG |
2569 | /* |
2570 | * Skip inaccessible VMAs to avoid any confusion between | |
2571 | * PROT_NONE and NUMA hinting ptes | |
2572 | */ | |
2573 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2574 | continue; | |
4591ce4f | 2575 | |
9f40604c MG |
2576 | do { |
2577 | start = max(start, vma->vm_start); | |
2578 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2579 | end = min(end, vma->vm_end); | |
4620f8c1 | 2580 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2581 | |
2582 | /* | |
4620f8c1 RR |
2583 | * Try to scan sysctl_numa_balancing_size worth of |
2584 | * hpages that have at least one present PTE that | |
2585 | * is not already pte-numa. If the VMA contains | |
2586 | * areas that are unused or already full of prot_numa | |
2587 | * PTEs, scan up to virtpages, to skip through those | |
2588 | * areas faster. | |
598f0ec0 MG |
2589 | */ |
2590 | if (nr_pte_updates) | |
2591 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2592 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2593 | |
9f40604c | 2594 | start = end; |
4620f8c1 | 2595 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2596 | goto out; |
3cf1962c RR |
2597 | |
2598 | cond_resched(); | |
9f40604c | 2599 | } while (end != vma->vm_end); |
cbee9f88 | 2600 | } |
6e5fb223 | 2601 | |
9f40604c | 2602 | out: |
6e5fb223 | 2603 | /* |
c69307d5 PZ |
2604 | * It is possible to reach the end of the VMA list but the last few |
2605 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2606 | * would find the !migratable VMA on the next scan but not reset the | |
2607 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2608 | */ |
2609 | if (vma) | |
9f40604c | 2610 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2611 | else |
2612 | reset_ptenuma_scan(p); | |
2613 | up_read(&mm->mmap_sem); | |
51170840 RR |
2614 | |
2615 | /* | |
2616 | * Make sure tasks use at least 32x as much time to run other code | |
2617 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2618 | * Usually update_task_scan_period slows down scanning enough; on an | |
2619 | * overloaded system we need to limit overhead on a per task basis. | |
2620 | */ | |
2621 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2622 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2623 | p->node_stamp += 32 * diff; | |
2624 | } | |
cbee9f88 PZ |
2625 | } |
2626 | ||
d35927a1 VS |
2627 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
2628 | { | |
2629 | int mm_users = 0; | |
2630 | struct mm_struct *mm = p->mm; | |
2631 | ||
2632 | if (mm) { | |
2633 | mm_users = atomic_read(&mm->mm_users); | |
2634 | if (mm_users == 1) { | |
2635 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
2636 | mm->numa_scan_seq = 0; | |
2637 | } | |
2638 | } | |
2639 | p->node_stamp = 0; | |
2640 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
2641 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
b34920d4 | 2642 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
2643 | p->numa_work.next = &p->numa_work; |
2644 | p->numa_faults = NULL; | |
2645 | RCU_INIT_POINTER(p->numa_group, NULL); | |
2646 | p->last_task_numa_placement = 0; | |
2647 | p->last_sum_exec_runtime = 0; | |
2648 | ||
b34920d4 VS |
2649 | init_task_work(&p->numa_work, task_numa_work); |
2650 | ||
d35927a1 VS |
2651 | /* New address space, reset the preferred nid */ |
2652 | if (!(clone_flags & CLONE_VM)) { | |
2653 | p->numa_preferred_nid = NUMA_NO_NODE; | |
2654 | return; | |
2655 | } | |
2656 | ||
2657 | /* | |
2658 | * New thread, keep existing numa_preferred_nid which should be copied | |
2659 | * already by arch_dup_task_struct but stagger when scans start. | |
2660 | */ | |
2661 | if (mm) { | |
2662 | unsigned int delay; | |
2663 | ||
2664 | delay = min_t(unsigned int, task_scan_max(current), | |
2665 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
2666 | delay += 2 * TICK_NSEC; | |
2667 | p->node_stamp = delay; | |
2668 | } | |
2669 | } | |
2670 | ||
cbee9f88 PZ |
2671 | /* |
2672 | * Drive the periodic memory faults.. | |
2673 | */ | |
b1546edc | 2674 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
2675 | { |
2676 | struct callback_head *work = &curr->numa_work; | |
2677 | u64 period, now; | |
2678 | ||
2679 | /* | |
2680 | * We don't care about NUMA placement if we don't have memory. | |
2681 | */ | |
2682 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2683 | return; | |
2684 | ||
2685 | /* | |
2686 | * Using runtime rather than walltime has the dual advantage that | |
2687 | * we (mostly) drive the selection from busy threads and that the | |
2688 | * task needs to have done some actual work before we bother with | |
2689 | * NUMA placement. | |
2690 | */ | |
2691 | now = curr->se.sum_exec_runtime; | |
2692 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2693 | ||
25b3e5a3 | 2694 | if (now > curr->node_stamp + period) { |
4b96a29b | 2695 | if (!curr->node_stamp) |
b5dd77c8 | 2696 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2697 | curr->node_stamp += period; |
cbee9f88 | 2698 | |
b34920d4 | 2699 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
cbee9f88 | 2700 | task_work_add(curr, work, true); |
cbee9f88 PZ |
2701 | } |
2702 | } | |
3fed382b | 2703 | |
3f9672ba SD |
2704 | static void update_scan_period(struct task_struct *p, int new_cpu) |
2705 | { | |
2706 | int src_nid = cpu_to_node(task_cpu(p)); | |
2707 | int dst_nid = cpu_to_node(new_cpu); | |
2708 | ||
05cbdf4f MG |
2709 | if (!static_branch_likely(&sched_numa_balancing)) |
2710 | return; | |
2711 | ||
3f9672ba SD |
2712 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
2713 | return; | |
2714 | ||
05cbdf4f MG |
2715 | if (src_nid == dst_nid) |
2716 | return; | |
2717 | ||
2718 | /* | |
2719 | * Allow resets if faults have been trapped before one scan | |
2720 | * has completed. This is most likely due to a new task that | |
2721 | * is pulled cross-node due to wakeups or load balancing. | |
2722 | */ | |
2723 | if (p->numa_scan_seq) { | |
2724 | /* | |
2725 | * Avoid scan adjustments if moving to the preferred | |
2726 | * node or if the task was not previously running on | |
2727 | * the preferred node. | |
2728 | */ | |
2729 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
2730 | (p->numa_preferred_nid != NUMA_NO_NODE && |
2731 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
2732 | return; |
2733 | } | |
2734 | ||
2735 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
2736 | } |
2737 | ||
cbee9f88 PZ |
2738 | #else |
2739 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2740 | { | |
2741 | } | |
0ec8aa00 PZ |
2742 | |
2743 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2744 | { | |
2745 | } | |
2746 | ||
2747 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2748 | { | |
2749 | } | |
3fed382b | 2750 | |
3f9672ba SD |
2751 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
2752 | { | |
2753 | } | |
2754 | ||
cbee9f88 PZ |
2755 | #endif /* CONFIG_NUMA_BALANCING */ |
2756 | ||
30cfdcfc DA |
2757 | static void |
2758 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2759 | { | |
2760 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 2761 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2762 | if (entity_is_task(se)) { |
2763 | struct rq *rq = rq_of(cfs_rq); | |
2764 | ||
2765 | account_numa_enqueue(rq, task_of(se)); | |
2766 | list_add(&se->group_node, &rq->cfs_tasks); | |
2767 | } | |
367456c7 | 2768 | #endif |
30cfdcfc | 2769 | cfs_rq->nr_running++; |
30cfdcfc DA |
2770 | } |
2771 | ||
2772 | static void | |
2773 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2774 | { | |
2775 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 2776 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2777 | if (entity_is_task(se)) { |
2778 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2779 | list_del_init(&se->group_node); |
0ec8aa00 | 2780 | } |
bfdb198c | 2781 | #endif |
30cfdcfc | 2782 | cfs_rq->nr_running--; |
30cfdcfc DA |
2783 | } |
2784 | ||
8d5b9025 PZ |
2785 | /* |
2786 | * Signed add and clamp on underflow. | |
2787 | * | |
2788 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2789 | * memory. This allows lockless observations without ever seeing the negative | |
2790 | * values. | |
2791 | */ | |
2792 | #define add_positive(_ptr, _val) do { \ | |
2793 | typeof(_ptr) ptr = (_ptr); \ | |
2794 | typeof(_val) val = (_val); \ | |
2795 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2796 | \ | |
2797 | res = var + val; \ | |
2798 | \ | |
2799 | if (val < 0 && res > var) \ | |
2800 | res = 0; \ | |
2801 | \ | |
2802 | WRITE_ONCE(*ptr, res); \ | |
2803 | } while (0) | |
2804 | ||
2805 | /* | |
2806 | * Unsigned subtract and clamp on underflow. | |
2807 | * | |
2808 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2809 | * memory. This allows lockless observations without ever seeing the negative | |
2810 | * values. | |
2811 | */ | |
2812 | #define sub_positive(_ptr, _val) do { \ | |
2813 | typeof(_ptr) ptr = (_ptr); \ | |
2814 | typeof(*ptr) val = (_val); \ | |
2815 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2816 | res = var - val; \ | |
2817 | if (res > var) \ | |
2818 | res = 0; \ | |
2819 | WRITE_ONCE(*ptr, res); \ | |
2820 | } while (0) | |
2821 | ||
b5c0ce7b PB |
2822 | /* |
2823 | * Remove and clamp on negative, from a local variable. | |
2824 | * | |
2825 | * A variant of sub_positive(), which does not use explicit load-store | |
2826 | * and is thus optimized for local variable updates. | |
2827 | */ | |
2828 | #define lsub_positive(_ptr, _val) do { \ | |
2829 | typeof(_ptr) ptr = (_ptr); \ | |
2830 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
2831 | } while (0) | |
2832 | ||
8d5b9025 | 2833 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
2834 | static inline void |
2835 | enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2836 | { | |
1ea6c46a PZ |
2837 | cfs_rq->runnable_weight += se->runnable_weight; |
2838 | ||
2839 | cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg; | |
2840 | cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum; | |
8d5b9025 PZ |
2841 | } |
2842 | ||
2843 | static inline void | |
2844 | dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2845 | { | |
1ea6c46a PZ |
2846 | cfs_rq->runnable_weight -= se->runnable_weight; |
2847 | ||
2848 | sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg); | |
2849 | sub_positive(&cfs_rq->avg.runnable_load_sum, | |
2850 | se_runnable(se) * se->avg.runnable_load_sum); | |
8d5b9025 PZ |
2851 | } |
2852 | ||
2853 | static inline void | |
2854 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2855 | { | |
2856 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
2857 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
2858 | } | |
2859 | ||
2860 | static inline void | |
2861 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2862 | { | |
2863 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
2864 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); | |
2865 | } | |
2866 | #else | |
2867 | static inline void | |
2868 | enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2869 | static inline void | |
2870 | dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2871 | static inline void | |
2872 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2873 | static inline void | |
2874 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2875 | #endif | |
2876 | ||
9059393e | 2877 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1ea6c46a | 2878 | unsigned long weight, unsigned long runnable) |
9059393e VG |
2879 | { |
2880 | if (se->on_rq) { | |
2881 | /* commit outstanding execution time */ | |
2882 | if (cfs_rq->curr == se) | |
2883 | update_curr(cfs_rq); | |
2884 | account_entity_dequeue(cfs_rq, se); | |
2885 | dequeue_runnable_load_avg(cfs_rq, se); | |
2886 | } | |
2887 | dequeue_load_avg(cfs_rq, se); | |
2888 | ||
1ea6c46a | 2889 | se->runnable_weight = runnable; |
9059393e VG |
2890 | update_load_set(&se->load, weight); |
2891 | ||
2892 | #ifdef CONFIG_SMP | |
1ea6c46a PZ |
2893 | do { |
2894 | u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib; | |
2895 | ||
2896 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
2897 | se->avg.runnable_load_avg = | |
2898 | div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider); | |
2899 | } while (0); | |
9059393e VG |
2900 | #endif |
2901 | ||
2902 | enqueue_load_avg(cfs_rq, se); | |
2903 | if (se->on_rq) { | |
2904 | account_entity_enqueue(cfs_rq, se); | |
2905 | enqueue_runnable_load_avg(cfs_rq, se); | |
2906 | } | |
2907 | } | |
2908 | ||
2909 | void reweight_task(struct task_struct *p, int prio) | |
2910 | { | |
2911 | struct sched_entity *se = &p->se; | |
2912 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2913 | struct load_weight *load = &se->load; | |
2914 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
2915 | ||
1ea6c46a | 2916 | reweight_entity(cfs_rq, se, weight, weight); |
9059393e VG |
2917 | load->inv_weight = sched_prio_to_wmult[prio]; |
2918 | } | |
2919 | ||
3ff6dcac | 2920 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 2921 | #ifdef CONFIG_SMP |
cef27403 PZ |
2922 | /* |
2923 | * All this does is approximate the hierarchical proportion which includes that | |
2924 | * global sum we all love to hate. | |
2925 | * | |
2926 | * That is, the weight of a group entity, is the proportional share of the | |
2927 | * group weight based on the group runqueue weights. That is: | |
2928 | * | |
2929 | * tg->weight * grq->load.weight | |
2930 | * ge->load.weight = ----------------------------- (1) | |
2931 | * \Sum grq->load.weight | |
2932 | * | |
2933 | * Now, because computing that sum is prohibitively expensive to compute (been | |
2934 | * there, done that) we approximate it with this average stuff. The average | |
2935 | * moves slower and therefore the approximation is cheaper and more stable. | |
2936 | * | |
2937 | * So instead of the above, we substitute: | |
2938 | * | |
2939 | * grq->load.weight -> grq->avg.load_avg (2) | |
2940 | * | |
2941 | * which yields the following: | |
2942 | * | |
2943 | * tg->weight * grq->avg.load_avg | |
2944 | * ge->load.weight = ------------------------------ (3) | |
2945 | * tg->load_avg | |
2946 | * | |
2947 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
2948 | * | |
2949 | * That is shares_avg, and it is right (given the approximation (2)). | |
2950 | * | |
2951 | * The problem with it is that because the average is slow -- it was designed | |
2952 | * to be exactly that of course -- this leads to transients in boundary | |
2953 | * conditions. In specific, the case where the group was idle and we start the | |
2954 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
2955 | * yielding bad latency etc.. | |
2956 | * | |
2957 | * Now, in that special case (1) reduces to: | |
2958 | * | |
2959 | * tg->weight * grq->load.weight | |
17de4ee0 | 2960 | * ge->load.weight = ----------------------------- = tg->weight (4) |
cef27403 PZ |
2961 | * grp->load.weight |
2962 | * | |
2963 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
2964 | * | |
2965 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
2966 | * UP case, like: | |
2967 | * | |
2968 | * ge->load.weight = | |
2969 | * | |
2970 | * tg->weight * grq->load.weight | |
2971 | * --------------------------------------------------- (5) | |
2972 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
2973 | * | |
17de4ee0 PZ |
2974 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
2975 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
2976 | * | |
2977 | * | |
2978 | * tg->weight * grq->load.weight | |
2979 | * ge->load.weight = ----------------------------- (6) | |
2980 | * tg_load_avg' | |
2981 | * | |
2982 | * Where: | |
2983 | * | |
2984 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
2985 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
2986 | * |
2987 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
2988 | * (4) while in the normal case it approaches (3). It consistently | |
2989 | * overestimates the ge->load.weight and therefore: | |
2990 | * | |
2991 | * \Sum ge->load.weight >= tg->weight | |
2992 | * | |
2993 | * hence icky! | |
2994 | */ | |
2c8e4dce | 2995 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 2996 | { |
7c80cfc9 PZ |
2997 | long tg_weight, tg_shares, load, shares; |
2998 | struct task_group *tg = cfs_rq->tg; | |
2999 | ||
3000 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 3001 | |
3d4b60d3 | 3002 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 3003 | |
ea1dc6fc | 3004 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 3005 | |
ea1dc6fc PZ |
3006 | /* Ensure tg_weight >= load */ |
3007 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
3008 | tg_weight += load; | |
3ff6dcac | 3009 | |
7c80cfc9 | 3010 | shares = (tg_shares * load); |
cf5f0acf PZ |
3011 | if (tg_weight) |
3012 | shares /= tg_weight; | |
3ff6dcac | 3013 | |
b8fd8423 DE |
3014 | /* |
3015 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
3016 | * of a group with small tg->shares value. It is a floor value which is | |
3017 | * assigned as a minimum load.weight to the sched_entity representing | |
3018 | * the group on a CPU. | |
3019 | * | |
3020 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
3021 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
3022 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
3023 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
3024 | * instead of 0. | |
3025 | */ | |
7c80cfc9 | 3026 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 3027 | } |
2c8e4dce JB |
3028 | |
3029 | /* | |
17de4ee0 PZ |
3030 | * This calculates the effective runnable weight for a group entity based on |
3031 | * the group entity weight calculated above. | |
3032 | * | |
3033 | * Because of the above approximation (2), our group entity weight is | |
3034 | * an load_avg based ratio (3). This means that it includes blocked load and | |
3035 | * does not represent the runnable weight. | |
3036 | * | |
3037 | * Approximate the group entity's runnable weight per ratio from the group | |
3038 | * runqueue: | |
3039 | * | |
3040 | * grq->avg.runnable_load_avg | |
3041 | * ge->runnable_weight = ge->load.weight * -------------------------- (7) | |
3042 | * grq->avg.load_avg | |
3043 | * | |
3044 | * However, analogous to above, since the avg numbers are slow, this leads to | |
3045 | * transients in the from-idle case. Instead we use: | |
3046 | * | |
3047 | * ge->runnable_weight = ge->load.weight * | |
3048 | * | |
3049 | * max(grq->avg.runnable_load_avg, grq->runnable_weight) | |
3050 | * ----------------------------------------------------- (8) | |
3051 | * max(grq->avg.load_avg, grq->load.weight) | |
3052 | * | |
3053 | * Where these max() serve both to use the 'instant' values to fix the slow | |
3054 | * from-idle and avoid the /0 on to-idle, similar to (6). | |
2c8e4dce JB |
3055 | */ |
3056 | static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares) | |
3057 | { | |
17de4ee0 PZ |
3058 | long runnable, load_avg; |
3059 | ||
3060 | load_avg = max(cfs_rq->avg.load_avg, | |
3061 | scale_load_down(cfs_rq->load.weight)); | |
3062 | ||
3063 | runnable = max(cfs_rq->avg.runnable_load_avg, | |
3064 | scale_load_down(cfs_rq->runnable_weight)); | |
2c8e4dce JB |
3065 | |
3066 | runnable *= shares; | |
3067 | if (load_avg) | |
3068 | runnable /= load_avg; | |
17de4ee0 | 3069 | |
2c8e4dce JB |
3070 | return clamp_t(long, runnable, MIN_SHARES, shares); |
3071 | } | |
387f77cc | 3072 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 3073 | |
82958366 PT |
3074 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
3075 | ||
1ea6c46a PZ |
3076 | /* |
3077 | * Recomputes the group entity based on the current state of its group | |
3078 | * runqueue. | |
3079 | */ | |
3080 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 3081 | { |
1ea6c46a PZ |
3082 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
3083 | long shares, runnable; | |
2069dd75 | 3084 | |
1ea6c46a | 3085 | if (!gcfs_rq) |
89ee048f VG |
3086 | return; |
3087 | ||
1ea6c46a | 3088 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3089 | return; |
89ee048f | 3090 | |
3ff6dcac | 3091 | #ifndef CONFIG_SMP |
1ea6c46a | 3092 | runnable = shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
3093 | |
3094 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 3095 | return; |
7c80cfc9 | 3096 | #else |
2c8e4dce JB |
3097 | shares = calc_group_shares(gcfs_rq); |
3098 | runnable = calc_group_runnable(gcfs_rq, shares); | |
3ff6dcac | 3099 | #endif |
2069dd75 | 3100 | |
1ea6c46a | 3101 | reweight_entity(cfs_rq_of(se), se, shares, runnable); |
2069dd75 | 3102 | } |
89ee048f | 3103 | |
2069dd75 | 3104 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3105 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3106 | { |
3107 | } | |
3108 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3109 | ||
ea14b57e | 3110 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3111 | { |
43964409 LT |
3112 | struct rq *rq = rq_of(cfs_rq); |
3113 | ||
ea14b57e | 3114 | if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) { |
a030d738 VK |
3115 | /* |
3116 | * There are a few boundary cases this might miss but it should | |
3117 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3118 | * a real problem. |
a030d738 VK |
3119 | * |
3120 | * It will not get called when we go idle, because the idle | |
3121 | * thread is a different class (!fair), nor will the utilization | |
3122 | * number include things like RT tasks. | |
3123 | * | |
3124 | * As is, the util number is not freq-invariant (we'd have to | |
3125 | * implement arch_scale_freq_capacity() for that). | |
3126 | * | |
3127 | * See cpu_util(). | |
3128 | */ | |
ea14b57e | 3129 | cpufreq_update_util(rq, flags); |
a030d738 VK |
3130 | } |
3131 | } | |
3132 | ||
141965c7 | 3133 | #ifdef CONFIG_SMP |
c566e8e9 | 3134 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7c3edd2c PZ |
3135 | /** |
3136 | * update_tg_load_avg - update the tg's load avg | |
3137 | * @cfs_rq: the cfs_rq whose avg changed | |
3138 | * @force: update regardless of how small the difference | |
3139 | * | |
3140 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3141 | * However, because tg->load_avg is a global value there are performance | |
3142 | * considerations. | |
3143 | * | |
3144 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3145 | * differential update where we store the last value we propagated. This in | |
3146 | * turn allows skipping updates if the differential is 'small'. | |
3147 | * | |
815abf5a | 3148 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3149 | */ |
9d89c257 | 3150 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 3151 | { |
9d89c257 | 3152 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3153 | |
aa0b7ae0 WL |
3154 | /* |
3155 | * No need to update load_avg for root_task_group as it is not used. | |
3156 | */ | |
3157 | if (cfs_rq->tg == &root_task_group) | |
3158 | return; | |
3159 | ||
9d89c257 YD |
3160 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
3161 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
3162 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3163 | } |
8165e145 | 3164 | } |
f5f9739d | 3165 | |
ad936d86 | 3166 | /* |
97fb7a0a | 3167 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3168 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3169 | * including the state of rq->lock, should be made. | |
3170 | */ | |
3171 | void set_task_rq_fair(struct sched_entity *se, | |
3172 | struct cfs_rq *prev, struct cfs_rq *next) | |
3173 | { | |
0ccb977f PZ |
3174 | u64 p_last_update_time; |
3175 | u64 n_last_update_time; | |
3176 | ||
ad936d86 BP |
3177 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3178 | return; | |
3179 | ||
3180 | /* | |
3181 | * We are supposed to update the task to "current" time, then its up to | |
3182 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3183 | * getting what current time is, so simply throw away the out-of-date | |
3184 | * time. This will result in the wakee task is less decayed, but giving | |
3185 | * the wakee more load sounds not bad. | |
3186 | */ | |
0ccb977f PZ |
3187 | if (!(se->avg.last_update_time && prev)) |
3188 | return; | |
ad936d86 BP |
3189 | |
3190 | #ifndef CONFIG_64BIT | |
0ccb977f | 3191 | { |
ad936d86 BP |
3192 | u64 p_last_update_time_copy; |
3193 | u64 n_last_update_time_copy; | |
3194 | ||
3195 | do { | |
3196 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3197 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3198 | ||
3199 | smp_rmb(); | |
3200 | ||
3201 | p_last_update_time = prev->avg.last_update_time; | |
3202 | n_last_update_time = next->avg.last_update_time; | |
3203 | ||
3204 | } while (p_last_update_time != p_last_update_time_copy || | |
3205 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3206 | } |
ad936d86 | 3207 | #else |
0ccb977f PZ |
3208 | p_last_update_time = prev->avg.last_update_time; |
3209 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3210 | #endif |
23127296 | 3211 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 3212 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 3213 | } |
09a43ace | 3214 | |
0e2d2aaa PZ |
3215 | |
3216 | /* | |
3217 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3218 | * propagate its contribution. The key to this propagation is the invariant | |
3219 | * that for each group: | |
3220 | * | |
3221 | * ge->avg == grq->avg (1) | |
3222 | * | |
3223 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3224 | * represent the very same entity, just at different points in the hierarchy. | |
3225 | * | |
a4c3c049 VG |
3226 | * Per the above update_tg_cfs_util() is trivial and simply copies the running |
3227 | * sum over (but still wrong, because the group entity and group rq do not have | |
3228 | * their PELT windows aligned). | |
0e2d2aaa PZ |
3229 | * |
3230 | * However, update_tg_cfs_runnable() is more complex. So we have: | |
3231 | * | |
3232 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3233 | * | |
3234 | * And since, like util, the runnable part should be directly transferable, | |
3235 | * the following would _appear_ to be the straight forward approach: | |
3236 | * | |
a4c3c049 | 3237 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3238 | * |
3239 | * And per (1) we have: | |
3240 | * | |
a4c3c049 | 3241 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3242 | * |
3243 | * Which gives: | |
3244 | * | |
3245 | * ge->load.weight * grq->avg.load_avg | |
3246 | * ge->avg.load_avg = ----------------------------------- (4) | |
3247 | * grq->load.weight | |
3248 | * | |
3249 | * Except that is wrong! | |
3250 | * | |
3251 | * Because while for entities historical weight is not important and we | |
3252 | * really only care about our future and therefore can consider a pure | |
3253 | * runnable sum, runqueues can NOT do this. | |
3254 | * | |
3255 | * We specifically want runqueues to have a load_avg that includes | |
3256 | * historical weights. Those represent the blocked load, the load we expect | |
3257 | * to (shortly) return to us. This only works by keeping the weights as | |
3258 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3259 | * | |
a4c3c049 VG |
3260 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3261 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3262 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3263 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3264 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3265 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3266 | * |
a4c3c049 | 3267 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3268 | * |
a4c3c049 | 3269 | * Given the constraint: |
0e2d2aaa | 3270 | * |
a4c3c049 | 3271 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3272 | * |
a4c3c049 VG |
3273 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3274 | * overlap. | |
0e2d2aaa | 3275 | * |
a4c3c049 | 3276 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3277 | * |
a4c3c049 | 3278 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3279 | * |
a4c3c049 | 3280 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3281 | * |
0e2d2aaa PZ |
3282 | */ |
3283 | ||
09a43ace | 3284 | static inline void |
0e2d2aaa | 3285 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3286 | { |
09a43ace VG |
3287 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; |
3288 | ||
3289 | /* Nothing to update */ | |
3290 | if (!delta) | |
3291 | return; | |
3292 | ||
a4c3c049 VG |
3293 | /* |
3294 | * The relation between sum and avg is: | |
3295 | * | |
3296 | * LOAD_AVG_MAX - 1024 + sa->period_contrib | |
3297 | * | |
3298 | * however, the PELT windows are not aligned between grq and gse. | |
3299 | */ | |
3300 | ||
09a43ace VG |
3301 | /* Set new sched_entity's utilization */ |
3302 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
3303 | se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX; | |
3304 | ||
3305 | /* Update parent cfs_rq utilization */ | |
3306 | add_positive(&cfs_rq->avg.util_avg, delta); | |
3307 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX; | |
3308 | } | |
3309 | ||
09a43ace | 3310 | static inline void |
0e2d2aaa | 3311 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3312 | { |
a4c3c049 VG |
3313 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
3314 | unsigned long runnable_load_avg, load_avg; | |
3315 | u64 runnable_load_sum, load_sum = 0; | |
3316 | s64 delta_sum; | |
09a43ace | 3317 | |
0e2d2aaa PZ |
3318 | if (!runnable_sum) |
3319 | return; | |
09a43ace | 3320 | |
0e2d2aaa | 3321 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3322 | |
a4c3c049 VG |
3323 | if (runnable_sum >= 0) { |
3324 | /* | |
3325 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3326 | * the CPU is saturated running == runnable. | |
3327 | */ | |
3328 | runnable_sum += se->avg.load_sum; | |
3329 | runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX); | |
3330 | } else { | |
3331 | /* | |
3332 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3333 | * assuming all tasks are equally runnable. | |
3334 | */ | |
3335 | if (scale_load_down(gcfs_rq->load.weight)) { | |
3336 | load_sum = div_s64(gcfs_rq->avg.load_sum, | |
3337 | scale_load_down(gcfs_rq->load.weight)); | |
3338 | } | |
3339 | ||
3340 | /* But make sure to not inflate se's runnable */ | |
3341 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3342 | } | |
3343 | ||
3344 | /* | |
3345 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
3346 | * Rescale running sum to be in the same range as runnable sum |
3347 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
3348 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 3349 | */ |
23127296 | 3350 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
3351 | runnable_sum = max(runnable_sum, running_sum); |
3352 | ||
0e2d2aaa PZ |
3353 | load_sum = (s64)se_weight(se) * runnable_sum; |
3354 | load_avg = div_s64(load_sum, LOAD_AVG_MAX); | |
09a43ace | 3355 | |
a4c3c049 VG |
3356 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
3357 | delta_avg = load_avg - se->avg.load_avg; | |
09a43ace | 3358 | |
a4c3c049 VG |
3359 | se->avg.load_sum = runnable_sum; |
3360 | se->avg.load_avg = load_avg; | |
3361 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
3362 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
09a43ace | 3363 | |
1ea6c46a PZ |
3364 | runnable_load_sum = (s64)se_runnable(se) * runnable_sum; |
3365 | runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX); | |
a4c3c049 VG |
3366 | delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum; |
3367 | delta_avg = runnable_load_avg - se->avg.runnable_load_avg; | |
1ea6c46a | 3368 | |
a4c3c049 VG |
3369 | se->avg.runnable_load_sum = runnable_sum; |
3370 | se->avg.runnable_load_avg = runnable_load_avg; | |
1ea6c46a | 3371 | |
09a43ace | 3372 | if (se->on_rq) { |
a4c3c049 VG |
3373 | add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg); |
3374 | add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum); | |
09a43ace VG |
3375 | } |
3376 | } | |
3377 | ||
0e2d2aaa | 3378 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3379 | { |
0e2d2aaa PZ |
3380 | cfs_rq->propagate = 1; |
3381 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3382 | } |
3383 | ||
3384 | /* Update task and its cfs_rq load average */ | |
3385 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3386 | { | |
0e2d2aaa | 3387 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3388 | |
3389 | if (entity_is_task(se)) | |
3390 | return 0; | |
3391 | ||
0e2d2aaa PZ |
3392 | gcfs_rq = group_cfs_rq(se); |
3393 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3394 | return 0; |
3395 | ||
0e2d2aaa PZ |
3396 | gcfs_rq->propagate = 0; |
3397 | ||
09a43ace VG |
3398 | cfs_rq = cfs_rq_of(se); |
3399 | ||
0e2d2aaa | 3400 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3401 | |
0e2d2aaa PZ |
3402 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
3403 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); | |
09a43ace | 3404 | |
ba19f51f | 3405 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 3406 | trace_pelt_se_tp(se); |
ba19f51f | 3407 | |
09a43ace VG |
3408 | return 1; |
3409 | } | |
3410 | ||
bc427898 VG |
3411 | /* |
3412 | * Check if we need to update the load and the utilization of a blocked | |
3413 | * group_entity: | |
3414 | */ | |
3415 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3416 | { | |
3417 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3418 | ||
3419 | /* | |
3420 | * If sched_entity still have not zero load or utilization, we have to | |
3421 | * decay it: | |
3422 | */ | |
3423 | if (se->avg.load_avg || se->avg.util_avg) | |
3424 | return false; | |
3425 | ||
3426 | /* | |
3427 | * If there is a pending propagation, we have to update the load and | |
3428 | * the utilization of the sched_entity: | |
3429 | */ | |
0e2d2aaa | 3430 | if (gcfs_rq->propagate) |
bc427898 VG |
3431 | return false; |
3432 | ||
3433 | /* | |
3434 | * Otherwise, the load and the utilization of the sched_entity is | |
3435 | * already zero and there is no pending propagation, so it will be a | |
3436 | * waste of time to try to decay it: | |
3437 | */ | |
3438 | return true; | |
3439 | } | |
3440 | ||
6e83125c | 3441 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3442 | |
9d89c257 | 3443 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
09a43ace VG |
3444 | |
3445 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3446 | { | |
3447 | return 0; | |
3448 | } | |
3449 | ||
0e2d2aaa | 3450 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3451 | |
6e83125c | 3452 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3453 | |
3d30544f PZ |
3454 | /** |
3455 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 3456 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 3457 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
3458 | * |
3459 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3460 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3461 | * post_init_entity_util_avg(). | |
3462 | * | |
3463 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3464 | * | |
7c3edd2c PZ |
3465 | * Returns true if the load decayed or we removed load. |
3466 | * | |
3467 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3468 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3469 | */ |
a2c6c91f | 3470 | static inline int |
3a123bbb | 3471 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3472 | { |
0e2d2aaa | 3473 | unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0; |
9d89c257 | 3474 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3475 | int decayed = 0; |
2dac754e | 3476 | |
2a2f5d4e PZ |
3477 | if (cfs_rq->removed.nr) { |
3478 | unsigned long r; | |
9a2dd585 | 3479 | u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib; |
2a2f5d4e PZ |
3480 | |
3481 | raw_spin_lock(&cfs_rq->removed.lock); | |
3482 | swap(cfs_rq->removed.util_avg, removed_util); | |
3483 | swap(cfs_rq->removed.load_avg, removed_load); | |
0e2d2aaa | 3484 | swap(cfs_rq->removed.runnable_sum, removed_runnable_sum); |
2a2f5d4e PZ |
3485 | cfs_rq->removed.nr = 0; |
3486 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3487 | ||
2a2f5d4e | 3488 | r = removed_load; |
89741892 | 3489 | sub_positive(&sa->load_avg, r); |
9a2dd585 | 3490 | sub_positive(&sa->load_sum, r * divider); |
2dac754e | 3491 | |
2a2f5d4e | 3492 | r = removed_util; |
89741892 | 3493 | sub_positive(&sa->util_avg, r); |
9a2dd585 | 3494 | sub_positive(&sa->util_sum, r * divider); |
2a2f5d4e | 3495 | |
0e2d2aaa | 3496 | add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum); |
2a2f5d4e PZ |
3497 | |
3498 | decayed = 1; | |
9d89c257 | 3499 | } |
36ee28e4 | 3500 | |
23127296 | 3501 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
36ee28e4 | 3502 | |
9d89c257 YD |
3503 | #ifndef CONFIG_64BIT |
3504 | smp_wmb(); | |
3505 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3506 | #endif | |
36ee28e4 | 3507 | |
2a2f5d4e | 3508 | if (decayed) |
ea14b57e | 3509 | cfs_rq_util_change(cfs_rq, 0); |
21e96f88 | 3510 | |
2a2f5d4e | 3511 | return decayed; |
21e96f88 SM |
3512 | } |
3513 | ||
3d30544f PZ |
3514 | /** |
3515 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3516 | * @cfs_rq: cfs_rq to attach to | |
3517 | * @se: sched_entity to attach | |
882a78a9 | 3518 | * @flags: migration hints |
3d30544f PZ |
3519 | * |
3520 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3521 | * cfs_rq->avg.last_update_time being current. | |
3522 | */ | |
ea14b57e | 3523 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
a05e8c51 | 3524 | { |
f207934f PZ |
3525 | u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; |
3526 | ||
3527 | /* | |
3528 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3529 | * window because without that, really weird and wonderful things can | |
3530 | * happen. | |
3531 | * | |
3532 | * XXX illustrate | |
3533 | */ | |
a05e8c51 | 3534 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3535 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3536 | ||
3537 | /* | |
3538 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3539 | * period_contrib. This isn't strictly correct, but since we're | |
3540 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3541 | * _sum a little. | |
3542 | */ | |
3543 | se->avg.util_sum = se->avg.util_avg * divider; | |
3544 | ||
3545 | se->avg.load_sum = divider; | |
3546 | if (se_weight(se)) { | |
3547 | se->avg.load_sum = | |
3548 | div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); | |
3549 | } | |
3550 | ||
3551 | se->avg.runnable_load_sum = se->avg.load_sum; | |
3552 | ||
8d5b9025 | 3553 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3554 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3555 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
0e2d2aaa PZ |
3556 | |
3557 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 3558 | |
ea14b57e | 3559 | cfs_rq_util_change(cfs_rq, flags); |
ba19f51f QY |
3560 | |
3561 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3562 | } |
3563 | ||
3d30544f PZ |
3564 | /** |
3565 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3566 | * @cfs_rq: cfs_rq to detach from | |
3567 | * @se: sched_entity to detach | |
3568 | * | |
3569 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3570 | * cfs_rq->avg.last_update_time being current. | |
3571 | */ | |
a05e8c51 BP |
3572 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3573 | { | |
8d5b9025 | 3574 | dequeue_load_avg(cfs_rq, se); |
89741892 PZ |
3575 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
3576 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
0e2d2aaa PZ |
3577 | |
3578 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 3579 | |
ea14b57e | 3580 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3581 | |
3582 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3583 | } |
3584 | ||
b382a531 PZ |
3585 | /* |
3586 | * Optional action to be done while updating the load average | |
3587 | */ | |
3588 | #define UPDATE_TG 0x1 | |
3589 | #define SKIP_AGE_LOAD 0x2 | |
3590 | #define DO_ATTACH 0x4 | |
3591 | ||
3592 | /* Update task and its cfs_rq load average */ | |
3593 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
3594 | { | |
23127296 | 3595 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
3596 | int decayed; |
3597 | ||
3598 | /* | |
3599 | * Track task load average for carrying it to new CPU after migrated, and | |
3600 | * track group sched_entity load average for task_h_load calc in migration | |
3601 | */ | |
3602 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 3603 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
3604 | |
3605 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
3606 | decayed |= propagate_entity_load_avg(se); | |
3607 | ||
3608 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
3609 | ||
ea14b57e PZ |
3610 | /* |
3611 | * DO_ATTACH means we're here from enqueue_entity(). | |
3612 | * !last_update_time means we've passed through | |
3613 | * migrate_task_rq_fair() indicating we migrated. | |
3614 | * | |
3615 | * IOW we're enqueueing a task on a new CPU. | |
3616 | */ | |
3617 | attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION); | |
b382a531 PZ |
3618 | update_tg_load_avg(cfs_rq, 0); |
3619 | ||
3620 | } else if (decayed && (flags & UPDATE_TG)) | |
3621 | update_tg_load_avg(cfs_rq, 0); | |
3622 | } | |
3623 | ||
9d89c257 | 3624 | #ifndef CONFIG_64BIT |
0905f04e YD |
3625 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3626 | { | |
9d89c257 | 3627 | u64 last_update_time_copy; |
0905f04e | 3628 | u64 last_update_time; |
9ee474f5 | 3629 | |
9d89c257 YD |
3630 | do { |
3631 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3632 | smp_rmb(); | |
3633 | last_update_time = cfs_rq->avg.last_update_time; | |
3634 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3635 | |
3636 | return last_update_time; | |
3637 | } | |
9d89c257 | 3638 | #else |
0905f04e YD |
3639 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3640 | { | |
3641 | return cfs_rq->avg.last_update_time; | |
3642 | } | |
9d89c257 YD |
3643 | #endif |
3644 | ||
104cb16d MR |
3645 | /* |
3646 | * Synchronize entity load avg of dequeued entity without locking | |
3647 | * the previous rq. | |
3648 | */ | |
71b47eaf | 3649 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
3650 | { |
3651 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3652 | u64 last_update_time; | |
3653 | ||
3654 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 3655 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
3656 | } |
3657 | ||
0905f04e YD |
3658 | /* |
3659 | * Task first catches up with cfs_rq, and then subtract | |
3660 | * itself from the cfs_rq (task must be off the queue now). | |
3661 | */ | |
71b47eaf | 3662 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
3663 | { |
3664 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 3665 | unsigned long flags; |
0905f04e YD |
3666 | |
3667 | /* | |
7dc603c9 PZ |
3668 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3669 | * post_init_entity_util_avg() which will have added things to the | |
3670 | * cfs_rq, so we can remove unconditionally. | |
0905f04e | 3671 | */ |
0905f04e | 3672 | |
104cb16d | 3673 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
3674 | |
3675 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
3676 | ++cfs_rq->removed.nr; | |
3677 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
3678 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
0e2d2aaa | 3679 | cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */ |
2a2f5d4e | 3680 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 3681 | } |
642dbc39 | 3682 | |
7ea241af YD |
3683 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) |
3684 | { | |
1ea6c46a | 3685 | return cfs_rq->avg.runnable_load_avg; |
7ea241af YD |
3686 | } |
3687 | ||
3688 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) | |
3689 | { | |
3690 | return cfs_rq->avg.load_avg; | |
3691 | } | |
3692 | ||
46f69fa3 | 3693 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf); |
6e83125c | 3694 | |
7f65ea42 PB |
3695 | static inline unsigned long task_util(struct task_struct *p) |
3696 | { | |
3697 | return READ_ONCE(p->se.avg.util_avg); | |
3698 | } | |
3699 | ||
3700 | static inline unsigned long _task_util_est(struct task_struct *p) | |
3701 | { | |
3702 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
3703 | ||
92a801e5 | 3704 | return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED); |
7f65ea42 PB |
3705 | } |
3706 | ||
3707 | static inline unsigned long task_util_est(struct task_struct *p) | |
3708 | { | |
3709 | return max(task_util(p), _task_util_est(p)); | |
3710 | } | |
3711 | ||
3712 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, | |
3713 | struct task_struct *p) | |
3714 | { | |
3715 | unsigned int enqueued; | |
3716 | ||
3717 | if (!sched_feat(UTIL_EST)) | |
3718 | return; | |
3719 | ||
3720 | /* Update root cfs_rq's estimated utilization */ | |
3721 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3722 | enqueued += _task_util_est(p); |
7f65ea42 PB |
3723 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
3724 | } | |
3725 | ||
3726 | /* | |
3727 | * Check if a (signed) value is within a specified (unsigned) margin, | |
3728 | * based on the observation that: | |
3729 | * | |
3730 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
3731 | * | |
3732 | * NOTE: this only works when value + maring < INT_MAX. | |
3733 | */ | |
3734 | static inline bool within_margin(int value, int margin) | |
3735 | { | |
3736 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
3737 | } | |
3738 | ||
3739 | static void | |
3740 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep) | |
3741 | { | |
3742 | long last_ewma_diff; | |
3743 | struct util_est ue; | |
10a35e68 | 3744 | int cpu; |
7f65ea42 PB |
3745 | |
3746 | if (!sched_feat(UTIL_EST)) | |
3747 | return; | |
3748 | ||
3482d98b VG |
3749 | /* Update root cfs_rq's estimated utilization */ |
3750 | ue.enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3751 | ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p)); |
7f65ea42 PB |
3752 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued); |
3753 | ||
3754 | /* | |
3755 | * Skip update of task's estimated utilization when the task has not | |
3756 | * yet completed an activation, e.g. being migrated. | |
3757 | */ | |
3758 | if (!task_sleep) | |
3759 | return; | |
3760 | ||
d519329f PB |
3761 | /* |
3762 | * If the PELT values haven't changed since enqueue time, | |
3763 | * skip the util_est update. | |
3764 | */ | |
3765 | ue = p->se.avg.util_est; | |
3766 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
3767 | return; | |
3768 | ||
7f65ea42 PB |
3769 | /* |
3770 | * Skip update of task's estimated utilization when its EWMA is | |
3771 | * already ~1% close to its last activation value. | |
3772 | */ | |
d519329f | 3773 | ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED); |
7f65ea42 PB |
3774 | last_ewma_diff = ue.enqueued - ue.ewma; |
3775 | if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100))) | |
3776 | return; | |
3777 | ||
10a35e68 VG |
3778 | /* |
3779 | * To avoid overestimation of actual task utilization, skip updates if | |
3780 | * we cannot grant there is idle time in this CPU. | |
3781 | */ | |
3782 | cpu = cpu_of(rq_of(cfs_rq)); | |
3783 | if (task_util(p) > capacity_orig_of(cpu)) | |
3784 | return; | |
3785 | ||
7f65ea42 PB |
3786 | /* |
3787 | * Update Task's estimated utilization | |
3788 | * | |
3789 | * When *p completes an activation we can consolidate another sample | |
3790 | * of the task size. This is done by storing the current PELT value | |
3791 | * as ue.enqueued and by using this value to update the Exponential | |
3792 | * Weighted Moving Average (EWMA): | |
3793 | * | |
3794 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
3795 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
3796 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
3797 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
3798 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
3799 | * | |
3800 | * Where 'w' is the weight of new samples, which is configured to be | |
3801 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
3802 | */ | |
3803 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
3804 | ue.ewma += last_ewma_diff; | |
3805 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
3806 | WRITE_ONCE(p->se.avg.util_est, ue); | |
3807 | } | |
3808 | ||
3b1baa64 MR |
3809 | static inline int task_fits_capacity(struct task_struct *p, long capacity) |
3810 | { | |
60e17f5c | 3811 | return fits_capacity(task_util_est(p), capacity); |
3b1baa64 MR |
3812 | } |
3813 | ||
3814 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
3815 | { | |
3816 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) | |
3817 | return; | |
3818 | ||
3819 | if (!p) { | |
3820 | rq->misfit_task_load = 0; | |
3821 | return; | |
3822 | } | |
3823 | ||
3824 | if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { | |
3825 | rq->misfit_task_load = 0; | |
3826 | return; | |
3827 | } | |
3828 | ||
3829 | rq->misfit_task_load = task_h_load(p); | |
3830 | } | |
3831 | ||
38033c37 PZ |
3832 | #else /* CONFIG_SMP */ |
3833 | ||
d31b1a66 VG |
3834 | #define UPDATE_TG 0x0 |
3835 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 3836 | #define DO_ATTACH 0x0 |
d31b1a66 | 3837 | |
88c0616e | 3838 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 3839 | { |
ea14b57e | 3840 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
3841 | } |
3842 | ||
9d89c257 | 3843 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 3844 | |
a05e8c51 | 3845 | static inline void |
ea14b57e | 3846 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {} |
a05e8c51 BP |
3847 | static inline void |
3848 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3849 | ||
46f69fa3 | 3850 | static inline int idle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
3851 | { |
3852 | return 0; | |
3853 | } | |
3854 | ||
7f65ea42 PB |
3855 | static inline void |
3856 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
3857 | ||
3858 | static inline void | |
3859 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, | |
3860 | bool task_sleep) {} | |
3b1baa64 | 3861 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 3862 | |
38033c37 | 3863 | #endif /* CONFIG_SMP */ |
9d85f21c | 3864 | |
ddc97297 PZ |
3865 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3866 | { | |
3867 | #ifdef CONFIG_SCHED_DEBUG | |
3868 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
3869 | ||
3870 | if (d < 0) | |
3871 | d = -d; | |
3872 | ||
3873 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 3874 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
3875 | #endif |
3876 | } | |
3877 | ||
aeb73b04 PZ |
3878 | static void |
3879 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
3880 | { | |
1af5f730 | 3881 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 3882 | |
2cb8600e PZ |
3883 | /* |
3884 | * The 'current' period is already promised to the current tasks, | |
3885 | * however the extra weight of the new task will slow them down a | |
3886 | * little, place the new task so that it fits in the slot that | |
3887 | * stays open at the end. | |
3888 | */ | |
94dfb5e7 | 3889 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 3890 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 3891 | |
a2e7a7eb | 3892 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 3893 | if (!initial) { |
a2e7a7eb | 3894 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 3895 | |
a2e7a7eb MG |
3896 | /* |
3897 | * Halve their sleep time's effect, to allow | |
3898 | * for a gentler effect of sleepers: | |
3899 | */ | |
3900 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
3901 | thresh >>= 1; | |
51e0304c | 3902 | |
a2e7a7eb | 3903 | vruntime -= thresh; |
aeb73b04 PZ |
3904 | } |
3905 | ||
b5d9d734 | 3906 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3907 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3908 | } |
3909 | ||
d3d9dc33 PT |
3910 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3911 | ||
cb251765 MG |
3912 | static inline void check_schedstat_required(void) |
3913 | { | |
3914 | #ifdef CONFIG_SCHEDSTATS | |
3915 | if (schedstat_enabled()) | |
3916 | return; | |
3917 | ||
3918 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
3919 | if (trace_sched_stat_wait_enabled() || | |
3920 | trace_sched_stat_sleep_enabled() || | |
3921 | trace_sched_stat_iowait_enabled() || | |
3922 | trace_sched_stat_blocked_enabled() || | |
3923 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 3924 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 | 3925 | "stat_blocked and stat_runtime require the " |
f67abed5 | 3926 | "kernel parameter schedstats=enable or " |
cb251765 MG |
3927 | "kernel.sched_schedstats=1\n"); |
3928 | } | |
3929 | #endif | |
3930 | } | |
3931 | ||
b5179ac7 PZ |
3932 | |
3933 | /* | |
3934 | * MIGRATION | |
3935 | * | |
3936 | * dequeue | |
3937 | * update_curr() | |
3938 | * update_min_vruntime() | |
3939 | * vruntime -= min_vruntime | |
3940 | * | |
3941 | * enqueue | |
3942 | * update_curr() | |
3943 | * update_min_vruntime() | |
3944 | * vruntime += min_vruntime | |
3945 | * | |
3946 | * this way the vruntime transition between RQs is done when both | |
3947 | * min_vruntime are up-to-date. | |
3948 | * | |
3949 | * WAKEUP (remote) | |
3950 | * | |
59efa0ba | 3951 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
3952 | * vruntime -= min_vruntime |
3953 | * | |
3954 | * enqueue | |
3955 | * update_curr() | |
3956 | * update_min_vruntime() | |
3957 | * vruntime += min_vruntime | |
3958 | * | |
3959 | * this way we don't have the most up-to-date min_vruntime on the originating | |
3960 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
3961 | */ | |
3962 | ||
bf0f6f24 | 3963 | static void |
88ec22d3 | 3964 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3965 | { |
2f950354 PZ |
3966 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
3967 | bool curr = cfs_rq->curr == se; | |
3968 | ||
88ec22d3 | 3969 | /* |
2f950354 PZ |
3970 | * If we're the current task, we must renormalise before calling |
3971 | * update_curr(). | |
88ec22d3 | 3972 | */ |
2f950354 | 3973 | if (renorm && curr) |
88ec22d3 PZ |
3974 | se->vruntime += cfs_rq->min_vruntime; |
3975 | ||
2f950354 PZ |
3976 | update_curr(cfs_rq); |
3977 | ||
bf0f6f24 | 3978 | /* |
2f950354 PZ |
3979 | * Otherwise, renormalise after, such that we're placed at the current |
3980 | * moment in time, instead of some random moment in the past. Being | |
3981 | * placed in the past could significantly boost this task to the | |
3982 | * fairness detriment of existing tasks. | |
bf0f6f24 | 3983 | */ |
2f950354 PZ |
3984 | if (renorm && !curr) |
3985 | se->vruntime += cfs_rq->min_vruntime; | |
3986 | ||
89ee048f VG |
3987 | /* |
3988 | * When enqueuing a sched_entity, we must: | |
3989 | * - Update loads to have both entity and cfs_rq synced with now. | |
3990 | * - Add its load to cfs_rq->runnable_avg | |
3991 | * - For group_entity, update its weight to reflect the new share of | |
3992 | * its group cfs_rq | |
3993 | * - Add its new weight to cfs_rq->load.weight | |
3994 | */ | |
b382a531 | 3995 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
1ea6c46a | 3996 | update_cfs_group(se); |
b5b3e35f | 3997 | enqueue_runnable_load_avg(cfs_rq, se); |
17bc14b7 | 3998 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 3999 | |
1a3d027c | 4000 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 4001 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 4002 | |
cb251765 | 4003 | check_schedstat_required(); |
4fa8d299 JP |
4004 | update_stats_enqueue(cfs_rq, se, flags); |
4005 | check_spread(cfs_rq, se); | |
2f950354 | 4006 | if (!curr) |
83b699ed | 4007 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4008 | se->on_rq = 1; |
3d4b47b4 | 4009 | |
d3d9dc33 | 4010 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 4011 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
4012 | check_enqueue_throttle(cfs_rq); |
4013 | } | |
bf0f6f24 IM |
4014 | } |
4015 | ||
2c13c919 | 4016 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4017 | { |
2c13c919 RR |
4018 | for_each_sched_entity(se) { |
4019 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4020 | if (cfs_rq->last != se) |
2c13c919 | 4021 | break; |
f1044799 PZ |
4022 | |
4023 | cfs_rq->last = NULL; | |
2c13c919 RR |
4024 | } |
4025 | } | |
2002c695 | 4026 | |
2c13c919 RR |
4027 | static void __clear_buddies_next(struct sched_entity *se) |
4028 | { | |
4029 | for_each_sched_entity(se) { | |
4030 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4031 | if (cfs_rq->next != se) |
2c13c919 | 4032 | break; |
f1044799 PZ |
4033 | |
4034 | cfs_rq->next = NULL; | |
2c13c919 | 4035 | } |
2002c695 PZ |
4036 | } |
4037 | ||
ac53db59 RR |
4038 | static void __clear_buddies_skip(struct sched_entity *se) |
4039 | { | |
4040 | for_each_sched_entity(se) { | |
4041 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4042 | if (cfs_rq->skip != se) |
ac53db59 | 4043 | break; |
f1044799 PZ |
4044 | |
4045 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4046 | } |
4047 | } | |
4048 | ||
a571bbea PZ |
4049 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4050 | { | |
2c13c919 RR |
4051 | if (cfs_rq->last == se) |
4052 | __clear_buddies_last(se); | |
4053 | ||
4054 | if (cfs_rq->next == se) | |
4055 | __clear_buddies_next(se); | |
ac53db59 RR |
4056 | |
4057 | if (cfs_rq->skip == se) | |
4058 | __clear_buddies_skip(se); | |
a571bbea PZ |
4059 | } |
4060 | ||
6c16a6dc | 4061 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4062 | |
bf0f6f24 | 4063 | static void |
371fd7e7 | 4064 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4065 | { |
a2a2d680 DA |
4066 | /* |
4067 | * Update run-time statistics of the 'current'. | |
4068 | */ | |
4069 | update_curr(cfs_rq); | |
89ee048f VG |
4070 | |
4071 | /* | |
4072 | * When dequeuing a sched_entity, we must: | |
4073 | * - Update loads to have both entity and cfs_rq synced with now. | |
dfcb245e IM |
4074 | * - Subtract its load from the cfs_rq->runnable_avg. |
4075 | * - Subtract its previous weight from cfs_rq->load.weight. | |
89ee048f VG |
4076 | * - For group entity, update its weight to reflect the new share |
4077 | * of its group cfs_rq. | |
4078 | */ | |
88c0616e | 4079 | update_load_avg(cfs_rq, se, UPDATE_TG); |
b5b3e35f | 4080 | dequeue_runnable_load_avg(cfs_rq, se); |
a2a2d680 | 4081 | |
4fa8d299 | 4082 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 4083 | |
2002c695 | 4084 | clear_buddies(cfs_rq, se); |
4793241b | 4085 | |
83b699ed | 4086 | if (se != cfs_rq->curr) |
30cfdcfc | 4087 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4088 | se->on_rq = 0; |
30cfdcfc | 4089 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4090 | |
4091 | /* | |
b60205c7 PZ |
4092 | * Normalize after update_curr(); which will also have moved |
4093 | * min_vruntime if @se is the one holding it back. But before doing | |
4094 | * update_min_vruntime() again, which will discount @se's position and | |
4095 | * can move min_vruntime forward still more. | |
88ec22d3 | 4096 | */ |
371fd7e7 | 4097 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4098 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4099 | |
d8b4986d PT |
4100 | /* return excess runtime on last dequeue */ |
4101 | return_cfs_rq_runtime(cfs_rq); | |
4102 | ||
1ea6c46a | 4103 | update_cfs_group(se); |
b60205c7 PZ |
4104 | |
4105 | /* | |
4106 | * Now advance min_vruntime if @se was the entity holding it back, | |
4107 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4108 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4109 | * further than we started -- ie. we'll be penalized. | |
4110 | */ | |
9845c49c | 4111 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4112 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
4113 | } |
4114 | ||
4115 | /* | |
4116 | * Preempt the current task with a newly woken task if needed: | |
4117 | */ | |
7c92e54f | 4118 | static void |
2e09bf55 | 4119 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4120 | { |
11697830 | 4121 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4122 | struct sched_entity *se; |
4123 | s64 delta; | |
11697830 | 4124 | |
6d0f0ebd | 4125 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4126 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4127 | if (delta_exec > ideal_runtime) { |
8875125e | 4128 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4129 | /* |
4130 | * The current task ran long enough, ensure it doesn't get | |
4131 | * re-elected due to buddy favours. | |
4132 | */ | |
4133 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4134 | return; |
4135 | } | |
4136 | ||
4137 | /* | |
4138 | * Ensure that a task that missed wakeup preemption by a | |
4139 | * narrow margin doesn't have to wait for a full slice. | |
4140 | * This also mitigates buddy induced latencies under load. | |
4141 | */ | |
f685ceac MG |
4142 | if (delta_exec < sysctl_sched_min_granularity) |
4143 | return; | |
4144 | ||
f4cfb33e WX |
4145 | se = __pick_first_entity(cfs_rq); |
4146 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4147 | |
f4cfb33e WX |
4148 | if (delta < 0) |
4149 | return; | |
d7d82944 | 4150 | |
f4cfb33e | 4151 | if (delta > ideal_runtime) |
8875125e | 4152 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4153 | } |
4154 | ||
83b699ed | 4155 | static void |
8494f412 | 4156 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4157 | { |
83b699ed SV |
4158 | /* 'current' is not kept within the tree. */ |
4159 | if (se->on_rq) { | |
4160 | /* | |
4161 | * Any task has to be enqueued before it get to execute on | |
4162 | * a CPU. So account for the time it spent waiting on the | |
4163 | * runqueue. | |
4164 | */ | |
4fa8d299 | 4165 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 4166 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4167 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4168 | } |
4169 | ||
79303e9e | 4170 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4171 | cfs_rq->curr = se; |
4fa8d299 | 4172 | |
eba1ed4b IM |
4173 | /* |
4174 | * Track our maximum slice length, if the CPU's load is at | |
4175 | * least twice that of our own weight (i.e. dont track it | |
4176 | * when there are only lesser-weight tasks around): | |
4177 | */ | |
f2bedc47 DE |
4178 | if (schedstat_enabled() && |
4179 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
4fa8d299 JP |
4180 | schedstat_set(se->statistics.slice_max, |
4181 | max((u64)schedstat_val(se->statistics.slice_max), | |
4182 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 4183 | } |
4fa8d299 | 4184 | |
4a55b450 | 4185 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4186 | } |
4187 | ||
3f3a4904 PZ |
4188 | static int |
4189 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4190 | ||
ac53db59 RR |
4191 | /* |
4192 | * Pick the next process, keeping these things in mind, in this order: | |
4193 | * 1) keep things fair between processes/task groups | |
4194 | * 2) pick the "next" process, since someone really wants that to run | |
4195 | * 3) pick the "last" process, for cache locality | |
4196 | * 4) do not run the "skip" process, if something else is available | |
4197 | */ | |
678d5718 PZ |
4198 | static struct sched_entity * |
4199 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4200 | { |
678d5718 PZ |
4201 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4202 | struct sched_entity *se; | |
4203 | ||
4204 | /* | |
4205 | * If curr is set we have to see if its left of the leftmost entity | |
4206 | * still in the tree, provided there was anything in the tree at all. | |
4207 | */ | |
4208 | if (!left || (curr && entity_before(curr, left))) | |
4209 | left = curr; | |
4210 | ||
4211 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4212 | |
ac53db59 RR |
4213 | /* |
4214 | * Avoid running the skip buddy, if running something else can | |
4215 | * be done without getting too unfair. | |
4216 | */ | |
4217 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
4218 | struct sched_entity *second; |
4219 | ||
4220 | if (se == curr) { | |
4221 | second = __pick_first_entity(cfs_rq); | |
4222 | } else { | |
4223 | second = __pick_next_entity(se); | |
4224 | if (!second || (curr && entity_before(curr, second))) | |
4225 | second = curr; | |
4226 | } | |
4227 | ||
ac53db59 RR |
4228 | if (second && wakeup_preempt_entity(second, left) < 1) |
4229 | se = second; | |
4230 | } | |
aa2ac252 | 4231 | |
f685ceac MG |
4232 | /* |
4233 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4234 | */ | |
4235 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
4236 | se = cfs_rq->last; | |
4237 | ||
ac53db59 RR |
4238 | /* |
4239 | * Someone really wants this to run. If it's not unfair, run it. | |
4240 | */ | |
4241 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
4242 | se = cfs_rq->next; | |
4243 | ||
f685ceac | 4244 | clear_buddies(cfs_rq, se); |
4793241b PZ |
4245 | |
4246 | return se; | |
aa2ac252 PZ |
4247 | } |
4248 | ||
678d5718 | 4249 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4250 | |
ab6cde26 | 4251 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4252 | { |
4253 | /* | |
4254 | * If still on the runqueue then deactivate_task() | |
4255 | * was not called and update_curr() has to be done: | |
4256 | */ | |
4257 | if (prev->on_rq) | |
b7cc0896 | 4258 | update_curr(cfs_rq); |
bf0f6f24 | 4259 | |
d3d9dc33 PT |
4260 | /* throttle cfs_rqs exceeding runtime */ |
4261 | check_cfs_rq_runtime(cfs_rq); | |
4262 | ||
4fa8d299 | 4263 | check_spread(cfs_rq, prev); |
cb251765 | 4264 | |
30cfdcfc | 4265 | if (prev->on_rq) { |
4fa8d299 | 4266 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
4267 | /* Put 'current' back into the tree. */ |
4268 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4269 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4270 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4271 | } |
429d43bc | 4272 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4273 | } |
4274 | ||
8f4d37ec PZ |
4275 | static void |
4276 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4277 | { |
bf0f6f24 | 4278 | /* |
30cfdcfc | 4279 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4280 | */ |
30cfdcfc | 4281 | update_curr(cfs_rq); |
bf0f6f24 | 4282 | |
9d85f21c PT |
4283 | /* |
4284 | * Ensure that runnable average is periodically updated. | |
4285 | */ | |
88c0616e | 4286 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4287 | update_cfs_group(curr); |
9d85f21c | 4288 | |
8f4d37ec PZ |
4289 | #ifdef CONFIG_SCHED_HRTICK |
4290 | /* | |
4291 | * queued ticks are scheduled to match the slice, so don't bother | |
4292 | * validating it and just reschedule. | |
4293 | */ | |
983ed7a6 | 4294 | if (queued) { |
8875125e | 4295 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4296 | return; |
4297 | } | |
8f4d37ec PZ |
4298 | /* |
4299 | * don't let the period tick interfere with the hrtick preemption | |
4300 | */ | |
4301 | if (!sched_feat(DOUBLE_TICK) && | |
4302 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4303 | return; | |
4304 | #endif | |
4305 | ||
2c2efaed | 4306 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4307 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4308 | } |
4309 | ||
ab84d31e PT |
4310 | |
4311 | /************************************************** | |
4312 | * CFS bandwidth control machinery | |
4313 | */ | |
4314 | ||
4315 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 4316 | |
e9666d10 | 4317 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 4318 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4319 | |
4320 | static inline bool cfs_bandwidth_used(void) | |
4321 | { | |
c5905afb | 4322 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4323 | } |
4324 | ||
1ee14e6c | 4325 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4326 | { |
ce48c146 | 4327 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
4328 | } |
4329 | ||
4330 | void cfs_bandwidth_usage_dec(void) | |
4331 | { | |
ce48c146 | 4332 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 4333 | } |
e9666d10 | 4334 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
4335 | static bool cfs_bandwidth_used(void) |
4336 | { | |
4337 | return true; | |
4338 | } | |
4339 | ||
1ee14e6c BS |
4340 | void cfs_bandwidth_usage_inc(void) {} |
4341 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 4342 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 4343 | |
ab84d31e PT |
4344 | /* |
4345 | * default period for cfs group bandwidth. | |
4346 | * default: 0.1s, units: nanoseconds | |
4347 | */ | |
4348 | static inline u64 default_cfs_period(void) | |
4349 | { | |
4350 | return 100000000ULL; | |
4351 | } | |
ec12cb7f PT |
4352 | |
4353 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4354 | { | |
4355 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4356 | } | |
4357 | ||
a9cf55b2 PT |
4358 | /* |
4359 | * Replenish runtime according to assigned quota and update expiration time. | |
4360 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
4361 | * additional synchronization around rq->lock. | |
4362 | * | |
4363 | * requires cfs_b->lock | |
4364 | */ | |
029632fb | 4365 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
4366 | { |
4367 | u64 now; | |
4368 | ||
4369 | if (cfs_b->quota == RUNTIME_INF) | |
4370 | return; | |
4371 | ||
4372 | now = sched_clock_cpu(smp_processor_id()); | |
4373 | cfs_b->runtime = cfs_b->quota; | |
a9cf55b2 PT |
4374 | } |
4375 | ||
029632fb PZ |
4376 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4377 | { | |
4378 | return &tg->cfs_bandwidth; | |
4379 | } | |
4380 | ||
f1b17280 PT |
4381 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
4382 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
4383 | { | |
4384 | if (unlikely(cfs_rq->throttle_count)) | |
1a99ae3f | 4385 | return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time; |
f1b17280 | 4386 | |
78becc27 | 4387 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
4388 | } |
4389 | ||
85dac906 PT |
4390 | /* returns 0 on failure to allocate runtime */ |
4391 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
4392 | { |
4393 | struct task_group *tg = cfs_rq->tg; | |
4394 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
de53fd7a | 4395 | u64 amount = 0, min_amount; |
ec12cb7f PT |
4396 | |
4397 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
4398 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
4399 | ||
4400 | raw_spin_lock(&cfs_b->lock); | |
4401 | if (cfs_b->quota == RUNTIME_INF) | |
4402 | amount = min_amount; | |
58088ad0 | 4403 | else { |
77a4d1a1 | 4404 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4405 | |
4406 | if (cfs_b->runtime > 0) { | |
4407 | amount = min(cfs_b->runtime, min_amount); | |
4408 | cfs_b->runtime -= amount; | |
4409 | cfs_b->idle = 0; | |
4410 | } | |
ec12cb7f PT |
4411 | } |
4412 | raw_spin_unlock(&cfs_b->lock); | |
4413 | ||
4414 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
4415 | |
4416 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4417 | } |
4418 | ||
9dbdb155 | 4419 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4420 | { |
4421 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4422 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4423 | |
4424 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4425 | return; |
4426 | ||
85dac906 PT |
4427 | /* |
4428 | * if we're unable to extend our runtime we resched so that the active | |
4429 | * hierarchy can be throttled | |
4430 | */ | |
4431 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4432 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4433 | } |
4434 | ||
6c16a6dc | 4435 | static __always_inline |
9dbdb155 | 4436 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4437 | { |
56f570e5 | 4438 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4439 | return; |
4440 | ||
4441 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4442 | } | |
4443 | ||
85dac906 PT |
4444 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4445 | { | |
56f570e5 | 4446 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4447 | } |
4448 | ||
64660c86 PT |
4449 | /* check whether cfs_rq, or any parent, is throttled */ |
4450 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4451 | { | |
56f570e5 | 4452 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4453 | } |
4454 | ||
4455 | /* | |
4456 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4457 | * dest_cpu are members of a throttled hierarchy when performing group | |
4458 | * load-balance operations. | |
4459 | */ | |
4460 | static inline int throttled_lb_pair(struct task_group *tg, | |
4461 | int src_cpu, int dest_cpu) | |
4462 | { | |
4463 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4464 | ||
4465 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4466 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4467 | ||
4468 | return throttled_hierarchy(src_cfs_rq) || | |
4469 | throttled_hierarchy(dest_cfs_rq); | |
4470 | } | |
4471 | ||
64660c86 PT |
4472 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
4473 | { | |
4474 | struct rq *rq = data; | |
4475 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4476 | ||
4477 | cfs_rq->throttle_count--; | |
64660c86 | 4478 | if (!cfs_rq->throttle_count) { |
f1b17280 | 4479 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 4480 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4481 | cfs_rq->throttled_clock_task; |
31bc6aea VG |
4482 | |
4483 | /* Add cfs_rq with already running entity in the list */ | |
4484 | if (cfs_rq->nr_running >= 1) | |
4485 | list_add_leaf_cfs_rq(cfs_rq); | |
64660c86 | 4486 | } |
64660c86 PT |
4487 | |
4488 | return 0; | |
4489 | } | |
4490 | ||
4491 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4492 | { | |
4493 | struct rq *rq = data; | |
4494 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4495 | ||
82958366 | 4496 | /* group is entering throttled state, stop time */ |
31bc6aea | 4497 | if (!cfs_rq->throttle_count) { |
78becc27 | 4498 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
31bc6aea VG |
4499 | list_del_leaf_cfs_rq(cfs_rq); |
4500 | } | |
64660c86 PT |
4501 | cfs_rq->throttle_count++; |
4502 | ||
4503 | return 0; | |
4504 | } | |
4505 | ||
d3d9dc33 | 4506 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4507 | { |
4508 | struct rq *rq = rq_of(cfs_rq); | |
4509 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4510 | struct sched_entity *se; | |
43e9f7f2 | 4511 | long task_delta, idle_task_delta, dequeue = 1; |
77a4d1a1 | 4512 | bool empty; |
85dac906 PT |
4513 | |
4514 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4515 | ||
f1b17280 | 4516 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4517 | rcu_read_lock(); |
4518 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4519 | rcu_read_unlock(); | |
85dac906 PT |
4520 | |
4521 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4522 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
4523 | for_each_sched_entity(se) { |
4524 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4525 | /* throttled entity or throttle-on-deactivate */ | |
4526 | if (!se->on_rq) | |
4527 | break; | |
4528 | ||
4529 | if (dequeue) | |
4530 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
4531 | qcfs_rq->h_nr_running -= task_delta; | |
43e9f7f2 | 4532 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 PT |
4533 | |
4534 | if (qcfs_rq->load.weight) | |
4535 | dequeue = 0; | |
4536 | } | |
4537 | ||
4538 | if (!se) | |
72465447 | 4539 | sub_nr_running(rq, task_delta); |
85dac906 PT |
4540 | |
4541 | cfs_rq->throttled = 1; | |
78becc27 | 4542 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 4543 | raw_spin_lock(&cfs_b->lock); |
d49db342 | 4544 | empty = list_empty(&cfs_b->throttled_cfs_rq); |
77a4d1a1 | 4545 | |
c06f04c7 BS |
4546 | /* |
4547 | * Add to the _head_ of the list, so that an already-started | |
baa9be4f PA |
4548 | * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is |
4549 | * not running add to the tail so that later runqueues don't get starved. | |
c06f04c7 | 4550 | */ |
baa9be4f PA |
4551 | if (cfs_b->distribute_running) |
4552 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
4553 | else | |
4554 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
77a4d1a1 PZ |
4555 | |
4556 | /* | |
4557 | * If we're the first throttled task, make sure the bandwidth | |
4558 | * timer is running. | |
4559 | */ | |
4560 | if (empty) | |
4561 | start_cfs_bandwidth(cfs_b); | |
4562 | ||
85dac906 PT |
4563 | raw_spin_unlock(&cfs_b->lock); |
4564 | } | |
4565 | ||
029632fb | 4566 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4567 | { |
4568 | struct rq *rq = rq_of(cfs_rq); | |
4569 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4570 | struct sched_entity *se; | |
4571 | int enqueue = 1; | |
43e9f7f2 | 4572 | long task_delta, idle_task_delta; |
671fd9da | 4573 | |
22b958d8 | 4574 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4575 | |
4576 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4577 | |
4578 | update_rq_clock(rq); | |
4579 | ||
671fd9da | 4580 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4581 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4582 | list_del_rcu(&cfs_rq->throttled_list); |
4583 | raw_spin_unlock(&cfs_b->lock); | |
4584 | ||
64660c86 PT |
4585 | /* update hierarchical throttle state */ |
4586 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4587 | ||
671fd9da PT |
4588 | if (!cfs_rq->load.weight) |
4589 | return; | |
4590 | ||
4591 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 4592 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da PT |
4593 | for_each_sched_entity(se) { |
4594 | if (se->on_rq) | |
4595 | enqueue = 0; | |
4596 | ||
4597 | cfs_rq = cfs_rq_of(se); | |
4598 | if (enqueue) | |
4599 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
4600 | cfs_rq->h_nr_running += task_delta; | |
43e9f7f2 | 4601 | cfs_rq->idle_h_nr_running += idle_task_delta; |
671fd9da PT |
4602 | |
4603 | if (cfs_rq_throttled(cfs_rq)) | |
4604 | break; | |
4605 | } | |
4606 | ||
31bc6aea VG |
4607 | assert_list_leaf_cfs_rq(rq); |
4608 | ||
671fd9da | 4609 | if (!se) |
72465447 | 4610 | add_nr_running(rq, task_delta); |
671fd9da | 4611 | |
97fb7a0a | 4612 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 4613 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 4614 | resched_curr(rq); |
671fd9da PT |
4615 | } |
4616 | ||
de53fd7a | 4617 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining) |
671fd9da PT |
4618 | { |
4619 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
4620 | u64 runtime; |
4621 | u64 starting_runtime = remaining; | |
671fd9da PT |
4622 | |
4623 | rcu_read_lock(); | |
4624 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4625 | throttled_list) { | |
4626 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 4627 | struct rq_flags rf; |
671fd9da | 4628 | |
c0ad4aa4 | 4629 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
4630 | if (!cfs_rq_throttled(cfs_rq)) |
4631 | goto next; | |
4632 | ||
4633 | runtime = -cfs_rq->runtime_remaining + 1; | |
4634 | if (runtime > remaining) | |
4635 | runtime = remaining; | |
4636 | remaining -= runtime; | |
4637 | ||
4638 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
4639 | |
4640 | /* we check whether we're throttled above */ | |
4641 | if (cfs_rq->runtime_remaining > 0) | |
4642 | unthrottle_cfs_rq(cfs_rq); | |
4643 | ||
4644 | next: | |
c0ad4aa4 | 4645 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
4646 | |
4647 | if (!remaining) | |
4648 | break; | |
4649 | } | |
4650 | rcu_read_unlock(); | |
4651 | ||
c06f04c7 | 4652 | return starting_runtime - remaining; |
671fd9da PT |
4653 | } |
4654 | ||
58088ad0 PT |
4655 | /* |
4656 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4657 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4658 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4659 | * used to track this state. | |
4660 | */ | |
c0ad4aa4 | 4661 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 4662 | { |
de53fd7a | 4663 | u64 runtime; |
51f2176d | 4664 | int throttled; |
58088ad0 | 4665 | |
58088ad0 PT |
4666 | /* no need to continue the timer with no bandwidth constraint */ |
4667 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4668 | goto out_deactivate; |
58088ad0 | 4669 | |
671fd9da | 4670 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4671 | cfs_b->nr_periods += overrun; |
671fd9da | 4672 | |
51f2176d BS |
4673 | /* |
4674 | * idle depends on !throttled (for the case of a large deficit), and if | |
4675 | * we're going inactive then everything else can be deferred | |
4676 | */ | |
4677 | if (cfs_b->idle && !throttled) | |
4678 | goto out_deactivate; | |
a9cf55b2 PT |
4679 | |
4680 | __refill_cfs_bandwidth_runtime(cfs_b); | |
4681 | ||
671fd9da PT |
4682 | if (!throttled) { |
4683 | /* mark as potentially idle for the upcoming period */ | |
4684 | cfs_b->idle = 1; | |
51f2176d | 4685 | return 0; |
671fd9da PT |
4686 | } |
4687 | ||
e8da1b18 NR |
4688 | /* account preceding periods in which throttling occurred */ |
4689 | cfs_b->nr_throttled += overrun; | |
4690 | ||
671fd9da | 4691 | /* |
c06f04c7 BS |
4692 | * This check is repeated as we are holding onto the new bandwidth while |
4693 | * we unthrottle. This can potentially race with an unthrottled group | |
4694 | * trying to acquire new bandwidth from the global pool. This can result | |
4695 | * in us over-using our runtime if it is all used during this loop, but | |
4696 | * only by limited amounts in that extreme case. | |
671fd9da | 4697 | */ |
baa9be4f | 4698 | while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) { |
c06f04c7 | 4699 | runtime = cfs_b->runtime; |
baa9be4f | 4700 | cfs_b->distribute_running = 1; |
c0ad4aa4 | 4701 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 4702 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
de53fd7a | 4703 | runtime = distribute_cfs_runtime(cfs_b, runtime); |
c0ad4aa4 | 4704 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da | 4705 | |
baa9be4f | 4706 | cfs_b->distribute_running = 0; |
671fd9da | 4707 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
c06f04c7 | 4708 | |
b5c0ce7b | 4709 | lsub_positive(&cfs_b->runtime, runtime); |
671fd9da | 4710 | } |
58088ad0 | 4711 | |
671fd9da PT |
4712 | /* |
4713 | * While we are ensured activity in the period following an | |
4714 | * unthrottle, this also covers the case in which the new bandwidth is | |
4715 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
4716 | * timer to remain active while there are any throttled entities.) | |
4717 | */ | |
4718 | cfs_b->idle = 0; | |
58088ad0 | 4719 | |
51f2176d BS |
4720 | return 0; |
4721 | ||
4722 | out_deactivate: | |
51f2176d | 4723 | return 1; |
58088ad0 | 4724 | } |
d3d9dc33 | 4725 | |
d8b4986d PT |
4726 | /* a cfs_rq won't donate quota below this amount */ |
4727 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
4728 | /* minimum remaining period time to redistribute slack quota */ | |
4729 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
4730 | /* how long we wait to gather additional slack before distributing */ | |
4731 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
4732 | ||
db06e78c BS |
4733 | /* |
4734 | * Are we near the end of the current quota period? | |
4735 | * | |
4736 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 4737 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
4738 | * migrate_hrtimers, base is never cleared, so we are fine. |
4739 | */ | |
d8b4986d PT |
4740 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
4741 | { | |
4742 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
4743 | u64 remaining; | |
4744 | ||
4745 | /* if the call-back is running a quota refresh is already occurring */ | |
4746 | if (hrtimer_callback_running(refresh_timer)) | |
4747 | return 1; | |
4748 | ||
4749 | /* is a quota refresh about to occur? */ | |
4750 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
4751 | if (remaining < min_expire) | |
4752 | return 1; | |
4753 | ||
4754 | return 0; | |
4755 | } | |
4756 | ||
4757 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
4758 | { | |
4759 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
4760 | ||
4761 | /* if there's a quota refresh soon don't bother with slack */ | |
4762 | if (runtime_refresh_within(cfs_b, min_left)) | |
4763 | return; | |
4764 | ||
66567fcb | 4765 | /* don't push forwards an existing deferred unthrottle */ |
4766 | if (cfs_b->slack_started) | |
4767 | return; | |
4768 | cfs_b->slack_started = true; | |
4769 | ||
4cfafd30 PZ |
4770 | hrtimer_start(&cfs_b->slack_timer, |
4771 | ns_to_ktime(cfs_bandwidth_slack_period), | |
4772 | HRTIMER_MODE_REL); | |
d8b4986d PT |
4773 | } |
4774 | ||
4775 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
4776 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4777 | { | |
4778 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4779 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
4780 | ||
4781 | if (slack_runtime <= 0) | |
4782 | return; | |
4783 | ||
4784 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 4785 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
4786 | cfs_b->runtime += slack_runtime; |
4787 | ||
4788 | /* we are under rq->lock, defer unthrottling using a timer */ | |
4789 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
4790 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
4791 | start_cfs_slack_bandwidth(cfs_b); | |
4792 | } | |
4793 | raw_spin_unlock(&cfs_b->lock); | |
4794 | ||
4795 | /* even if it's not valid for return we don't want to try again */ | |
4796 | cfs_rq->runtime_remaining -= slack_runtime; | |
4797 | } | |
4798 | ||
4799 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4800 | { | |
56f570e5 PT |
4801 | if (!cfs_bandwidth_used()) |
4802 | return; | |
4803 | ||
fccfdc6f | 4804 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
4805 | return; |
4806 | ||
4807 | __return_cfs_rq_runtime(cfs_rq); | |
4808 | } | |
4809 | ||
4810 | /* | |
4811 | * This is done with a timer (instead of inline with bandwidth return) since | |
4812 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
4813 | */ | |
4814 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
4815 | { | |
4816 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 4817 | unsigned long flags; |
d8b4986d PT |
4818 | |
4819 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 4820 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 4821 | cfs_b->slack_started = false; |
baa9be4f | 4822 | if (cfs_b->distribute_running) { |
c0ad4aa4 | 4823 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
baa9be4f PA |
4824 | return; |
4825 | } | |
4826 | ||
db06e78c | 4827 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 4828 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 4829 | return; |
db06e78c | 4830 | } |
d8b4986d | 4831 | |
c06f04c7 | 4832 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 4833 | runtime = cfs_b->runtime; |
c06f04c7 | 4834 | |
baa9be4f PA |
4835 | if (runtime) |
4836 | cfs_b->distribute_running = 1; | |
4837 | ||
c0ad4aa4 | 4838 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
4839 | |
4840 | if (!runtime) | |
4841 | return; | |
4842 | ||
de53fd7a | 4843 | runtime = distribute_cfs_runtime(cfs_b, runtime); |
d8b4986d | 4844 | |
c0ad4aa4 | 4845 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
de53fd7a | 4846 | lsub_positive(&cfs_b->runtime, runtime); |
baa9be4f | 4847 | cfs_b->distribute_running = 0; |
c0ad4aa4 | 4848 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
4849 | } |
4850 | ||
d3d9dc33 PT |
4851 | /* |
4852 | * When a group wakes up we want to make sure that its quota is not already | |
4853 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
4854 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
4855 | */ | |
4856 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
4857 | { | |
56f570e5 PT |
4858 | if (!cfs_bandwidth_used()) |
4859 | return; | |
4860 | ||
d3d9dc33 PT |
4861 | /* an active group must be handled by the update_curr()->put() path */ |
4862 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
4863 | return; | |
4864 | ||
4865 | /* ensure the group is not already throttled */ | |
4866 | if (cfs_rq_throttled(cfs_rq)) | |
4867 | return; | |
4868 | ||
4869 | /* update runtime allocation */ | |
4870 | account_cfs_rq_runtime(cfs_rq, 0); | |
4871 | if (cfs_rq->runtime_remaining <= 0) | |
4872 | throttle_cfs_rq(cfs_rq); | |
4873 | } | |
4874 | ||
55e16d30 PZ |
4875 | static void sync_throttle(struct task_group *tg, int cpu) |
4876 | { | |
4877 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
4878 | ||
4879 | if (!cfs_bandwidth_used()) | |
4880 | return; | |
4881 | ||
4882 | if (!tg->parent) | |
4883 | return; | |
4884 | ||
4885 | cfs_rq = tg->cfs_rq[cpu]; | |
4886 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
4887 | ||
4888 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 4889 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
4890 | } |
4891 | ||
d3d9dc33 | 4892 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 4893 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 4894 | { |
56f570e5 | 4895 | if (!cfs_bandwidth_used()) |
678d5718 | 4896 | return false; |
56f570e5 | 4897 | |
d3d9dc33 | 4898 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 4899 | return false; |
d3d9dc33 PT |
4900 | |
4901 | /* | |
4902 | * it's possible for a throttled entity to be forced into a running | |
4903 | * state (e.g. set_curr_task), in this case we're finished. | |
4904 | */ | |
4905 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 4906 | return true; |
d3d9dc33 PT |
4907 | |
4908 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 4909 | return true; |
d3d9dc33 | 4910 | } |
029632fb | 4911 | |
029632fb PZ |
4912 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
4913 | { | |
4914 | struct cfs_bandwidth *cfs_b = | |
4915 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 4916 | |
029632fb PZ |
4917 | do_sched_cfs_slack_timer(cfs_b); |
4918 | ||
4919 | return HRTIMER_NORESTART; | |
4920 | } | |
4921 | ||
2e8e1922 PA |
4922 | extern const u64 max_cfs_quota_period; |
4923 | ||
029632fb PZ |
4924 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
4925 | { | |
4926 | struct cfs_bandwidth *cfs_b = | |
4927 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 4928 | unsigned long flags; |
029632fb PZ |
4929 | int overrun; |
4930 | int idle = 0; | |
2e8e1922 | 4931 | int count = 0; |
029632fb | 4932 | |
c0ad4aa4 | 4933 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 4934 | for (;;) { |
77a4d1a1 | 4935 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
4936 | if (!overrun) |
4937 | break; | |
4938 | ||
2e8e1922 PA |
4939 | if (++count > 3) { |
4940 | u64 new, old = ktime_to_ns(cfs_b->period); | |
4941 | ||
4942 | new = (old * 147) / 128; /* ~115% */ | |
4943 | new = min(new, max_cfs_quota_period); | |
4944 | ||
4945 | cfs_b->period = ns_to_ktime(new); | |
4946 | ||
4947 | /* since max is 1s, this is limited to 1e9^2, which fits in u64 */ | |
4948 | cfs_b->quota *= new; | |
4949 | cfs_b->quota = div64_u64(cfs_b->quota, old); | |
4950 | ||
4951 | pr_warn_ratelimited( | |
4952 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n", | |
4953 | smp_processor_id(), | |
4954 | div_u64(new, NSEC_PER_USEC), | |
4955 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
4956 | ||
4957 | /* reset count so we don't come right back in here */ | |
4958 | count = 0; | |
4959 | } | |
4960 | ||
c0ad4aa4 | 4961 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
029632fb | 4962 | } |
4cfafd30 PZ |
4963 | if (idle) |
4964 | cfs_b->period_active = 0; | |
c0ad4aa4 | 4965 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
4966 | |
4967 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
4968 | } | |
4969 | ||
4970 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4971 | { | |
4972 | raw_spin_lock_init(&cfs_b->lock); | |
4973 | cfs_b->runtime = 0; | |
4974 | cfs_b->quota = RUNTIME_INF; | |
4975 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
4976 | ||
4977 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 4978 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
4979 | cfs_b->period_timer.function = sched_cfs_period_timer; |
4980 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
4981 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
baa9be4f | 4982 | cfs_b->distribute_running = 0; |
66567fcb | 4983 | cfs_b->slack_started = false; |
029632fb PZ |
4984 | } |
4985 | ||
4986 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4987 | { | |
4988 | cfs_rq->runtime_enabled = 0; | |
4989 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
4990 | } | |
4991 | ||
77a4d1a1 | 4992 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 4993 | { |
f1d1be8a XP |
4994 | u64 overrun; |
4995 | ||
4cfafd30 | 4996 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 4997 | |
f1d1be8a XP |
4998 | if (cfs_b->period_active) |
4999 | return; | |
5000 | ||
5001 | cfs_b->period_active = 1; | |
5002 | overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); | |
f1d1be8a | 5003 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5004 | } |
5005 | ||
5006 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5007 | { | |
7f1a169b TH |
5008 | /* init_cfs_bandwidth() was not called */ |
5009 | if (!cfs_b->throttled_cfs_rq.next) | |
5010 | return; | |
5011 | ||
029632fb PZ |
5012 | hrtimer_cancel(&cfs_b->period_timer); |
5013 | hrtimer_cancel(&cfs_b->slack_timer); | |
5014 | } | |
5015 | ||
502ce005 | 5016 | /* |
97fb7a0a | 5017 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
5018 | * |
5019 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5020 | * bits doesn't do much. | |
5021 | */ | |
5022 | ||
5023 | /* cpu online calback */ | |
0e59bdae KT |
5024 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5025 | { | |
502ce005 | 5026 | struct task_group *tg; |
0e59bdae | 5027 | |
502ce005 PZ |
5028 | lockdep_assert_held(&rq->lock); |
5029 | ||
5030 | rcu_read_lock(); | |
5031 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5032 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5033 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5034 | |
5035 | raw_spin_lock(&cfs_b->lock); | |
5036 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5037 | raw_spin_unlock(&cfs_b->lock); | |
5038 | } | |
502ce005 | 5039 | rcu_read_unlock(); |
0e59bdae KT |
5040 | } |
5041 | ||
502ce005 | 5042 | /* cpu offline callback */ |
38dc3348 | 5043 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5044 | { |
502ce005 PZ |
5045 | struct task_group *tg; |
5046 | ||
5047 | lockdep_assert_held(&rq->lock); | |
5048 | ||
5049 | rcu_read_lock(); | |
5050 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5051 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5052 | |
029632fb PZ |
5053 | if (!cfs_rq->runtime_enabled) |
5054 | continue; | |
5055 | ||
5056 | /* | |
5057 | * clock_task is not advancing so we just need to make sure | |
5058 | * there's some valid quota amount | |
5059 | */ | |
51f2176d | 5060 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 5061 | /* |
97fb7a0a | 5062 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
5063 | * in take_cpu_down(), so we prevent new cfs throttling here. |
5064 | */ | |
5065 | cfs_rq->runtime_enabled = 0; | |
5066 | ||
029632fb PZ |
5067 | if (cfs_rq_throttled(cfs_rq)) |
5068 | unthrottle_cfs_rq(cfs_rq); | |
5069 | } | |
502ce005 | 5070 | rcu_read_unlock(); |
029632fb PZ |
5071 | } |
5072 | ||
5073 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f6783319 VG |
5074 | |
5075 | static inline bool cfs_bandwidth_used(void) | |
5076 | { | |
5077 | return false; | |
5078 | } | |
5079 | ||
f1b17280 PT |
5080 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
5081 | { | |
78becc27 | 5082 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
5083 | } |
5084 | ||
9dbdb155 | 5085 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5086 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5087 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5088 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5089 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5090 | |
5091 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5092 | { | |
5093 | return 0; | |
5094 | } | |
64660c86 PT |
5095 | |
5096 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5097 | { | |
5098 | return 0; | |
5099 | } | |
5100 | ||
5101 | static inline int throttled_lb_pair(struct task_group *tg, | |
5102 | int src_cpu, int dest_cpu) | |
5103 | { | |
5104 | return 0; | |
5105 | } | |
029632fb PZ |
5106 | |
5107 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5108 | ||
5109 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5110 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5111 | #endif |
5112 | ||
029632fb PZ |
5113 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5114 | { | |
5115 | return NULL; | |
5116 | } | |
5117 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5118 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5119 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5120 | |
5121 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5122 | ||
bf0f6f24 IM |
5123 | /************************************************** |
5124 | * CFS operations on tasks: | |
5125 | */ | |
5126 | ||
8f4d37ec PZ |
5127 | #ifdef CONFIG_SCHED_HRTICK |
5128 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5129 | { | |
8f4d37ec PZ |
5130 | struct sched_entity *se = &p->se; |
5131 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5132 | ||
9148a3a1 | 5133 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5134 | |
8bf46a39 | 5135 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5136 | u64 slice = sched_slice(cfs_rq, se); |
5137 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5138 | s64 delta = slice - ran; | |
5139 | ||
5140 | if (delta < 0) { | |
5141 | if (rq->curr == p) | |
8875125e | 5142 | resched_curr(rq); |
8f4d37ec PZ |
5143 | return; |
5144 | } | |
31656519 | 5145 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5146 | } |
5147 | } | |
a4c2f00f PZ |
5148 | |
5149 | /* | |
5150 | * called from enqueue/dequeue and updates the hrtick when the | |
5151 | * current task is from our class and nr_running is low enough | |
5152 | * to matter. | |
5153 | */ | |
5154 | static void hrtick_update(struct rq *rq) | |
5155 | { | |
5156 | struct task_struct *curr = rq->curr; | |
5157 | ||
b39e66ea | 5158 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5159 | return; |
5160 | ||
5161 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5162 | hrtick_start_fair(rq, curr); | |
5163 | } | |
55e12e5e | 5164 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5165 | static inline void |
5166 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5167 | { | |
5168 | } | |
a4c2f00f PZ |
5169 | |
5170 | static inline void hrtick_update(struct rq *rq) | |
5171 | { | |
5172 | } | |
8f4d37ec PZ |
5173 | #endif |
5174 | ||
2802bf3c MR |
5175 | #ifdef CONFIG_SMP |
5176 | static inline unsigned long cpu_util(int cpu); | |
2802bf3c MR |
5177 | |
5178 | static inline bool cpu_overutilized(int cpu) | |
5179 | { | |
60e17f5c | 5180 | return !fits_capacity(cpu_util(cpu), capacity_of(cpu)); |
2802bf3c MR |
5181 | } |
5182 | ||
5183 | static inline void update_overutilized_status(struct rq *rq) | |
5184 | { | |
f9f240f9 | 5185 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { |
2802bf3c | 5186 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); |
f9f240f9 QY |
5187 | trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); |
5188 | } | |
2802bf3c MR |
5189 | } |
5190 | #else | |
5191 | static inline void update_overutilized_status(struct rq *rq) { } | |
5192 | #endif | |
5193 | ||
bf0f6f24 IM |
5194 | /* |
5195 | * The enqueue_task method is called before nr_running is | |
5196 | * increased. Here we update the fair scheduling stats and | |
5197 | * then put the task into the rbtree: | |
5198 | */ | |
ea87bb78 | 5199 | static void |
371fd7e7 | 5200 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5201 | { |
5202 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5203 | struct sched_entity *se = &p->se; |
43e9f7f2 | 5204 | int idle_h_nr_running = task_has_idle_policy(p); |
bf0f6f24 | 5205 | |
2539fc82 PB |
5206 | /* |
5207 | * The code below (indirectly) updates schedutil which looks at | |
5208 | * the cfs_rq utilization to select a frequency. | |
5209 | * Let's add the task's estimated utilization to the cfs_rq's | |
5210 | * estimated utilization, before we update schedutil. | |
5211 | */ | |
5212 | util_est_enqueue(&rq->cfs, p); | |
5213 | ||
8c34ab19 RW |
5214 | /* |
5215 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5216 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5217 | * passed. | |
5218 | */ | |
5219 | if (p->in_iowait) | |
674e7541 | 5220 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5221 | |
bf0f6f24 | 5222 | for_each_sched_entity(se) { |
62fb1851 | 5223 | if (se->on_rq) |
bf0f6f24 IM |
5224 | break; |
5225 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5226 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
5227 | |
5228 | /* | |
5229 | * end evaluation on encountering a throttled cfs_rq | |
5230 | * | |
5231 | * note: in the case of encountering a throttled cfs_rq we will | |
5232 | * post the final h_nr_running increment below. | |
e210bffd | 5233 | */ |
85dac906 PT |
5234 | if (cfs_rq_throttled(cfs_rq)) |
5235 | break; | |
953bfcd1 | 5236 | cfs_rq->h_nr_running++; |
43e9f7f2 | 5237 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 5238 | |
88ec22d3 | 5239 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5240 | } |
8f4d37ec | 5241 | |
2069dd75 | 5242 | for_each_sched_entity(se) { |
0f317143 | 5243 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 5244 | cfs_rq->h_nr_running++; |
43e9f7f2 | 5245 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
2069dd75 | 5246 | |
85dac906 PT |
5247 | if (cfs_rq_throttled(cfs_rq)) |
5248 | break; | |
5249 | ||
88c0616e | 5250 | update_load_avg(cfs_rq, se, UPDATE_TG); |
1ea6c46a | 5251 | update_cfs_group(se); |
2069dd75 PZ |
5252 | } |
5253 | ||
2802bf3c | 5254 | if (!se) { |
72465447 | 5255 | add_nr_running(rq, 1); |
2802bf3c MR |
5256 | /* |
5257 | * Since new tasks are assigned an initial util_avg equal to | |
5258 | * half of the spare capacity of their CPU, tiny tasks have the | |
5259 | * ability to cross the overutilized threshold, which will | |
5260 | * result in the load balancer ruining all the task placement | |
5261 | * done by EAS. As a way to mitigate that effect, do not account | |
5262 | * for the first enqueue operation of new tasks during the | |
5263 | * overutilized flag detection. | |
5264 | * | |
5265 | * A better way of solving this problem would be to wait for | |
5266 | * the PELT signals of tasks to converge before taking them | |
5267 | * into account, but that is not straightforward to implement, | |
5268 | * and the following generally works well enough in practice. | |
5269 | */ | |
5270 | if (flags & ENQUEUE_WAKEUP) | |
5271 | update_overutilized_status(rq); | |
5272 | ||
5273 | } | |
cd126afe | 5274 | |
f6783319 VG |
5275 | if (cfs_bandwidth_used()) { |
5276 | /* | |
5277 | * When bandwidth control is enabled; the cfs_rq_throttled() | |
5278 | * breaks in the above iteration can result in incomplete | |
5279 | * leaf list maintenance, resulting in triggering the assertion | |
5280 | * below. | |
5281 | */ | |
5282 | for_each_sched_entity(se) { | |
5283 | cfs_rq = cfs_rq_of(se); | |
5284 | ||
5285 | if (list_add_leaf_cfs_rq(cfs_rq)) | |
5286 | break; | |
5287 | } | |
5288 | } | |
5289 | ||
5d299eab PZ |
5290 | assert_list_leaf_cfs_rq(rq); |
5291 | ||
a4c2f00f | 5292 | hrtick_update(rq); |
bf0f6f24 IM |
5293 | } |
5294 | ||
2f36825b VP |
5295 | static void set_next_buddy(struct sched_entity *se); |
5296 | ||
bf0f6f24 IM |
5297 | /* |
5298 | * The dequeue_task method is called before nr_running is | |
5299 | * decreased. We remove the task from the rbtree and | |
5300 | * update the fair scheduling stats: | |
5301 | */ | |
371fd7e7 | 5302 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5303 | { |
5304 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5305 | struct sched_entity *se = &p->se; |
2f36825b | 5306 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 5307 | int idle_h_nr_running = task_has_idle_policy(p); |
bf0f6f24 IM |
5308 | |
5309 | for_each_sched_entity(se) { | |
5310 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5311 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
5312 | |
5313 | /* | |
5314 | * end evaluation on encountering a throttled cfs_rq | |
5315 | * | |
5316 | * note: in the case of encountering a throttled cfs_rq we will | |
5317 | * post the final h_nr_running decrement below. | |
5318 | */ | |
5319 | if (cfs_rq_throttled(cfs_rq)) | |
5320 | break; | |
953bfcd1 | 5321 | cfs_rq->h_nr_running--; |
43e9f7f2 | 5322 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 5323 | |
bf0f6f24 | 5324 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5325 | if (cfs_rq->load.weight) { |
754bd598 KK |
5326 | /* Avoid re-evaluating load for this entity: */ |
5327 | se = parent_entity(se); | |
2f36825b VP |
5328 | /* |
5329 | * Bias pick_next to pick a task from this cfs_rq, as | |
5330 | * p is sleeping when it is within its sched_slice. | |
5331 | */ | |
754bd598 KK |
5332 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5333 | set_next_buddy(se); | |
bf0f6f24 | 5334 | break; |
2f36825b | 5335 | } |
371fd7e7 | 5336 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5337 | } |
8f4d37ec | 5338 | |
2069dd75 | 5339 | for_each_sched_entity(se) { |
0f317143 | 5340 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 5341 | cfs_rq->h_nr_running--; |
43e9f7f2 | 5342 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 5343 | |
85dac906 PT |
5344 | if (cfs_rq_throttled(cfs_rq)) |
5345 | break; | |
5346 | ||
88c0616e | 5347 | update_load_avg(cfs_rq, se, UPDATE_TG); |
1ea6c46a | 5348 | update_cfs_group(se); |
2069dd75 PZ |
5349 | } |
5350 | ||
cd126afe | 5351 | if (!se) |
72465447 | 5352 | sub_nr_running(rq, 1); |
cd126afe | 5353 | |
7f65ea42 | 5354 | util_est_dequeue(&rq->cfs, p, task_sleep); |
a4c2f00f | 5355 | hrtick_update(rq); |
bf0f6f24 IM |
5356 | } |
5357 | ||
e7693a36 | 5358 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5359 | |
5360 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
5361 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5362 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5363 | ||
9fd81dd5 | 5364 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
5365 | |
5366 | static struct { | |
5367 | cpumask_var_t idle_cpus_mask; | |
5368 | atomic_t nr_cpus; | |
f643ea22 | 5369 | int has_blocked; /* Idle CPUS has blocked load */ |
e022e0d3 | 5370 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 5371 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
5372 | } nohz ____cacheline_aligned; |
5373 | ||
9fd81dd5 | 5374 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5375 | |
3c29e651 VK |
5376 | /* CPU only has SCHED_IDLE tasks enqueued */ |
5377 | static int sched_idle_cpu(int cpu) | |
5378 | { | |
5379 | struct rq *rq = cpu_rq(cpu); | |
5380 | ||
5381 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
5382 | rq->nr_running); | |
5383 | } | |
5384 | ||
a3df0679 | 5385 | static unsigned long cpu_runnable_load(struct rq *rq) |
7ea241af | 5386 | { |
c7132dd6 | 5387 | return cfs_rq_runnable_load_avg(&rq->cfs); |
7ea241af YD |
5388 | } |
5389 | ||
ced549fa | 5390 | static unsigned long capacity_of(int cpu) |
029632fb | 5391 | { |
ced549fa | 5392 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5393 | } |
5394 | ||
5395 | static unsigned long cpu_avg_load_per_task(int cpu) | |
5396 | { | |
5397 | struct rq *rq = cpu_rq(cpu); | |
316c1608 | 5398 | unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); |
a3df0679 | 5399 | unsigned long load_avg = cpu_runnable_load(rq); |
029632fb PZ |
5400 | |
5401 | if (nr_running) | |
b92486cb | 5402 | return load_avg / nr_running; |
029632fb PZ |
5403 | |
5404 | return 0; | |
5405 | } | |
5406 | ||
c58d25f3 PZ |
5407 | static void record_wakee(struct task_struct *p) |
5408 | { | |
5409 | /* | |
5410 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5411 | * jiffy will not have built up many flips. | |
5412 | */ | |
5413 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5414 | current->wakee_flips >>= 1; | |
5415 | current->wakee_flip_decay_ts = jiffies; | |
5416 | } | |
5417 | ||
5418 | if (current->last_wakee != p) { | |
5419 | current->last_wakee = p; | |
5420 | current->wakee_flips++; | |
5421 | } | |
5422 | } | |
5423 | ||
63b0e9ed MG |
5424 | /* |
5425 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5426 | * |
63b0e9ed | 5427 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5428 | * at a frequency roughly N times higher than one of its wakees. |
5429 | * | |
5430 | * In order to determine whether we should let the load spread vs consolidating | |
5431 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5432 | * partner, and a factor of lls_size higher frequency in the other. | |
5433 | * | |
5434 | * With both conditions met, we can be relatively sure that the relationship is | |
5435 | * non-monogamous, with partner count exceeding socket size. | |
5436 | * | |
5437 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5438 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5439 | * socket size. | |
63b0e9ed | 5440 | */ |
62470419 MW |
5441 | static int wake_wide(struct task_struct *p) |
5442 | { | |
63b0e9ed MG |
5443 | unsigned int master = current->wakee_flips; |
5444 | unsigned int slave = p->wakee_flips; | |
7d9ffa89 | 5445 | int factor = this_cpu_read(sd_llc_size); |
62470419 | 5446 | |
63b0e9ed MG |
5447 | if (master < slave) |
5448 | swap(master, slave); | |
5449 | if (slave < factor || master < slave * factor) | |
5450 | return 0; | |
5451 | return 1; | |
62470419 MW |
5452 | } |
5453 | ||
90001d67 | 5454 | /* |
d153b153 PZ |
5455 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
5456 | * soonest. For the purpose of speed we only consider the waking and previous | |
5457 | * CPU. | |
90001d67 | 5458 | * |
7332dec0 MG |
5459 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
5460 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
5461 | * |
5462 | * wake_affine_weight() - considers the weight to reflect the average | |
5463 | * scheduling latency of the CPUs. This seems to work | |
5464 | * for the overloaded case. | |
90001d67 | 5465 | */ |
3b76c4a3 | 5466 | static int |
89a55f56 | 5467 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 5468 | { |
7332dec0 MG |
5469 | /* |
5470 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
5471 | * context. Only allow the move if cache is shared. Otherwise an | |
5472 | * interrupt intensive workload could force all tasks onto one | |
5473 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
5474 | * |
5475 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
5476 | * There is no guarantee that the cache hot data from an interrupt | |
5477 | * is more important than cache hot data on the prev_cpu and from | |
5478 | * a cpufreq perspective, it's better to have higher utilisation | |
5479 | * on one CPU. | |
7332dec0 | 5480 | */ |
943d355d RJ |
5481 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
5482 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 5483 | |
d153b153 | 5484 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 5485 | return this_cpu; |
90001d67 | 5486 | |
3b76c4a3 | 5487 | return nr_cpumask_bits; |
90001d67 PZ |
5488 | } |
5489 | ||
3b76c4a3 | 5490 | static int |
f2cdd9cc PZ |
5491 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
5492 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5493 | { |
90001d67 PZ |
5494 | s64 this_eff_load, prev_eff_load; |
5495 | unsigned long task_load; | |
5496 | ||
a3df0679 | 5497 | this_eff_load = cpu_runnable_load(cpu_rq(this_cpu)); |
90001d67 | 5498 | |
90001d67 PZ |
5499 | if (sync) { |
5500 | unsigned long current_load = task_h_load(current); | |
5501 | ||
f2cdd9cc | 5502 | if (current_load > this_eff_load) |
3b76c4a3 | 5503 | return this_cpu; |
90001d67 | 5504 | |
f2cdd9cc | 5505 | this_eff_load -= current_load; |
90001d67 PZ |
5506 | } |
5507 | ||
90001d67 PZ |
5508 | task_load = task_h_load(p); |
5509 | ||
f2cdd9cc PZ |
5510 | this_eff_load += task_load; |
5511 | if (sched_feat(WA_BIAS)) | |
5512 | this_eff_load *= 100; | |
5513 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 5514 | |
a3df0679 | 5515 | prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
5516 | prev_eff_load -= task_load; |
5517 | if (sched_feat(WA_BIAS)) | |
5518 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
5519 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 5520 | |
082f764a MG |
5521 | /* |
5522 | * If sync, adjust the weight of prev_eff_load such that if | |
5523 | * prev_eff == this_eff that select_idle_sibling() will consider | |
5524 | * stacking the wakee on top of the waker if no other CPU is | |
5525 | * idle. | |
5526 | */ | |
5527 | if (sync) | |
5528 | prev_eff_load += 1; | |
5529 | ||
5530 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
5531 | } |
5532 | ||
772bd008 | 5533 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 5534 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 5535 | { |
3b76c4a3 | 5536 | int target = nr_cpumask_bits; |
098fb9db | 5537 | |
89a55f56 | 5538 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 5539 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 5540 | |
3b76c4a3 MG |
5541 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
5542 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 5543 | |
ae92882e | 5544 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
3b76c4a3 MG |
5545 | if (target == nr_cpumask_bits) |
5546 | return prev_cpu; | |
098fb9db | 5547 | |
3b76c4a3 MG |
5548 | schedstat_inc(sd->ttwu_move_affine); |
5549 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
5550 | return target; | |
098fb9db IM |
5551 | } |
5552 | ||
c469933e | 5553 | static unsigned long cpu_util_without(int cpu, struct task_struct *p); |
6a0b19c0 | 5554 | |
c469933e | 5555 | static unsigned long capacity_spare_without(int cpu, struct task_struct *p) |
6a0b19c0 | 5556 | { |
c469933e | 5557 | return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0); |
6a0b19c0 MR |
5558 | } |
5559 | ||
aaee1203 PZ |
5560 | /* |
5561 | * find_idlest_group finds and returns the least busy CPU group within the | |
5562 | * domain. | |
6fee85cc BJ |
5563 | * |
5564 | * Assumes p is allowed on at least one CPU in sd. | |
aaee1203 PZ |
5565 | */ |
5566 | static struct sched_group * | |
78e7ed53 | 5567 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 5568 | int this_cpu, int sd_flag) |
e7693a36 | 5569 | { |
b3bd3de6 | 5570 | struct sched_group *idlest = NULL, *group = sd->groups; |
6a0b19c0 | 5571 | struct sched_group *most_spare_sg = NULL; |
0d10ab95 BJ |
5572 | unsigned long min_runnable_load = ULONG_MAX; |
5573 | unsigned long this_runnable_load = ULONG_MAX; | |
5574 | unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX; | |
6a0b19c0 | 5575 | unsigned long most_spare = 0, this_spare = 0; |
6b94780e VG |
5576 | int imbalance_scale = 100 + (sd->imbalance_pct-100)/2; |
5577 | unsigned long imbalance = scale_load_down(NICE_0_LOAD) * | |
5578 | (sd->imbalance_pct-100) / 100; | |
e7693a36 | 5579 | |
aaee1203 | 5580 | do { |
6b94780e VG |
5581 | unsigned long load, avg_load, runnable_load; |
5582 | unsigned long spare_cap, max_spare_cap; | |
aaee1203 PZ |
5583 | int local_group; |
5584 | int i; | |
e7693a36 | 5585 | |
aaee1203 | 5586 | /* Skip over this group if it has no CPUs allowed */ |
ae4df9d6 | 5587 | if (!cpumask_intersects(sched_group_span(group), |
3bd37062 | 5588 | p->cpus_ptr)) |
aaee1203 PZ |
5589 | continue; |
5590 | ||
5591 | local_group = cpumask_test_cpu(this_cpu, | |
ae4df9d6 | 5592 | sched_group_span(group)); |
aaee1203 | 5593 | |
6a0b19c0 MR |
5594 | /* |
5595 | * Tally up the load of all CPUs in the group and find | |
5596 | * the group containing the CPU with most spare capacity. | |
5597 | */ | |
aaee1203 | 5598 | avg_load = 0; |
6b94780e | 5599 | runnable_load = 0; |
6a0b19c0 | 5600 | max_spare_cap = 0; |
aaee1203 | 5601 | |
ae4df9d6 | 5602 | for_each_cpu(i, sched_group_span(group)) { |
a3df0679 | 5603 | load = cpu_runnable_load(cpu_rq(i)); |
6b94780e VG |
5604 | runnable_load += load; |
5605 | ||
5606 | avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs); | |
6a0b19c0 | 5607 | |
c469933e | 5608 | spare_cap = capacity_spare_without(i, p); |
6a0b19c0 MR |
5609 | |
5610 | if (spare_cap > max_spare_cap) | |
5611 | max_spare_cap = spare_cap; | |
aaee1203 PZ |
5612 | } |
5613 | ||
63b2ca30 | 5614 | /* Adjust by relative CPU capacity of the group */ |
6b94780e VG |
5615 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / |
5616 | group->sgc->capacity; | |
5617 | runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) / | |
5618 | group->sgc->capacity; | |
aaee1203 PZ |
5619 | |
5620 | if (local_group) { | |
6b94780e VG |
5621 | this_runnable_load = runnable_load; |
5622 | this_avg_load = avg_load; | |
6a0b19c0 MR |
5623 | this_spare = max_spare_cap; |
5624 | } else { | |
6b94780e VG |
5625 | if (min_runnable_load > (runnable_load + imbalance)) { |
5626 | /* | |
5627 | * The runnable load is significantly smaller | |
97fb7a0a | 5628 | * so we can pick this new CPU: |
6b94780e VG |
5629 | */ |
5630 | min_runnable_load = runnable_load; | |
5631 | min_avg_load = avg_load; | |
5632 | idlest = group; | |
5633 | } else if ((runnable_load < (min_runnable_load + imbalance)) && | |
5634 | (100*min_avg_load > imbalance_scale*avg_load)) { | |
5635 | /* | |
5636 | * The runnable loads are close so take the | |
97fb7a0a | 5637 | * blocked load into account through avg_load: |
6b94780e VG |
5638 | */ |
5639 | min_avg_load = avg_load; | |
6a0b19c0 MR |
5640 | idlest = group; |
5641 | } | |
5642 | ||
5643 | if (most_spare < max_spare_cap) { | |
5644 | most_spare = max_spare_cap; | |
5645 | most_spare_sg = group; | |
5646 | } | |
aaee1203 PZ |
5647 | } |
5648 | } while (group = group->next, group != sd->groups); | |
5649 | ||
6a0b19c0 MR |
5650 | /* |
5651 | * The cross-over point between using spare capacity or least load | |
5652 | * is too conservative for high utilization tasks on partially | |
5653 | * utilized systems if we require spare_capacity > task_util(p), | |
5654 | * so we allow for some task stuffing by using | |
5655 | * spare_capacity > task_util(p)/2. | |
f519a3f1 VG |
5656 | * |
5657 | * Spare capacity can't be used for fork because the utilization has | |
5658 | * not been set yet, we must first select a rq to compute the initial | |
5659 | * utilization. | |
6a0b19c0 | 5660 | */ |
f519a3f1 VG |
5661 | if (sd_flag & SD_BALANCE_FORK) |
5662 | goto skip_spare; | |
5663 | ||
6a0b19c0 | 5664 | if (this_spare > task_util(p) / 2 && |
6b94780e | 5665 | imbalance_scale*this_spare > 100*most_spare) |
6a0b19c0 | 5666 | return NULL; |
6b94780e VG |
5667 | |
5668 | if (most_spare > task_util(p) / 2) | |
6a0b19c0 MR |
5669 | return most_spare_sg; |
5670 | ||
f519a3f1 | 5671 | skip_spare: |
6b94780e VG |
5672 | if (!idlest) |
5673 | return NULL; | |
5674 | ||
2c833627 MG |
5675 | /* |
5676 | * When comparing groups across NUMA domains, it's possible for the | |
5677 | * local domain to be very lightly loaded relative to the remote | |
5678 | * domains but "imbalance" skews the comparison making remote CPUs | |
5679 | * look much more favourable. When considering cross-domain, add | |
5680 | * imbalance to the runnable load on the remote node and consider | |
5681 | * staying local. | |
5682 | */ | |
5683 | if ((sd->flags & SD_NUMA) && | |
5684 | min_runnable_load + imbalance >= this_runnable_load) | |
5685 | return NULL; | |
5686 | ||
6b94780e | 5687 | if (min_runnable_load > (this_runnable_load + imbalance)) |
aaee1203 | 5688 | return NULL; |
6b94780e VG |
5689 | |
5690 | if ((this_runnable_load < (min_runnable_load + imbalance)) && | |
5691 | (100*this_avg_load < imbalance_scale*min_avg_load)) | |
5692 | return NULL; | |
5693 | ||
aaee1203 PZ |
5694 | return idlest; |
5695 | } | |
5696 | ||
5697 | /* | |
97fb7a0a | 5698 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
5699 | */ |
5700 | static int | |
18bd1b4b | 5701 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
5702 | { |
5703 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5704 | unsigned int min_exit_latency = UINT_MAX; |
5705 | u64 latest_idle_timestamp = 0; | |
5706 | int least_loaded_cpu = this_cpu; | |
3c29e651 | 5707 | int shallowest_idle_cpu = -1, si_cpu = -1; |
aaee1203 PZ |
5708 | int i; |
5709 | ||
eaecf41f MR |
5710 | /* Check if we have any choice: */ |
5711 | if (group->group_weight == 1) | |
ae4df9d6 | 5712 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 5713 | |
aaee1203 | 5714 | /* Traverse only the allowed CPUs */ |
3bd37062 | 5715 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
943d355d | 5716 | if (available_idle_cpu(i)) { |
83a0a96a NP |
5717 | struct rq *rq = cpu_rq(i); |
5718 | struct cpuidle_state *idle = idle_get_state(rq); | |
5719 | if (idle && idle->exit_latency < min_exit_latency) { | |
5720 | /* | |
5721 | * We give priority to a CPU whose idle state | |
5722 | * has the smallest exit latency irrespective | |
5723 | * of any idle timestamp. | |
5724 | */ | |
5725 | min_exit_latency = idle->exit_latency; | |
5726 | latest_idle_timestamp = rq->idle_stamp; | |
5727 | shallowest_idle_cpu = i; | |
5728 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5729 | rq->idle_stamp > latest_idle_timestamp) { | |
5730 | /* | |
5731 | * If equal or no active idle state, then | |
5732 | * the most recently idled CPU might have | |
5733 | * a warmer cache. | |
5734 | */ | |
5735 | latest_idle_timestamp = rq->idle_stamp; | |
5736 | shallowest_idle_cpu = i; | |
5737 | } | |
3c29e651 VK |
5738 | } else if (shallowest_idle_cpu == -1 && si_cpu == -1) { |
5739 | if (sched_idle_cpu(i)) { | |
5740 | si_cpu = i; | |
5741 | continue; | |
5742 | } | |
5743 | ||
a3df0679 | 5744 | load = cpu_runnable_load(cpu_rq(i)); |
18cec7e0 | 5745 | if (load < min_load) { |
83a0a96a NP |
5746 | min_load = load; |
5747 | least_loaded_cpu = i; | |
5748 | } | |
e7693a36 GH |
5749 | } |
5750 | } | |
5751 | ||
3c29e651 VK |
5752 | if (shallowest_idle_cpu != -1) |
5753 | return shallowest_idle_cpu; | |
5754 | if (si_cpu != -1) | |
5755 | return si_cpu; | |
5756 | return least_loaded_cpu; | |
aaee1203 | 5757 | } |
e7693a36 | 5758 | |
18bd1b4b BJ |
5759 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
5760 | int cpu, int prev_cpu, int sd_flag) | |
5761 | { | |
93f50f90 | 5762 | int new_cpu = cpu; |
18bd1b4b | 5763 | |
3bd37062 | 5764 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
5765 | return prev_cpu; |
5766 | ||
c976a862 | 5767 | /* |
c469933e PB |
5768 | * We need task's util for capacity_spare_without, sync it up to |
5769 | * prev_cpu's last_update_time. | |
c976a862 VK |
5770 | */ |
5771 | if (!(sd_flag & SD_BALANCE_FORK)) | |
5772 | sync_entity_load_avg(&p->se); | |
5773 | ||
18bd1b4b BJ |
5774 | while (sd) { |
5775 | struct sched_group *group; | |
5776 | struct sched_domain *tmp; | |
5777 | int weight; | |
5778 | ||
5779 | if (!(sd->flags & sd_flag)) { | |
5780 | sd = sd->child; | |
5781 | continue; | |
5782 | } | |
5783 | ||
5784 | group = find_idlest_group(sd, p, cpu, sd_flag); | |
5785 | if (!group) { | |
5786 | sd = sd->child; | |
5787 | continue; | |
5788 | } | |
5789 | ||
5790 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 5791 | if (new_cpu == cpu) { |
97fb7a0a | 5792 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
5793 | sd = sd->child; |
5794 | continue; | |
5795 | } | |
5796 | ||
97fb7a0a | 5797 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
5798 | cpu = new_cpu; |
5799 | weight = sd->span_weight; | |
5800 | sd = NULL; | |
5801 | for_each_domain(cpu, tmp) { | |
5802 | if (weight <= tmp->span_weight) | |
5803 | break; | |
5804 | if (tmp->flags & sd_flag) | |
5805 | sd = tmp; | |
5806 | } | |
18bd1b4b BJ |
5807 | } |
5808 | ||
5809 | return new_cpu; | |
5810 | } | |
5811 | ||
10e2f1ac | 5812 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 5813 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 5814 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
5815 | |
5816 | static inline void set_idle_cores(int cpu, int val) | |
5817 | { | |
5818 | struct sched_domain_shared *sds; | |
5819 | ||
5820 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5821 | if (sds) | |
5822 | WRITE_ONCE(sds->has_idle_cores, val); | |
5823 | } | |
5824 | ||
5825 | static inline bool test_idle_cores(int cpu, bool def) | |
5826 | { | |
5827 | struct sched_domain_shared *sds; | |
5828 | ||
5829 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5830 | if (sds) | |
5831 | return READ_ONCE(sds->has_idle_cores); | |
5832 | ||
5833 | return def; | |
5834 | } | |
5835 | ||
5836 | /* | |
5837 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
5838 | * information in sd_llc_shared->has_idle_cores. | |
5839 | * | |
5840 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
5841 | * state should be fairly cheap. | |
5842 | */ | |
1b568f0a | 5843 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
5844 | { |
5845 | int core = cpu_of(rq); | |
5846 | int cpu; | |
5847 | ||
5848 | rcu_read_lock(); | |
5849 | if (test_idle_cores(core, true)) | |
5850 | goto unlock; | |
5851 | ||
5852 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
5853 | if (cpu == core) | |
5854 | continue; | |
5855 | ||
943d355d | 5856 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
5857 | goto unlock; |
5858 | } | |
5859 | ||
5860 | set_idle_cores(core, 1); | |
5861 | unlock: | |
5862 | rcu_read_unlock(); | |
5863 | } | |
5864 | ||
5865 | /* | |
5866 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
5867 | * there are no idle cores left in the system; tracked through | |
5868 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
5869 | */ | |
5870 | static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
5871 | { | |
5872 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); | |
c743f0a5 | 5873 | int core, cpu; |
10e2f1ac | 5874 | |
1b568f0a PZ |
5875 | if (!static_branch_likely(&sched_smt_present)) |
5876 | return -1; | |
5877 | ||
10e2f1ac PZ |
5878 | if (!test_idle_cores(target, false)) |
5879 | return -1; | |
5880 | ||
3bd37062 | 5881 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
10e2f1ac | 5882 | |
c743f0a5 | 5883 | for_each_cpu_wrap(core, cpus, target) { |
10e2f1ac PZ |
5884 | bool idle = true; |
5885 | ||
5886 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
c89d92ed | 5887 | __cpumask_clear_cpu(cpu, cpus); |
943d355d | 5888 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
5889 | idle = false; |
5890 | } | |
5891 | ||
5892 | if (idle) | |
5893 | return core; | |
5894 | } | |
5895 | ||
5896 | /* | |
5897 | * Failed to find an idle core; stop looking for one. | |
5898 | */ | |
5899 | set_idle_cores(target, 0); | |
5900 | ||
5901 | return -1; | |
5902 | } | |
5903 | ||
5904 | /* | |
5905 | * Scan the local SMT mask for idle CPUs. | |
5906 | */ | |
1b5500d7 | 5907 | static int select_idle_smt(struct task_struct *p, int target) |
10e2f1ac | 5908 | { |
3c29e651 | 5909 | int cpu, si_cpu = -1; |
10e2f1ac | 5910 | |
1b568f0a PZ |
5911 | if (!static_branch_likely(&sched_smt_present)) |
5912 | return -1; | |
5913 | ||
10e2f1ac | 5914 | for_each_cpu(cpu, cpu_smt_mask(target)) { |
3bd37062 | 5915 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
10e2f1ac | 5916 | continue; |
943d355d | 5917 | if (available_idle_cpu(cpu)) |
10e2f1ac | 5918 | return cpu; |
3c29e651 VK |
5919 | if (si_cpu == -1 && sched_idle_cpu(cpu)) |
5920 | si_cpu = cpu; | |
10e2f1ac PZ |
5921 | } |
5922 | ||
3c29e651 | 5923 | return si_cpu; |
10e2f1ac PZ |
5924 | } |
5925 | ||
5926 | #else /* CONFIG_SCHED_SMT */ | |
5927 | ||
5928 | static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
5929 | { | |
5930 | return -1; | |
5931 | } | |
5932 | ||
1b5500d7 | 5933 | static inline int select_idle_smt(struct task_struct *p, int target) |
10e2f1ac PZ |
5934 | { |
5935 | return -1; | |
5936 | } | |
5937 | ||
5938 | #endif /* CONFIG_SCHED_SMT */ | |
5939 | ||
5940 | /* | |
5941 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
5942 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
5943 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 5944 | */ |
10e2f1ac PZ |
5945 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target) |
5946 | { | |
9cfb38a7 | 5947 | struct sched_domain *this_sd; |
1ad3aaf3 | 5948 | u64 avg_cost, avg_idle; |
10e2f1ac PZ |
5949 | u64 time, cost; |
5950 | s64 delta; | |
8dc2d993 | 5951 | int this = smp_processor_id(); |
3c29e651 | 5952 | int cpu, nr = INT_MAX, si_cpu = -1; |
10e2f1ac | 5953 | |
9cfb38a7 WL |
5954 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
5955 | if (!this_sd) | |
5956 | return -1; | |
5957 | ||
10e2f1ac PZ |
5958 | /* |
5959 | * Due to large variance we need a large fuzz factor; hackbench in | |
5960 | * particularly is sensitive here. | |
5961 | */ | |
1ad3aaf3 PZ |
5962 | avg_idle = this_rq()->avg_idle / 512; |
5963 | avg_cost = this_sd->avg_scan_cost + 1; | |
5964 | ||
5965 | if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost) | |
10e2f1ac PZ |
5966 | return -1; |
5967 | ||
1ad3aaf3 PZ |
5968 | if (sched_feat(SIS_PROP)) { |
5969 | u64 span_avg = sd->span_weight * avg_idle; | |
5970 | if (span_avg > 4*avg_cost) | |
5971 | nr = div_u64(span_avg, avg_cost); | |
5972 | else | |
5973 | nr = 4; | |
5974 | } | |
5975 | ||
8dc2d993 | 5976 | time = cpu_clock(this); |
10e2f1ac | 5977 | |
c743f0a5 | 5978 | for_each_cpu_wrap(cpu, sched_domain_span(sd), target) { |
1ad3aaf3 | 5979 | if (!--nr) |
3c29e651 | 5980 | return si_cpu; |
3bd37062 | 5981 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
10e2f1ac | 5982 | continue; |
943d355d | 5983 | if (available_idle_cpu(cpu)) |
10e2f1ac | 5984 | break; |
3c29e651 VK |
5985 | if (si_cpu == -1 && sched_idle_cpu(cpu)) |
5986 | si_cpu = cpu; | |
10e2f1ac PZ |
5987 | } |
5988 | ||
8dc2d993 | 5989 | time = cpu_clock(this) - time; |
10e2f1ac PZ |
5990 | cost = this_sd->avg_scan_cost; |
5991 | delta = (s64)(time - cost) / 8; | |
5992 | this_sd->avg_scan_cost += delta; | |
5993 | ||
5994 | return cpu; | |
5995 | } | |
5996 | ||
5997 | /* | |
5998 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 5999 | */ |
772bd008 | 6000 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6001 | { |
99bd5e2f | 6002 | struct sched_domain *sd; |
32e839dd | 6003 | int i, recent_used_cpu; |
a50bde51 | 6004 | |
3c29e651 | 6005 | if (available_idle_cpu(target) || sched_idle_cpu(target)) |
e0a79f52 | 6006 | return target; |
99bd5e2f SS |
6007 | |
6008 | /* | |
97fb7a0a | 6009 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6010 | */ |
3c29e651 VK |
6011 | if (prev != target && cpus_share_cache(prev, target) && |
6012 | (available_idle_cpu(prev) || sched_idle_cpu(prev))) | |
772bd008 | 6013 | return prev; |
a50bde51 | 6014 | |
97fb7a0a | 6015 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd MG |
6016 | recent_used_cpu = p->recent_used_cpu; |
6017 | if (recent_used_cpu != prev && | |
6018 | recent_used_cpu != target && | |
6019 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 6020 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
3bd37062 | 6021 | cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) { |
32e839dd MG |
6022 | /* |
6023 | * Replace recent_used_cpu with prev as it is a potential | |
97fb7a0a | 6024 | * candidate for the next wake: |
32e839dd MG |
6025 | */ |
6026 | p->recent_used_cpu = prev; | |
6027 | return recent_used_cpu; | |
6028 | } | |
6029 | ||
518cd623 | 6030 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6031 | if (!sd) |
6032 | return target; | |
772bd008 | 6033 | |
10e2f1ac PZ |
6034 | i = select_idle_core(p, sd, target); |
6035 | if ((unsigned)i < nr_cpumask_bits) | |
6036 | return i; | |
37407ea7 | 6037 | |
10e2f1ac PZ |
6038 | i = select_idle_cpu(p, sd, target); |
6039 | if ((unsigned)i < nr_cpumask_bits) | |
6040 | return i; | |
6041 | ||
1b5500d7 | 6042 | i = select_idle_smt(p, target); |
10e2f1ac PZ |
6043 | if ((unsigned)i < nr_cpumask_bits) |
6044 | return i; | |
970e1789 | 6045 | |
a50bde51 PZ |
6046 | return target; |
6047 | } | |
231678b7 | 6048 | |
f9be3e59 PB |
6049 | /** |
6050 | * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks | |
6051 | * @cpu: the CPU to get the utilization of | |
6052 | * | |
6053 | * The unit of the return value must be the one of capacity so we can compare | |
6054 | * the utilization with the capacity of the CPU that is available for CFS task | |
6055 | * (ie cpu_capacity). | |
231678b7 DE |
6056 | * |
6057 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
6058 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
6059 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
6060 | * capacity_orig is the cpu_capacity available at the highest frequency | |
6061 | * (arch_scale_freq_capacity()). | |
6062 | * The utilization of a CPU converges towards a sum equal to or less than the | |
6063 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
6064 | * the running time on this CPU scaled by capacity_curr. | |
6065 | * | |
f9be3e59 PB |
6066 | * The estimated utilization of a CPU is defined to be the maximum between its |
6067 | * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks | |
6068 | * currently RUNNABLE on that CPU. | |
6069 | * This allows to properly represent the expected utilization of a CPU which | |
6070 | * has just got a big task running since a long sleep period. At the same time | |
6071 | * however it preserves the benefits of the "blocked utilization" in | |
6072 | * describing the potential for other tasks waking up on the same CPU. | |
6073 | * | |
231678b7 DE |
6074 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even |
6075 | * higher than capacity_orig because of unfortunate rounding in | |
6076 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
6077 | * the average stabilizes with the new running time. We need to check that the | |
6078 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
6079 | * necessary. Without utilization capping, a group could be seen as overloaded | |
6080 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
6081 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
6082 | * capacity_orig) as it useful for predicting the capacity required after task | |
6083 | * migrations (scheduler-driven DVFS). | |
f9be3e59 PB |
6084 | * |
6085 | * Return: the (estimated) utilization for the specified CPU | |
8bb5b00c | 6086 | */ |
f9be3e59 | 6087 | static inline unsigned long cpu_util(int cpu) |
8bb5b00c | 6088 | { |
f9be3e59 PB |
6089 | struct cfs_rq *cfs_rq; |
6090 | unsigned int util; | |
6091 | ||
6092 | cfs_rq = &cpu_rq(cpu)->cfs; | |
6093 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6094 | ||
6095 | if (sched_feat(UTIL_EST)) | |
6096 | util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); | |
8bb5b00c | 6097 | |
f9be3e59 | 6098 | return min_t(unsigned long, util, capacity_orig_of(cpu)); |
8bb5b00c | 6099 | } |
a50bde51 | 6100 | |
104cb16d | 6101 | /* |
c469933e PB |
6102 | * cpu_util_without: compute cpu utilization without any contributions from *p |
6103 | * @cpu: the CPU which utilization is requested | |
6104 | * @p: the task which utilization should be discounted | |
6105 | * | |
6106 | * The utilization of a CPU is defined by the utilization of tasks currently | |
6107 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
6108 | * execution on that CPU. | |
6109 | * | |
6110 | * This method returns the utilization of the specified CPU by discounting the | |
6111 | * utilization of the specified task, whenever the task is currently | |
6112 | * contributing to the CPU utilization. | |
104cb16d | 6113 | */ |
c469933e | 6114 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) |
104cb16d | 6115 | { |
f9be3e59 PB |
6116 | struct cfs_rq *cfs_rq; |
6117 | unsigned int util; | |
104cb16d MR |
6118 | |
6119 | /* Task has no contribution or is new */ | |
f9be3e59 | 6120 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) |
104cb16d MR |
6121 | return cpu_util(cpu); |
6122 | ||
f9be3e59 PB |
6123 | cfs_rq = &cpu_rq(cpu)->cfs; |
6124 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6125 | ||
c469933e | 6126 | /* Discount task's util from CPU's util */ |
b5c0ce7b | 6127 | lsub_positive(&util, task_util(p)); |
104cb16d | 6128 | |
f9be3e59 PB |
6129 | /* |
6130 | * Covered cases: | |
6131 | * | |
6132 | * a) if *p is the only task sleeping on this CPU, then: | |
6133 | * cpu_util (== task_util) > util_est (== 0) | |
6134 | * and thus we return: | |
c469933e | 6135 | * cpu_util_without = (cpu_util - task_util) = 0 |
f9be3e59 PB |
6136 | * |
6137 | * b) if other tasks are SLEEPING on this CPU, which is now exiting | |
6138 | * IDLE, then: | |
6139 | * cpu_util >= task_util | |
6140 | * cpu_util > util_est (== 0) | |
6141 | * and thus we discount *p's blocked utilization to return: | |
c469933e | 6142 | * cpu_util_without = (cpu_util - task_util) >= 0 |
f9be3e59 PB |
6143 | * |
6144 | * c) if other tasks are RUNNABLE on that CPU and | |
6145 | * util_est > cpu_util | |
6146 | * then we use util_est since it returns a more restrictive | |
6147 | * estimation of the spare capacity on that CPU, by just | |
6148 | * considering the expected utilization of tasks already | |
6149 | * runnable on that CPU. | |
6150 | * | |
6151 | * Cases a) and b) are covered by the above code, while case c) is | |
6152 | * covered by the following code when estimated utilization is | |
6153 | * enabled. | |
6154 | */ | |
c469933e PB |
6155 | if (sched_feat(UTIL_EST)) { |
6156 | unsigned int estimated = | |
6157 | READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6158 | ||
6159 | /* | |
6160 | * Despite the following checks we still have a small window | |
6161 | * for a possible race, when an execl's select_task_rq_fair() | |
6162 | * races with LB's detach_task(): | |
6163 | * | |
6164 | * detach_task() | |
6165 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
6166 | * ---------------------------------- A | |
6167 | * deactivate_task() \ | |
6168 | * dequeue_task() + RaceTime | |
6169 | * util_est_dequeue() / | |
6170 | * ---------------------------------- B | |
6171 | * | |
6172 | * The additional check on "current == p" it's required to | |
6173 | * properly fix the execl regression and it helps in further | |
6174 | * reducing the chances for the above race. | |
6175 | */ | |
b5c0ce7b PB |
6176 | if (unlikely(task_on_rq_queued(p) || current == p)) |
6177 | lsub_positive(&estimated, _task_util_est(p)); | |
6178 | ||
c469933e PB |
6179 | util = max(util, estimated); |
6180 | } | |
f9be3e59 PB |
6181 | |
6182 | /* | |
6183 | * Utilization (estimated) can exceed the CPU capacity, thus let's | |
6184 | * clamp to the maximum CPU capacity to ensure consistency with | |
6185 | * the cpu_util call. | |
6186 | */ | |
6187 | return min_t(unsigned long, util, capacity_orig_of(cpu)); | |
104cb16d MR |
6188 | } |
6189 | ||
3273163c MR |
6190 | /* |
6191 | * Disable WAKE_AFFINE in the case where task @p doesn't fit in the | |
6192 | * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu. | |
6193 | * | |
6194 | * In that case WAKE_AFFINE doesn't make sense and we'll let | |
6195 | * BALANCE_WAKE sort things out. | |
6196 | */ | |
6197 | static int wake_cap(struct task_struct *p, int cpu, int prev_cpu) | |
6198 | { | |
6199 | long min_cap, max_cap; | |
6200 | ||
df054e84 MR |
6201 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) |
6202 | return 0; | |
6203 | ||
3273163c MR |
6204 | min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu)); |
6205 | max_cap = cpu_rq(cpu)->rd->max_cpu_capacity; | |
6206 | ||
6207 | /* Minimum capacity is close to max, no need to abort wake_affine */ | |
6208 | if (max_cap - min_cap < max_cap >> 3) | |
6209 | return 0; | |
6210 | ||
104cb16d MR |
6211 | /* Bring task utilization in sync with prev_cpu */ |
6212 | sync_entity_load_avg(&p->se); | |
6213 | ||
3b1baa64 | 6214 | return !task_fits_capacity(p, min_cap); |
3273163c MR |
6215 | } |
6216 | ||
390031e4 QP |
6217 | /* |
6218 | * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) | |
6219 | * to @dst_cpu. | |
6220 | */ | |
6221 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
6222 | { | |
6223 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
6224 | unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); | |
6225 | ||
6226 | /* | |
6227 | * If @p migrates from @cpu to another, remove its contribution. Or, | |
6228 | * if @p migrates from another CPU to @cpu, add its contribution. In | |
6229 | * the other cases, @cpu is not impacted by the migration, so the | |
6230 | * util_avg should already be correct. | |
6231 | */ | |
6232 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
6233 | sub_positive(&util, task_util(p)); | |
6234 | else if (task_cpu(p) != cpu && dst_cpu == cpu) | |
6235 | util += task_util(p); | |
6236 | ||
6237 | if (sched_feat(UTIL_EST)) { | |
6238 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6239 | ||
6240 | /* | |
6241 | * During wake-up, the task isn't enqueued yet and doesn't | |
6242 | * appear in the cfs_rq->avg.util_est.enqueued of any rq, | |
6243 | * so just add it (if needed) to "simulate" what will be | |
6244 | * cpu_util() after the task has been enqueued. | |
6245 | */ | |
6246 | if (dst_cpu == cpu) | |
6247 | util_est += _task_util_est(p); | |
6248 | ||
6249 | util = max(util, util_est); | |
6250 | } | |
6251 | ||
6252 | return min(util, capacity_orig_of(cpu)); | |
6253 | } | |
6254 | ||
6255 | /* | |
6256 | * compute_energy(): Estimates the energy that would be consumed if @p was | |
6257 | * migrated to @dst_cpu. compute_energy() predicts what will be the utilization | |
6258 | * landscape of the * CPUs after the task migration, and uses the Energy Model | |
6259 | * to compute what would be the energy if we decided to actually migrate that | |
6260 | * task. | |
6261 | */ | |
6262 | static long | |
6263 | compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) | |
6264 | { | |
af24bde8 PB |
6265 | unsigned int max_util, util_cfs, cpu_util, cpu_cap; |
6266 | unsigned long sum_util, energy = 0; | |
6267 | struct task_struct *tsk; | |
390031e4 QP |
6268 | int cpu; |
6269 | ||
6270 | for (; pd; pd = pd->next) { | |
af24bde8 PB |
6271 | struct cpumask *pd_mask = perf_domain_span(pd); |
6272 | ||
6273 | /* | |
6274 | * The energy model mandates all the CPUs of a performance | |
6275 | * domain have the same capacity. | |
6276 | */ | |
6277 | cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); | |
390031e4 | 6278 | max_util = sum_util = 0; |
af24bde8 | 6279 | |
390031e4 QP |
6280 | /* |
6281 | * The capacity state of CPUs of the current rd can be driven by | |
6282 | * CPUs of another rd if they belong to the same performance | |
6283 | * domain. So, account for the utilization of these CPUs too | |
6284 | * by masking pd with cpu_online_mask instead of the rd span. | |
6285 | * | |
6286 | * If an entire performance domain is outside of the current rd, | |
6287 | * it will not appear in its pd list and will not be accounted | |
6288 | * by compute_energy(). | |
6289 | */ | |
af24bde8 PB |
6290 | for_each_cpu_and(cpu, pd_mask, cpu_online_mask) { |
6291 | util_cfs = cpu_util_next(cpu, p, dst_cpu); | |
6292 | ||
6293 | /* | |
6294 | * Busy time computation: utilization clamping is not | |
6295 | * required since the ratio (sum_util / cpu_capacity) | |
6296 | * is already enough to scale the EM reported power | |
6297 | * consumption at the (eventually clamped) cpu_capacity. | |
6298 | */ | |
6299 | sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap, | |
6300 | ENERGY_UTIL, NULL); | |
6301 | ||
6302 | /* | |
6303 | * Performance domain frequency: utilization clamping | |
6304 | * must be considered since it affects the selection | |
6305 | * of the performance domain frequency. | |
6306 | * NOTE: in case RT tasks are running, by default the | |
6307 | * FREQUENCY_UTIL's utilization can be max OPP. | |
6308 | */ | |
6309 | tsk = cpu == dst_cpu ? p : NULL; | |
6310 | cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap, | |
6311 | FREQUENCY_UTIL, tsk); | |
6312 | max_util = max(max_util, cpu_util); | |
390031e4 QP |
6313 | } |
6314 | ||
6315 | energy += em_pd_energy(pd->em_pd, max_util, sum_util); | |
6316 | } | |
6317 | ||
6318 | return energy; | |
6319 | } | |
6320 | ||
732cd75b QP |
6321 | /* |
6322 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
6323 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
6324 | * spare capacity in each performance domain and uses it as a potential | |
6325 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
6326 | * out which of the CPU candidates is the most energy-efficient. | |
6327 | * | |
6328 | * The rationale for this heuristic is as follows. In a performance domain, | |
6329 | * all the most energy efficient CPU candidates (according to the Energy | |
6330 | * Model) are those for which we'll request a low frequency. When there are | |
6331 | * several CPUs for which the frequency request will be the same, we don't | |
6332 | * have enough data to break the tie between them, because the Energy Model | |
6333 | * only includes active power costs. With this model, if we assume that | |
6334 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
6335 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
6336 | * the best candidates of the performance domain. | |
6337 | * | |
6338 | * In practice, it could be preferable from an energy standpoint to pack | |
6339 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
6340 | * but that could also hurt our chances to go cluster idle, and we have no | |
6341 | * ways to tell with the current Energy Model if this is actually a good | |
6342 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
6343 | * cluster-packing, and spreading inside a cluster. That should at least be | |
6344 | * a good thing for latency, and this is consistent with the idea that most | |
6345 | * of the energy savings of EAS come from the asymmetry of the system, and | |
6346 | * not so much from breaking the tie between identical CPUs. That's also the | |
6347 | * reason why EAS is enabled in the topology code only for systems where | |
6348 | * SD_ASYM_CPUCAPACITY is set. | |
6349 | * | |
6350 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
6351 | * they don't have any useful utilization data yet and it's not possible to | |
6352 | * forecast their impact on energy consumption. Consequently, they will be | |
6353 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
6354 | * to be energy-inefficient in some use-cases. The alternative would be to | |
6355 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
6356 | * their util_avg from the parent task, but those heuristics could hurt | |
6357 | * other use-cases too. So, until someone finds a better way to solve this, | |
6358 | * let's keep things simple by re-using the existing slow path. | |
6359 | */ | |
6360 | ||
6361 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) | |
6362 | { | |
6363 | unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX; | |
6364 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; | |
6365 | int cpu, best_energy_cpu = prev_cpu; | |
6366 | struct perf_domain *head, *pd; | |
6367 | unsigned long cpu_cap, util; | |
6368 | struct sched_domain *sd; | |
6369 | ||
6370 | rcu_read_lock(); | |
6371 | pd = rcu_dereference(rd->pd); | |
6372 | if (!pd || READ_ONCE(rd->overutilized)) | |
6373 | goto fail; | |
6374 | head = pd; | |
6375 | ||
6376 | /* | |
6377 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
6378 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
6379 | */ | |
6380 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
6381 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
6382 | sd = sd->parent; | |
6383 | if (!sd) | |
6384 | goto fail; | |
6385 | ||
6386 | sync_entity_load_avg(&p->se); | |
6387 | if (!task_util_est(p)) | |
6388 | goto unlock; | |
6389 | ||
6390 | for (; pd; pd = pd->next) { | |
6391 | unsigned long cur_energy, spare_cap, max_spare_cap = 0; | |
6392 | int max_spare_cap_cpu = -1; | |
6393 | ||
6394 | for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { | |
3bd37062 | 6395 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
6396 | continue; |
6397 | ||
6398 | /* Skip CPUs that will be overutilized. */ | |
6399 | util = cpu_util_next(cpu, p, cpu); | |
6400 | cpu_cap = capacity_of(cpu); | |
60e17f5c | 6401 | if (!fits_capacity(util, cpu_cap)) |
732cd75b QP |
6402 | continue; |
6403 | ||
6404 | /* Always use prev_cpu as a candidate. */ | |
6405 | if (cpu == prev_cpu) { | |
6406 | prev_energy = compute_energy(p, prev_cpu, head); | |
6407 | best_energy = min(best_energy, prev_energy); | |
6408 | continue; | |
6409 | } | |
6410 | ||
6411 | /* | |
6412 | * Find the CPU with the maximum spare capacity in | |
6413 | * the performance domain | |
6414 | */ | |
6415 | spare_cap = cpu_cap - util; | |
6416 | if (spare_cap > max_spare_cap) { | |
6417 | max_spare_cap = spare_cap; | |
6418 | max_spare_cap_cpu = cpu; | |
6419 | } | |
6420 | } | |
6421 | ||
6422 | /* Evaluate the energy impact of using this CPU. */ | |
6423 | if (max_spare_cap_cpu >= 0) { | |
6424 | cur_energy = compute_energy(p, max_spare_cap_cpu, head); | |
6425 | if (cur_energy < best_energy) { | |
6426 | best_energy = cur_energy; | |
6427 | best_energy_cpu = max_spare_cap_cpu; | |
6428 | } | |
6429 | } | |
6430 | } | |
6431 | unlock: | |
6432 | rcu_read_unlock(); | |
6433 | ||
6434 | /* | |
6435 | * Pick the best CPU if prev_cpu cannot be used, or if it saves at | |
6436 | * least 6% of the energy used by prev_cpu. | |
6437 | */ | |
6438 | if (prev_energy == ULONG_MAX) | |
6439 | return best_energy_cpu; | |
6440 | ||
6441 | if ((prev_energy - best_energy) > (prev_energy >> 4)) | |
6442 | return best_energy_cpu; | |
6443 | ||
6444 | return prev_cpu; | |
6445 | ||
6446 | fail: | |
6447 | rcu_read_unlock(); | |
6448 | ||
6449 | return -1; | |
6450 | } | |
6451 | ||
aaee1203 | 6452 | /* |
de91b9cb MR |
6453 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
6454 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
6455 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 6456 | * |
97fb7a0a IM |
6457 | * Balances load by selecting the idlest CPU in the idlest group, or under |
6458 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 6459 | * |
97fb7a0a | 6460 | * Returns the target CPU number. |
aaee1203 PZ |
6461 | * |
6462 | * preempt must be disabled. | |
6463 | */ | |
0017d735 | 6464 | static int |
ac66f547 | 6465 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 6466 | { |
f1d88b44 | 6467 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 6468 | int cpu = smp_processor_id(); |
63b0e9ed | 6469 | int new_cpu = prev_cpu; |
99bd5e2f | 6470 | int want_affine = 0; |
24d0c1d6 | 6471 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
c88d5910 | 6472 | |
c58d25f3 PZ |
6473 | if (sd_flag & SD_BALANCE_WAKE) { |
6474 | record_wakee(p); | |
732cd75b | 6475 | |
f8a696f2 | 6476 | if (sched_energy_enabled()) { |
732cd75b QP |
6477 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
6478 | if (new_cpu >= 0) | |
6479 | return new_cpu; | |
6480 | new_cpu = prev_cpu; | |
6481 | } | |
6482 | ||
6483 | want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) && | |
3bd37062 | 6484 | cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 6485 | } |
aaee1203 | 6486 | |
dce840a0 | 6487 | rcu_read_lock(); |
aaee1203 | 6488 | for_each_domain(cpu, tmp) { |
e4f42888 | 6489 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
63b0e9ed | 6490 | break; |
e4f42888 | 6491 | |
fe3bcfe1 | 6492 | /* |
97fb7a0a | 6493 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 6494 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 6495 | */ |
99bd5e2f SS |
6496 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6497 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
6498 | if (cpu != prev_cpu) |
6499 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
6500 | ||
6501 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 6502 | break; |
f03542a7 | 6503 | } |
29cd8bae | 6504 | |
f03542a7 | 6505 | if (tmp->flags & sd_flag) |
29cd8bae | 6506 | sd = tmp; |
63b0e9ed MG |
6507 | else if (!want_affine) |
6508 | break; | |
29cd8bae PZ |
6509 | } |
6510 | ||
f1d88b44 VK |
6511 | if (unlikely(sd)) { |
6512 | /* Slow path */ | |
18bd1b4b | 6513 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
f1d88b44 VK |
6514 | } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */ |
6515 | /* Fast path */ | |
6516 | ||
6517 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); | |
6518 | ||
6519 | if (want_affine) | |
6520 | current->recent_used_cpu = cpu; | |
e7693a36 | 6521 | } |
dce840a0 | 6522 | rcu_read_unlock(); |
e7693a36 | 6523 | |
c88d5910 | 6524 | return new_cpu; |
e7693a36 | 6525 | } |
0a74bef8 | 6526 | |
144d8487 PZ |
6527 | static void detach_entity_cfs_rq(struct sched_entity *se); |
6528 | ||
0a74bef8 | 6529 | /* |
97fb7a0a | 6530 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 6531 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 6532 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6533 | */ |
3f9672ba | 6534 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 6535 | { |
59efa0ba PZ |
6536 | /* |
6537 | * As blocked tasks retain absolute vruntime the migration needs to | |
6538 | * deal with this by subtracting the old and adding the new | |
6539 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6540 | * the task on the new runqueue. | |
6541 | */ | |
6542 | if (p->state == TASK_WAKING) { | |
6543 | struct sched_entity *se = &p->se; | |
6544 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6545 | u64 min_vruntime; | |
6546 | ||
6547 | #ifndef CONFIG_64BIT | |
6548 | u64 min_vruntime_copy; | |
6549 | ||
6550 | do { | |
6551 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
6552 | smp_rmb(); | |
6553 | min_vruntime = cfs_rq->min_vruntime; | |
6554 | } while (min_vruntime != min_vruntime_copy); | |
6555 | #else | |
6556 | min_vruntime = cfs_rq->min_vruntime; | |
6557 | #endif | |
6558 | ||
6559 | se->vruntime -= min_vruntime; | |
6560 | } | |
6561 | ||
144d8487 PZ |
6562 | if (p->on_rq == TASK_ON_RQ_MIGRATING) { |
6563 | /* | |
6564 | * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' | |
6565 | * rq->lock and can modify state directly. | |
6566 | */ | |
6567 | lockdep_assert_held(&task_rq(p)->lock); | |
6568 | detach_entity_cfs_rq(&p->se); | |
6569 | ||
6570 | } else { | |
6571 | /* | |
6572 | * We are supposed to update the task to "current" time, then | |
6573 | * its up to date and ready to go to new CPU/cfs_rq. But we | |
6574 | * have difficulty in getting what current time is, so simply | |
6575 | * throw away the out-of-date time. This will result in the | |
6576 | * wakee task is less decayed, but giving the wakee more load | |
6577 | * sounds not bad. | |
6578 | */ | |
6579 | remove_entity_load_avg(&p->se); | |
6580 | } | |
9d89c257 YD |
6581 | |
6582 | /* Tell new CPU we are migrated */ | |
6583 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
6584 | |
6585 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 6586 | p->se.exec_start = 0; |
3f9672ba SD |
6587 | |
6588 | update_scan_period(p, new_cpu); | |
0a74bef8 | 6589 | } |
12695578 YD |
6590 | |
6591 | static void task_dead_fair(struct task_struct *p) | |
6592 | { | |
6593 | remove_entity_load_avg(&p->se); | |
6594 | } | |
e7693a36 GH |
6595 | #endif /* CONFIG_SMP */ |
6596 | ||
a555e9d8 | 6597 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
6598 | { |
6599 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
6600 | ||
6601 | /* | |
e52fb7c0 PZ |
6602 | * Since its curr running now, convert the gran from real-time |
6603 | * to virtual-time in his units. | |
13814d42 MG |
6604 | * |
6605 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
6606 | * they get preempted easier. That is, if 'se' < 'curr' then | |
6607 | * the resulting gran will be larger, therefore penalizing the | |
6608 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
6609 | * be smaller, again penalizing the lighter task. | |
6610 | * | |
6611 | * This is especially important for buddies when the leftmost | |
6612 | * task is higher priority than the buddy. | |
0bbd3336 | 6613 | */ |
f4ad9bd2 | 6614 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
6615 | } |
6616 | ||
464b7527 PZ |
6617 | /* |
6618 | * Should 'se' preempt 'curr'. | |
6619 | * | |
6620 | * |s1 | |
6621 | * |s2 | |
6622 | * |s3 | |
6623 | * g | |
6624 | * |<--->|c | |
6625 | * | |
6626 | * w(c, s1) = -1 | |
6627 | * w(c, s2) = 0 | |
6628 | * w(c, s3) = 1 | |
6629 | * | |
6630 | */ | |
6631 | static int | |
6632 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
6633 | { | |
6634 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
6635 | ||
6636 | if (vdiff <= 0) | |
6637 | return -1; | |
6638 | ||
a555e9d8 | 6639 | gran = wakeup_gran(se); |
464b7527 PZ |
6640 | if (vdiff > gran) |
6641 | return 1; | |
6642 | ||
6643 | return 0; | |
6644 | } | |
6645 | ||
02479099 PZ |
6646 | static void set_last_buddy(struct sched_entity *se) |
6647 | { | |
1da1843f | 6648 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6649 | return; |
6650 | ||
c5ae366e DA |
6651 | for_each_sched_entity(se) { |
6652 | if (SCHED_WARN_ON(!se->on_rq)) | |
6653 | return; | |
69c80f3e | 6654 | cfs_rq_of(se)->last = se; |
c5ae366e | 6655 | } |
02479099 PZ |
6656 | } |
6657 | ||
6658 | static void set_next_buddy(struct sched_entity *se) | |
6659 | { | |
1da1843f | 6660 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6661 | return; |
6662 | ||
c5ae366e DA |
6663 | for_each_sched_entity(se) { |
6664 | if (SCHED_WARN_ON(!se->on_rq)) | |
6665 | return; | |
69c80f3e | 6666 | cfs_rq_of(se)->next = se; |
c5ae366e | 6667 | } |
02479099 PZ |
6668 | } |
6669 | ||
ac53db59 RR |
6670 | static void set_skip_buddy(struct sched_entity *se) |
6671 | { | |
69c80f3e VP |
6672 | for_each_sched_entity(se) |
6673 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
6674 | } |
6675 | ||
bf0f6f24 IM |
6676 | /* |
6677 | * Preempt the current task with a newly woken task if needed: | |
6678 | */ | |
5a9b86f6 | 6679 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
6680 | { |
6681 | struct task_struct *curr = rq->curr; | |
8651a86c | 6682 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 6683 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 6684 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 6685 | int next_buddy_marked = 0; |
bf0f6f24 | 6686 | |
4ae7d5ce IM |
6687 | if (unlikely(se == pse)) |
6688 | return; | |
6689 | ||
5238cdd3 | 6690 | /* |
163122b7 | 6691 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
6692 | * unconditionally check_prempt_curr() after an enqueue (which may have |
6693 | * lead to a throttle). This both saves work and prevents false | |
6694 | * next-buddy nomination below. | |
6695 | */ | |
6696 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
6697 | return; | |
6698 | ||
2f36825b | 6699 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 6700 | set_next_buddy(pse); |
2f36825b VP |
6701 | next_buddy_marked = 1; |
6702 | } | |
57fdc26d | 6703 | |
aec0a514 BR |
6704 | /* |
6705 | * We can come here with TIF_NEED_RESCHED already set from new task | |
6706 | * wake up path. | |
5238cdd3 PT |
6707 | * |
6708 | * Note: this also catches the edge-case of curr being in a throttled | |
6709 | * group (e.g. via set_curr_task), since update_curr() (in the | |
6710 | * enqueue of curr) will have resulted in resched being set. This | |
6711 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
6712 | * below. | |
aec0a514 BR |
6713 | */ |
6714 | if (test_tsk_need_resched(curr)) | |
6715 | return; | |
6716 | ||
a2f5c9ab | 6717 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
6718 | if (unlikely(task_has_idle_policy(curr)) && |
6719 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
6720 | goto preempt; |
6721 | ||
91c234b4 | 6722 | /* |
a2f5c9ab DH |
6723 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
6724 | * is driven by the tick): | |
91c234b4 | 6725 | */ |
8ed92e51 | 6726 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 6727 | return; |
bf0f6f24 | 6728 | |
464b7527 | 6729 | find_matching_se(&se, &pse); |
9bbd7374 | 6730 | update_curr(cfs_rq_of(se)); |
002f128b | 6731 | BUG_ON(!pse); |
2f36825b VP |
6732 | if (wakeup_preempt_entity(se, pse) == 1) { |
6733 | /* | |
6734 | * Bias pick_next to pick the sched entity that is | |
6735 | * triggering this preemption. | |
6736 | */ | |
6737 | if (!next_buddy_marked) | |
6738 | set_next_buddy(pse); | |
3a7e73a2 | 6739 | goto preempt; |
2f36825b | 6740 | } |
464b7527 | 6741 | |
3a7e73a2 | 6742 | return; |
a65ac745 | 6743 | |
3a7e73a2 | 6744 | preempt: |
8875125e | 6745 | resched_curr(rq); |
3a7e73a2 PZ |
6746 | /* |
6747 | * Only set the backward buddy when the current task is still | |
6748 | * on the rq. This can happen when a wakeup gets interleaved | |
6749 | * with schedule on the ->pre_schedule() or idle_balance() | |
6750 | * point, either of which can * drop the rq lock. | |
6751 | * | |
6752 | * Also, during early boot the idle thread is in the fair class, | |
6753 | * for obvious reasons its a bad idea to schedule back to it. | |
6754 | */ | |
6755 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
6756 | return; | |
6757 | ||
6758 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
6759 | set_last_buddy(se); | |
bf0f6f24 IM |
6760 | } |
6761 | ||
606dba2e | 6762 | static struct task_struct * |
d8ac8971 | 6763 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
6764 | { |
6765 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6766 | struct sched_entity *se; | |
678d5718 | 6767 | struct task_struct *p; |
37e117c0 | 6768 | int new_tasks; |
678d5718 | 6769 | |
6e83125c | 6770 | again: |
678d5718 | 6771 | if (!cfs_rq->nr_running) |
38033c37 | 6772 | goto idle; |
678d5718 | 6773 | |
9674f5ca | 6774 | #ifdef CONFIG_FAIR_GROUP_SCHED |
3f1d2a31 | 6775 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
6776 | goto simple; |
6777 | ||
6778 | /* | |
6779 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
6780 | * likely that a next task is from the same cgroup as the current. | |
6781 | * | |
6782 | * Therefore attempt to avoid putting and setting the entire cgroup | |
6783 | * hierarchy, only change the part that actually changes. | |
6784 | */ | |
6785 | ||
6786 | do { | |
6787 | struct sched_entity *curr = cfs_rq->curr; | |
6788 | ||
6789 | /* | |
6790 | * Since we got here without doing put_prev_entity() we also | |
6791 | * have to consider cfs_rq->curr. If it is still a runnable | |
6792 | * entity, update_curr() will update its vruntime, otherwise | |
6793 | * forget we've ever seen it. | |
6794 | */ | |
54d27365 BS |
6795 | if (curr) { |
6796 | if (curr->on_rq) | |
6797 | update_curr(cfs_rq); | |
6798 | else | |
6799 | curr = NULL; | |
678d5718 | 6800 | |
54d27365 BS |
6801 | /* |
6802 | * This call to check_cfs_rq_runtime() will do the | |
6803 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 6804 | * Therefore the nr_running test will indeed |
54d27365 BS |
6805 | * be correct. |
6806 | */ | |
9674f5ca VK |
6807 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
6808 | cfs_rq = &rq->cfs; | |
6809 | ||
6810 | if (!cfs_rq->nr_running) | |
6811 | goto idle; | |
6812 | ||
54d27365 | 6813 | goto simple; |
9674f5ca | 6814 | } |
54d27365 | 6815 | } |
678d5718 PZ |
6816 | |
6817 | se = pick_next_entity(cfs_rq, curr); | |
6818 | cfs_rq = group_cfs_rq(se); | |
6819 | } while (cfs_rq); | |
6820 | ||
6821 | p = task_of(se); | |
6822 | ||
6823 | /* | |
6824 | * Since we haven't yet done put_prev_entity and if the selected task | |
6825 | * is a different task than we started out with, try and touch the | |
6826 | * least amount of cfs_rqs. | |
6827 | */ | |
6828 | if (prev != p) { | |
6829 | struct sched_entity *pse = &prev->se; | |
6830 | ||
6831 | while (!(cfs_rq = is_same_group(se, pse))) { | |
6832 | int se_depth = se->depth; | |
6833 | int pse_depth = pse->depth; | |
6834 | ||
6835 | if (se_depth <= pse_depth) { | |
6836 | put_prev_entity(cfs_rq_of(pse), pse); | |
6837 | pse = parent_entity(pse); | |
6838 | } | |
6839 | if (se_depth >= pse_depth) { | |
6840 | set_next_entity(cfs_rq_of(se), se); | |
6841 | se = parent_entity(se); | |
6842 | } | |
6843 | } | |
6844 | ||
6845 | put_prev_entity(cfs_rq, pse); | |
6846 | set_next_entity(cfs_rq, se); | |
6847 | } | |
6848 | ||
93824900 | 6849 | goto done; |
678d5718 | 6850 | simple: |
678d5718 | 6851 | #endif |
bf0f6f24 | 6852 | |
3f1d2a31 | 6853 | put_prev_task(rq, prev); |
606dba2e | 6854 | |
bf0f6f24 | 6855 | do { |
678d5718 | 6856 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 6857 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
6858 | cfs_rq = group_cfs_rq(se); |
6859 | } while (cfs_rq); | |
6860 | ||
8f4d37ec | 6861 | p = task_of(se); |
678d5718 | 6862 | |
13a453c2 | 6863 | done: __maybe_unused; |
93824900 UR |
6864 | #ifdef CONFIG_SMP |
6865 | /* | |
6866 | * Move the next running task to the front of | |
6867 | * the list, so our cfs_tasks list becomes MRU | |
6868 | * one. | |
6869 | */ | |
6870 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
6871 | #endif | |
6872 | ||
b39e66ea MG |
6873 | if (hrtick_enabled(rq)) |
6874 | hrtick_start_fair(rq, p); | |
8f4d37ec | 6875 | |
3b1baa64 MR |
6876 | update_misfit_status(p, rq); |
6877 | ||
8f4d37ec | 6878 | return p; |
38033c37 PZ |
6879 | |
6880 | idle: | |
3b1baa64 | 6881 | update_misfit_status(NULL, rq); |
46f69fa3 MF |
6882 | new_tasks = idle_balance(rq, rf); |
6883 | ||
37e117c0 PZ |
6884 | /* |
6885 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
6886 | * possible for any higher priority task to appear. In that case we | |
6887 | * must re-start the pick_next_entity() loop. | |
6888 | */ | |
e4aa358b | 6889 | if (new_tasks < 0) |
37e117c0 PZ |
6890 | return RETRY_TASK; |
6891 | ||
e4aa358b | 6892 | if (new_tasks > 0) |
38033c37 | 6893 | goto again; |
38033c37 | 6894 | |
23127296 VG |
6895 | /* |
6896 | * rq is about to be idle, check if we need to update the | |
6897 | * lost_idle_time of clock_pelt | |
6898 | */ | |
6899 | update_idle_rq_clock_pelt(rq); | |
6900 | ||
38033c37 | 6901 | return NULL; |
bf0f6f24 IM |
6902 | } |
6903 | ||
6904 | /* | |
6905 | * Account for a descheduled task: | |
6906 | */ | |
31ee529c | 6907 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
6908 | { |
6909 | struct sched_entity *se = &prev->se; | |
6910 | struct cfs_rq *cfs_rq; | |
6911 | ||
6912 | for_each_sched_entity(se) { | |
6913 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 6914 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
6915 | } |
6916 | } | |
6917 | ||
ac53db59 RR |
6918 | /* |
6919 | * sched_yield() is very simple | |
6920 | * | |
6921 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
6922 | */ | |
6923 | static void yield_task_fair(struct rq *rq) | |
6924 | { | |
6925 | struct task_struct *curr = rq->curr; | |
6926 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
6927 | struct sched_entity *se = &curr->se; | |
6928 | ||
6929 | /* | |
6930 | * Are we the only task in the tree? | |
6931 | */ | |
6932 | if (unlikely(rq->nr_running == 1)) | |
6933 | return; | |
6934 | ||
6935 | clear_buddies(cfs_rq, se); | |
6936 | ||
6937 | if (curr->policy != SCHED_BATCH) { | |
6938 | update_rq_clock(rq); | |
6939 | /* | |
6940 | * Update run-time statistics of the 'current'. | |
6941 | */ | |
6942 | update_curr(cfs_rq); | |
916671c0 MG |
6943 | /* |
6944 | * Tell update_rq_clock() that we've just updated, | |
6945 | * so we don't do microscopic update in schedule() | |
6946 | * and double the fastpath cost. | |
6947 | */ | |
adcc8da8 | 6948 | rq_clock_skip_update(rq); |
ac53db59 RR |
6949 | } |
6950 | ||
6951 | set_skip_buddy(se); | |
6952 | } | |
6953 | ||
d95f4122 MG |
6954 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
6955 | { | |
6956 | struct sched_entity *se = &p->se; | |
6957 | ||
5238cdd3 PT |
6958 | /* throttled hierarchies are not runnable */ |
6959 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
6960 | return false; |
6961 | ||
6962 | /* Tell the scheduler that we'd really like pse to run next. */ | |
6963 | set_next_buddy(se); | |
6964 | ||
d95f4122 MG |
6965 | yield_task_fair(rq); |
6966 | ||
6967 | return true; | |
6968 | } | |
6969 | ||
681f3e68 | 6970 | #ifdef CONFIG_SMP |
bf0f6f24 | 6971 | /************************************************** |
e9c84cb8 PZ |
6972 | * Fair scheduling class load-balancing methods. |
6973 | * | |
6974 | * BASICS | |
6975 | * | |
6976 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 6977 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
6978 | * time to each task. This is expressed in the following equation: |
6979 | * | |
6980 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
6981 | * | |
97fb7a0a | 6982 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
6983 | * W_i,0 is defined as: |
6984 | * | |
6985 | * W_i,0 = \Sum_j w_i,j (2) | |
6986 | * | |
97fb7a0a | 6987 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 6988 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
6989 | * |
6990 | * The weight average is an exponential decay average of the instantaneous | |
6991 | * weight: | |
6992 | * | |
6993 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
6994 | * | |
97fb7a0a | 6995 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
6996 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
6997 | * can also include other factors [XXX]. | |
6998 | * | |
6999 | * To achieve this balance we define a measure of imbalance which follows | |
7000 | * directly from (1): | |
7001 | * | |
ced549fa | 7002 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
7003 | * |
7004 | * We them move tasks around to minimize the imbalance. In the continuous | |
7005 | * function space it is obvious this converges, in the discrete case we get | |
7006 | * a few fun cases generally called infeasible weight scenarios. | |
7007 | * | |
7008 | * [XXX expand on: | |
7009 | * - infeasible weights; | |
7010 | * - local vs global optima in the discrete case. ] | |
7011 | * | |
7012 | * | |
7013 | * SCHED DOMAINS | |
7014 | * | |
7015 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 7016 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 7017 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 7018 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 7019 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 7020 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
7021 | * the groups. |
7022 | * | |
7023 | * This yields: | |
7024 | * | |
7025 | * log_2 n 1 n | |
7026 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
7027 | * i = 0 2^i 2^i | |
7028 | * `- size of each group | |
97fb7a0a | 7029 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
7030 | * | `- freq |
7031 | * `- sum over all levels | |
7032 | * | |
7033 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
7034 | * this makes (5) the runtime complexity of the balancer. | |
7035 | * | |
7036 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 7037 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
7038 | * |
7039 | * The adjacency matrix of the resulting graph is given by: | |
7040 | * | |
97a7142f | 7041 | * log_2 n |
e9c84cb8 PZ |
7042 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
7043 | * k = 0 | |
7044 | * | |
7045 | * And you'll find that: | |
7046 | * | |
7047 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
7048 | * | |
97fb7a0a | 7049 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
7050 | * The task movement gives a factor of O(m), giving a convergence complexity |
7051 | * of: | |
7052 | * | |
7053 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
7054 | * | |
7055 | * | |
7056 | * WORK CONSERVING | |
7057 | * | |
7058 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 7059 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
7060 | * tree itself instead of relying on other CPUs to bring it work. |
7061 | * | |
7062 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
7063 | * time. | |
7064 | * | |
7065 | * [XXX more?] | |
7066 | * | |
7067 | * | |
7068 | * CGROUPS | |
7069 | * | |
7070 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
7071 | * | |
7072 | * s_k,i | |
7073 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
7074 | * S_k | |
7075 | * | |
7076 | * Where | |
7077 | * | |
7078 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
7079 | * | |
97fb7a0a | 7080 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
7081 | * |
7082 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
7083 | * property. | |
7084 | * | |
7085 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
7086 | * rewrite all of this once again.] | |
97a7142f | 7087 | */ |
bf0f6f24 | 7088 | |
ed387b78 HS |
7089 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
7090 | ||
0ec8aa00 PZ |
7091 | enum fbq_type { regular, remote, all }; |
7092 | ||
3b1baa64 MR |
7093 | enum group_type { |
7094 | group_other = 0, | |
7095 | group_misfit_task, | |
7096 | group_imbalanced, | |
7097 | group_overloaded, | |
7098 | }; | |
7099 | ||
ddcdf6e7 | 7100 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 7101 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
7102 | #define LBF_DST_PINNED 0x04 |
7103 | #define LBF_SOME_PINNED 0x08 | |
e022e0d3 | 7104 | #define LBF_NOHZ_STATS 0x10 |
f643ea22 | 7105 | #define LBF_NOHZ_AGAIN 0x20 |
ddcdf6e7 PZ |
7106 | |
7107 | struct lb_env { | |
7108 | struct sched_domain *sd; | |
7109 | ||
ddcdf6e7 | 7110 | struct rq *src_rq; |
85c1e7da | 7111 | int src_cpu; |
ddcdf6e7 PZ |
7112 | |
7113 | int dst_cpu; | |
7114 | struct rq *dst_rq; | |
7115 | ||
88b8dac0 SV |
7116 | struct cpumask *dst_grpmask; |
7117 | int new_dst_cpu; | |
ddcdf6e7 | 7118 | enum cpu_idle_type idle; |
bd939f45 | 7119 | long imbalance; |
b9403130 MW |
7120 | /* The set of CPUs under consideration for load-balancing */ |
7121 | struct cpumask *cpus; | |
7122 | ||
ddcdf6e7 | 7123 | unsigned int flags; |
367456c7 PZ |
7124 | |
7125 | unsigned int loop; | |
7126 | unsigned int loop_break; | |
7127 | unsigned int loop_max; | |
0ec8aa00 PZ |
7128 | |
7129 | enum fbq_type fbq_type; | |
cad68e55 | 7130 | enum group_type src_grp_type; |
163122b7 | 7131 | struct list_head tasks; |
ddcdf6e7 PZ |
7132 | }; |
7133 | ||
029632fb PZ |
7134 | /* |
7135 | * Is this task likely cache-hot: | |
7136 | */ | |
5d5e2b1b | 7137 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
7138 | { |
7139 | s64 delta; | |
7140 | ||
e5673f28 KT |
7141 | lockdep_assert_held(&env->src_rq->lock); |
7142 | ||
029632fb PZ |
7143 | if (p->sched_class != &fair_sched_class) |
7144 | return 0; | |
7145 | ||
1da1843f | 7146 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
7147 | return 0; |
7148 | ||
7149 | /* | |
7150 | * Buddy candidates are cache hot: | |
7151 | */ | |
5d5e2b1b | 7152 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
7153 | (&p->se == cfs_rq_of(&p->se)->next || |
7154 | &p->se == cfs_rq_of(&p->se)->last)) | |
7155 | return 1; | |
7156 | ||
7157 | if (sysctl_sched_migration_cost == -1) | |
7158 | return 1; | |
7159 | if (sysctl_sched_migration_cost == 0) | |
7160 | return 0; | |
7161 | ||
5d5e2b1b | 7162 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
7163 | |
7164 | return delta < (s64)sysctl_sched_migration_cost; | |
7165 | } | |
7166 | ||
3a7053b3 | 7167 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 7168 | /* |
2a1ed24c SD |
7169 | * Returns 1, if task migration degrades locality |
7170 | * Returns 0, if task migration improves locality i.e migration preferred. | |
7171 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 7172 | */ |
2a1ed24c | 7173 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 7174 | { |
b1ad065e | 7175 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
7176 | unsigned long src_weight, dst_weight; |
7177 | int src_nid, dst_nid, dist; | |
3a7053b3 | 7178 | |
2a595721 | 7179 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
7180 | return -1; |
7181 | ||
c3b9bc5b | 7182 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 7183 | return -1; |
7a0f3083 MG |
7184 | |
7185 | src_nid = cpu_to_node(env->src_cpu); | |
7186 | dst_nid = cpu_to_node(env->dst_cpu); | |
7187 | ||
83e1d2cd | 7188 | if (src_nid == dst_nid) |
2a1ed24c | 7189 | return -1; |
7a0f3083 | 7190 | |
2a1ed24c SD |
7191 | /* Migrating away from the preferred node is always bad. */ |
7192 | if (src_nid == p->numa_preferred_nid) { | |
7193 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
7194 | return 1; | |
7195 | else | |
7196 | return -1; | |
7197 | } | |
b1ad065e | 7198 | |
c1ceac62 RR |
7199 | /* Encourage migration to the preferred node. */ |
7200 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 7201 | return 0; |
b1ad065e | 7202 | |
739294fb | 7203 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 7204 | if (env->idle == CPU_IDLE) |
739294fb RR |
7205 | return -1; |
7206 | ||
f35678b6 | 7207 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 7208 | if (numa_group) { |
f35678b6 SD |
7209 | src_weight = group_weight(p, src_nid, dist); |
7210 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 7211 | } else { |
f35678b6 SD |
7212 | src_weight = task_weight(p, src_nid, dist); |
7213 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
7214 | } |
7215 | ||
f35678b6 | 7216 | return dst_weight < src_weight; |
7a0f3083 MG |
7217 | } |
7218 | ||
3a7053b3 | 7219 | #else |
2a1ed24c | 7220 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
7221 | struct lb_env *env) |
7222 | { | |
2a1ed24c | 7223 | return -1; |
7a0f3083 | 7224 | } |
3a7053b3 MG |
7225 | #endif |
7226 | ||
1e3c88bd PZ |
7227 | /* |
7228 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
7229 | */ | |
7230 | static | |
8e45cb54 | 7231 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 7232 | { |
2a1ed24c | 7233 | int tsk_cache_hot; |
e5673f28 KT |
7234 | |
7235 | lockdep_assert_held(&env->src_rq->lock); | |
7236 | ||
1e3c88bd PZ |
7237 | /* |
7238 | * We do not migrate tasks that are: | |
d3198084 | 7239 | * 1) throttled_lb_pair, or |
3bd37062 | 7240 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
7241 | * 3) running (obviously), or |
7242 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7243 | */ |
d3198084 JK |
7244 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7245 | return 0; | |
7246 | ||
3bd37062 | 7247 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 7248 | int cpu; |
88b8dac0 | 7249 | |
ae92882e | 7250 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 7251 | |
6263322c PZ |
7252 | env->flags |= LBF_SOME_PINNED; |
7253 | ||
88b8dac0 | 7254 | /* |
97fb7a0a | 7255 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
7256 | * our sched_group. We may want to revisit it if we couldn't |
7257 | * meet load balance goals by pulling other tasks on src_cpu. | |
7258 | * | |
65a4433a JH |
7259 | * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have |
7260 | * already computed one in current iteration. | |
88b8dac0 | 7261 | */ |
65a4433a | 7262 | if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
7263 | return 0; |
7264 | ||
97fb7a0a | 7265 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 7266 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 7267 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 7268 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7269 | env->new_dst_cpu = cpu; |
7270 | break; | |
7271 | } | |
88b8dac0 | 7272 | } |
e02e60c1 | 7273 | |
1e3c88bd PZ |
7274 | return 0; |
7275 | } | |
88b8dac0 SV |
7276 | |
7277 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 7278 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7279 | |
ddcdf6e7 | 7280 | if (task_running(env->src_rq, p)) { |
ae92882e | 7281 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
7282 | return 0; |
7283 | } | |
7284 | ||
7285 | /* | |
7286 | * Aggressive migration if: | |
3a7053b3 MG |
7287 | * 1) destination numa is preferred |
7288 | * 2) task is cache cold, or | |
7289 | * 3) too many balance attempts have failed. | |
1e3c88bd | 7290 | */ |
2a1ed24c SD |
7291 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7292 | if (tsk_cache_hot == -1) | |
7293 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7294 | |
2a1ed24c | 7295 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7296 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7297 | if (tsk_cache_hot == 1) { |
ae92882e JP |
7298 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
7299 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 7300 | } |
1e3c88bd PZ |
7301 | return 1; |
7302 | } | |
7303 | ||
ae92882e | 7304 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 7305 | return 0; |
1e3c88bd PZ |
7306 | } |
7307 | ||
897c395f | 7308 | /* |
163122b7 KT |
7309 | * detach_task() -- detach the task for the migration specified in env |
7310 | */ | |
7311 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7312 | { | |
7313 | lockdep_assert_held(&env->src_rq->lock); | |
7314 | ||
5704ac0a | 7315 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7316 | set_task_cpu(p, env->dst_cpu); |
7317 | } | |
7318 | ||
897c395f | 7319 | /* |
e5673f28 | 7320 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7321 | * part of active balancing operations within "domain". |
897c395f | 7322 | * |
e5673f28 | 7323 | * Returns a task if successful and NULL otherwise. |
897c395f | 7324 | */ |
e5673f28 | 7325 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7326 | { |
93824900 | 7327 | struct task_struct *p; |
897c395f | 7328 | |
e5673f28 KT |
7329 | lockdep_assert_held(&env->src_rq->lock); |
7330 | ||
93824900 UR |
7331 | list_for_each_entry_reverse(p, |
7332 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7333 | if (!can_migrate_task(p, env)) |
7334 | continue; | |
897c395f | 7335 | |
163122b7 | 7336 | detach_task(p, env); |
e5673f28 | 7337 | |
367456c7 | 7338 | /* |
e5673f28 | 7339 | * Right now, this is only the second place where |
163122b7 | 7340 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 7341 | * so we can safely collect stats here rather than |
163122b7 | 7342 | * inside detach_tasks(). |
367456c7 | 7343 | */ |
ae92882e | 7344 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 7345 | return p; |
897c395f | 7346 | } |
e5673f28 | 7347 | return NULL; |
897c395f PZ |
7348 | } |
7349 | ||
eb95308e PZ |
7350 | static const unsigned int sched_nr_migrate_break = 32; |
7351 | ||
5d6523eb | 7352 | /* |
a3df0679 | 7353 | * detach_tasks() -- tries to detach up to imbalance runnable load from |
163122b7 | 7354 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 7355 | * |
163122b7 | 7356 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 7357 | */ |
163122b7 | 7358 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 7359 | { |
5d6523eb PZ |
7360 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
7361 | struct task_struct *p; | |
367456c7 | 7362 | unsigned long load; |
163122b7 KT |
7363 | int detached = 0; |
7364 | ||
7365 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 7366 | |
bd939f45 | 7367 | if (env->imbalance <= 0) |
5d6523eb | 7368 | return 0; |
1e3c88bd | 7369 | |
5d6523eb | 7370 | while (!list_empty(tasks)) { |
985d3a4c YD |
7371 | /* |
7372 | * We don't want to steal all, otherwise we may be treated likewise, | |
7373 | * which could at worst lead to a livelock crash. | |
7374 | */ | |
7375 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
7376 | break; | |
7377 | ||
93824900 | 7378 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 7379 | |
367456c7 PZ |
7380 | env->loop++; |
7381 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 7382 | if (env->loop > env->loop_max) |
367456c7 | 7383 | break; |
5d6523eb PZ |
7384 | |
7385 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 7386 | if (env->loop > env->loop_break) { |
eb95308e | 7387 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 7388 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 7389 | break; |
a195f004 | 7390 | } |
1e3c88bd | 7391 | |
d3198084 | 7392 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
7393 | goto next; |
7394 | ||
7395 | load = task_h_load(p); | |
5d6523eb | 7396 | |
eb95308e | 7397 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
7398 | goto next; |
7399 | ||
bd939f45 | 7400 | if ((load / 2) > env->imbalance) |
367456c7 | 7401 | goto next; |
1e3c88bd | 7402 | |
163122b7 KT |
7403 | detach_task(p, env); |
7404 | list_add(&p->se.group_node, &env->tasks); | |
7405 | ||
7406 | detached++; | |
bd939f45 | 7407 | env->imbalance -= load; |
1e3c88bd PZ |
7408 | |
7409 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
7410 | /* |
7411 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 7412 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
7413 | * the critical section. |
7414 | */ | |
5d6523eb | 7415 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 7416 | break; |
1e3c88bd PZ |
7417 | #endif |
7418 | ||
ee00e66f PZ |
7419 | /* |
7420 | * We only want to steal up to the prescribed amount of | |
a3df0679 | 7421 | * runnable load. |
ee00e66f | 7422 | */ |
bd939f45 | 7423 | if (env->imbalance <= 0) |
ee00e66f | 7424 | break; |
367456c7 PZ |
7425 | |
7426 | continue; | |
7427 | next: | |
93824900 | 7428 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 7429 | } |
5d6523eb | 7430 | |
1e3c88bd | 7431 | /* |
163122b7 KT |
7432 | * Right now, this is one of only two places we collect this stat |
7433 | * so we can safely collect detach_one_task() stats here rather | |
7434 | * than inside detach_one_task(). | |
1e3c88bd | 7435 | */ |
ae92882e | 7436 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 7437 | |
163122b7 KT |
7438 | return detached; |
7439 | } | |
7440 | ||
7441 | /* | |
7442 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
7443 | */ | |
7444 | static void attach_task(struct rq *rq, struct task_struct *p) | |
7445 | { | |
7446 | lockdep_assert_held(&rq->lock); | |
7447 | ||
7448 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 7449 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
163122b7 KT |
7450 | check_preempt_curr(rq, p, 0); |
7451 | } | |
7452 | ||
7453 | /* | |
7454 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
7455 | * its new rq. | |
7456 | */ | |
7457 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
7458 | { | |
8a8c69c3 PZ |
7459 | struct rq_flags rf; |
7460 | ||
7461 | rq_lock(rq, &rf); | |
5704ac0a | 7462 | update_rq_clock(rq); |
163122b7 | 7463 | attach_task(rq, p); |
8a8c69c3 | 7464 | rq_unlock(rq, &rf); |
163122b7 KT |
7465 | } |
7466 | ||
7467 | /* | |
7468 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
7469 | * new rq. | |
7470 | */ | |
7471 | static void attach_tasks(struct lb_env *env) | |
7472 | { | |
7473 | struct list_head *tasks = &env->tasks; | |
7474 | struct task_struct *p; | |
8a8c69c3 | 7475 | struct rq_flags rf; |
163122b7 | 7476 | |
8a8c69c3 | 7477 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 7478 | update_rq_clock(env->dst_rq); |
163122b7 KT |
7479 | |
7480 | while (!list_empty(tasks)) { | |
7481 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
7482 | list_del_init(&p->se.group_node); | |
1e3c88bd | 7483 | |
163122b7 KT |
7484 | attach_task(env->dst_rq, p); |
7485 | } | |
7486 | ||
8a8c69c3 | 7487 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
7488 | } |
7489 | ||
b0c79224 | 7490 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
7491 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
7492 | { | |
7493 | if (cfs_rq->avg.load_avg) | |
7494 | return true; | |
7495 | ||
7496 | if (cfs_rq->avg.util_avg) | |
7497 | return true; | |
7498 | ||
7499 | return false; | |
7500 | } | |
7501 | ||
91c27493 | 7502 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
7503 | { |
7504 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
7505 | return true; | |
7506 | ||
3727e0e1 VG |
7507 | if (READ_ONCE(rq->avg_dl.util_avg)) |
7508 | return true; | |
7509 | ||
11d4afd4 | 7510 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
7511 | if (READ_ONCE(rq->avg_irq.util_avg)) |
7512 | return true; | |
7513 | #endif | |
7514 | ||
371bf427 VG |
7515 | return false; |
7516 | } | |
7517 | ||
b0c79224 VS |
7518 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
7519 | { | |
7520 | rq->last_blocked_load_update_tick = jiffies; | |
7521 | ||
7522 | if (!has_blocked) | |
7523 | rq->has_blocked_load = 0; | |
7524 | } | |
7525 | #else | |
7526 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
7527 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
7528 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} | |
7529 | #endif | |
7530 | ||
1936c53c VG |
7531 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7532 | ||
039ae8bc VG |
7533 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
7534 | { | |
7535 | if (cfs_rq->load.weight) | |
7536 | return false; | |
7537 | ||
7538 | if (cfs_rq->avg.load_sum) | |
7539 | return false; | |
7540 | ||
7541 | if (cfs_rq->avg.util_sum) | |
7542 | return false; | |
7543 | ||
7544 | if (cfs_rq->avg.runnable_load_sum) | |
7545 | return false; | |
7546 | ||
7547 | return true; | |
7548 | } | |
7549 | ||
48a16753 | 7550 | static void update_blocked_averages(int cpu) |
9e3081ca | 7551 | { |
9e3081ca | 7552 | struct rq *rq = cpu_rq(cpu); |
039ae8bc | 7553 | struct cfs_rq *cfs_rq, *pos; |
12b04875 | 7554 | const struct sched_class *curr_class; |
8a8c69c3 | 7555 | struct rq_flags rf; |
f643ea22 | 7556 | bool done = true; |
9e3081ca | 7557 | |
8a8c69c3 | 7558 | rq_lock_irqsave(rq, &rf); |
48a16753 | 7559 | update_rq_clock(rq); |
9d89c257 | 7560 | |
9763b67f PZ |
7561 | /* |
7562 | * Iterates the task_group tree in a bottom up fashion, see | |
7563 | * list_add_leaf_cfs_rq() for details. | |
7564 | */ | |
039ae8bc | 7565 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
7566 | struct sched_entity *se; |
7567 | ||
23127296 | 7568 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) |
9d89c257 | 7569 | update_tg_load_avg(cfs_rq, 0); |
4e516076 | 7570 | |
bc427898 VG |
7571 | /* Propagate pending load changes to the parent, if any: */ |
7572 | se = cfs_rq->tg->se[cpu]; | |
7573 | if (se && !skip_blocked_update(se)) | |
88c0616e | 7574 | update_load_avg(cfs_rq_of(se), se, 0); |
a9e7f654 | 7575 | |
039ae8bc VG |
7576 | /* |
7577 | * There can be a lot of idle CPU cgroups. Don't let fully | |
7578 | * decayed cfs_rqs linger on the list. | |
7579 | */ | |
7580 | if (cfs_rq_is_decayed(cfs_rq)) | |
7581 | list_del_leaf_cfs_rq(cfs_rq); | |
7582 | ||
1936c53c VG |
7583 | /* Don't need periodic decay once load/util_avg are null */ |
7584 | if (cfs_rq_has_blocked(cfs_rq)) | |
f643ea22 | 7585 | done = false; |
9d89c257 | 7586 | } |
12b04875 VG |
7587 | |
7588 | curr_class = rq->curr->sched_class; | |
23127296 VG |
7589 | update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class); |
7590 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class); | |
91c27493 | 7591 | update_irq_load_avg(rq, 0); |
371bf427 | 7592 | /* Don't need periodic decay once load/util_avg are null */ |
91c27493 | 7593 | if (others_have_blocked(rq)) |
371bf427 | 7594 | done = false; |
e022e0d3 | 7595 | |
b0c79224 | 7596 | update_blocked_load_status(rq, !done); |
8a8c69c3 | 7597 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7598 | } |
7599 | ||
9763b67f | 7600 | /* |
68520796 | 7601 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
7602 | * This needs to be done in a top-down fashion because the load of a child |
7603 | * group is a fraction of its parents load. | |
7604 | */ | |
68520796 | 7605 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 7606 | { |
68520796 VD |
7607 | struct rq *rq = rq_of(cfs_rq); |
7608 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 7609 | unsigned long now = jiffies; |
68520796 | 7610 | unsigned long load; |
a35b6466 | 7611 | |
68520796 | 7612 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
7613 | return; |
7614 | ||
0e9f0245 | 7615 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
7616 | for_each_sched_entity(se) { |
7617 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 7618 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
7619 | if (cfs_rq->last_h_load_update == now) |
7620 | break; | |
7621 | } | |
a35b6466 | 7622 | |
68520796 | 7623 | if (!se) { |
7ea241af | 7624 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
7625 | cfs_rq->last_h_load_update = now; |
7626 | } | |
7627 | ||
0e9f0245 | 7628 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 7629 | load = cfs_rq->h_load; |
7ea241af YD |
7630 | load = div64_ul(load * se->avg.load_avg, |
7631 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
7632 | cfs_rq = group_cfs_rq(se); |
7633 | cfs_rq->h_load = load; | |
7634 | cfs_rq->last_h_load_update = now; | |
7635 | } | |
9763b67f PZ |
7636 | } |
7637 | ||
367456c7 | 7638 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 7639 | { |
367456c7 | 7640 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 7641 | |
68520796 | 7642 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 7643 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 7644 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
7645 | } |
7646 | #else | |
48a16753 | 7647 | static inline void update_blocked_averages(int cpu) |
9e3081ca | 7648 | { |
6c1d47c0 VG |
7649 | struct rq *rq = cpu_rq(cpu); |
7650 | struct cfs_rq *cfs_rq = &rq->cfs; | |
12b04875 | 7651 | const struct sched_class *curr_class; |
8a8c69c3 | 7652 | struct rq_flags rf; |
6c1d47c0 | 7653 | |
8a8c69c3 | 7654 | rq_lock_irqsave(rq, &rf); |
6c1d47c0 | 7655 | update_rq_clock(rq); |
23127296 | 7656 | update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
12b04875 VG |
7657 | |
7658 | curr_class = rq->curr->sched_class; | |
23127296 VG |
7659 | update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class); |
7660 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class); | |
91c27493 | 7661 | update_irq_load_avg(rq, 0); |
b0c79224 | 7662 | update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq)); |
8a8c69c3 | 7663 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7664 | } |
7665 | ||
367456c7 | 7666 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 7667 | { |
9d89c257 | 7668 | return p->se.avg.load_avg; |
1e3c88bd | 7669 | } |
230059de | 7670 | #endif |
1e3c88bd | 7671 | |
1e3c88bd | 7672 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 7673 | |
1e3c88bd PZ |
7674 | /* |
7675 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
7676 | */ | |
7677 | struct sg_lb_stats { | |
7678 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
7679 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
56cf515b | 7680 | unsigned long load_per_task; |
63b2ca30 | 7681 | unsigned long group_capacity; |
9e91d61d | 7682 | unsigned long group_util; /* Total utilization of the group */ |
147c5fc2 | 7683 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
147c5fc2 PZ |
7684 | unsigned int idle_cpus; |
7685 | unsigned int group_weight; | |
caeb178c | 7686 | enum group_type group_type; |
ea67821b | 7687 | int group_no_capacity; |
3b1baa64 | 7688 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
7689 | #ifdef CONFIG_NUMA_BALANCING |
7690 | unsigned int nr_numa_running; | |
7691 | unsigned int nr_preferred_running; | |
7692 | #endif | |
1e3c88bd PZ |
7693 | }; |
7694 | ||
56cf515b JK |
7695 | /* |
7696 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
7697 | * during load balancing. | |
7698 | */ | |
7699 | struct sd_lb_stats { | |
7700 | struct sched_group *busiest; /* Busiest group in this sd */ | |
7701 | struct sched_group *local; /* Local group in this sd */ | |
90001d67 | 7702 | unsigned long total_running; |
56cf515b | 7703 | unsigned long total_load; /* Total load of all groups in sd */ |
63b2ca30 | 7704 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
7705 | unsigned long avg_load; /* Average load across all groups in sd */ |
7706 | ||
56cf515b | 7707 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 7708 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
7709 | }; |
7710 | ||
147c5fc2 PZ |
7711 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
7712 | { | |
7713 | /* | |
7714 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
7715 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
7716 | * We must however clear busiest_stat::avg_load because | |
7717 | * update_sd_pick_busiest() reads this before assignment. | |
7718 | */ | |
7719 | *sds = (struct sd_lb_stats){ | |
7720 | .busiest = NULL, | |
7721 | .local = NULL, | |
90001d67 | 7722 | .total_running = 0UL, |
147c5fc2 | 7723 | .total_load = 0UL, |
63b2ca30 | 7724 | .total_capacity = 0UL, |
147c5fc2 PZ |
7725 | .busiest_stat = { |
7726 | .avg_load = 0UL, | |
caeb178c RR |
7727 | .sum_nr_running = 0, |
7728 | .group_type = group_other, | |
147c5fc2 PZ |
7729 | }, |
7730 | }; | |
7731 | } | |
7732 | ||
287cdaac | 7733 | static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
7734 | { |
7735 | struct rq *rq = cpu_rq(cpu); | |
8ec59c0f | 7736 | unsigned long max = arch_scale_cpu_capacity(cpu); |
523e979d | 7737 | unsigned long used, free; |
523e979d | 7738 | unsigned long irq; |
b654f7de | 7739 | |
2e62c474 | 7740 | irq = cpu_util_irq(rq); |
cadefd3d | 7741 | |
523e979d VG |
7742 | if (unlikely(irq >= max)) |
7743 | return 1; | |
aa483808 | 7744 | |
523e979d VG |
7745 | used = READ_ONCE(rq->avg_rt.util_avg); |
7746 | used += READ_ONCE(rq->avg_dl.util_avg); | |
1e3c88bd | 7747 | |
523e979d VG |
7748 | if (unlikely(used >= max)) |
7749 | return 1; | |
1e3c88bd | 7750 | |
523e979d | 7751 | free = max - used; |
2e62c474 VG |
7752 | |
7753 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
7754 | } |
7755 | ||
ced549fa | 7756 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 7757 | { |
287cdaac | 7758 | unsigned long capacity = scale_rt_capacity(sd, cpu); |
1e3c88bd PZ |
7759 | struct sched_group *sdg = sd->groups; |
7760 | ||
8ec59c0f | 7761 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); |
1e3c88bd | 7762 | |
ced549fa NP |
7763 | if (!capacity) |
7764 | capacity = 1; | |
1e3c88bd | 7765 | |
ced549fa NP |
7766 | cpu_rq(cpu)->cpu_capacity = capacity; |
7767 | sdg->sgc->capacity = capacity; | |
bf475ce0 | 7768 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 7769 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
7770 | } |
7771 | ||
63b2ca30 | 7772 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
7773 | { |
7774 | struct sched_domain *child = sd->child; | |
7775 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 7776 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
7777 | unsigned long interval; |
7778 | ||
7779 | interval = msecs_to_jiffies(sd->balance_interval); | |
7780 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 7781 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
7782 | |
7783 | if (!child) { | |
ced549fa | 7784 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
7785 | return; |
7786 | } | |
7787 | ||
dc7ff76e | 7788 | capacity = 0; |
bf475ce0 | 7789 | min_capacity = ULONG_MAX; |
e3d6d0cb | 7790 | max_capacity = 0; |
1e3c88bd | 7791 | |
74a5ce20 PZ |
7792 | if (child->flags & SD_OVERLAP) { |
7793 | /* | |
7794 | * SD_OVERLAP domains cannot assume that child groups | |
7795 | * span the current group. | |
7796 | */ | |
7797 | ||
ae4df9d6 | 7798 | for_each_cpu(cpu, sched_group_span(sdg)) { |
63b2ca30 | 7799 | struct sched_group_capacity *sgc; |
9abf24d4 | 7800 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 7801 | |
9abf24d4 | 7802 | /* |
63b2ca30 | 7803 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
7804 | * gets here before we've attached the domains to the |
7805 | * runqueues. | |
7806 | * | |
ced549fa NP |
7807 | * Use capacity_of(), which is set irrespective of domains |
7808 | * in update_cpu_capacity(). | |
9abf24d4 | 7809 | * |
dc7ff76e | 7810 | * This avoids capacity from being 0 and |
9abf24d4 | 7811 | * causing divide-by-zero issues on boot. |
9abf24d4 SD |
7812 | */ |
7813 | if (unlikely(!rq->sd)) { | |
ced549fa | 7814 | capacity += capacity_of(cpu); |
bf475ce0 MR |
7815 | } else { |
7816 | sgc = rq->sd->groups->sgc; | |
7817 | capacity += sgc->capacity; | |
9abf24d4 | 7818 | } |
863bffc8 | 7819 | |
bf475ce0 | 7820 | min_capacity = min(capacity, min_capacity); |
e3d6d0cb | 7821 | max_capacity = max(capacity, max_capacity); |
863bffc8 | 7822 | } |
74a5ce20 PZ |
7823 | } else { |
7824 | /* | |
7825 | * !SD_OVERLAP domains can assume that child groups | |
7826 | * span the current group. | |
97a7142f | 7827 | */ |
74a5ce20 PZ |
7828 | |
7829 | group = child->groups; | |
7830 | do { | |
bf475ce0 MR |
7831 | struct sched_group_capacity *sgc = group->sgc; |
7832 | ||
7833 | capacity += sgc->capacity; | |
7834 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 7835 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
7836 | group = group->next; |
7837 | } while (group != child->groups); | |
7838 | } | |
1e3c88bd | 7839 | |
63b2ca30 | 7840 | sdg->sgc->capacity = capacity; |
bf475ce0 | 7841 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 7842 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
7843 | } |
7844 | ||
9d5efe05 | 7845 | /* |
ea67821b VG |
7846 | * Check whether the capacity of the rq has been noticeably reduced by side |
7847 | * activity. The imbalance_pct is used for the threshold. | |
7848 | * Return true is the capacity is reduced | |
9d5efe05 SV |
7849 | */ |
7850 | static inline int | |
ea67821b | 7851 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 7852 | { |
ea67821b VG |
7853 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
7854 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
7855 | } |
7856 | ||
a0fe2cf0 VS |
7857 | /* |
7858 | * Check whether a rq has a misfit task and if it looks like we can actually | |
7859 | * help that task: we can migrate the task to a CPU of higher capacity, or | |
7860 | * the task's current CPU is heavily pressured. | |
7861 | */ | |
7862 | static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) | |
7863 | { | |
7864 | return rq->misfit_task_load && | |
7865 | (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || | |
7866 | check_cpu_capacity(rq, sd)); | |
7867 | } | |
7868 | ||
30ce5dab PZ |
7869 | /* |
7870 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 7871 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 7872 | * |
97fb7a0a IM |
7873 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
7874 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
7875 | * Something like: |
7876 | * | |
2b4d5b25 IM |
7877 | * { 0 1 2 3 } { 4 5 6 7 } |
7878 | * * * * * | |
30ce5dab PZ |
7879 | * |
7880 | * If we were to balance group-wise we'd place two tasks in the first group and | |
7881 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 7882 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
7883 | * |
7884 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
7885 | * by noticing the lower domain failed to reach balance and had difficulty |
7886 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
7887 | * |
7888 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 7889 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 7890 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
7891 | * to create an effective group imbalance. |
7892 | * | |
7893 | * This is a somewhat tricky proposition since the next run might not find the | |
7894 | * group imbalance and decide the groups need to be balanced again. A most | |
7895 | * subtle and fragile situation. | |
7896 | */ | |
7897 | ||
6263322c | 7898 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 7899 | { |
63b2ca30 | 7900 | return group->sgc->imbalance; |
30ce5dab PZ |
7901 | } |
7902 | ||
b37d9316 | 7903 | /* |
ea67821b VG |
7904 | * group_has_capacity returns true if the group has spare capacity that could |
7905 | * be used by some tasks. | |
7906 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
7907 | * smaller than the number of CPUs or if the utilization is lower than the |
7908 | * available capacity for CFS tasks. | |
ea67821b VG |
7909 | * For the latter, we use a threshold to stabilize the state, to take into |
7910 | * account the variance of the tasks' load and to return true if the available | |
7911 | * capacity in meaningful for the load balancer. | |
7912 | * As an example, an available capacity of 1% can appear but it doesn't make | |
7913 | * any benefit for the load balance. | |
b37d9316 | 7914 | */ |
ea67821b VG |
7915 | static inline bool |
7916 | group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) | |
b37d9316 | 7917 | { |
ea67821b VG |
7918 | if (sgs->sum_nr_running < sgs->group_weight) |
7919 | return true; | |
c61037e9 | 7920 | |
ea67821b | 7921 | if ((sgs->group_capacity * 100) > |
9e91d61d | 7922 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 7923 | return true; |
b37d9316 | 7924 | |
ea67821b VG |
7925 | return false; |
7926 | } | |
7927 | ||
7928 | /* | |
7929 | * group_is_overloaded returns true if the group has more tasks than it can | |
7930 | * handle. | |
7931 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
7932 | * with the exact right number of tasks, has no more spare capacity but is not | |
7933 | * overloaded so both group_has_capacity and group_is_overloaded return | |
7934 | * false. | |
7935 | */ | |
7936 | static inline bool | |
7937 | group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) | |
7938 | { | |
7939 | if (sgs->sum_nr_running <= sgs->group_weight) | |
7940 | return false; | |
b37d9316 | 7941 | |
ea67821b | 7942 | if ((sgs->group_capacity * 100) < |
9e91d61d | 7943 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 7944 | return true; |
b37d9316 | 7945 | |
ea67821b | 7946 | return false; |
b37d9316 PZ |
7947 | } |
7948 | ||
9e0994c0 | 7949 | /* |
e3d6d0cb | 7950 | * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller |
9e0994c0 MR |
7951 | * per-CPU capacity than sched_group ref. |
7952 | */ | |
7953 | static inline bool | |
e3d6d0cb | 7954 | group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref) |
9e0994c0 | 7955 | { |
60e17f5c | 7956 | return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity); |
9e0994c0 MR |
7957 | } |
7958 | ||
e3d6d0cb MR |
7959 | /* |
7960 | * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller | |
7961 | * per-CPU capacity_orig than sched_group ref. | |
7962 | */ | |
7963 | static inline bool | |
7964 | group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref) | |
7965 | { | |
60e17f5c | 7966 | return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity); |
e3d6d0cb MR |
7967 | } |
7968 | ||
79a89f92 LY |
7969 | static inline enum |
7970 | group_type group_classify(struct sched_group *group, | |
7971 | struct sg_lb_stats *sgs) | |
caeb178c | 7972 | { |
ea67821b | 7973 | if (sgs->group_no_capacity) |
caeb178c RR |
7974 | return group_overloaded; |
7975 | ||
7976 | if (sg_imbalanced(group)) | |
7977 | return group_imbalanced; | |
7978 | ||
3b1baa64 MR |
7979 | if (sgs->group_misfit_task_load) |
7980 | return group_misfit_task; | |
7981 | ||
caeb178c RR |
7982 | return group_other; |
7983 | } | |
7984 | ||
63928384 | 7985 | static bool update_nohz_stats(struct rq *rq, bool force) |
e022e0d3 PZ |
7986 | { |
7987 | #ifdef CONFIG_NO_HZ_COMMON | |
7988 | unsigned int cpu = rq->cpu; | |
7989 | ||
f643ea22 VG |
7990 | if (!rq->has_blocked_load) |
7991 | return false; | |
7992 | ||
e022e0d3 | 7993 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
f643ea22 | 7994 | return false; |
e022e0d3 | 7995 | |
63928384 | 7996 | if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick)) |
f643ea22 | 7997 | return true; |
e022e0d3 PZ |
7998 | |
7999 | update_blocked_averages(cpu); | |
f643ea22 VG |
8000 | |
8001 | return rq->has_blocked_load; | |
8002 | #else | |
8003 | return false; | |
e022e0d3 PZ |
8004 | #endif |
8005 | } | |
8006 | ||
1e3c88bd PZ |
8007 | /** |
8008 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 8009 | * @env: The load balancing environment. |
1e3c88bd | 8010 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 8011 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 8012 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 8013 | */ |
bd939f45 | 8014 | static inline void update_sg_lb_stats(struct lb_env *env, |
630246a0 QP |
8015 | struct sched_group *group, |
8016 | struct sg_lb_stats *sgs, | |
8017 | int *sg_status) | |
1e3c88bd | 8018 | { |
a426f99c | 8019 | int i, nr_running; |
1e3c88bd | 8020 | |
b72ff13c PZ |
8021 | memset(sgs, 0, sizeof(*sgs)); |
8022 | ||
ae4df9d6 | 8023 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd PZ |
8024 | struct rq *rq = cpu_rq(i); |
8025 | ||
63928384 | 8026 | if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false)) |
f643ea22 | 8027 | env->flags |= LBF_NOHZ_AGAIN; |
e022e0d3 | 8028 | |
a3df0679 | 8029 | sgs->group_load += cpu_runnable_load(rq); |
9e91d61d | 8030 | sgs->group_util += cpu_util(i); |
65fdac08 | 8031 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 8032 | |
a426f99c WL |
8033 | nr_running = rq->nr_running; |
8034 | if (nr_running > 1) | |
630246a0 | 8035 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 8036 | |
2802bf3c MR |
8037 | if (cpu_overutilized(i)) |
8038 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 8039 | |
0ec8aa00 PZ |
8040 | #ifdef CONFIG_NUMA_BALANCING |
8041 | sgs->nr_numa_running += rq->nr_numa_running; | |
8042 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
8043 | #endif | |
a426f99c WL |
8044 | /* |
8045 | * No need to call idle_cpu() if nr_running is not 0 | |
8046 | */ | |
8047 | if (!nr_running && idle_cpu(i)) | |
aae6d3dd | 8048 | sgs->idle_cpus++; |
3b1baa64 MR |
8049 | |
8050 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
757ffdd7 | 8051 | sgs->group_misfit_task_load < rq->misfit_task_load) { |
3b1baa64 | 8052 | sgs->group_misfit_task_load = rq->misfit_task_load; |
630246a0 | 8053 | *sg_status |= SG_OVERLOAD; |
757ffdd7 | 8054 | } |
1e3c88bd PZ |
8055 | } |
8056 | ||
63b2ca30 NP |
8057 | /* Adjust by relative CPU capacity of the group */ |
8058 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 8059 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 8060 | |
dd5feea1 | 8061 | if (sgs->sum_nr_running) |
af75d1a9 | 8062 | sgs->load_per_task = sgs->group_load / sgs->sum_nr_running; |
1e3c88bd | 8063 | |
aae6d3dd | 8064 | sgs->group_weight = group->group_weight; |
b37d9316 | 8065 | |
ea67821b | 8066 | sgs->group_no_capacity = group_is_overloaded(env, sgs); |
79a89f92 | 8067 | sgs->group_type = group_classify(group, sgs); |
1e3c88bd PZ |
8068 | } |
8069 | ||
532cb4c4 MN |
8070 | /** |
8071 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 8072 | * @env: The load balancing environment. |
532cb4c4 MN |
8073 | * @sds: sched_domain statistics |
8074 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 8075 | * @sgs: sched_group statistics |
532cb4c4 MN |
8076 | * |
8077 | * Determine if @sg is a busier group than the previously selected | |
8078 | * busiest group. | |
e69f6186 YB |
8079 | * |
8080 | * Return: %true if @sg is a busier group than the previously selected | |
8081 | * busiest group. %false otherwise. | |
532cb4c4 | 8082 | */ |
bd939f45 | 8083 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
8084 | struct sd_lb_stats *sds, |
8085 | struct sched_group *sg, | |
bd939f45 | 8086 | struct sg_lb_stats *sgs) |
532cb4c4 | 8087 | { |
caeb178c | 8088 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 8089 | |
cad68e55 MR |
8090 | /* |
8091 | * Don't try to pull misfit tasks we can't help. | |
8092 | * We can use max_capacity here as reduction in capacity on some | |
8093 | * CPUs in the group should either be possible to resolve | |
8094 | * internally or be covered by avg_load imbalance (eventually). | |
8095 | */ | |
8096 | if (sgs->group_type == group_misfit_task && | |
8097 | (!group_smaller_max_cpu_capacity(sg, sds->local) || | |
8098 | !group_has_capacity(env, &sds->local_stat))) | |
8099 | return false; | |
8100 | ||
caeb178c | 8101 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
8102 | return true; |
8103 | ||
caeb178c RR |
8104 | if (sgs->group_type < busiest->group_type) |
8105 | return false; | |
8106 | ||
8107 | if (sgs->avg_load <= busiest->avg_load) | |
8108 | return false; | |
8109 | ||
9e0994c0 MR |
8110 | if (!(env->sd->flags & SD_ASYM_CPUCAPACITY)) |
8111 | goto asym_packing; | |
8112 | ||
8113 | /* | |
8114 | * Candidate sg has no more than one task per CPU and | |
8115 | * has higher per-CPU capacity. Migrating tasks to less | |
8116 | * capable CPUs may harm throughput. Maximize throughput, | |
8117 | * power/energy consequences are not considered. | |
8118 | */ | |
8119 | if (sgs->sum_nr_running <= sgs->group_weight && | |
e3d6d0cb | 8120 | group_smaller_min_cpu_capacity(sds->local, sg)) |
9e0994c0 MR |
8121 | return false; |
8122 | ||
cad68e55 MR |
8123 | /* |
8124 | * If we have more than one misfit sg go with the biggest misfit. | |
8125 | */ | |
8126 | if (sgs->group_type == group_misfit_task && | |
8127 | sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
9e0994c0 MR |
8128 | return false; |
8129 | ||
8130 | asym_packing: | |
caeb178c RR |
8131 | /* This is the busiest node in its class. */ |
8132 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
8133 | return true; |
8134 | ||
97fb7a0a | 8135 | /* No ASYM_PACKING if target CPU is already busy */ |
1f621e02 SD |
8136 | if (env->idle == CPU_NOT_IDLE) |
8137 | return true; | |
532cb4c4 | 8138 | /* |
afe06efd TC |
8139 | * ASYM_PACKING needs to move all the work to the highest |
8140 | * prority CPUs in the group, therefore mark all groups | |
8141 | * of lower priority than ourself as busy. | |
532cb4c4 | 8142 | */ |
afe06efd TC |
8143 | if (sgs->sum_nr_running && |
8144 | sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) { | |
532cb4c4 MN |
8145 | if (!sds->busiest) |
8146 | return true; | |
8147 | ||
97fb7a0a | 8148 | /* Prefer to move from lowest priority CPU's work */ |
afe06efd TC |
8149 | if (sched_asym_prefer(sds->busiest->asym_prefer_cpu, |
8150 | sg->asym_prefer_cpu)) | |
532cb4c4 MN |
8151 | return true; |
8152 | } | |
8153 | ||
8154 | return false; | |
8155 | } | |
8156 | ||
0ec8aa00 PZ |
8157 | #ifdef CONFIG_NUMA_BALANCING |
8158 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8159 | { | |
8160 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
8161 | return regular; | |
8162 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
8163 | return remote; | |
8164 | return all; | |
8165 | } | |
8166 | ||
8167 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8168 | { | |
8169 | if (rq->nr_running > rq->nr_numa_running) | |
8170 | return regular; | |
8171 | if (rq->nr_running > rq->nr_preferred_running) | |
8172 | return remote; | |
8173 | return all; | |
8174 | } | |
8175 | #else | |
8176 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8177 | { | |
8178 | return all; | |
8179 | } | |
8180 | ||
8181 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8182 | { | |
8183 | return regular; | |
8184 | } | |
8185 | #endif /* CONFIG_NUMA_BALANCING */ | |
8186 | ||
1e3c88bd | 8187 | /** |
461819ac | 8188 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 8189 | * @env: The load balancing environment. |
1e3c88bd PZ |
8190 | * @sds: variable to hold the statistics for this sched_domain. |
8191 | */ | |
0ec8aa00 | 8192 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8193 | { |
bd939f45 PZ |
8194 | struct sched_domain *child = env->sd->child; |
8195 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 8196 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 8197 | struct sg_lb_stats tmp_sgs; |
dbbad719 | 8198 | bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING; |
630246a0 | 8199 | int sg_status = 0; |
1e3c88bd | 8200 | |
e022e0d3 | 8201 | #ifdef CONFIG_NO_HZ_COMMON |
f643ea22 | 8202 | if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked)) |
e022e0d3 | 8203 | env->flags |= LBF_NOHZ_STATS; |
e022e0d3 PZ |
8204 | #endif |
8205 | ||
1e3c88bd | 8206 | do { |
56cf515b | 8207 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
8208 | int local_group; |
8209 | ||
ae4df9d6 | 8210 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
8211 | if (local_group) { |
8212 | sds->local = sg; | |
05b40e05 | 8213 | sgs = local; |
b72ff13c PZ |
8214 | |
8215 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
8216 | time_after_eq(jiffies, sg->sgc->next_update)) |
8217 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 8218 | } |
1e3c88bd | 8219 | |
630246a0 | 8220 | update_sg_lb_stats(env, sg, sgs, &sg_status); |
1e3c88bd | 8221 | |
b72ff13c PZ |
8222 | if (local_group) |
8223 | goto next_group; | |
8224 | ||
1e3c88bd PZ |
8225 | /* |
8226 | * In case the child domain prefers tasks go to siblings | |
ea67821b | 8227 | * first, lower the sg capacity so that we'll try |
75dd321d NR |
8228 | * and move all the excess tasks away. We lower the capacity |
8229 | * of a group only if the local group has the capacity to fit | |
ea67821b VG |
8230 | * these excess tasks. The extra check prevents the case where |
8231 | * you always pull from the heaviest group when it is already | |
8232 | * under-utilized (possible with a large weight task outweighs | |
8233 | * the tasks on the system). | |
1e3c88bd | 8234 | */ |
b72ff13c | 8235 | if (prefer_sibling && sds->local && |
05b40e05 SD |
8236 | group_has_capacity(env, local) && |
8237 | (sgs->sum_nr_running > local->sum_nr_running + 1)) { | |
ea67821b | 8238 | sgs->group_no_capacity = 1; |
79a89f92 | 8239 | sgs->group_type = group_classify(sg, sgs); |
cb0b9f24 | 8240 | } |
1e3c88bd | 8241 | |
b72ff13c | 8242 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 8243 | sds->busiest = sg; |
56cf515b | 8244 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
8245 | } |
8246 | ||
b72ff13c PZ |
8247 | next_group: |
8248 | /* Now, start updating sd_lb_stats */ | |
90001d67 | 8249 | sds->total_running += sgs->sum_nr_running; |
b72ff13c | 8250 | sds->total_load += sgs->group_load; |
63b2ca30 | 8251 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 8252 | |
532cb4c4 | 8253 | sg = sg->next; |
bd939f45 | 8254 | } while (sg != env->sd->groups); |
0ec8aa00 | 8255 | |
f643ea22 VG |
8256 | #ifdef CONFIG_NO_HZ_COMMON |
8257 | if ((env->flags & LBF_NOHZ_AGAIN) && | |
8258 | cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) { | |
8259 | ||
8260 | WRITE_ONCE(nohz.next_blocked, | |
8261 | jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
8262 | } | |
8263 | #endif | |
8264 | ||
0ec8aa00 PZ |
8265 | if (env->sd->flags & SD_NUMA) |
8266 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
8267 | |
8268 | if (!env->sd->parent) { | |
2802bf3c MR |
8269 | struct root_domain *rd = env->dst_rq->rd; |
8270 | ||
4486edd1 | 8271 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
8272 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
8273 | ||
8274 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
8275 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
f9f240f9 | 8276 | trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); |
2802bf3c | 8277 | } else if (sg_status & SG_OVERUTILIZED) { |
f9f240f9 QY |
8278 | struct root_domain *rd = env->dst_rq->rd; |
8279 | ||
8280 | WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); | |
8281 | trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); | |
4486edd1 | 8282 | } |
532cb4c4 MN |
8283 | } |
8284 | ||
532cb4c4 MN |
8285 | /** |
8286 | * check_asym_packing - Check to see if the group is packed into the | |
0ba42a59 | 8287 | * sched domain. |
532cb4c4 MN |
8288 | * |
8289 | * This is primarily intended to used at the sibling level. Some | |
8290 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
8291 | * case of POWER7, it can move to lower SMT modes only when higher | |
8292 | * threads are idle. When in lower SMT modes, the threads will | |
8293 | * perform better since they share less core resources. Hence when we | |
8294 | * have idle threads, we want them to be the higher ones. | |
8295 | * | |
8296 | * This packing function is run on idle threads. It checks to see if | |
8297 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
8298 | * CPU number than the packing function is being run on. Here we are | |
8299 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
8300 | * number. | |
8301 | * | |
e69f6186 | 8302 | * Return: 1 when packing is required and a task should be moved to |
46123355 | 8303 | * this CPU. The amount of the imbalance is returned in env->imbalance. |
b6b12294 | 8304 | * |
cd96891d | 8305 | * @env: The load balancing environment. |
532cb4c4 | 8306 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 8307 | */ |
bd939f45 | 8308 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
8309 | { |
8310 | int busiest_cpu; | |
8311 | ||
bd939f45 | 8312 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
8313 | return 0; |
8314 | ||
1f621e02 SD |
8315 | if (env->idle == CPU_NOT_IDLE) |
8316 | return 0; | |
8317 | ||
532cb4c4 MN |
8318 | if (!sds->busiest) |
8319 | return 0; | |
8320 | ||
afe06efd TC |
8321 | busiest_cpu = sds->busiest->asym_prefer_cpu; |
8322 | if (sched_asym_prefer(busiest_cpu, env->dst_cpu)) | |
532cb4c4 MN |
8323 | return 0; |
8324 | ||
4ad4e481 | 8325 | env->imbalance = sds->busiest_stat.group_load; |
bd939f45 | 8326 | |
532cb4c4 | 8327 | return 1; |
1e3c88bd PZ |
8328 | } |
8329 | ||
8330 | /** | |
8331 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
8332 | * amongst the groups of a sched_domain, during | |
8333 | * load balancing. | |
cd96891d | 8334 | * @env: The load balancing environment. |
1e3c88bd | 8335 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 8336 | */ |
bd939f45 PZ |
8337 | static inline |
8338 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 8339 | { |
63b2ca30 | 8340 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 8341 | unsigned int imbn = 2; |
dd5feea1 | 8342 | unsigned long scaled_busy_load_per_task; |
56cf515b | 8343 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 8344 | |
56cf515b JK |
8345 | local = &sds->local_stat; |
8346 | busiest = &sds->busiest_stat; | |
1e3c88bd | 8347 | |
56cf515b JK |
8348 | if (!local->sum_nr_running) |
8349 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
8350 | else if (busiest->load_per_task > local->load_per_task) | |
8351 | imbn = 1; | |
dd5feea1 | 8352 | |
56cf515b | 8353 | scaled_busy_load_per_task = |
ca8ce3d0 | 8354 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 8355 | busiest->group_capacity; |
56cf515b | 8356 | |
3029ede3 VD |
8357 | if (busiest->avg_load + scaled_busy_load_per_task >= |
8358 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 8359 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
8360 | return; |
8361 | } | |
8362 | ||
8363 | /* | |
8364 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 8365 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
8366 | * moving them. |
8367 | */ | |
8368 | ||
63b2ca30 | 8369 | capa_now += busiest->group_capacity * |
56cf515b | 8370 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 8371 | capa_now += local->group_capacity * |
56cf515b | 8372 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 8373 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8374 | |
8375 | /* Amount of load we'd subtract */ | |
a2cd4260 | 8376 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 8377 | capa_move += busiest->group_capacity * |
56cf515b | 8378 | min(busiest->load_per_task, |
a2cd4260 | 8379 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 8380 | } |
1e3c88bd PZ |
8381 | |
8382 | /* Amount of load we'd add */ | |
63b2ca30 | 8383 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 8384 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
8385 | tmp = (busiest->avg_load * busiest->group_capacity) / |
8386 | local->group_capacity; | |
56cf515b | 8387 | } else { |
ca8ce3d0 | 8388 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 8389 | local->group_capacity; |
56cf515b | 8390 | } |
63b2ca30 | 8391 | capa_move += local->group_capacity * |
3ae11c90 | 8392 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 8393 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8394 | |
8395 | /* Move if we gain throughput */ | |
63b2ca30 | 8396 | if (capa_move > capa_now) |
56cf515b | 8397 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
8398 | } |
8399 | ||
8400 | /** | |
8401 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
8402 | * groups of a given sched_domain during load balance. | |
bd939f45 | 8403 | * @env: load balance environment |
1e3c88bd | 8404 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 8405 | */ |
bd939f45 | 8406 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8407 | { |
dd5feea1 | 8408 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
8409 | struct sg_lb_stats *local, *busiest; |
8410 | ||
8411 | local = &sds->local_stat; | |
56cf515b | 8412 | busiest = &sds->busiest_stat; |
dd5feea1 | 8413 | |
caeb178c | 8414 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
8415 | /* |
8416 | * In the group_imb case we cannot rely on group-wide averages | |
97fb7a0a | 8417 | * to ensure CPU-load equilibrium, look at wider averages. XXX |
30ce5dab | 8418 | */ |
56cf515b JK |
8419 | busiest->load_per_task = |
8420 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
8421 | } |
8422 | ||
1e3c88bd | 8423 | /* |
885e542c DE |
8424 | * Avg load of busiest sg can be less and avg load of local sg can |
8425 | * be greater than avg load across all sgs of sd because avg load | |
8426 | * factors in sg capacity and sgs with smaller group_type are | |
8427 | * skipped when updating the busiest sg: | |
1e3c88bd | 8428 | */ |
cad68e55 MR |
8429 | if (busiest->group_type != group_misfit_task && |
8430 | (busiest->avg_load <= sds->avg_load || | |
8431 | local->avg_load >= sds->avg_load)) { | |
bd939f45 PZ |
8432 | env->imbalance = 0; |
8433 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
8434 | } |
8435 | ||
9a5d9ba6 | 8436 | /* |
97fb7a0a | 8437 | * If there aren't any idle CPUs, avoid creating some. |
9a5d9ba6 PZ |
8438 | */ |
8439 | if (busiest->group_type == group_overloaded && | |
8440 | local->group_type == group_overloaded) { | |
1be0eb2a | 8441 | load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE; |
cfa10334 | 8442 | if (load_above_capacity > busiest->group_capacity) { |
ea67821b | 8443 | load_above_capacity -= busiest->group_capacity; |
26656215 | 8444 | load_above_capacity *= scale_load_down(NICE_0_LOAD); |
cfa10334 MR |
8445 | load_above_capacity /= busiest->group_capacity; |
8446 | } else | |
ea67821b | 8447 | load_above_capacity = ~0UL; |
dd5feea1 SS |
8448 | } |
8449 | ||
8450 | /* | |
97fb7a0a | 8451 | * We're trying to get all the CPUs to the average_load, so we don't |
dd5feea1 | 8452 | * want to push ourselves above the average load, nor do we wish to |
97fb7a0a | 8453 | * reduce the max loaded CPU below the average load. At the same time, |
0a9b23ce DE |
8454 | * we also don't want to reduce the group load below the group |
8455 | * capacity. Thus we look for the minimum possible imbalance. | |
dd5feea1 | 8456 | */ |
30ce5dab | 8457 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
8458 | |
8459 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 8460 | env->imbalance = min( |
63b2ca30 NP |
8461 | max_pull * busiest->group_capacity, |
8462 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 8463 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 8464 | |
cad68e55 MR |
8465 | /* Boost imbalance to allow misfit task to be balanced. */ |
8466 | if (busiest->group_type == group_misfit_task) { | |
8467 | env->imbalance = max_t(long, env->imbalance, | |
8468 | busiest->group_misfit_task_load); | |
8469 | } | |
8470 | ||
1e3c88bd PZ |
8471 | /* |
8472 | * if *imbalance is less than the average load per runnable task | |
25985edc | 8473 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
8474 | * a think about bumping its value to force at least one task to be |
8475 | * moved | |
8476 | */ | |
56cf515b | 8477 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 8478 | return fix_small_imbalance(env, sds); |
1e3c88bd | 8479 | } |
fab47622 | 8480 | |
1e3c88bd PZ |
8481 | /******* find_busiest_group() helpers end here *********************/ |
8482 | ||
8483 | /** | |
8484 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 8485 | * if there is an imbalance. |
1e3c88bd | 8486 | * |
a3df0679 | 8487 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
8488 | * to restore balance. |
8489 | * | |
cd96891d | 8490 | * @env: The load balancing environment. |
1e3c88bd | 8491 | * |
e69f6186 | 8492 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 8493 | */ |
56cf515b | 8494 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 8495 | { |
56cf515b | 8496 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
8497 | struct sd_lb_stats sds; |
8498 | ||
147c5fc2 | 8499 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
8500 | |
8501 | /* | |
8502 | * Compute the various statistics relavent for load balancing at | |
8503 | * this level. | |
8504 | */ | |
23f0d209 | 8505 | update_sd_lb_stats(env, &sds); |
2802bf3c | 8506 | |
f8a696f2 | 8507 | if (sched_energy_enabled()) { |
2802bf3c MR |
8508 | struct root_domain *rd = env->dst_rq->rd; |
8509 | ||
8510 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
8511 | goto out_balanced; | |
8512 | } | |
8513 | ||
56cf515b JK |
8514 | local = &sds.local_stat; |
8515 | busiest = &sds.busiest_stat; | |
1e3c88bd | 8516 | |
ea67821b | 8517 | /* ASYM feature bypasses nice load balance check */ |
1f621e02 | 8518 | if (check_asym_packing(env, &sds)) |
532cb4c4 MN |
8519 | return sds.busiest; |
8520 | ||
cc57aa8f | 8521 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 8522 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
8523 | goto out_balanced; |
8524 | ||
90001d67 | 8525 | /* XXX broken for overlapping NUMA groups */ |
ca8ce3d0 NP |
8526 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
8527 | / sds.total_capacity; | |
b0432d8f | 8528 | |
866ab43e PZ |
8529 | /* |
8530 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 8531 | * work because they assume all things are equal, which typically |
3bd37062 | 8532 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 8533 | */ |
caeb178c | 8534 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
8535 | goto force_balance; |
8536 | ||
583ffd99 BJ |
8537 | /* |
8538 | * When dst_cpu is idle, prevent SMP nice and/or asymmetric group | |
8539 | * capacities from resulting in underutilization due to avg_load. | |
8540 | */ | |
8541 | if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) && | |
ea67821b | 8542 | busiest->group_no_capacity) |
fab47622 NR |
8543 | goto force_balance; |
8544 | ||
cad68e55 MR |
8545 | /* Misfit tasks should be dealt with regardless of the avg load */ |
8546 | if (busiest->group_type == group_misfit_task) | |
8547 | goto force_balance; | |
8548 | ||
cc57aa8f | 8549 | /* |
9c58c79a | 8550 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
8551 | * don't try and pull any tasks. |
8552 | */ | |
56cf515b | 8553 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
8554 | goto out_balanced; |
8555 | ||
cc57aa8f PZ |
8556 | /* |
8557 | * Don't pull any tasks if this group is already above the domain | |
8558 | * average load. | |
8559 | */ | |
56cf515b | 8560 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
8561 | goto out_balanced; |
8562 | ||
bd939f45 | 8563 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 8564 | /* |
97fb7a0a | 8565 | * This CPU is idle. If the busiest group is not overloaded |
43f4d666 | 8566 | * and there is no imbalance between this and busiest group |
97fb7a0a | 8567 | * wrt idle CPUs, it is balanced. The imbalance becomes |
43f4d666 VG |
8568 | * significant if the diff is greater than 1 otherwise we |
8569 | * might end up to just move the imbalance on another group | |
aae6d3dd | 8570 | */ |
43f4d666 VG |
8571 | if ((busiest->group_type != group_overloaded) && |
8572 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 8573 | goto out_balanced; |
c186fafe PZ |
8574 | } else { |
8575 | /* | |
8576 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
8577 | * imbalance_pct to be conservative. | |
8578 | */ | |
56cf515b JK |
8579 | if (100 * busiest->avg_load <= |
8580 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 8581 | goto out_balanced; |
aae6d3dd | 8582 | } |
1e3c88bd | 8583 | |
fab47622 | 8584 | force_balance: |
1e3c88bd | 8585 | /* Looks like there is an imbalance. Compute it */ |
cad68e55 | 8586 | env->src_grp_type = busiest->group_type; |
bd939f45 | 8587 | calculate_imbalance(env, &sds); |
bb3485c8 | 8588 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
8589 | |
8590 | out_balanced: | |
bd939f45 | 8591 | env->imbalance = 0; |
1e3c88bd PZ |
8592 | return NULL; |
8593 | } | |
8594 | ||
8595 | /* | |
97fb7a0a | 8596 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 8597 | */ |
bd939f45 | 8598 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 8599 | struct sched_group *group) |
1e3c88bd PZ |
8600 | { |
8601 | struct rq *busiest = NULL, *rq; | |
ced549fa | 8602 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
8603 | int i; |
8604 | ||
ae4df9d6 | 8605 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
a3df0679 | 8606 | unsigned long capacity, load; |
0ec8aa00 PZ |
8607 | enum fbq_type rt; |
8608 | ||
8609 | rq = cpu_rq(i); | |
8610 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 8611 | |
0ec8aa00 PZ |
8612 | /* |
8613 | * We classify groups/runqueues into three groups: | |
8614 | * - regular: there are !numa tasks | |
8615 | * - remote: there are numa tasks that run on the 'wrong' node | |
8616 | * - all: there is no distinction | |
8617 | * | |
8618 | * In order to avoid migrating ideally placed numa tasks, | |
8619 | * ignore those when there's better options. | |
8620 | * | |
8621 | * If we ignore the actual busiest queue to migrate another | |
8622 | * task, the next balance pass can still reduce the busiest | |
8623 | * queue by moving tasks around inside the node. | |
8624 | * | |
8625 | * If we cannot move enough load due to this classification | |
8626 | * the next pass will adjust the group classification and | |
8627 | * allow migration of more tasks. | |
8628 | * | |
8629 | * Both cases only affect the total convergence complexity. | |
8630 | */ | |
8631 | if (rt > env->fbq_type) | |
8632 | continue; | |
8633 | ||
cad68e55 MR |
8634 | /* |
8635 | * For ASYM_CPUCAPACITY domains with misfit tasks we simply | |
8636 | * seek the "biggest" misfit task. | |
8637 | */ | |
8638 | if (env->src_grp_type == group_misfit_task) { | |
8639 | if (rq->misfit_task_load > busiest_load) { | |
8640 | busiest_load = rq->misfit_task_load; | |
8641 | busiest = rq; | |
8642 | } | |
8643 | ||
8644 | continue; | |
8645 | } | |
8646 | ||
ced549fa | 8647 | capacity = capacity_of(i); |
9d5efe05 | 8648 | |
4ad3831a CR |
8649 | /* |
8650 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
8651 | * eventually lead to active_balancing high->low capacity. | |
8652 | * Higher per-CPU capacity is considered better than balancing | |
8653 | * average load. | |
8654 | */ | |
8655 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
8656 | capacity_of(env->dst_cpu) < capacity && | |
8657 | rq->nr_running == 1) | |
8658 | continue; | |
8659 | ||
a3df0679 | 8660 | load = cpu_runnable_load(rq); |
1e3c88bd | 8661 | |
6e40f5bb | 8662 | /* |
a3df0679 | 8663 | * When comparing with imbalance, use cpu_runnable_load() |
97fb7a0a | 8664 | * which is not scaled with the CPU capacity. |
6e40f5bb | 8665 | */ |
ea67821b | 8666 | |
a3df0679 | 8667 | if (rq->nr_running == 1 && load > env->imbalance && |
ea67821b | 8668 | !check_cpu_capacity(rq, env->sd)) |
1e3c88bd PZ |
8669 | continue; |
8670 | ||
6e40f5bb | 8671 | /* |
97fb7a0a | 8672 | * For the load comparisons with the other CPU's, consider |
a3df0679 | 8673 | * the cpu_runnable_load() scaled with the CPU capacity, so |
97fb7a0a | 8674 | * that the load can be moved away from the CPU that is |
ced549fa | 8675 | * potentially running at a lower capacity. |
95a79b80 | 8676 | * |
a3df0679 | 8677 | * Thus we're looking for max(load_i / capacity_i), crosswise |
95a79b80 | 8678 | * multiplication to rid ourselves of the division works out |
a3df0679 | 8679 | * to: load_i * capacity_j > load_j * capacity_i; where j is |
ced549fa | 8680 | * our previous maximum. |
6e40f5bb | 8681 | */ |
a3df0679 DE |
8682 | if (load * busiest_capacity > busiest_load * capacity) { |
8683 | busiest_load = load; | |
ced549fa | 8684 | busiest_capacity = capacity; |
1e3c88bd PZ |
8685 | busiest = rq; |
8686 | } | |
8687 | } | |
8688 | ||
8689 | return busiest; | |
8690 | } | |
8691 | ||
8692 | /* | |
8693 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
8694 | * so long as it is large enough. | |
8695 | */ | |
8696 | #define MAX_PINNED_INTERVAL 512 | |
8697 | ||
46a745d9 VG |
8698 | static inline bool |
8699 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 8700 | { |
46a745d9 VG |
8701 | /* |
8702 | * ASYM_PACKING needs to force migrate tasks from busy but | |
8703 | * lower priority CPUs in order to pack all tasks in the | |
8704 | * highest priority CPUs. | |
8705 | */ | |
8706 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
8707 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
8708 | } | |
bd939f45 | 8709 | |
46a745d9 VG |
8710 | static inline bool |
8711 | voluntary_active_balance(struct lb_env *env) | |
8712 | { | |
8713 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 8714 | |
46a745d9 VG |
8715 | if (asym_active_balance(env)) |
8716 | return 1; | |
1af3ed3d | 8717 | |
1aaf90a4 VG |
8718 | /* |
8719 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
8720 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
8721 | * because of other sched_class or IRQs if more capacity stays | |
8722 | * available on dst_cpu. | |
8723 | */ | |
8724 | if ((env->idle != CPU_NOT_IDLE) && | |
8725 | (env->src_rq->cfs.h_nr_running == 1)) { | |
8726 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
8727 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
8728 | return 1; | |
8729 | } | |
8730 | ||
cad68e55 MR |
8731 | if (env->src_grp_type == group_misfit_task) |
8732 | return 1; | |
8733 | ||
46a745d9 VG |
8734 | return 0; |
8735 | } | |
8736 | ||
8737 | static int need_active_balance(struct lb_env *env) | |
8738 | { | |
8739 | struct sched_domain *sd = env->sd; | |
8740 | ||
8741 | if (voluntary_active_balance(env)) | |
8742 | return 1; | |
8743 | ||
1af3ed3d PZ |
8744 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
8745 | } | |
8746 | ||
969c7921 TH |
8747 | static int active_load_balance_cpu_stop(void *data); |
8748 | ||
23f0d209 JK |
8749 | static int should_we_balance(struct lb_env *env) |
8750 | { | |
8751 | struct sched_group *sg = env->sd->groups; | |
23f0d209 JK |
8752 | int cpu, balance_cpu = -1; |
8753 | ||
024c9d2f PZ |
8754 | /* |
8755 | * Ensure the balancing environment is consistent; can happen | |
8756 | * when the softirq triggers 'during' hotplug. | |
8757 | */ | |
8758 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
8759 | return 0; | |
8760 | ||
23f0d209 | 8761 | /* |
97fb7a0a | 8762 | * In the newly idle case, we will allow all the CPUs |
23f0d209 JK |
8763 | * to do the newly idle load balance. |
8764 | */ | |
8765 | if (env->idle == CPU_NEWLY_IDLE) | |
8766 | return 1; | |
8767 | ||
97fb7a0a | 8768 | /* Try to find first idle CPU */ |
e5c14b1f | 8769 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 8770 | if (!idle_cpu(cpu)) |
23f0d209 JK |
8771 | continue; |
8772 | ||
8773 | balance_cpu = cpu; | |
8774 | break; | |
8775 | } | |
8776 | ||
8777 | if (balance_cpu == -1) | |
8778 | balance_cpu = group_balance_cpu(sg); | |
8779 | ||
8780 | /* | |
97fb7a0a | 8781 | * First idle CPU or the first CPU(busiest) in this sched group |
23f0d209 JK |
8782 | * is eligible for doing load balancing at this and above domains. |
8783 | */ | |
b0cff9d8 | 8784 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
8785 | } |
8786 | ||
1e3c88bd PZ |
8787 | /* |
8788 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
8789 | * tasks if there is an imbalance. | |
8790 | */ | |
8791 | static int load_balance(int this_cpu, struct rq *this_rq, | |
8792 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 8793 | int *continue_balancing) |
1e3c88bd | 8794 | { |
88b8dac0 | 8795 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 8796 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 8797 | struct sched_group *group; |
1e3c88bd | 8798 | struct rq *busiest; |
8a8c69c3 | 8799 | struct rq_flags rf; |
4ba29684 | 8800 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 8801 | |
8e45cb54 PZ |
8802 | struct lb_env env = { |
8803 | .sd = sd, | |
ddcdf6e7 PZ |
8804 | .dst_cpu = this_cpu, |
8805 | .dst_rq = this_rq, | |
ae4df9d6 | 8806 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 8807 | .idle = idle, |
eb95308e | 8808 | .loop_break = sched_nr_migrate_break, |
b9403130 | 8809 | .cpus = cpus, |
0ec8aa00 | 8810 | .fbq_type = all, |
163122b7 | 8811 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
8812 | }; |
8813 | ||
65a4433a | 8814 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 8815 | |
ae92882e | 8816 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
8817 | |
8818 | redo: | |
23f0d209 JK |
8819 | if (!should_we_balance(&env)) { |
8820 | *continue_balancing = 0; | |
1e3c88bd | 8821 | goto out_balanced; |
23f0d209 | 8822 | } |
1e3c88bd | 8823 | |
23f0d209 | 8824 | group = find_busiest_group(&env); |
1e3c88bd | 8825 | if (!group) { |
ae92882e | 8826 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
8827 | goto out_balanced; |
8828 | } | |
8829 | ||
b9403130 | 8830 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 8831 | if (!busiest) { |
ae92882e | 8832 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
8833 | goto out_balanced; |
8834 | } | |
8835 | ||
78feefc5 | 8836 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 8837 | |
ae92882e | 8838 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 8839 | |
1aaf90a4 VG |
8840 | env.src_cpu = busiest->cpu; |
8841 | env.src_rq = busiest; | |
8842 | ||
1e3c88bd PZ |
8843 | ld_moved = 0; |
8844 | if (busiest->nr_running > 1) { | |
8845 | /* | |
8846 | * Attempt to move tasks. If find_busiest_group has found | |
8847 | * an imbalance but busiest->nr_running <= 1, the group is | |
8848 | * still unbalanced. ld_moved simply stays zero, so it is | |
8849 | * correctly treated as an imbalance. | |
8850 | */ | |
8e45cb54 | 8851 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 8852 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 8853 | |
5d6523eb | 8854 | more_balance: |
8a8c69c3 | 8855 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 8856 | update_rq_clock(busiest); |
88b8dac0 SV |
8857 | |
8858 | /* | |
8859 | * cur_ld_moved - load moved in current iteration | |
8860 | * ld_moved - cumulative load moved across iterations | |
8861 | */ | |
163122b7 | 8862 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
8863 | |
8864 | /* | |
163122b7 KT |
8865 | * We've detached some tasks from busiest_rq. Every |
8866 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
8867 | * unlock busiest->lock, and we are able to be sure | |
8868 | * that nobody can manipulate the tasks in parallel. | |
8869 | * See task_rq_lock() family for the details. | |
1e3c88bd | 8870 | */ |
163122b7 | 8871 | |
8a8c69c3 | 8872 | rq_unlock(busiest, &rf); |
163122b7 KT |
8873 | |
8874 | if (cur_ld_moved) { | |
8875 | attach_tasks(&env); | |
8876 | ld_moved += cur_ld_moved; | |
8877 | } | |
8878 | ||
8a8c69c3 | 8879 | local_irq_restore(rf.flags); |
88b8dac0 | 8880 | |
f1cd0858 JK |
8881 | if (env.flags & LBF_NEED_BREAK) { |
8882 | env.flags &= ~LBF_NEED_BREAK; | |
8883 | goto more_balance; | |
8884 | } | |
8885 | ||
88b8dac0 SV |
8886 | /* |
8887 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
8888 | * us and move them to an alternate dst_cpu in our sched_group | |
8889 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 8890 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
8891 | * sched_group. |
8892 | * | |
8893 | * This changes load balance semantics a bit on who can move | |
8894 | * load to a given_cpu. In addition to the given_cpu itself | |
8895 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
8896 | * nohz-idle), we now have balance_cpu in a position to move | |
8897 | * load to given_cpu. In rare situations, this may cause | |
8898 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
8899 | * _independently_ and at _same_ time to move some load to | |
8900 | * given_cpu) causing exceess load to be moved to given_cpu. | |
8901 | * This however should not happen so much in practice and | |
8902 | * moreover subsequent load balance cycles should correct the | |
8903 | * excess load moved. | |
8904 | */ | |
6263322c | 8905 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 8906 | |
97fb7a0a | 8907 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 8908 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 8909 | |
78feefc5 | 8910 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 8911 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 8912 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
8913 | env.loop = 0; |
8914 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 8915 | |
88b8dac0 SV |
8916 | /* |
8917 | * Go back to "more_balance" rather than "redo" since we | |
8918 | * need to continue with same src_cpu. | |
8919 | */ | |
8920 | goto more_balance; | |
8921 | } | |
1e3c88bd | 8922 | |
6263322c PZ |
8923 | /* |
8924 | * We failed to reach balance because of affinity. | |
8925 | */ | |
8926 | if (sd_parent) { | |
63b2ca30 | 8927 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 8928 | |
afdeee05 | 8929 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 8930 | *group_imbalance = 1; |
6263322c PZ |
8931 | } |
8932 | ||
1e3c88bd | 8933 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 8934 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 8935 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
8936 | /* |
8937 | * Attempting to continue load balancing at the current | |
8938 | * sched_domain level only makes sense if there are | |
8939 | * active CPUs remaining as possible busiest CPUs to | |
8940 | * pull load from which are not contained within the | |
8941 | * destination group that is receiving any migrated | |
8942 | * load. | |
8943 | */ | |
8944 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
8945 | env.loop = 0; |
8946 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 8947 | goto redo; |
bbf18b19 | 8948 | } |
afdeee05 | 8949 | goto out_all_pinned; |
1e3c88bd PZ |
8950 | } |
8951 | } | |
8952 | ||
8953 | if (!ld_moved) { | |
ae92882e | 8954 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
8955 | /* |
8956 | * Increment the failure counter only on periodic balance. | |
8957 | * We do not want newidle balance, which can be very | |
8958 | * frequent, pollute the failure counter causing | |
8959 | * excessive cache_hot migrations and active balances. | |
8960 | */ | |
8961 | if (idle != CPU_NEWLY_IDLE) | |
8962 | sd->nr_balance_failed++; | |
1e3c88bd | 8963 | |
bd939f45 | 8964 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
8965 | unsigned long flags; |
8966 | ||
1e3c88bd PZ |
8967 | raw_spin_lock_irqsave(&busiest->lock, flags); |
8968 | ||
97fb7a0a IM |
8969 | /* |
8970 | * Don't kick the active_load_balance_cpu_stop, | |
8971 | * if the curr task on busiest CPU can't be | |
8972 | * moved to this_cpu: | |
1e3c88bd | 8973 | */ |
3bd37062 | 8974 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
1e3c88bd PZ |
8975 | raw_spin_unlock_irqrestore(&busiest->lock, |
8976 | flags); | |
8e45cb54 | 8977 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
8978 | goto out_one_pinned; |
8979 | } | |
8980 | ||
969c7921 TH |
8981 | /* |
8982 | * ->active_balance synchronizes accesses to | |
8983 | * ->active_balance_work. Once set, it's cleared | |
8984 | * only after active load balance is finished. | |
8985 | */ | |
1e3c88bd PZ |
8986 | if (!busiest->active_balance) { |
8987 | busiest->active_balance = 1; | |
8988 | busiest->push_cpu = this_cpu; | |
8989 | active_balance = 1; | |
8990 | } | |
8991 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 8992 | |
bd939f45 | 8993 | if (active_balance) { |
969c7921 TH |
8994 | stop_one_cpu_nowait(cpu_of(busiest), |
8995 | active_load_balance_cpu_stop, busiest, | |
8996 | &busiest->active_balance_work); | |
bd939f45 | 8997 | } |
1e3c88bd | 8998 | |
d02c0711 | 8999 | /* We've kicked active balancing, force task migration. */ |
1e3c88bd PZ |
9000 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
9001 | } | |
9002 | } else | |
9003 | sd->nr_balance_failed = 0; | |
9004 | ||
46a745d9 | 9005 | if (likely(!active_balance) || voluntary_active_balance(&env)) { |
1e3c88bd PZ |
9006 | /* We were unbalanced, so reset the balancing interval */ |
9007 | sd->balance_interval = sd->min_interval; | |
9008 | } else { | |
9009 | /* | |
9010 | * If we've begun active balancing, start to back off. This | |
9011 | * case may not be covered by the all_pinned logic if there | |
9012 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 9013 | * detach_tasks). |
1e3c88bd PZ |
9014 | */ |
9015 | if (sd->balance_interval < sd->max_interval) | |
9016 | sd->balance_interval *= 2; | |
9017 | } | |
9018 | ||
1e3c88bd PZ |
9019 | goto out; |
9020 | ||
9021 | out_balanced: | |
afdeee05 VG |
9022 | /* |
9023 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
9024 | * constraints. Clear the imbalance flag only if other tasks got |
9025 | * a chance to move and fix the imbalance. | |
afdeee05 | 9026 | */ |
f6cad8df | 9027 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
9028 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
9029 | ||
9030 | if (*group_imbalance) | |
9031 | *group_imbalance = 0; | |
9032 | } | |
9033 | ||
9034 | out_all_pinned: | |
9035 | /* | |
9036 | * We reach balance because all tasks are pinned at this level so | |
9037 | * we can't migrate them. Let the imbalance flag set so parent level | |
9038 | * can try to migrate them. | |
9039 | */ | |
ae92882e | 9040 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
9041 | |
9042 | sd->nr_balance_failed = 0; | |
9043 | ||
9044 | out_one_pinned: | |
3f130a37 VS |
9045 | ld_moved = 0; |
9046 | ||
9047 | /* | |
9048 | * idle_balance() disregards balance intervals, so we could repeatedly | |
9049 | * reach this code, which would lead to balance_interval skyrocketting | |
9050 | * in a short amount of time. Skip the balance_interval increase logic | |
9051 | * to avoid that. | |
9052 | */ | |
9053 | if (env.idle == CPU_NEWLY_IDLE) | |
9054 | goto out; | |
9055 | ||
1e3c88bd | 9056 | /* tune up the balancing interval */ |
47b7aee1 VS |
9057 | if ((env.flags & LBF_ALL_PINNED && |
9058 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
9059 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 9060 | sd->balance_interval *= 2; |
1e3c88bd | 9061 | out: |
1e3c88bd PZ |
9062 | return ld_moved; |
9063 | } | |
9064 | ||
52a08ef1 JL |
9065 | static inline unsigned long |
9066 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
9067 | { | |
9068 | unsigned long interval = sd->balance_interval; | |
9069 | ||
9070 | if (cpu_busy) | |
9071 | interval *= sd->busy_factor; | |
9072 | ||
9073 | /* scale ms to jiffies */ | |
9074 | interval = msecs_to_jiffies(interval); | |
9075 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
9076 | ||
9077 | return interval; | |
9078 | } | |
9079 | ||
9080 | static inline void | |
31851a98 | 9081 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
9082 | { |
9083 | unsigned long interval, next; | |
9084 | ||
31851a98 LY |
9085 | /* used by idle balance, so cpu_busy = 0 */ |
9086 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
9087 | next = sd->last_balance + interval; |
9088 | ||
9089 | if (time_after(*next_balance, next)) | |
9090 | *next_balance = next; | |
9091 | } | |
9092 | ||
1e3c88bd | 9093 | /* |
97fb7a0a | 9094 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
9095 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
9096 | * least 1 task to be running on each physical CPU where possible, and | |
9097 | * avoids physical / logical imbalances. | |
1e3c88bd | 9098 | */ |
969c7921 | 9099 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 9100 | { |
969c7921 TH |
9101 | struct rq *busiest_rq = data; |
9102 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 9103 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 9104 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 9105 | struct sched_domain *sd; |
e5673f28 | 9106 | struct task_struct *p = NULL; |
8a8c69c3 | 9107 | struct rq_flags rf; |
969c7921 | 9108 | |
8a8c69c3 | 9109 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
9110 | /* |
9111 | * Between queueing the stop-work and running it is a hole in which | |
9112 | * CPUs can become inactive. We should not move tasks from or to | |
9113 | * inactive CPUs. | |
9114 | */ | |
9115 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
9116 | goto out_unlock; | |
969c7921 | 9117 | |
97fb7a0a | 9118 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
9119 | if (unlikely(busiest_cpu != smp_processor_id() || |
9120 | !busiest_rq->active_balance)) | |
9121 | goto out_unlock; | |
1e3c88bd PZ |
9122 | |
9123 | /* Is there any task to move? */ | |
9124 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 9125 | goto out_unlock; |
1e3c88bd PZ |
9126 | |
9127 | /* | |
9128 | * This condition is "impossible", if it occurs | |
9129 | * we need to fix it. Originally reported by | |
97fb7a0a | 9130 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd PZ |
9131 | */ |
9132 | BUG_ON(busiest_rq == target_rq); | |
9133 | ||
1e3c88bd | 9134 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 9135 | rcu_read_lock(); |
1e3c88bd PZ |
9136 | for_each_domain(target_cpu, sd) { |
9137 | if ((sd->flags & SD_LOAD_BALANCE) && | |
9138 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
9139 | break; | |
9140 | } | |
9141 | ||
9142 | if (likely(sd)) { | |
8e45cb54 PZ |
9143 | struct lb_env env = { |
9144 | .sd = sd, | |
ddcdf6e7 PZ |
9145 | .dst_cpu = target_cpu, |
9146 | .dst_rq = target_rq, | |
9147 | .src_cpu = busiest_rq->cpu, | |
9148 | .src_rq = busiest_rq, | |
8e45cb54 | 9149 | .idle = CPU_IDLE, |
65a4433a JH |
9150 | /* |
9151 | * can_migrate_task() doesn't need to compute new_dst_cpu | |
9152 | * for active balancing. Since we have CPU_IDLE, but no | |
9153 | * @dst_grpmask we need to make that test go away with lying | |
9154 | * about DST_PINNED. | |
9155 | */ | |
9156 | .flags = LBF_DST_PINNED, | |
8e45cb54 PZ |
9157 | }; |
9158 | ||
ae92882e | 9159 | schedstat_inc(sd->alb_count); |
3bed5e21 | 9160 | update_rq_clock(busiest_rq); |
1e3c88bd | 9161 | |
e5673f28 | 9162 | p = detach_one_task(&env); |
d02c0711 | 9163 | if (p) { |
ae92882e | 9164 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
9165 | /* Active balancing done, reset the failure counter. */ |
9166 | sd->nr_balance_failed = 0; | |
9167 | } else { | |
ae92882e | 9168 | schedstat_inc(sd->alb_failed); |
d02c0711 | 9169 | } |
1e3c88bd | 9170 | } |
dce840a0 | 9171 | rcu_read_unlock(); |
969c7921 TH |
9172 | out_unlock: |
9173 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 9174 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
9175 | |
9176 | if (p) | |
9177 | attach_one_task(target_rq, p); | |
9178 | ||
9179 | local_irq_enable(); | |
9180 | ||
969c7921 | 9181 | return 0; |
1e3c88bd PZ |
9182 | } |
9183 | ||
af3fe03c PZ |
9184 | static DEFINE_SPINLOCK(balancing); |
9185 | ||
9186 | /* | |
9187 | * Scale the max load_balance interval with the number of CPUs in the system. | |
9188 | * This trades load-balance latency on larger machines for less cross talk. | |
9189 | */ | |
9190 | void update_max_interval(void) | |
9191 | { | |
9192 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
9193 | } | |
9194 | ||
9195 | /* | |
9196 | * It checks each scheduling domain to see if it is due to be balanced, | |
9197 | * and initiates a balancing operation if so. | |
9198 | * | |
9199 | * Balancing parameters are set up in init_sched_domains. | |
9200 | */ | |
9201 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
9202 | { | |
9203 | int continue_balancing = 1; | |
9204 | int cpu = rq->cpu; | |
9205 | unsigned long interval; | |
9206 | struct sched_domain *sd; | |
9207 | /* Earliest time when we have to do rebalance again */ | |
9208 | unsigned long next_balance = jiffies + 60*HZ; | |
9209 | int update_next_balance = 0; | |
9210 | int need_serialize, need_decay = 0; | |
9211 | u64 max_cost = 0; | |
9212 | ||
9213 | rcu_read_lock(); | |
9214 | for_each_domain(cpu, sd) { | |
9215 | /* | |
9216 | * Decay the newidle max times here because this is a regular | |
9217 | * visit to all the domains. Decay ~1% per second. | |
9218 | */ | |
9219 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
9220 | sd->max_newidle_lb_cost = | |
9221 | (sd->max_newidle_lb_cost * 253) / 256; | |
9222 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
9223 | need_decay = 1; | |
9224 | } | |
9225 | max_cost += sd->max_newidle_lb_cost; | |
9226 | ||
9227 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
9228 | continue; | |
9229 | ||
9230 | /* | |
9231 | * Stop the load balance at this level. There is another | |
9232 | * CPU in our sched group which is doing load balancing more | |
9233 | * actively. | |
9234 | */ | |
9235 | if (!continue_balancing) { | |
9236 | if (need_decay) | |
9237 | continue; | |
9238 | break; | |
9239 | } | |
9240 | ||
9241 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); | |
9242 | ||
9243 | need_serialize = sd->flags & SD_SERIALIZE; | |
9244 | if (need_serialize) { | |
9245 | if (!spin_trylock(&balancing)) | |
9246 | goto out; | |
9247 | } | |
9248 | ||
9249 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
9250 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
9251 | /* | |
9252 | * The LBF_DST_PINNED logic could have changed | |
9253 | * env->dst_cpu, so we can't know our idle | |
9254 | * state even if we migrated tasks. Update it. | |
9255 | */ | |
9256 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
9257 | } | |
9258 | sd->last_balance = jiffies; | |
9259 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); | |
9260 | } | |
9261 | if (need_serialize) | |
9262 | spin_unlock(&balancing); | |
9263 | out: | |
9264 | if (time_after(next_balance, sd->last_balance + interval)) { | |
9265 | next_balance = sd->last_balance + interval; | |
9266 | update_next_balance = 1; | |
9267 | } | |
9268 | } | |
9269 | if (need_decay) { | |
9270 | /* | |
9271 | * Ensure the rq-wide value also decays but keep it at a | |
9272 | * reasonable floor to avoid funnies with rq->avg_idle. | |
9273 | */ | |
9274 | rq->max_idle_balance_cost = | |
9275 | max((u64)sysctl_sched_migration_cost, max_cost); | |
9276 | } | |
9277 | rcu_read_unlock(); | |
9278 | ||
9279 | /* | |
9280 | * next_balance will be updated only when there is a need. | |
9281 | * When the cpu is attached to null domain for ex, it will not be | |
9282 | * updated. | |
9283 | */ | |
9284 | if (likely(update_next_balance)) { | |
9285 | rq->next_balance = next_balance; | |
9286 | ||
9287 | #ifdef CONFIG_NO_HZ_COMMON | |
9288 | /* | |
9289 | * If this CPU has been elected to perform the nohz idle | |
9290 | * balance. Other idle CPUs have already rebalanced with | |
9291 | * nohz_idle_balance() and nohz.next_balance has been | |
9292 | * updated accordingly. This CPU is now running the idle load | |
9293 | * balance for itself and we need to update the | |
9294 | * nohz.next_balance accordingly. | |
9295 | */ | |
9296 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
9297 | nohz.next_balance = rq->next_balance; | |
9298 | #endif | |
9299 | } | |
9300 | } | |
9301 | ||
d987fc7f MG |
9302 | static inline int on_null_domain(struct rq *rq) |
9303 | { | |
9304 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
9305 | } | |
9306 | ||
3451d024 | 9307 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
9308 | /* |
9309 | * idle load balancing details | |
83cd4fe2 VP |
9310 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
9311 | * needed, they will kick the idle load balancer, which then does idle | |
9312 | * load balancing for all the idle CPUs. | |
9b019acb NP |
9313 | * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set |
9314 | * anywhere yet. | |
83cd4fe2 | 9315 | */ |
1e3c88bd | 9316 | |
3dd0337d | 9317 | static inline int find_new_ilb(void) |
1e3c88bd | 9318 | { |
9b019acb | 9319 | int ilb; |
1e3c88bd | 9320 | |
9b019acb NP |
9321 | for_each_cpu_and(ilb, nohz.idle_cpus_mask, |
9322 | housekeeping_cpumask(HK_FLAG_MISC)) { | |
9323 | if (idle_cpu(ilb)) | |
9324 | return ilb; | |
9325 | } | |
786d6dc7 SS |
9326 | |
9327 | return nr_cpu_ids; | |
1e3c88bd | 9328 | } |
1e3c88bd | 9329 | |
83cd4fe2 | 9330 | /* |
9b019acb NP |
9331 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick any |
9332 | * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one). | |
83cd4fe2 | 9333 | */ |
a4064fb6 | 9334 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
9335 | { |
9336 | int ilb_cpu; | |
9337 | ||
9338 | nohz.next_balance++; | |
9339 | ||
3dd0337d | 9340 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 9341 | |
0b005cf5 SS |
9342 | if (ilb_cpu >= nr_cpu_ids) |
9343 | return; | |
83cd4fe2 | 9344 | |
a4064fb6 | 9345 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 9346 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 9347 | return; |
4550487a | 9348 | |
1c792db7 SS |
9349 | /* |
9350 | * Use smp_send_reschedule() instead of resched_cpu(). | |
97fb7a0a | 9351 | * This way we generate a sched IPI on the target CPU which |
1c792db7 SS |
9352 | * is idle. And the softirq performing nohz idle load balance |
9353 | * will be run before returning from the IPI. | |
9354 | */ | |
9355 | smp_send_reschedule(ilb_cpu); | |
4550487a PZ |
9356 | } |
9357 | ||
9358 | /* | |
9f132742 VS |
9359 | * Current decision point for kicking the idle load balancer in the presence |
9360 | * of idle CPUs in the system. | |
4550487a PZ |
9361 | */ |
9362 | static void nohz_balancer_kick(struct rq *rq) | |
9363 | { | |
9364 | unsigned long now = jiffies; | |
9365 | struct sched_domain_shared *sds; | |
9366 | struct sched_domain *sd; | |
9367 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 9368 | unsigned int flags = 0; |
4550487a PZ |
9369 | |
9370 | if (unlikely(rq->idle_balance)) | |
9371 | return; | |
9372 | ||
9373 | /* | |
9374 | * We may be recently in ticked or tickless idle mode. At the first | |
9375 | * busy tick after returning from idle, we will update the busy stats. | |
9376 | */ | |
00357f5e | 9377 | nohz_balance_exit_idle(rq); |
4550487a PZ |
9378 | |
9379 | /* | |
9380 | * None are in tickless mode and hence no need for NOHZ idle load | |
9381 | * balancing. | |
9382 | */ | |
9383 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
9384 | return; | |
9385 | ||
f643ea22 VG |
9386 | if (READ_ONCE(nohz.has_blocked) && |
9387 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
9388 | flags = NOHZ_STATS_KICK; |
9389 | ||
4550487a | 9390 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 9391 | goto out; |
4550487a | 9392 | |
a0fe2cf0 | 9393 | if (rq->nr_running >= 2) { |
a4064fb6 | 9394 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9395 | goto out; |
9396 | } | |
9397 | ||
9398 | rcu_read_lock(); | |
4550487a PZ |
9399 | |
9400 | sd = rcu_dereference(rq->sd); | |
9401 | if (sd) { | |
e25a7a94 VS |
9402 | /* |
9403 | * If there's a CFS task and the current CPU has reduced | |
9404 | * capacity; kick the ILB to see if there's a better CPU to run | |
9405 | * on. | |
9406 | */ | |
9407 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
a4064fb6 | 9408 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9409 | goto unlock; |
9410 | } | |
9411 | } | |
9412 | ||
011b27bb | 9413 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 9414 | if (sd) { |
b9a7b883 VS |
9415 | /* |
9416 | * When ASYM_PACKING; see if there's a more preferred CPU | |
9417 | * currently idle; in which case, kick the ILB to move tasks | |
9418 | * around. | |
9419 | */ | |
7edab78d | 9420 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
4550487a | 9421 | if (sched_asym_prefer(i, cpu)) { |
a4064fb6 | 9422 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9423 | goto unlock; |
9424 | } | |
9425 | } | |
9426 | } | |
b9a7b883 | 9427 | |
a0fe2cf0 VS |
9428 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
9429 | if (sd) { | |
9430 | /* | |
9431 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
9432 | * to run the misfit task on. | |
9433 | */ | |
9434 | if (check_misfit_status(rq, sd)) { | |
9435 | flags = NOHZ_KICK_MASK; | |
9436 | goto unlock; | |
9437 | } | |
b9a7b883 VS |
9438 | |
9439 | /* | |
9440 | * For asymmetric systems, we do not want to nicely balance | |
9441 | * cache use, instead we want to embrace asymmetry and only | |
9442 | * ensure tasks have enough CPU capacity. | |
9443 | * | |
9444 | * Skip the LLC logic because it's not relevant in that case. | |
9445 | */ | |
9446 | goto unlock; | |
a0fe2cf0 VS |
9447 | } |
9448 | ||
b9a7b883 VS |
9449 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
9450 | if (sds) { | |
e25a7a94 | 9451 | /* |
b9a7b883 VS |
9452 | * If there is an imbalance between LLC domains (IOW we could |
9453 | * increase the overall cache use), we need some less-loaded LLC | |
9454 | * domain to pull some load. Likewise, we may need to spread | |
9455 | * load within the current LLC domain (e.g. packed SMT cores but | |
9456 | * other CPUs are idle). We can't really know from here how busy | |
9457 | * the others are - so just get a nohz balance going if it looks | |
9458 | * like this LLC domain has tasks we could move. | |
e25a7a94 | 9459 | */ |
b9a7b883 VS |
9460 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
9461 | if (nr_busy > 1) { | |
9462 | flags = NOHZ_KICK_MASK; | |
9463 | goto unlock; | |
4550487a PZ |
9464 | } |
9465 | } | |
9466 | unlock: | |
9467 | rcu_read_unlock(); | |
9468 | out: | |
a4064fb6 PZ |
9469 | if (flags) |
9470 | kick_ilb(flags); | |
83cd4fe2 VP |
9471 | } |
9472 | ||
00357f5e | 9473 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 9474 | { |
00357f5e | 9475 | struct sched_domain *sd; |
a22e47a4 | 9476 | |
00357f5e PZ |
9477 | rcu_read_lock(); |
9478 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 9479 | |
00357f5e PZ |
9480 | if (!sd || !sd->nohz_idle) |
9481 | goto unlock; | |
9482 | sd->nohz_idle = 0; | |
9483 | ||
9484 | atomic_inc(&sd->shared->nr_busy_cpus); | |
9485 | unlock: | |
9486 | rcu_read_unlock(); | |
71325960 SS |
9487 | } |
9488 | ||
00357f5e PZ |
9489 | void nohz_balance_exit_idle(struct rq *rq) |
9490 | { | |
9491 | SCHED_WARN_ON(rq != this_rq()); | |
9492 | ||
9493 | if (likely(!rq->nohz_tick_stopped)) | |
9494 | return; | |
9495 | ||
9496 | rq->nohz_tick_stopped = 0; | |
9497 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
9498 | atomic_dec(&nohz.nr_cpus); | |
9499 | ||
9500 | set_cpu_sd_state_busy(rq->cpu); | |
9501 | } | |
9502 | ||
9503 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
9504 | { |
9505 | struct sched_domain *sd; | |
69e1e811 | 9506 | |
69e1e811 | 9507 | rcu_read_lock(); |
0e369d75 | 9508 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
9509 | |
9510 | if (!sd || sd->nohz_idle) | |
9511 | goto unlock; | |
9512 | sd->nohz_idle = 1; | |
9513 | ||
0e369d75 | 9514 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 9515 | unlock: |
69e1e811 SS |
9516 | rcu_read_unlock(); |
9517 | } | |
9518 | ||
1e3c88bd | 9519 | /* |
97fb7a0a | 9520 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 9521 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 9522 | */ |
c1cc017c | 9523 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 9524 | { |
00357f5e PZ |
9525 | struct rq *rq = cpu_rq(cpu); |
9526 | ||
9527 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
9528 | ||
97fb7a0a | 9529 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
9530 | if (!cpu_active(cpu)) |
9531 | return; | |
9532 | ||
387bc8b5 | 9533 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
de201559 | 9534 | if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) |
387bc8b5 FW |
9535 | return; |
9536 | ||
f643ea22 VG |
9537 | /* |
9538 | * Can be set safely without rq->lock held | |
9539 | * If a clear happens, it will have evaluated last additions because | |
9540 | * rq->lock is held during the check and the clear | |
9541 | */ | |
9542 | rq->has_blocked_load = 1; | |
9543 | ||
9544 | /* | |
9545 | * The tick is still stopped but load could have been added in the | |
9546 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
9547 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
9548 | * of nohz.has_blocked can only happen after checking the new load | |
9549 | */ | |
00357f5e | 9550 | if (rq->nohz_tick_stopped) |
f643ea22 | 9551 | goto out; |
1e3c88bd | 9552 | |
97fb7a0a | 9553 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 9554 | if (on_null_domain(rq)) |
d987fc7f MG |
9555 | return; |
9556 | ||
00357f5e PZ |
9557 | rq->nohz_tick_stopped = 1; |
9558 | ||
c1cc017c AS |
9559 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
9560 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 9561 | |
f643ea22 VG |
9562 | /* |
9563 | * Ensures that if nohz_idle_balance() fails to observe our | |
9564 | * @idle_cpus_mask store, it must observe the @has_blocked | |
9565 | * store. | |
9566 | */ | |
9567 | smp_mb__after_atomic(); | |
9568 | ||
00357f5e | 9569 | set_cpu_sd_state_idle(cpu); |
f643ea22 VG |
9570 | |
9571 | out: | |
9572 | /* | |
9573 | * Each time a cpu enter idle, we assume that it has blocked load and | |
9574 | * enable the periodic update of the load of idle cpus | |
9575 | */ | |
9576 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 9577 | } |
1e3c88bd | 9578 | |
1e3c88bd | 9579 | /* |
31e77c93 VG |
9580 | * Internal function that runs load balance for all idle cpus. The load balance |
9581 | * can be a simple update of blocked load or a complete load balance with | |
9582 | * tasks movement depending of flags. | |
9583 | * The function returns false if the loop has stopped before running | |
9584 | * through all idle CPUs. | |
1e3c88bd | 9585 | */ |
31e77c93 VG |
9586 | static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags, |
9587 | enum cpu_idle_type idle) | |
83cd4fe2 | 9588 | { |
c5afb6a8 | 9589 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
9590 | unsigned long now = jiffies; |
9591 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 9592 | bool has_blocked_load = false; |
c5afb6a8 | 9593 | int update_next_balance = 0; |
b7031a02 | 9594 | int this_cpu = this_rq->cpu; |
b7031a02 | 9595 | int balance_cpu; |
31e77c93 | 9596 | int ret = false; |
b7031a02 | 9597 | struct rq *rq; |
83cd4fe2 | 9598 | |
b7031a02 | 9599 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 9600 | |
f643ea22 VG |
9601 | /* |
9602 | * We assume there will be no idle load after this update and clear | |
9603 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
9604 | * set the has_blocked flag and trig another update of idle load. | |
9605 | * Because a cpu that becomes idle, is added to idle_cpus_mask before | |
9606 | * setting the flag, we are sure to not clear the state and not | |
9607 | * check the load of an idle cpu. | |
9608 | */ | |
9609 | WRITE_ONCE(nohz.has_blocked, 0); | |
9610 | ||
9611 | /* | |
9612 | * Ensures that if we miss the CPU, we must see the has_blocked | |
9613 | * store from nohz_balance_enter_idle(). | |
9614 | */ | |
9615 | smp_mb(); | |
9616 | ||
83cd4fe2 | 9617 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { |
8a6d42d1 | 9618 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
9619 | continue; |
9620 | ||
9621 | /* | |
97fb7a0a IM |
9622 | * If this CPU gets work to do, stop the load balancing |
9623 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
9624 | * balancing owner will pick it up. |
9625 | */ | |
f643ea22 VG |
9626 | if (need_resched()) { |
9627 | has_blocked_load = true; | |
9628 | goto abort; | |
9629 | } | |
83cd4fe2 | 9630 | |
5ed4f1d9 VG |
9631 | rq = cpu_rq(balance_cpu); |
9632 | ||
63928384 | 9633 | has_blocked_load |= update_nohz_stats(rq, true); |
f643ea22 | 9634 | |
ed61bbc6 TC |
9635 | /* |
9636 | * If time for next balance is due, | |
9637 | * do the balance. | |
9638 | */ | |
9639 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
9640 | struct rq_flags rf; |
9641 | ||
31e77c93 | 9642 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 9643 | update_rq_clock(rq); |
31e77c93 | 9644 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 9645 | |
b7031a02 PZ |
9646 | if (flags & NOHZ_BALANCE_KICK) |
9647 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 9648 | } |
83cd4fe2 | 9649 | |
c5afb6a8 VG |
9650 | if (time_after(next_balance, rq->next_balance)) { |
9651 | next_balance = rq->next_balance; | |
9652 | update_next_balance = 1; | |
9653 | } | |
83cd4fe2 | 9654 | } |
c5afb6a8 | 9655 | |
31e77c93 VG |
9656 | /* Newly idle CPU doesn't need an update */ |
9657 | if (idle != CPU_NEWLY_IDLE) { | |
9658 | update_blocked_averages(this_cpu); | |
9659 | has_blocked_load |= this_rq->has_blocked_load; | |
9660 | } | |
9661 | ||
b7031a02 PZ |
9662 | if (flags & NOHZ_BALANCE_KICK) |
9663 | rebalance_domains(this_rq, CPU_IDLE); | |
9664 | ||
f643ea22 VG |
9665 | WRITE_ONCE(nohz.next_blocked, |
9666 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
9667 | ||
31e77c93 VG |
9668 | /* The full idle balance loop has been done */ |
9669 | ret = true; | |
9670 | ||
f643ea22 VG |
9671 | abort: |
9672 | /* There is still blocked load, enable periodic update */ | |
9673 | if (has_blocked_load) | |
9674 | WRITE_ONCE(nohz.has_blocked, 1); | |
a4064fb6 | 9675 | |
c5afb6a8 VG |
9676 | /* |
9677 | * next_balance will be updated only when there is a need. | |
9678 | * When the CPU is attached to null domain for ex, it will not be | |
9679 | * updated. | |
9680 | */ | |
9681 | if (likely(update_next_balance)) | |
9682 | nohz.next_balance = next_balance; | |
b7031a02 | 9683 | |
31e77c93 VG |
9684 | return ret; |
9685 | } | |
9686 | ||
9687 | /* | |
9688 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
9689 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
9690 | */ | |
9691 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
9692 | { | |
9693 | int this_cpu = this_rq->cpu; | |
9694 | unsigned int flags; | |
9695 | ||
9696 | if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK)) | |
9697 | return false; | |
9698 | ||
9699 | if (idle != CPU_IDLE) { | |
9700 | atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu)); | |
9701 | return false; | |
9702 | } | |
9703 | ||
80eb8657 | 9704 | /* could be _relaxed() */ |
31e77c93 VG |
9705 | flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu)); |
9706 | if (!(flags & NOHZ_KICK_MASK)) | |
9707 | return false; | |
9708 | ||
9709 | _nohz_idle_balance(this_rq, flags, idle); | |
9710 | ||
b7031a02 | 9711 | return true; |
83cd4fe2 | 9712 | } |
31e77c93 VG |
9713 | |
9714 | static void nohz_newidle_balance(struct rq *this_rq) | |
9715 | { | |
9716 | int this_cpu = this_rq->cpu; | |
9717 | ||
9718 | /* | |
9719 | * This CPU doesn't want to be disturbed by scheduler | |
9720 | * housekeeping | |
9721 | */ | |
9722 | if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) | |
9723 | return; | |
9724 | ||
9725 | /* Will wake up very soon. No time for doing anything else*/ | |
9726 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
9727 | return; | |
9728 | ||
9729 | /* Don't need to update blocked load of idle CPUs*/ | |
9730 | if (!READ_ONCE(nohz.has_blocked) || | |
9731 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
9732 | return; | |
9733 | ||
9734 | raw_spin_unlock(&this_rq->lock); | |
9735 | /* | |
9736 | * This CPU is going to be idle and blocked load of idle CPUs | |
9737 | * need to be updated. Run the ilb locally as it is a good | |
9738 | * candidate for ilb instead of waking up another idle CPU. | |
9739 | * Kick an normal ilb if we failed to do the update. | |
9740 | */ | |
9741 | if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE)) | |
9742 | kick_ilb(NOHZ_STATS_KICK); | |
9743 | raw_spin_lock(&this_rq->lock); | |
9744 | } | |
9745 | ||
dd707247 PZ |
9746 | #else /* !CONFIG_NO_HZ_COMMON */ |
9747 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
9748 | ||
31e77c93 | 9749 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
9750 | { |
9751 | return false; | |
9752 | } | |
31e77c93 VG |
9753 | |
9754 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 9755 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 9756 | |
47ea5412 PZ |
9757 | /* |
9758 | * idle_balance is called by schedule() if this_cpu is about to become | |
9759 | * idle. Attempts to pull tasks from other CPUs. | |
9760 | */ | |
9761 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf) | |
9762 | { | |
9763 | unsigned long next_balance = jiffies + HZ; | |
9764 | int this_cpu = this_rq->cpu; | |
9765 | struct sched_domain *sd; | |
9766 | int pulled_task = 0; | |
9767 | u64 curr_cost = 0; | |
9768 | ||
9769 | /* | |
9770 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
9771 | * measure the duration of idle_balance() as idle time. | |
9772 | */ | |
9773 | this_rq->idle_stamp = rq_clock(this_rq); | |
9774 | ||
9775 | /* | |
9776 | * Do not pull tasks towards !active CPUs... | |
9777 | */ | |
9778 | if (!cpu_active(this_cpu)) | |
9779 | return 0; | |
9780 | ||
9781 | /* | |
9782 | * This is OK, because current is on_cpu, which avoids it being picked | |
9783 | * for load-balance and preemption/IRQs are still disabled avoiding | |
9784 | * further scheduler activity on it and we're being very careful to | |
9785 | * re-start the picking loop. | |
9786 | */ | |
9787 | rq_unpin_lock(this_rq, rf); | |
9788 | ||
9789 | if (this_rq->avg_idle < sysctl_sched_migration_cost || | |
e90c8fe1 | 9790 | !READ_ONCE(this_rq->rd->overload)) { |
31e77c93 | 9791 | |
47ea5412 PZ |
9792 | rcu_read_lock(); |
9793 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
9794 | if (sd) | |
9795 | update_next_balance(sd, &next_balance); | |
9796 | rcu_read_unlock(); | |
9797 | ||
31e77c93 VG |
9798 | nohz_newidle_balance(this_rq); |
9799 | ||
47ea5412 PZ |
9800 | goto out; |
9801 | } | |
9802 | ||
9803 | raw_spin_unlock(&this_rq->lock); | |
9804 | ||
9805 | update_blocked_averages(this_cpu); | |
9806 | rcu_read_lock(); | |
9807 | for_each_domain(this_cpu, sd) { | |
9808 | int continue_balancing = 1; | |
9809 | u64 t0, domain_cost; | |
9810 | ||
9811 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
9812 | continue; | |
9813 | ||
9814 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { | |
9815 | update_next_balance(sd, &next_balance); | |
9816 | break; | |
9817 | } | |
9818 | ||
9819 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
9820 | t0 = sched_clock_cpu(this_cpu); | |
9821 | ||
9822 | pulled_task = load_balance(this_cpu, this_rq, | |
9823 | sd, CPU_NEWLY_IDLE, | |
9824 | &continue_balancing); | |
9825 | ||
9826 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
9827 | if (domain_cost > sd->max_newidle_lb_cost) | |
9828 | sd->max_newidle_lb_cost = domain_cost; | |
9829 | ||
9830 | curr_cost += domain_cost; | |
9831 | } | |
9832 | ||
9833 | update_next_balance(sd, &next_balance); | |
9834 | ||
9835 | /* | |
9836 | * Stop searching for tasks to pull if there are | |
9837 | * now runnable tasks on this rq. | |
9838 | */ | |
9839 | if (pulled_task || this_rq->nr_running > 0) | |
9840 | break; | |
9841 | } | |
9842 | rcu_read_unlock(); | |
9843 | ||
9844 | raw_spin_lock(&this_rq->lock); | |
9845 | ||
9846 | if (curr_cost > this_rq->max_idle_balance_cost) | |
9847 | this_rq->max_idle_balance_cost = curr_cost; | |
9848 | ||
457be908 | 9849 | out: |
47ea5412 PZ |
9850 | /* |
9851 | * While browsing the domains, we released the rq lock, a task could | |
9852 | * have been enqueued in the meantime. Since we're not going idle, | |
9853 | * pretend we pulled a task. | |
9854 | */ | |
9855 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
9856 | pulled_task = 1; | |
9857 | ||
47ea5412 PZ |
9858 | /* Move the next balance forward */ |
9859 | if (time_after(this_rq->next_balance, next_balance)) | |
9860 | this_rq->next_balance = next_balance; | |
9861 | ||
9862 | /* Is there a task of a high priority class? */ | |
9863 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
9864 | pulled_task = -1; | |
9865 | ||
9866 | if (pulled_task) | |
9867 | this_rq->idle_stamp = 0; | |
9868 | ||
9869 | rq_repin_lock(this_rq, rf); | |
9870 | ||
9871 | return pulled_task; | |
9872 | } | |
9873 | ||
83cd4fe2 VP |
9874 | /* |
9875 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
9876 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
9877 | */ | |
0766f788 | 9878 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 9879 | { |
208cb16b | 9880 | struct rq *this_rq = this_rq(); |
6eb57e0d | 9881 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
9882 | CPU_IDLE : CPU_NOT_IDLE; |
9883 | ||
1e3c88bd | 9884 | /* |
97fb7a0a IM |
9885 | * If this CPU has a pending nohz_balance_kick, then do the |
9886 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 9887 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 9888 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
9889 | * load balance only within the local sched_domain hierarchy |
9890 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 9891 | */ |
b7031a02 PZ |
9892 | if (nohz_idle_balance(this_rq, idle)) |
9893 | return; | |
9894 | ||
9895 | /* normal load balance */ | |
9896 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 9897 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
9898 | } |
9899 | ||
1e3c88bd PZ |
9900 | /* |
9901 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 9902 | */ |
7caff66f | 9903 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 9904 | { |
1e3c88bd | 9905 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
9906 | if (unlikely(on_null_domain(rq))) |
9907 | return; | |
9908 | ||
9909 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 9910 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
9911 | |
9912 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
9913 | } |
9914 | ||
0bcdcf28 CE |
9915 | static void rq_online_fair(struct rq *rq) |
9916 | { | |
9917 | update_sysctl(); | |
0e59bdae KT |
9918 | |
9919 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
9920 | } |
9921 | ||
9922 | static void rq_offline_fair(struct rq *rq) | |
9923 | { | |
9924 | update_sysctl(); | |
a4c96ae3 PB |
9925 | |
9926 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
9927 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
9928 | } |
9929 | ||
55e12e5e | 9930 | #endif /* CONFIG_SMP */ |
e1d1484f | 9931 | |
bf0f6f24 | 9932 | /* |
d84b3131 FW |
9933 | * scheduler tick hitting a task of our scheduling class. |
9934 | * | |
9935 | * NOTE: This function can be called remotely by the tick offload that | |
9936 | * goes along full dynticks. Therefore no local assumption can be made | |
9937 | * and everything must be accessed through the @rq and @curr passed in | |
9938 | * parameters. | |
bf0f6f24 | 9939 | */ |
8f4d37ec | 9940 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
9941 | { |
9942 | struct cfs_rq *cfs_rq; | |
9943 | struct sched_entity *se = &curr->se; | |
9944 | ||
9945 | for_each_sched_entity(se) { | |
9946 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 9947 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 9948 | } |
18bf2805 | 9949 | |
b52da86e | 9950 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 9951 | task_tick_numa(rq, curr); |
3b1baa64 MR |
9952 | |
9953 | update_misfit_status(curr, rq); | |
2802bf3c | 9954 | update_overutilized_status(task_rq(curr)); |
bf0f6f24 IM |
9955 | } |
9956 | ||
9957 | /* | |
cd29fe6f PZ |
9958 | * called on fork with the child task as argument from the parent's context |
9959 | * - child not yet on the tasklist | |
9960 | * - preemption disabled | |
bf0f6f24 | 9961 | */ |
cd29fe6f | 9962 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 9963 | { |
4fc420c9 DN |
9964 | struct cfs_rq *cfs_rq; |
9965 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 9966 | struct rq *rq = this_rq(); |
8a8c69c3 | 9967 | struct rq_flags rf; |
bf0f6f24 | 9968 | |
8a8c69c3 | 9969 | rq_lock(rq, &rf); |
861d034e PZ |
9970 | update_rq_clock(rq); |
9971 | ||
4fc420c9 DN |
9972 | cfs_rq = task_cfs_rq(current); |
9973 | curr = cfs_rq->curr; | |
e210bffd PZ |
9974 | if (curr) { |
9975 | update_curr(cfs_rq); | |
b5d9d734 | 9976 | se->vruntime = curr->vruntime; |
e210bffd | 9977 | } |
aeb73b04 | 9978 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 9979 | |
cd29fe6f | 9980 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 9981 | /* |
edcb60a3 IM |
9982 | * Upon rescheduling, sched_class::put_prev_task() will place |
9983 | * 'current' within the tree based on its new key value. | |
9984 | */ | |
4d78e7b6 | 9985 | swap(curr->vruntime, se->vruntime); |
8875125e | 9986 | resched_curr(rq); |
4d78e7b6 | 9987 | } |
bf0f6f24 | 9988 | |
88ec22d3 | 9989 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 9990 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
9991 | } |
9992 | ||
cb469845 SR |
9993 | /* |
9994 | * Priority of the task has changed. Check to see if we preempt | |
9995 | * the current task. | |
9996 | */ | |
da7a735e PZ |
9997 | static void |
9998 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 9999 | { |
da0c1e65 | 10000 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
10001 | return; |
10002 | ||
cb469845 SR |
10003 | /* |
10004 | * Reschedule if we are currently running on this runqueue and | |
10005 | * our priority decreased, or if we are not currently running on | |
10006 | * this runqueue and our priority is higher than the current's | |
10007 | */ | |
da7a735e | 10008 | if (rq->curr == p) { |
cb469845 | 10009 | if (p->prio > oldprio) |
8875125e | 10010 | resched_curr(rq); |
cb469845 | 10011 | } else |
15afe09b | 10012 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
10013 | } |
10014 | ||
daa59407 | 10015 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
10016 | { |
10017 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
10018 | |
10019 | /* | |
daa59407 BP |
10020 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
10021 | * the dequeue_entity(.flags=0) will already have normalized the | |
10022 | * vruntime. | |
10023 | */ | |
10024 | if (p->on_rq) | |
10025 | return true; | |
10026 | ||
10027 | /* | |
10028 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
10029 | * But there are some cases where it has already been normalized: | |
da7a735e | 10030 | * |
daa59407 BP |
10031 | * - A forked child which is waiting for being woken up by |
10032 | * wake_up_new_task(). | |
10033 | * - A task which has been woken up by try_to_wake_up() and | |
10034 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 10035 | */ |
d0cdb3ce SM |
10036 | if (!se->sum_exec_runtime || |
10037 | (p->state == TASK_WAKING && p->sched_remote_wakeup)) | |
daa59407 BP |
10038 | return true; |
10039 | ||
10040 | return false; | |
10041 | } | |
10042 | ||
09a43ace VG |
10043 | #ifdef CONFIG_FAIR_GROUP_SCHED |
10044 | /* | |
10045 | * Propagate the changes of the sched_entity across the tg tree to make it | |
10046 | * visible to the root | |
10047 | */ | |
10048 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
10049 | { | |
10050 | struct cfs_rq *cfs_rq; | |
10051 | ||
10052 | /* Start to propagate at parent */ | |
10053 | se = se->parent; | |
10054 | ||
10055 | for_each_sched_entity(se) { | |
10056 | cfs_rq = cfs_rq_of(se); | |
10057 | ||
10058 | if (cfs_rq_throttled(cfs_rq)) | |
10059 | break; | |
10060 | ||
88c0616e | 10061 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace VG |
10062 | } |
10063 | } | |
10064 | #else | |
10065 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
10066 | #endif | |
10067 | ||
df217913 | 10068 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 10069 | { |
daa59407 BP |
10070 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
10071 | ||
9d89c257 | 10072 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 10073 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 10074 | detach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 10075 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 10076 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
10077 | } |
10078 | ||
df217913 | 10079 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 10080 | { |
daa59407 | 10081 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
10082 | |
10083 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
10084 | /* |
10085 | * Since the real-depth could have been changed (only FAIR | |
10086 | * class maintain depth value), reset depth properly. | |
10087 | */ | |
10088 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10089 | #endif | |
7855a35a | 10090 | |
df217913 | 10091 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 10092 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
ea14b57e | 10093 | attach_entity_load_avg(cfs_rq, se, 0); |
7c3edd2c | 10094 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 10095 | propagate_entity_cfs_rq(se); |
df217913 VG |
10096 | } |
10097 | ||
10098 | static void detach_task_cfs_rq(struct task_struct *p) | |
10099 | { | |
10100 | struct sched_entity *se = &p->se; | |
10101 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10102 | ||
10103 | if (!vruntime_normalized(p)) { | |
10104 | /* | |
10105 | * Fix up our vruntime so that the current sleep doesn't | |
10106 | * cause 'unlimited' sleep bonus. | |
10107 | */ | |
10108 | place_entity(cfs_rq, se, 0); | |
10109 | se->vruntime -= cfs_rq->min_vruntime; | |
10110 | } | |
10111 | ||
10112 | detach_entity_cfs_rq(se); | |
10113 | } | |
10114 | ||
10115 | static void attach_task_cfs_rq(struct task_struct *p) | |
10116 | { | |
10117 | struct sched_entity *se = &p->se; | |
10118 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10119 | ||
10120 | attach_entity_cfs_rq(se); | |
daa59407 BP |
10121 | |
10122 | if (!vruntime_normalized(p)) | |
10123 | se->vruntime += cfs_rq->min_vruntime; | |
10124 | } | |
6efdb105 | 10125 | |
daa59407 BP |
10126 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
10127 | { | |
10128 | detach_task_cfs_rq(p); | |
10129 | } | |
10130 | ||
10131 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
10132 | { | |
10133 | attach_task_cfs_rq(p); | |
7855a35a | 10134 | |
daa59407 | 10135 | if (task_on_rq_queued(p)) { |
7855a35a | 10136 | /* |
daa59407 BP |
10137 | * We were most likely switched from sched_rt, so |
10138 | * kick off the schedule if running, otherwise just see | |
10139 | * if we can still preempt the current task. | |
7855a35a | 10140 | */ |
daa59407 BP |
10141 | if (rq->curr == p) |
10142 | resched_curr(rq); | |
10143 | else | |
10144 | check_preempt_curr(rq, p, 0); | |
7855a35a | 10145 | } |
cb469845 SR |
10146 | } |
10147 | ||
83b699ed SV |
10148 | /* Account for a task changing its policy or group. |
10149 | * | |
10150 | * This routine is mostly called to set cfs_rq->curr field when a task | |
10151 | * migrates between groups/classes. | |
10152 | */ | |
10153 | static void set_curr_task_fair(struct rq *rq) | |
10154 | { | |
10155 | struct sched_entity *se = &rq->curr->se; | |
10156 | ||
ec12cb7f PT |
10157 | for_each_sched_entity(se) { |
10158 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10159 | ||
10160 | set_next_entity(cfs_rq, se); | |
10161 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
10162 | account_cfs_rq_runtime(cfs_rq, 0); | |
10163 | } | |
83b699ed SV |
10164 | } |
10165 | ||
029632fb PZ |
10166 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
10167 | { | |
bfb06889 | 10168 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
029632fb PZ |
10169 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
10170 | #ifndef CONFIG_64BIT | |
10171 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
10172 | #endif | |
141965c7 | 10173 | #ifdef CONFIG_SMP |
2a2f5d4e | 10174 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 10175 | #endif |
029632fb PZ |
10176 | } |
10177 | ||
810b3817 | 10178 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
10179 | static void task_set_group_fair(struct task_struct *p) |
10180 | { | |
10181 | struct sched_entity *se = &p->se; | |
10182 | ||
10183 | set_task_rq(p, task_cpu(p)); | |
10184 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10185 | } | |
10186 | ||
bc54da21 | 10187 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 10188 | { |
daa59407 | 10189 | detach_task_cfs_rq(p); |
b2b5ce02 | 10190 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
10191 | |
10192 | #ifdef CONFIG_SMP | |
10193 | /* Tell se's cfs_rq has been changed -- migrated */ | |
10194 | p->se.avg.last_update_time = 0; | |
10195 | #endif | |
daa59407 | 10196 | attach_task_cfs_rq(p); |
810b3817 | 10197 | } |
029632fb | 10198 | |
ea86cb4b VG |
10199 | static void task_change_group_fair(struct task_struct *p, int type) |
10200 | { | |
10201 | switch (type) { | |
10202 | case TASK_SET_GROUP: | |
10203 | task_set_group_fair(p); | |
10204 | break; | |
10205 | ||
10206 | case TASK_MOVE_GROUP: | |
10207 | task_move_group_fair(p); | |
10208 | break; | |
10209 | } | |
10210 | } | |
10211 | ||
029632fb PZ |
10212 | void free_fair_sched_group(struct task_group *tg) |
10213 | { | |
10214 | int i; | |
10215 | ||
10216 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
10217 | ||
10218 | for_each_possible_cpu(i) { | |
10219 | if (tg->cfs_rq) | |
10220 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 10221 | if (tg->se) |
029632fb PZ |
10222 | kfree(tg->se[i]); |
10223 | } | |
10224 | ||
10225 | kfree(tg->cfs_rq); | |
10226 | kfree(tg->se); | |
10227 | } | |
10228 | ||
10229 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
10230 | { | |
029632fb | 10231 | struct sched_entity *se; |
b7fa30c9 | 10232 | struct cfs_rq *cfs_rq; |
029632fb PZ |
10233 | int i; |
10234 | ||
6396bb22 | 10235 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
10236 | if (!tg->cfs_rq) |
10237 | goto err; | |
6396bb22 | 10238 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
10239 | if (!tg->se) |
10240 | goto err; | |
10241 | ||
10242 | tg->shares = NICE_0_LOAD; | |
10243 | ||
10244 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
10245 | ||
10246 | for_each_possible_cpu(i) { | |
10247 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
10248 | GFP_KERNEL, cpu_to_node(i)); | |
10249 | if (!cfs_rq) | |
10250 | goto err; | |
10251 | ||
10252 | se = kzalloc_node(sizeof(struct sched_entity), | |
10253 | GFP_KERNEL, cpu_to_node(i)); | |
10254 | if (!se) | |
10255 | goto err_free_rq; | |
10256 | ||
10257 | init_cfs_rq(cfs_rq); | |
10258 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 10259 | init_entity_runnable_average(se); |
029632fb PZ |
10260 | } |
10261 | ||
10262 | return 1; | |
10263 | ||
10264 | err_free_rq: | |
10265 | kfree(cfs_rq); | |
10266 | err: | |
10267 | return 0; | |
10268 | } | |
10269 | ||
8663e24d PZ |
10270 | void online_fair_sched_group(struct task_group *tg) |
10271 | { | |
10272 | struct sched_entity *se; | |
10273 | struct rq *rq; | |
10274 | int i; | |
10275 | ||
10276 | for_each_possible_cpu(i) { | |
10277 | rq = cpu_rq(i); | |
10278 | se = tg->se[i]; | |
10279 | ||
10280 | raw_spin_lock_irq(&rq->lock); | |
4126bad6 | 10281 | update_rq_clock(rq); |
d0326691 | 10282 | attach_entity_cfs_rq(se); |
55e16d30 | 10283 | sync_throttle(tg, i); |
8663e24d PZ |
10284 | raw_spin_unlock_irq(&rq->lock); |
10285 | } | |
10286 | } | |
10287 | ||
6fe1f348 | 10288 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 10289 | { |
029632fb | 10290 | unsigned long flags; |
6fe1f348 PZ |
10291 | struct rq *rq; |
10292 | int cpu; | |
029632fb | 10293 | |
6fe1f348 PZ |
10294 | for_each_possible_cpu(cpu) { |
10295 | if (tg->se[cpu]) | |
10296 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 10297 | |
6fe1f348 PZ |
10298 | /* |
10299 | * Only empty task groups can be destroyed; so we can speculatively | |
10300 | * check on_list without danger of it being re-added. | |
10301 | */ | |
10302 | if (!tg->cfs_rq[cpu]->on_list) | |
10303 | continue; | |
10304 | ||
10305 | rq = cpu_rq(cpu); | |
10306 | ||
10307 | raw_spin_lock_irqsave(&rq->lock, flags); | |
10308 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
10309 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
10310 | } | |
029632fb PZ |
10311 | } |
10312 | ||
10313 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
10314 | struct sched_entity *se, int cpu, | |
10315 | struct sched_entity *parent) | |
10316 | { | |
10317 | struct rq *rq = cpu_rq(cpu); | |
10318 | ||
10319 | cfs_rq->tg = tg; | |
10320 | cfs_rq->rq = rq; | |
029632fb PZ |
10321 | init_cfs_rq_runtime(cfs_rq); |
10322 | ||
10323 | tg->cfs_rq[cpu] = cfs_rq; | |
10324 | tg->se[cpu] = se; | |
10325 | ||
10326 | /* se could be NULL for root_task_group */ | |
10327 | if (!se) | |
10328 | return; | |
10329 | ||
fed14d45 | 10330 | if (!parent) { |
029632fb | 10331 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
10332 | se->depth = 0; |
10333 | } else { | |
029632fb | 10334 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
10335 | se->depth = parent->depth + 1; |
10336 | } | |
029632fb PZ |
10337 | |
10338 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
10339 | /* guarantee group entities always have weight */ |
10340 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
10341 | se->parent = parent; |
10342 | } | |
10343 | ||
10344 | static DEFINE_MUTEX(shares_mutex); | |
10345 | ||
10346 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
10347 | { | |
10348 | int i; | |
029632fb PZ |
10349 | |
10350 | /* | |
10351 | * We can't change the weight of the root cgroup. | |
10352 | */ | |
10353 | if (!tg->se[0]) | |
10354 | return -EINVAL; | |
10355 | ||
10356 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
10357 | ||
10358 | mutex_lock(&shares_mutex); | |
10359 | if (tg->shares == shares) | |
10360 | goto done; | |
10361 | ||
10362 | tg->shares = shares; | |
10363 | for_each_possible_cpu(i) { | |
10364 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
10365 | struct sched_entity *se = tg->se[i]; |
10366 | struct rq_flags rf; | |
029632fb | 10367 | |
029632fb | 10368 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 10369 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 10370 | update_rq_clock(rq); |
89ee048f | 10371 | for_each_sched_entity(se) { |
88c0616e | 10372 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 10373 | update_cfs_group(se); |
89ee048f | 10374 | } |
8a8c69c3 | 10375 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
10376 | } |
10377 | ||
10378 | done: | |
10379 | mutex_unlock(&shares_mutex); | |
10380 | return 0; | |
10381 | } | |
10382 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
10383 | ||
10384 | void free_fair_sched_group(struct task_group *tg) { } | |
10385 | ||
10386 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
10387 | { | |
10388 | return 1; | |
10389 | } | |
10390 | ||
8663e24d PZ |
10391 | void online_fair_sched_group(struct task_group *tg) { } |
10392 | ||
6fe1f348 | 10393 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
10394 | |
10395 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
10396 | ||
810b3817 | 10397 | |
6d686f45 | 10398 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
10399 | { |
10400 | struct sched_entity *se = &task->se; | |
0d721cea PW |
10401 | unsigned int rr_interval = 0; |
10402 | ||
10403 | /* | |
10404 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
10405 | * idle runqueue: | |
10406 | */ | |
0d721cea | 10407 | if (rq->cfs.load.weight) |
a59f4e07 | 10408 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
10409 | |
10410 | return rr_interval; | |
10411 | } | |
10412 | ||
bf0f6f24 IM |
10413 | /* |
10414 | * All the scheduling class methods: | |
10415 | */ | |
029632fb | 10416 | const struct sched_class fair_sched_class = { |
5522d5d5 | 10417 | .next = &idle_sched_class, |
bf0f6f24 IM |
10418 | .enqueue_task = enqueue_task_fair, |
10419 | .dequeue_task = dequeue_task_fair, | |
10420 | .yield_task = yield_task_fair, | |
d95f4122 | 10421 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 10422 | |
2e09bf55 | 10423 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
10424 | |
10425 | .pick_next_task = pick_next_task_fair, | |
10426 | .put_prev_task = put_prev_task_fair, | |
10427 | ||
681f3e68 | 10428 | #ifdef CONFIG_SMP |
4ce72a2c | 10429 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 10430 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 10431 | |
0bcdcf28 CE |
10432 | .rq_online = rq_online_fair, |
10433 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 10434 | |
12695578 | 10435 | .task_dead = task_dead_fair, |
c5b28038 | 10436 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 10437 | #endif |
bf0f6f24 | 10438 | |
83b699ed | 10439 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 10440 | .task_tick = task_tick_fair, |
cd29fe6f | 10441 | .task_fork = task_fork_fair, |
cb469845 SR |
10442 | |
10443 | .prio_changed = prio_changed_fair, | |
da7a735e | 10444 | .switched_from = switched_from_fair, |
cb469845 | 10445 | .switched_to = switched_to_fair, |
810b3817 | 10446 | |
0d721cea PW |
10447 | .get_rr_interval = get_rr_interval_fair, |
10448 | ||
6e998916 SG |
10449 | .update_curr = update_curr_fair, |
10450 | ||
810b3817 | 10451 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 10452 | .task_change_group = task_change_group_fair, |
810b3817 | 10453 | #endif |
982d9cdc PB |
10454 | |
10455 | #ifdef CONFIG_UCLAMP_TASK | |
10456 | .uclamp_enabled = 1, | |
10457 | #endif | |
bf0f6f24 IM |
10458 | }; |
10459 | ||
10460 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 10461 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 10462 | { |
039ae8bc | 10463 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 10464 | |
5973e5b9 | 10465 | rcu_read_lock(); |
039ae8bc | 10466 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 10467 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 10468 | rcu_read_unlock(); |
bf0f6f24 | 10469 | } |
397f2378 SD |
10470 | |
10471 | #ifdef CONFIG_NUMA_BALANCING | |
10472 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
10473 | { | |
10474 | int node; | |
10475 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 10476 | struct numa_group *ng; |
397f2378 | 10477 | |
cb361d8c JH |
10478 | rcu_read_lock(); |
10479 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
10480 | for_each_online_node(node) { |
10481 | if (p->numa_faults) { | |
10482 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
10483 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
10484 | } | |
cb361d8c JH |
10485 | if (ng) { |
10486 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
10487 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
10488 | } |
10489 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
10490 | } | |
cb361d8c | 10491 | rcu_read_unlock(); |
397f2378 SD |
10492 | } |
10493 | #endif /* CONFIG_NUMA_BALANCING */ | |
10494 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
10495 | |
10496 | __init void init_sched_fair_class(void) | |
10497 | { | |
10498 | #ifdef CONFIG_SMP | |
10499 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
10500 | ||
3451d024 | 10501 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 10502 | nohz.next_balance = jiffies; |
f643ea22 | 10503 | nohz.next_blocked = jiffies; |
029632fb | 10504 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
10505 | #endif |
10506 | #endif /* SMP */ | |
10507 | ||
10508 | } | |
3c93a0c0 QY |
10509 | |
10510 | /* | |
10511 | * Helper functions to facilitate extracting info from tracepoints. | |
10512 | */ | |
10513 | ||
10514 | const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq) | |
10515 | { | |
10516 | #ifdef CONFIG_SMP | |
10517 | return cfs_rq ? &cfs_rq->avg : NULL; | |
10518 | #else | |
10519 | return NULL; | |
10520 | #endif | |
10521 | } | |
10522 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg); | |
10523 | ||
10524 | char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len) | |
10525 | { | |
10526 | if (!cfs_rq) { | |
10527 | if (str) | |
10528 | strlcpy(str, "(null)", len); | |
10529 | else | |
10530 | return NULL; | |
10531 | } | |
10532 | ||
10533 | cfs_rq_tg_path(cfs_rq, str, len); | |
10534 | return str; | |
10535 | } | |
10536 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path); | |
10537 | ||
10538 | int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq) | |
10539 | { | |
10540 | return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; | |
10541 | } | |
10542 | EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu); | |
10543 | ||
10544 | const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq) | |
10545 | { | |
10546 | #ifdef CONFIG_SMP | |
10547 | return rq ? &rq->avg_rt : NULL; | |
10548 | #else | |
10549 | return NULL; | |
10550 | #endif | |
10551 | } | |
10552 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt); | |
10553 | ||
10554 | const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq) | |
10555 | { | |
10556 | #ifdef CONFIG_SMP | |
10557 | return rq ? &rq->avg_dl : NULL; | |
10558 | #else | |
10559 | return NULL; | |
10560 | #endif | |
10561 | } | |
10562 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl); | |
10563 | ||
10564 | const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq) | |
10565 | { | |
10566 | #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ) | |
10567 | return rq ? &rq->avg_irq : NULL; | |
10568 | #else | |
10569 | return NULL; | |
10570 | #endif | |
10571 | } | |
10572 | EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq); | |
10573 | ||
10574 | int sched_trace_rq_cpu(struct rq *rq) | |
10575 | { | |
10576 | return rq ? cpu_of(rq) : -1; | |
10577 | } | |
10578 | EXPORT_SYMBOL_GPL(sched_trace_rq_cpu); | |
10579 | ||
10580 | const struct cpumask *sched_trace_rd_span(struct root_domain *rd) | |
10581 | { | |
10582 | #ifdef CONFIG_SMP | |
10583 | return rd ? rd->span : NULL; | |
10584 | #else | |
10585 | return NULL; | |
10586 | #endif | |
10587 | } | |
10588 | EXPORT_SYMBOL_GPL(sched_trace_rd_span); |