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 | */ |
c4ad6fcb IM |
23 | #include <linux/energy_model.h> |
24 | #include <linux/mmap_lock.h> | |
25 | #include <linux/hugetlb_inline.h> | |
26 | #include <linux/jiffies.h> | |
27 | #include <linux/mm_api.h> | |
28 | #include <linux/highmem.h> | |
29 | #include <linux/spinlock_api.h> | |
30 | #include <linux/cpumask_api.h> | |
31 | #include <linux/lockdep_api.h> | |
32 | #include <linux/softirq.h> | |
33 | #include <linux/refcount_api.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/sched/clock.h> | |
36 | #include <linux/sched/cond_resched.h> | |
37 | #include <linux/sched/cputime.h> | |
38 | #include <linux/sched/isolation.h> | |
d664e399 | 39 | #include <linux/sched/nohz.h> |
c4ad6fcb IM |
40 | |
41 | #include <linux/cpuidle.h> | |
42 | #include <linux/interrupt.h> | |
467b171a | 43 | #include <linux/memory-tiers.h> |
c4ad6fcb IM |
44 | #include <linux/mempolicy.h> |
45 | #include <linux/mutex_api.h> | |
46 | #include <linux/profile.h> | |
47 | #include <linux/psi.h> | |
48 | #include <linux/ratelimit.h> | |
1930a6e7 | 49 | #include <linux/task_work.h> |
147f3efa | 50 | #include <linux/rbtree_augmented.h> |
c4ad6fcb IM |
51 | |
52 | #include <asm/switch_to.h> | |
53 | ||
325ea10c | 54 | #include "sched.h" |
b9e9c6ca IM |
55 | #include "stats.h" |
56 | #include "autogroup.h" | |
029632fb | 57 | |
1983a922 CE |
58 | /* |
59 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
60 | * |
61 | * Options are: | |
2b4d5b25 IM |
62 | * |
63 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
64 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
65 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
66 | * | |
67 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 68 | */ |
8a99b683 | 69 | unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 70 | |
2bd8e6d4 | 71 | /* |
b2be5e96 | 72 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 73 | * |
864616ee | 74 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 75 | */ |
e4ec3318 PZ |
76 | unsigned int sysctl_sched_base_slice = 750000ULL; |
77 | static unsigned int normalized_sysctl_sched_base_slice = 750000ULL; | |
b2be5e96 | 78 | |
2b4d5b25 | 79 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 80 | |
05289b90 TG |
81 | static int __init setup_sched_thermal_decay_shift(char *str) |
82 | { | |
97450eb9 | 83 | pr_warn("Ignoring the deprecated sched_thermal_decay_shift= option\n"); |
05289b90 TG |
84 | return 1; |
85 | } | |
86 | __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift); | |
87 | ||
afe06efd TC |
88 | #ifdef CONFIG_SMP |
89 | /* | |
97fb7a0a | 90 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
91 | */ |
92 | int __weak arch_asym_cpu_priority(int cpu) | |
93 | { | |
94 | return -cpu; | |
95 | } | |
6d101ba6 OJ |
96 | |
97 | /* | |
60e17f5c | 98 | * The margin used when comparing utilization with CPU capacity. |
6d101ba6 OJ |
99 | * |
100 | * (default: ~20%) | |
101 | */ | |
60e17f5c VK |
102 | #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) |
103 | ||
4aed8aa4 VS |
104 | /* |
105 | * The margin used when comparing CPU capacities. | |
106 | * is 'cap1' noticeably greater than 'cap2' | |
107 | * | |
108 | * (default: ~5%) | |
109 | */ | |
110 | #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078) | |
afe06efd TC |
111 | #endif |
112 | ||
ec12cb7f PT |
113 | #ifdef CONFIG_CFS_BANDWIDTH |
114 | /* | |
115 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
116 | * each time a cfs_rq requests quota. | |
117 | * | |
118 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
119 | * to consumption or the quota being specified to be smaller than the slice) | |
120 | * we will always only issue the remaining available time. | |
121 | * | |
2b4d5b25 IM |
122 | * (default: 5 msec, units: microseconds) |
123 | */ | |
d4ae80ff ZN |
124 | static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; |
125 | #endif | |
126 | ||
0dff89c4 KW |
127 | #ifdef CONFIG_NUMA_BALANCING |
128 | /* Restrict the NUMA promotion throughput (MB/s) for each target node. */ | |
129 | static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536; | |
130 | #endif | |
131 | ||
d4ae80ff ZN |
132 | #ifdef CONFIG_SYSCTL |
133 | static struct ctl_table sched_fair_sysctls[] = { | |
d4ae80ff ZN |
134 | #ifdef CONFIG_CFS_BANDWIDTH |
135 | { | |
136 | .procname = "sched_cfs_bandwidth_slice_us", | |
137 | .data = &sysctl_sched_cfs_bandwidth_slice, | |
138 | .maxlen = sizeof(unsigned int), | |
139 | .mode = 0644, | |
140 | .proc_handler = proc_dointvec_minmax, | |
141 | .extra1 = SYSCTL_ONE, | |
142 | }, | |
143 | #endif | |
0dff89c4 KW |
144 | #ifdef CONFIG_NUMA_BALANCING |
145 | { | |
146 | .procname = "numa_balancing_promote_rate_limit_MBps", | |
147 | .data = &sysctl_numa_balancing_promote_rate_limit, | |
148 | .maxlen = sizeof(unsigned int), | |
149 | .mode = 0644, | |
150 | .proc_handler = proc_dointvec_minmax, | |
151 | .extra1 = SYSCTL_ZERO, | |
152 | }, | |
153 | #endif /* CONFIG_NUMA_BALANCING */ | |
d4ae80ff ZN |
154 | }; |
155 | ||
156 | static int __init sched_fair_sysctl_init(void) | |
157 | { | |
158 | register_sysctl_init("kernel", sched_fair_sysctls); | |
159 | return 0; | |
160 | } | |
161 | late_initcall(sched_fair_sysctl_init); | |
ec12cb7f PT |
162 | #endif |
163 | ||
8527632d PG |
164 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
165 | { | |
166 | lw->weight += inc; | |
167 | lw->inv_weight = 0; | |
168 | } | |
169 | ||
170 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
171 | { | |
172 | lw->weight -= dec; | |
173 | lw->inv_weight = 0; | |
174 | } | |
175 | ||
176 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
177 | { | |
178 | lw->weight = w; | |
179 | lw->inv_weight = 0; | |
180 | } | |
181 | ||
029632fb PZ |
182 | /* |
183 | * Increase the granularity value when there are more CPUs, | |
184 | * because with more CPUs the 'effective latency' as visible | |
185 | * to users decreases. But the relationship is not linear, | |
186 | * so pick a second-best guess by going with the log2 of the | |
187 | * number of CPUs. | |
188 | * | |
189 | * This idea comes from the SD scheduler of Con Kolivas: | |
190 | */ | |
58ac93e4 | 191 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 192 | { |
58ac93e4 | 193 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
194 | unsigned int factor; |
195 | ||
196 | switch (sysctl_sched_tunable_scaling) { | |
197 | case SCHED_TUNABLESCALING_NONE: | |
198 | factor = 1; | |
199 | break; | |
200 | case SCHED_TUNABLESCALING_LINEAR: | |
201 | factor = cpus; | |
202 | break; | |
203 | case SCHED_TUNABLESCALING_LOG: | |
204 | default: | |
205 | factor = 1 + ilog2(cpus); | |
206 | break; | |
207 | } | |
208 | ||
209 | return factor; | |
210 | } | |
211 | ||
212 | static void update_sysctl(void) | |
213 | { | |
214 | unsigned int factor = get_update_sysctl_factor(); | |
215 | ||
216 | #define SET_SYSCTL(name) \ | |
217 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
e4ec3318 | 218 | SET_SYSCTL(sched_base_slice); |
029632fb PZ |
219 | #undef SET_SYSCTL |
220 | } | |
221 | ||
f38f12d1 | 222 | void __init sched_init_granularity(void) |
029632fb PZ |
223 | { |
224 | update_sysctl(); | |
225 | } | |
226 | ||
9dbdb155 | 227 | #define WMULT_CONST (~0U) |
029632fb PZ |
228 | #define WMULT_SHIFT 32 |
229 | ||
9dbdb155 PZ |
230 | static void __update_inv_weight(struct load_weight *lw) |
231 | { | |
232 | unsigned long w; | |
233 | ||
234 | if (likely(lw->inv_weight)) | |
235 | return; | |
236 | ||
237 | w = scale_load_down(lw->weight); | |
238 | ||
239 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
240 | lw->inv_weight = 1; | |
241 | else if (unlikely(!w)) | |
242 | lw->inv_weight = WMULT_CONST; | |
243 | else | |
244 | lw->inv_weight = WMULT_CONST / w; | |
245 | } | |
029632fb PZ |
246 | |
247 | /* | |
9dbdb155 PZ |
248 | * delta_exec * weight / lw.weight |
249 | * OR | |
250 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
251 | * | |
1c3de5e1 | 252 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
253 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
254 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
255 | * | |
256 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
257 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 258 | */ |
9dbdb155 | 259 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 260 | { |
9dbdb155 | 261 | u64 fact = scale_load_down(weight); |
1e17fb8e | 262 | u32 fact_hi = (u32)(fact >> 32); |
9dbdb155 | 263 | int shift = WMULT_SHIFT; |
1e17fb8e | 264 | int fs; |
029632fb | 265 | |
9dbdb155 | 266 | __update_inv_weight(lw); |
029632fb | 267 | |
1e17fb8e CC |
268 | if (unlikely(fact_hi)) { |
269 | fs = fls(fact_hi); | |
270 | shift -= fs; | |
271 | fact >>= fs; | |
029632fb PZ |
272 | } |
273 | ||
2eeb01a2 | 274 | fact = mul_u32_u32(fact, lw->inv_weight); |
029632fb | 275 | |
1e17fb8e CC |
276 | fact_hi = (u32)(fact >> 32); |
277 | if (fact_hi) { | |
278 | fs = fls(fact_hi); | |
279 | shift -= fs; | |
280 | fact >>= fs; | |
9dbdb155 | 281 | } |
029632fb | 282 | |
9dbdb155 | 283 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
284 | } |
285 | ||
147f3efa PZ |
286 | /* |
287 | * delta /= w | |
288 | */ | |
289 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) | |
290 | { | |
291 | if (unlikely(se->load.weight != NICE_0_LOAD)) | |
292 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); | |
293 | ||
294 | return delta; | |
295 | } | |
029632fb PZ |
296 | |
297 | const struct sched_class fair_sched_class; | |
a4c2f00f | 298 | |
bf0f6f24 IM |
299 | /************************************************************** |
300 | * CFS operations on generic schedulable entities: | |
301 | */ | |
302 | ||
62160e3f | 303 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8f48894f | 304 | |
b758149c PZ |
305 | /* Walk up scheduling entities hierarchy */ |
306 | #define for_each_sched_entity(se) \ | |
307 | for (; se; se = se->parent) | |
308 | ||
f6783319 | 309 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 310 | { |
5d299eab PZ |
311 | struct rq *rq = rq_of(cfs_rq); |
312 | int cpu = cpu_of(rq); | |
313 | ||
314 | if (cfs_rq->on_list) | |
f6783319 | 315 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
316 | |
317 | cfs_rq->on_list = 1; | |
318 | ||
319 | /* | |
320 | * Ensure we either appear before our parent (if already | |
321 | * enqueued) or force our parent to appear after us when it is | |
322 | * enqueued. The fact that we always enqueue bottom-up | |
323 | * reduces this to two cases and a special case for the root | |
324 | * cfs_rq. Furthermore, it also means that we will always reset | |
325 | * tmp_alone_branch either when the branch is connected | |
326 | * to a tree or when we reach the top of the tree | |
327 | */ | |
328 | if (cfs_rq->tg->parent && | |
329 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 330 | /* |
5d299eab PZ |
331 | * If parent is already on the list, we add the child |
332 | * just before. Thanks to circular linked property of | |
333 | * the list, this means to put the child at the tail | |
334 | * of the list that starts by parent. | |
67e86250 | 335 | */ |
5d299eab PZ |
336 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
337 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
338 | /* | |
339 | * The branch is now connected to its tree so we can | |
340 | * reset tmp_alone_branch to the beginning of the | |
341 | * list. | |
342 | */ | |
343 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 344 | return true; |
5d299eab | 345 | } |
3d4b47b4 | 346 | |
5d299eab PZ |
347 | if (!cfs_rq->tg->parent) { |
348 | /* | |
349 | * cfs rq without parent should be put | |
350 | * at the tail of the list. | |
351 | */ | |
352 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
353 | &rq->leaf_cfs_rq_list); | |
354 | /* | |
355 | * We have reach the top of a tree so we can reset | |
356 | * tmp_alone_branch to the beginning of the list. | |
357 | */ | |
358 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 359 | return true; |
3d4b47b4 | 360 | } |
5d299eab PZ |
361 | |
362 | /* | |
363 | * The parent has not already been added so we want to | |
364 | * make sure that it will be put after us. | |
365 | * tmp_alone_branch points to the begin of the branch | |
366 | * where we will add parent. | |
367 | */ | |
368 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
369 | /* | |
370 | * update tmp_alone_branch to points to the new begin | |
371 | * of the branch | |
372 | */ | |
373 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 374 | return false; |
3d4b47b4 PZ |
375 | } |
376 | ||
377 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
378 | { | |
379 | if (cfs_rq->on_list) { | |
31bc6aea VG |
380 | struct rq *rq = rq_of(cfs_rq); |
381 | ||
382 | /* | |
383 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
b9e6e286 IM |
384 | * it can happen the tmp_alone_branch points to the leaf that |
385 | * we finally want to delete. In this case, tmp_alone_branch moves | |
31bc6aea VG |
386 | * to the prev element but it will point to rq->leaf_cfs_rq_list |
387 | * at the end of the enqueue. | |
388 | */ | |
389 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
390 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
391 | ||
3d4b47b4 PZ |
392 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
393 | cfs_rq->on_list = 0; | |
394 | } | |
395 | } | |
396 | ||
5d299eab PZ |
397 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
398 | { | |
399 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
400 | } | |
401 | ||
b9e6e286 | 402 | /* Iterate through all leaf cfs_rq's on a runqueue */ |
039ae8bc VG |
403 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
404 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
405 | leaf_cfs_rq_list) | |
b758149c PZ |
406 | |
407 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 408 | static inline struct cfs_rq * |
b758149c PZ |
409 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
410 | { | |
411 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 412 | return se->cfs_rq; |
b758149c | 413 | |
fed14d45 | 414 | return NULL; |
b758149c PZ |
415 | } |
416 | ||
904cbab7 | 417 | static inline struct sched_entity *parent_entity(const struct sched_entity *se) |
b758149c PZ |
418 | { |
419 | return se->parent; | |
420 | } | |
421 | ||
464b7527 PZ |
422 | static void |
423 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
424 | { | |
425 | int se_depth, pse_depth; | |
426 | ||
427 | /* | |
428 | * preemption test can be made between sibling entities who are in the | |
429 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
430 | * both tasks until we find their ancestors who are siblings of common | |
431 | * parent. | |
432 | */ | |
433 | ||
434 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
435 | se_depth = (*se)->depth; |
436 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
437 | |
438 | while (se_depth > pse_depth) { | |
439 | se_depth--; | |
440 | *se = parent_entity(*se); | |
441 | } | |
442 | ||
443 | while (pse_depth > se_depth) { | |
444 | pse_depth--; | |
445 | *pse = parent_entity(*pse); | |
446 | } | |
447 | ||
448 | while (!is_same_group(*se, *pse)) { | |
449 | *se = parent_entity(*se); | |
450 | *pse = parent_entity(*pse); | |
451 | } | |
452 | } | |
453 | ||
30400039 JD |
454 | static int tg_is_idle(struct task_group *tg) |
455 | { | |
456 | return tg->idle > 0; | |
457 | } | |
458 | ||
459 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
460 | { | |
461 | return cfs_rq->idle > 0; | |
462 | } | |
463 | ||
464 | static int se_is_idle(struct sched_entity *se) | |
465 | { | |
466 | if (entity_is_task(se)) | |
467 | return task_has_idle_policy(task_of(se)); | |
468 | return cfs_rq_is_idle(group_cfs_rq(se)); | |
469 | } | |
470 | ||
8f48894f PZ |
471 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
472 | ||
b758149c PZ |
473 | #define for_each_sched_entity(se) \ |
474 | for (; se; se = NULL) | |
bf0f6f24 | 475 | |
f6783319 | 476 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 477 | { |
f6783319 | 478 | return true; |
3d4b47b4 PZ |
479 | } |
480 | ||
481 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
482 | { | |
483 | } | |
484 | ||
5d299eab PZ |
485 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
486 | { | |
487 | } | |
488 | ||
039ae8bc VG |
489 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
490 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 491 | |
b758149c PZ |
492 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
493 | { | |
494 | return NULL; | |
495 | } | |
496 | ||
464b7527 PZ |
497 | static inline void |
498 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
499 | { | |
500 | } | |
501 | ||
366e7ad6 | 502 | static inline int tg_is_idle(struct task_group *tg) |
30400039 JD |
503 | { |
504 | return 0; | |
505 | } | |
506 | ||
507 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
508 | { | |
509 | return 0; | |
510 | } | |
511 | ||
512 | static int se_is_idle(struct sched_entity *se) | |
513 | { | |
514 | return 0; | |
515 | } | |
516 | ||
b758149c PZ |
517 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
518 | ||
6c16a6dc | 519 | static __always_inline |
9dbdb155 | 520 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
521 | |
522 | /************************************************************** | |
523 | * Scheduling class tree data structure manipulation methods: | |
524 | */ | |
525 | ||
1bf08230 | 526 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 527 | { |
1bf08230 | 528 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 529 | if (delta > 0) |
1bf08230 | 530 | max_vruntime = vruntime; |
02e0431a | 531 | |
1bf08230 | 532 | return max_vruntime; |
02e0431a PZ |
533 | } |
534 | ||
0702e3eb | 535 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
536 | { |
537 | s64 delta = (s64)(vruntime - min_vruntime); | |
538 | if (delta < 0) | |
539 | min_vruntime = vruntime; | |
540 | ||
541 | return min_vruntime; | |
542 | } | |
543 | ||
904cbab7 MWO |
544 | static inline bool entity_before(const struct sched_entity *a, |
545 | const struct sched_entity *b) | |
54fdc581 | 546 | { |
2227a957 AW |
547 | /* |
548 | * Tiebreak on vruntime seems unnecessary since it can | |
549 | * hardly happen. | |
550 | */ | |
551 | return (s64)(a->deadline - b->deadline) < 0; | |
54fdc581 FC |
552 | } |
553 | ||
af4cf404 PZ |
554 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
555 | { | |
556 | return (s64)(se->vruntime - cfs_rq->min_vruntime); | |
557 | } | |
558 | ||
bf9be9a1 PZ |
559 | #define __node_2_se(node) \ |
560 | rb_entry((node), struct sched_entity, run_node) | |
561 | ||
af4cf404 PZ |
562 | /* |
563 | * Compute virtual time from the per-task service numbers: | |
564 | * | |
565 | * Fair schedulers conserve lag: | |
566 | * | |
567 | * \Sum lag_i = 0 | |
568 | * | |
569 | * Where lag_i is given by: | |
570 | * | |
571 | * lag_i = S - s_i = w_i * (V - v_i) | |
572 | * | |
573 | * Where S is the ideal service time and V is it's virtual time counterpart. | |
574 | * Therefore: | |
575 | * | |
576 | * \Sum lag_i = 0 | |
577 | * \Sum w_i * (V - v_i) = 0 | |
578 | * \Sum w_i * V - w_i * v_i = 0 | |
579 | * | |
580 | * From which we can solve an expression for V in v_i (which we have in | |
581 | * se->vruntime): | |
582 | * | |
583 | * \Sum v_i * w_i \Sum v_i * w_i | |
584 | * V = -------------- = -------------- | |
585 | * \Sum w_i W | |
586 | * | |
587 | * Specifically, this is the weighted average of all entity virtual runtimes. | |
588 | * | |
589 | * [[ NOTE: this is only equal to the ideal scheduler under the condition | |
590 | * that join/leave operations happen at lag_i = 0, otherwise the | |
b9e6e286 | 591 | * virtual time has non-contiguous motion equivalent to: |
af4cf404 PZ |
592 | * |
593 | * V +-= lag_i / W | |
594 | * | |
595 | * Also see the comment in place_entity() that deals with this. ]] | |
596 | * | |
b9e6e286 | 597 | * However, since v_i is u64, and the multiplication could easily overflow |
af4cf404 PZ |
598 | * transform it into a relative form that uses smaller quantities: |
599 | * | |
600 | * Substitute: v_i == (v_i - v0) + v0 | |
601 | * | |
602 | * \Sum ((v_i - v0) + v0) * w_i \Sum (v_i - v0) * w_i | |
603 | * V = ---------------------------- = --------------------- + v0 | |
604 | * W W | |
605 | * | |
606 | * Which we track using: | |
607 | * | |
608 | * v0 := cfs_rq->min_vruntime | |
609 | * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime | |
610 | * \Sum w_i := cfs_rq->avg_load | |
611 | * | |
612 | * Since min_vruntime is a monotonic increasing variable that closely tracks | |
613 | * the per-task service, these deltas: (v_i - v), will be in the order of the | |
614 | * maximal (virtual) lag induced in the system due to quantisation. | |
615 | * | |
616 | * Also, we use scale_load_down() to reduce the size. | |
617 | * | |
618 | * As measured, the max (key * weight) value was ~44 bits for a kernel build. | |
619 | */ | |
620 | static void | |
621 | avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
622 | { | |
623 | unsigned long weight = scale_load_down(se->load.weight); | |
624 | s64 key = entity_key(cfs_rq, se); | |
625 | ||
626 | cfs_rq->avg_vruntime += key * weight; | |
627 | cfs_rq->avg_load += weight; | |
628 | } | |
629 | ||
630 | static void | |
631 | avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
632 | { | |
633 | unsigned long weight = scale_load_down(se->load.weight); | |
634 | s64 key = entity_key(cfs_rq, se); | |
635 | ||
636 | cfs_rq->avg_vruntime -= key * weight; | |
637 | cfs_rq->avg_load -= weight; | |
638 | } | |
639 | ||
640 | static inline | |
641 | void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta) | |
642 | { | |
643 | /* | |
644 | * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load | |
645 | */ | |
646 | cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta; | |
647 | } | |
648 | ||
650cad56 PZ |
649 | /* |
650 | * Specifically: avg_runtime() + 0 must result in entity_eligible() := true | |
651 | * For this to be so, the result of this function must have a left bias. | |
652 | */ | |
af4cf404 PZ |
653 | u64 avg_vruntime(struct cfs_rq *cfs_rq) |
654 | { | |
655 | struct sched_entity *curr = cfs_rq->curr; | |
656 | s64 avg = cfs_rq->avg_vruntime; | |
657 | long load = cfs_rq->avg_load; | |
658 | ||
659 | if (curr && curr->on_rq) { | |
660 | unsigned long weight = scale_load_down(curr->load.weight); | |
661 | ||
662 | avg += entity_key(cfs_rq, curr) * weight; | |
663 | load += weight; | |
664 | } | |
665 | ||
650cad56 | 666 | if (load) { |
b9e6e286 | 667 | /* sign flips effective floor / ceiling */ |
650cad56 PZ |
668 | if (avg < 0) |
669 | avg -= (load - 1); | |
af4cf404 | 670 | avg = div_s64(avg, load); |
650cad56 | 671 | } |
af4cf404 PZ |
672 | |
673 | return cfs_rq->min_vruntime + avg; | |
674 | } | |
675 | ||
86bfbb7c PZ |
676 | /* |
677 | * lag_i = S - s_i = w_i * (V - v_i) | |
147f3efa PZ |
678 | * |
679 | * However, since V is approximated by the weighted average of all entities it | |
680 | * is possible -- by addition/removal/reweight to the tree -- to move V around | |
681 | * and end up with a larger lag than we started with. | |
682 | * | |
683 | * Limit this to either double the slice length with a minimum of TICK_NSEC | |
684 | * since that is the timing granularity. | |
685 | * | |
686 | * EEVDF gives the following limit for a steady state system: | |
687 | * | |
688 | * -r_max < lag < max(r_max, q) | |
689 | * | |
690 | * XXX could add max_slice to the augmented data to track this. | |
86bfbb7c | 691 | */ |
1560d1f6 | 692 | static s64 entity_lag(u64 avruntime, struct sched_entity *se) |
86bfbb7c | 693 | { |
1560d1f6 XY |
694 | s64 vlag, limit; |
695 | ||
696 | vlag = avruntime - se->vruntime; | |
697 | limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se); | |
698 | ||
699 | return clamp(vlag, -limit, limit); | |
700 | } | |
147f3efa | 701 | |
1560d1f6 XY |
702 | static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se) |
703 | { | |
86bfbb7c | 704 | SCHED_WARN_ON(!se->on_rq); |
147f3efa | 705 | |
1560d1f6 | 706 | se->vlag = entity_lag(avg_vruntime(cfs_rq), se); |
147f3efa PZ |
707 | } |
708 | ||
709 | /* | |
710 | * Entity is eligible once it received less service than it ought to have, | |
711 | * eg. lag >= 0. | |
712 | * | |
713 | * lag_i = S - s_i = w_i*(V - v_i) | |
714 | * | |
715 | * lag_i >= 0 -> V >= v_i | |
716 | * | |
717 | * \Sum (v_i - v)*w_i | |
718 | * V = ------------------ + v | |
719 | * \Sum w_i | |
720 | * | |
721 | * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i) | |
722 | * | |
b9e6e286 | 723 | * Note: using 'avg_vruntime() > se->vruntime' is inaccurate due |
147f3efa PZ |
724 | * to the loss in precision caused by the division. |
725 | */ | |
2227a957 | 726 | static int vruntime_eligible(struct cfs_rq *cfs_rq, u64 vruntime) |
147f3efa PZ |
727 | { |
728 | struct sched_entity *curr = cfs_rq->curr; | |
729 | s64 avg = cfs_rq->avg_vruntime; | |
730 | long load = cfs_rq->avg_load; | |
731 | ||
732 | if (curr && curr->on_rq) { | |
733 | unsigned long weight = scale_load_down(curr->load.weight); | |
734 | ||
735 | avg += entity_key(cfs_rq, curr) * weight; | |
736 | load += weight; | |
737 | } | |
738 | ||
2227a957 AW |
739 | return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load; |
740 | } | |
741 | ||
742 | int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
743 | { | |
744 | return vruntime_eligible(cfs_rq, se->vruntime); | |
86bfbb7c PZ |
745 | } |
746 | ||
af4cf404 PZ |
747 | static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime) |
748 | { | |
749 | u64 min_vruntime = cfs_rq->min_vruntime; | |
750 | /* | |
751 | * open coded max_vruntime() to allow updating avg_vruntime | |
752 | */ | |
753 | s64 delta = (s64)(vruntime - min_vruntime); | |
754 | if (delta > 0) { | |
755 | avg_vruntime_update(cfs_rq, delta); | |
756 | min_vruntime = vruntime; | |
757 | } | |
758 | return min_vruntime; | |
759 | } | |
760 | ||
1af5f730 PZ |
761 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
762 | { | |
2227a957 | 763 | struct sched_entity *se = __pick_root_entity(cfs_rq); |
b60205c7 | 764 | struct sched_entity *curr = cfs_rq->curr; |
1af5f730 PZ |
765 | u64 vruntime = cfs_rq->min_vruntime; |
766 | ||
b60205c7 PZ |
767 | if (curr) { |
768 | if (curr->on_rq) | |
769 | vruntime = curr->vruntime; | |
770 | else | |
771 | curr = NULL; | |
772 | } | |
1af5f730 | 773 | |
147f3efa | 774 | if (se) { |
b60205c7 | 775 | if (!curr) |
2227a957 | 776 | vruntime = se->min_vruntime; |
1af5f730 | 777 | else |
2227a957 | 778 | vruntime = min_vruntime(vruntime, se->min_vruntime); |
1af5f730 PZ |
779 | } |
780 | ||
1bf08230 | 781 | /* ensure we never gain time by being placed backwards. */ |
d05b4305 | 782 | u64_u32_store(cfs_rq->min_vruntime, |
af4cf404 | 783 | __update_min_vruntime(cfs_rq, vruntime)); |
1af5f730 PZ |
784 | } |
785 | ||
bf9be9a1 PZ |
786 | static inline bool __entity_less(struct rb_node *a, const struct rb_node *b) |
787 | { | |
788 | return entity_before(__node_2_se(a), __node_2_se(b)); | |
789 | } | |
790 | ||
2227a957 | 791 | #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; }) |
147f3efa | 792 | |
2227a957 | 793 | static inline void __min_vruntime_update(struct sched_entity *se, struct rb_node *node) |
147f3efa PZ |
794 | { |
795 | if (node) { | |
796 | struct sched_entity *rse = __node_2_se(node); | |
2227a957 AW |
797 | if (vruntime_gt(min_vruntime, se, rse)) |
798 | se->min_vruntime = rse->min_vruntime; | |
147f3efa PZ |
799 | } |
800 | } | |
801 | ||
802 | /* | |
2227a957 | 803 | * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime) |
147f3efa | 804 | */ |
2227a957 | 805 | static inline bool min_vruntime_update(struct sched_entity *se, bool exit) |
147f3efa | 806 | { |
2227a957 | 807 | u64 old_min_vruntime = se->min_vruntime; |
147f3efa PZ |
808 | struct rb_node *node = &se->run_node; |
809 | ||
2227a957 AW |
810 | se->min_vruntime = se->vruntime; |
811 | __min_vruntime_update(se, node->rb_right); | |
812 | __min_vruntime_update(se, node->rb_left); | |
147f3efa | 813 | |
2227a957 | 814 | return se->min_vruntime == old_min_vruntime; |
147f3efa PZ |
815 | } |
816 | ||
2227a957 AW |
817 | RB_DECLARE_CALLBACKS(static, min_vruntime_cb, struct sched_entity, |
818 | run_node, min_vruntime, min_vruntime_update); | |
147f3efa | 819 | |
bf0f6f24 IM |
820 | /* |
821 | * Enqueue an entity into the rb-tree: | |
822 | */ | |
0702e3eb | 823 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 824 | { |
af4cf404 | 825 | avg_vruntime_add(cfs_rq, se); |
2227a957 | 826 | se->min_vruntime = se->vruntime; |
147f3efa | 827 | rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline, |
2227a957 | 828 | __entity_less, &min_vruntime_cb); |
bf0f6f24 IM |
829 | } |
830 | ||
0702e3eb | 831 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 832 | { |
147f3efa | 833 | rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline, |
2227a957 | 834 | &min_vruntime_cb); |
af4cf404 | 835 | avg_vruntime_sub(cfs_rq, se); |
bf0f6f24 IM |
836 | } |
837 | ||
2227a957 AW |
838 | struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq) |
839 | { | |
840 | struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node; | |
841 | ||
842 | if (!root) | |
843 | return NULL; | |
844 | ||
845 | return __node_2_se(root); | |
846 | } | |
847 | ||
029632fb | 848 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 849 | { |
bfb06889 | 850 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
851 | |
852 | if (!left) | |
853 | return NULL; | |
854 | ||
bf9be9a1 | 855 | return __node_2_se(left); |
bf0f6f24 IM |
856 | } |
857 | ||
147f3efa PZ |
858 | /* |
859 | * Earliest Eligible Virtual Deadline First | |
860 | * | |
861 | * In order to provide latency guarantees for different request sizes | |
862 | * EEVDF selects the best runnable task from two criteria: | |
863 | * | |
864 | * 1) the task must be eligible (must be owed service) | |
865 | * | |
866 | * 2) from those tasks that meet 1), we select the one | |
867 | * with the earliest virtual deadline. | |
868 | * | |
869 | * We can do this in O(log n) time due to an augmented RB-tree. The | |
2227a957 AW |
870 | * tree keeps the entries sorted on deadline, but also functions as a |
871 | * heap based on the vruntime by keeping: | |
147f3efa | 872 | * |
2227a957 | 873 | * se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime) |
147f3efa | 874 | * |
2227a957 | 875 | * Which allows tree pruning through eligibility. |
147f3efa | 876 | */ |
2227a957 | 877 | static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq) |
ac53db59 | 878 | { |
147f3efa | 879 | struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node; |
ee4373dc | 880 | struct sched_entity *se = __pick_first_entity(cfs_rq); |
147f3efa PZ |
881 | struct sched_entity *curr = cfs_rq->curr; |
882 | struct sched_entity *best = NULL; | |
2227a957 AW |
883 | |
884 | /* | |
885 | * We can safely skip eligibility check if there is only one entity | |
886 | * in this cfs_rq, saving some cycles. | |
887 | */ | |
888 | if (cfs_rq->nr_running == 1) | |
ee4373dc | 889 | return curr && curr->on_rq ? curr : se; |
ac53db59 | 890 | |
147f3efa PZ |
891 | if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr))) |
892 | curr = NULL; | |
893 | ||
63304558 PZ |
894 | /* |
895 | * Once selected, run a task until it either becomes non-eligible or | |
896 | * until it gets a new slice. See the HACK in set_next_entity(). | |
897 | */ | |
898 | if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline) | |
899 | return curr; | |
900 | ||
ee4373dc AW |
901 | /* Pick the leftmost entity if it's eligible */ |
902 | if (se && entity_eligible(cfs_rq, se)) { | |
903 | best = se; | |
904 | goto found; | |
905 | } | |
906 | ||
2227a957 | 907 | /* Heap search for the EEVD entity */ |
147f3efa | 908 | while (node) { |
2227a957 | 909 | struct rb_node *left = node->rb_left; |
ac53db59 | 910 | |
147f3efa | 911 | /* |
2227a957 AW |
912 | * Eligible entities in left subtree are always better |
913 | * choices, since they have earlier deadlines. | |
147f3efa | 914 | */ |
2227a957 AW |
915 | if (left && vruntime_eligible(cfs_rq, |
916 | __node_2_se(left)->min_vruntime)) { | |
917 | node = left; | |
147f3efa PZ |
918 | continue; |
919 | } | |
920 | ||
ee4373dc AW |
921 | se = __node_2_se(node); |
922 | ||
147f3efa | 923 | /* |
2227a957 AW |
924 | * The left subtree either is empty or has no eligible |
925 | * entity, so check the current node since it is the one | |
926 | * with earliest deadline that might be eligible. | |
147f3efa | 927 | */ |
2227a957 | 928 | if (entity_eligible(cfs_rq, se)) { |
147f3efa | 929 | best = se; |
b01db23d | 930 | break; |
147f3efa PZ |
931 | } |
932 | ||
933 | node = node->rb_right; | |
934 | } | |
ee4373dc | 935 | found: |
2227a957 AW |
936 | if (!best || (curr && entity_before(curr, best))) |
937 | best = curr; | |
147f3efa | 938 | |
2227a957 | 939 | return best; |
ac53db59 RR |
940 | } |
941 | ||
942 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 943 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 944 | { |
bfb06889 | 945 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 946 | |
70eee74b BS |
947 | if (!last) |
948 | return NULL; | |
7eee3e67 | 949 | |
bf9be9a1 | 950 | return __node_2_se(last); |
aeb73b04 PZ |
951 | } |
952 | ||
bf0f6f24 IM |
953 | /************************************************************** |
954 | * Scheduling class statistics methods: | |
955 | */ | |
22dc02f8 | 956 | #ifdef CONFIG_SMP |
8a99b683 | 957 | int sched_update_scaling(void) |
b2be5e96 | 958 | { |
58ac93e4 | 959 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 | 960 | |
acb4a848 CE |
961 | #define WRT_SYSCTL(name) \ |
962 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
e4ec3318 | 963 | WRT_SYSCTL(sched_base_slice); |
acb4a848 CE |
964 | #undef WRT_SYSCTL |
965 | ||
b2be5e96 PZ |
966 | return 0; |
967 | } | |
968 | #endif | |
22dc02f8 | 969 | #endif |
647e7cac | 970 | |
147f3efa | 971 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se); |
51ce83ed | 972 | |
647e7cac | 973 | /* |
147f3efa PZ |
974 | * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i |
975 | * this is probably good enough. | |
647e7cac | 976 | */ |
147f3efa | 977 | static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 978 | { |
147f3efa PZ |
979 | if ((s64)(se->vruntime - se->deadline) < 0) |
980 | return; | |
0a582440 | 981 | |
5e963f2b PZ |
982 | /* |
983 | * For EEVDF the virtual time slope is determined by w_i (iow. | |
984 | * nice) while the request time r_i is determined by | |
e4ec3318 | 985 | * sysctl_sched_base_slice. |
5e963f2b | 986 | */ |
e4ec3318 | 987 | se->slice = sysctl_sched_base_slice; |
0c2de3f0 | 988 | |
147f3efa PZ |
989 | /* |
990 | * EEVDF: vd_i = ve_i + r_i / w_i | |
991 | */ | |
992 | se->deadline = se->vruntime + calc_delta_fair(se->slice, se); | |
51ce83ed | 993 | |
5e963f2b PZ |
994 | /* |
995 | * The task has consumed its request, reschedule. | |
996 | */ | |
997 | if (cfs_rq->nr_running > 1) { | |
998 | resched_curr(rq_of(cfs_rq)); | |
999 | clear_buddies(cfs_rq, se); | |
51ce83ed | 1000 | } |
a7be37ac PZ |
1001 | } |
1002 | ||
c0796298 | 1003 | #include "pelt.h" |
23127296 | 1004 | #ifdef CONFIG_SMP |
283e2ed3 | 1005 | |
772bd008 | 1006 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 1007 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 1008 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 1009 | |
540247fb YD |
1010 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
1011 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 1012 | { |
540247fb | 1013 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 1014 | |
f207934f PZ |
1015 | memset(sa, 0, sizeof(*sa)); |
1016 | ||
b5a9b340 | 1017 | /* |
dfcb245e | 1018 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 1019 | * they get a chance to stabilize to their real load level. |
dfcb245e | 1020 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
1021 | * nothing has been attached to the task group yet. |
1022 | */ | |
1023 | if (entity_is_task(se)) | |
0dacee1b | 1024 | sa->load_avg = scale_load_down(se->load.weight); |
f207934f | 1025 | |
b9e6e286 | 1026 | /* when this task is enqueued, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 1027 | } |
7ea241af | 1028 | |
2b8c41da YD |
1029 | /* |
1030 | * With new tasks being created, their initial util_avgs are extrapolated | |
1031 | * based on the cfs_rq's current util_avg: | |
1032 | * | |
72bffbf5 DL |
1033 | * util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1) |
1034 | * * se_weight(se) | |
2b8c41da YD |
1035 | * |
1036 | * However, in many cases, the above util_avg does not give a desired | |
1037 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
1038 | * as when the series is a harmonic series. | |
1039 | * | |
1040 | * To solve this problem, we also cap the util_avg of successive tasks to | |
1041 | * only 1/2 of the left utilization budget: | |
1042 | * | |
8fe5c5a9 | 1043 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 1044 | * |
8fe5c5a9 | 1045 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 1046 | * |
8fe5c5a9 QP |
1047 | * For example, for a CPU with 1024 of capacity, a simplest series from |
1048 | * the beginning would be like: | |
2b8c41da YD |
1049 | * |
1050 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
1051 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
1052 | * | |
1053 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
1054 | * if util_avg > util_avg_cap. | |
1055 | */ | |
d0fe0b9c | 1056 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 1057 | { |
d0fe0b9c | 1058 | struct sched_entity *se = &p->se; |
2b8c41da YD |
1059 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
1060 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 1061 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 1062 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da | 1063 | |
d0fe0b9c DE |
1064 | if (p->sched_class != &fair_sched_class) { |
1065 | /* | |
1066 | * For !fair tasks do: | |
1067 | * | |
1068 | update_cfs_rq_load_avg(now, cfs_rq); | |
a4f9a0e5 | 1069 | attach_entity_load_avg(cfs_rq, se); |
d0fe0b9c DE |
1070 | switched_from_fair(rq, p); |
1071 | * | |
1072 | * such that the next switched_to_fair() has the | |
1073 | * expected state. | |
1074 | */ | |
1075 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
1076 | return; | |
7dc603c9 | 1077 | } |
e4fe074d CZ |
1078 | |
1079 | if (cap > 0) { | |
1080 | if (cfs_rq->avg.util_avg != 0) { | |
72bffbf5 | 1081 | sa->util_avg = cfs_rq->avg.util_avg * se_weight(se); |
e4fe074d CZ |
1082 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); |
1083 | ||
1084 | if (sa->util_avg > cap) | |
1085 | sa->util_avg = cap; | |
1086 | } else { | |
1087 | sa->util_avg = cap; | |
1088 | } | |
1089 | } | |
1090 | ||
1091 | sa->runnable_avg = sa->util_avg; | |
2b8c41da YD |
1092 | } |
1093 | ||
7dc603c9 | 1094 | #else /* !CONFIG_SMP */ |
540247fb | 1095 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
1096 | { |
1097 | } | |
d0fe0b9c | 1098 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
1099 | { |
1100 | } | |
fe749158 | 1101 | static void update_tg_load_avg(struct cfs_rq *cfs_rq) |
3d30544f PZ |
1102 | { |
1103 | } | |
7dc603c9 | 1104 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 1105 | |
5d69eca5 | 1106 | static s64 update_curr_se(struct rq *rq, struct sched_entity *curr) |
bf0f6f24 | 1107 | { |
5d69eca5 PZ |
1108 | u64 now = rq_clock_task(rq); |
1109 | s64 delta_exec; | |
bf0f6f24 | 1110 | |
9dbdb155 | 1111 | delta_exec = now - curr->exec_start; |
5d69eca5 PZ |
1112 | if (unlikely(delta_exec <= 0)) |
1113 | return delta_exec; | |
bf0f6f24 | 1114 | |
8ebc91d9 | 1115 | curr->exec_start = now; |
5d69eca5 | 1116 | curr->sum_exec_runtime += delta_exec; |
d842de87 | 1117 | |
ceeadb83 YS |
1118 | if (schedstat_enabled()) { |
1119 | struct sched_statistics *stats; | |
1120 | ||
1121 | stats = __schedstats_from_se(curr); | |
1122 | __schedstat_set(stats->exec_max, | |
1123 | max(delta_exec, stats->exec_max)); | |
1124 | } | |
9dbdb155 | 1125 | |
5d69eca5 PZ |
1126 | return delta_exec; |
1127 | } | |
1128 | ||
c708a4dc PZ |
1129 | static inline void update_curr_task(struct task_struct *p, s64 delta_exec) |
1130 | { | |
1131 | trace_sched_stat_runtime(p, delta_exec); | |
1132 | account_group_exec_runtime(p, delta_exec); | |
1133 | cgroup_account_cputime(p, delta_exec); | |
63ba8422 PZ |
1134 | if (p->dl_server) |
1135 | dl_server_update(p->dl_server, delta_exec); | |
c708a4dc PZ |
1136 | } |
1137 | ||
5d69eca5 PZ |
1138 | /* |
1139 | * Used by other classes to account runtime. | |
1140 | */ | |
1141 | s64 update_curr_common(struct rq *rq) | |
1142 | { | |
1143 | struct task_struct *curr = rq->curr; | |
1144 | s64 delta_exec; | |
1145 | ||
1146 | delta_exec = update_curr_se(rq, &curr->se); | |
c708a4dc PZ |
1147 | if (likely(delta_exec > 0)) |
1148 | update_curr_task(curr, delta_exec); | |
5d69eca5 PZ |
1149 | |
1150 | return delta_exec; | |
1151 | } | |
1152 | ||
1153 | /* | |
1154 | * Update the current task's runtime statistics. | |
1155 | */ | |
1156 | static void update_curr(struct cfs_rq *cfs_rq) | |
1157 | { | |
1158 | struct sched_entity *curr = cfs_rq->curr; | |
1159 | s64 delta_exec; | |
1160 | ||
1161 | if (unlikely(!curr)) | |
1162 | return; | |
1163 | ||
1164 | delta_exec = update_curr_se(rq_of(cfs_rq), curr); | |
1165 | if (unlikely(delta_exec <= 0)) | |
1166 | return; | |
9dbdb155 PZ |
1167 | |
1168 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
147f3efa | 1169 | update_deadline(cfs_rq, curr); |
9dbdb155 PZ |
1170 | update_min_vruntime(cfs_rq); |
1171 | ||
c708a4dc PZ |
1172 | if (entity_is_task(curr)) |
1173 | update_curr_task(task_of(curr), delta_exec); | |
ec12cb7f PT |
1174 | |
1175 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
1176 | } |
1177 | ||
6e998916 SG |
1178 | static void update_curr_fair(struct rq *rq) |
1179 | { | |
1180 | update_curr(cfs_rq_of(&rq->curr->se)); | |
1181 | } | |
1182 | ||
bf0f6f24 | 1183 | static inline void |
60f2415e | 1184 | update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 1185 | { |
ceeadb83 | 1186 | struct sched_statistics *stats; |
60f2415e | 1187 | struct task_struct *p = NULL; |
4fa8d299 JP |
1188 | |
1189 | if (!schedstat_enabled()) | |
1190 | return; | |
1191 | ||
ceeadb83 YS |
1192 | stats = __schedstats_from_se(se); |
1193 | ||
60f2415e YS |
1194 | if (entity_is_task(se)) |
1195 | p = task_of(se); | |
3ea94de1 | 1196 | |
60f2415e | 1197 | __update_stats_wait_start(rq_of(cfs_rq), p, stats); |
bf0f6f24 IM |
1198 | } |
1199 | ||
4fa8d299 | 1200 | static inline void |
60f2415e | 1201 | update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3ea94de1 | 1202 | { |
ceeadb83 YS |
1203 | struct sched_statistics *stats; |
1204 | struct task_struct *p = NULL; | |
cb251765 | 1205 | |
4fa8d299 JP |
1206 | if (!schedstat_enabled()) |
1207 | return; | |
1208 | ||
ceeadb83 YS |
1209 | stats = __schedstats_from_se(se); |
1210 | ||
b9c88f75 | 1211 | /* |
1212 | * When the sched_schedstat changes from 0 to 1, some sched se | |
1213 | * maybe already in the runqueue, the se->statistics.wait_start | |
1214 | * will be 0.So it will let the delta wrong. We need to avoid this | |
1215 | * scenario. | |
1216 | */ | |
ceeadb83 | 1217 | if (unlikely(!schedstat_val(stats->wait_start))) |
b9c88f75 | 1218 | return; |
1219 | ||
60f2415e | 1220 | if (entity_is_task(se)) |
3ea94de1 | 1221 | p = task_of(se); |
3ea94de1 | 1222 | |
60f2415e | 1223 | __update_stats_wait_end(rq_of(cfs_rq), p, stats); |
3ea94de1 | 1224 | } |
3ea94de1 | 1225 | |
4fa8d299 | 1226 | static inline void |
60f2415e | 1227 | update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1a3d027c | 1228 | { |
ceeadb83 | 1229 | struct sched_statistics *stats; |
1a3d027c | 1230 | struct task_struct *tsk = NULL; |
4fa8d299 JP |
1231 | |
1232 | if (!schedstat_enabled()) | |
1233 | return; | |
1234 | ||
ceeadb83 YS |
1235 | stats = __schedstats_from_se(se); |
1236 | ||
1a3d027c JP |
1237 | if (entity_is_task(se)) |
1238 | tsk = task_of(se); | |
1239 | ||
60f2415e | 1240 | __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats); |
3ea94de1 | 1241 | } |
3ea94de1 | 1242 | |
bf0f6f24 IM |
1243 | /* |
1244 | * Task is being enqueued - update stats: | |
1245 | */ | |
cb251765 | 1246 | static inline void |
60f2415e | 1247 | update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1248 | { |
4fa8d299 JP |
1249 | if (!schedstat_enabled()) |
1250 | return; | |
1251 | ||
bf0f6f24 IM |
1252 | /* |
1253 | * Are we enqueueing a waiting task? (for current tasks | |
1254 | * a dequeue/enqueue event is a NOP) | |
1255 | */ | |
429d43bc | 1256 | if (se != cfs_rq->curr) |
60f2415e | 1257 | update_stats_wait_start_fair(cfs_rq, se); |
1a3d027c JP |
1258 | |
1259 | if (flags & ENQUEUE_WAKEUP) | |
60f2415e | 1260 | update_stats_enqueue_sleeper_fair(cfs_rq, se); |
bf0f6f24 IM |
1261 | } |
1262 | ||
bf0f6f24 | 1263 | static inline void |
60f2415e | 1264 | update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1265 | { |
4fa8d299 JP |
1266 | |
1267 | if (!schedstat_enabled()) | |
1268 | return; | |
1269 | ||
bf0f6f24 IM |
1270 | /* |
1271 | * Mark the end of the wait period if dequeueing a | |
1272 | * waiting task: | |
1273 | */ | |
429d43bc | 1274 | if (se != cfs_rq->curr) |
60f2415e | 1275 | update_stats_wait_end_fair(cfs_rq, se); |
cb251765 | 1276 | |
4fa8d299 JP |
1277 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1278 | struct task_struct *tsk = task_of(se); | |
2f064a59 | 1279 | unsigned int state; |
cb251765 | 1280 | |
2f064a59 PZ |
1281 | /* XXX racy against TTWU */ |
1282 | state = READ_ONCE(tsk->__state); | |
1283 | if (state & TASK_INTERRUPTIBLE) | |
ceeadb83 | 1284 | __schedstat_set(tsk->stats.sleep_start, |
4fa8d299 | 1285 | rq_clock(rq_of(cfs_rq))); |
2f064a59 | 1286 | if (state & TASK_UNINTERRUPTIBLE) |
ceeadb83 | 1287 | __schedstat_set(tsk->stats.block_start, |
4fa8d299 | 1288 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1289 | } |
cb251765 MG |
1290 | } |
1291 | ||
bf0f6f24 IM |
1292 | /* |
1293 | * We are picking a new current task - update its stats: | |
1294 | */ | |
1295 | static inline void | |
79303e9e | 1296 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1297 | { |
1298 | /* | |
1299 | * We are starting a new run period: | |
1300 | */ | |
78becc27 | 1301 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1302 | } |
1303 | ||
bf0f6f24 IM |
1304 | /************************************************** |
1305 | * Scheduling class queueing methods: | |
1306 | */ | |
1307 | ||
8b36d07f RN |
1308 | static inline bool is_core_idle(int cpu) |
1309 | { | |
1310 | #ifdef CONFIG_SCHED_SMT | |
1311 | int sibling; | |
1312 | ||
1313 | for_each_cpu(sibling, cpu_smt_mask(cpu)) { | |
1314 | if (cpu == sibling) | |
1315 | continue; | |
1316 | ||
1317 | if (!idle_cpu(sibling)) | |
1318 | return false; | |
1319 | } | |
1320 | #endif | |
1321 | ||
1322 | return true; | |
1323 | } | |
1324 | ||
cb29a5c1 MG |
1325 | #ifdef CONFIG_NUMA |
1326 | #define NUMA_IMBALANCE_MIN 2 | |
1327 | ||
1328 | static inline long | |
1329 | adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr) | |
1330 | { | |
1331 | /* | |
1332 | * Allow a NUMA imbalance if busy CPUs is less than the maximum | |
1333 | * threshold. Above this threshold, individual tasks may be contending | |
1334 | * for both memory bandwidth and any shared HT resources. This is an | |
1335 | * approximation as the number of running tasks may not be related to | |
1336 | * the number of busy CPUs due to sched_setaffinity. | |
1337 | */ | |
1338 | if (dst_running > imb_numa_nr) | |
1339 | return imbalance; | |
1340 | ||
1341 | /* | |
1342 | * Allow a small imbalance based on a simple pair of communicating | |
1343 | * tasks that remain local when the destination is lightly loaded. | |
1344 | */ | |
1345 | if (imbalance <= NUMA_IMBALANCE_MIN) | |
1346 | return 0; | |
1347 | ||
1348 | return imbalance; | |
1349 | } | |
1350 | #endif /* CONFIG_NUMA */ | |
1351 | ||
cbee9f88 PZ |
1352 | #ifdef CONFIG_NUMA_BALANCING |
1353 | /* | |
598f0ec0 MG |
1354 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1355 | * calculated based on the tasks virtual memory size and | |
1356 | * numa_balancing_scan_size. | |
cbee9f88 | 1357 | */ |
598f0ec0 MG |
1358 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1359 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1360 | |
1361 | /* Portion of address space to scan in MB */ | |
1362 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1363 | |
4b96a29b PZ |
1364 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1365 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1366 | ||
33024536 HY |
1367 | /* The page with hint page fault latency < threshold in ms is considered hot */ |
1368 | unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC; | |
1369 | ||
b5dd77c8 | 1370 | struct numa_group { |
c45a7795 | 1371 | refcount_t refcount; |
b5dd77c8 RR |
1372 | |
1373 | spinlock_t lock; /* nr_tasks, tasks */ | |
1374 | int nr_tasks; | |
1375 | pid_t gid; | |
1376 | int active_nodes; | |
1377 | ||
1378 | struct rcu_head rcu; | |
1379 | unsigned long total_faults; | |
1380 | unsigned long max_faults_cpu; | |
1381 | /* | |
5b763a14 BR |
1382 | * faults[] array is split into two regions: faults_mem and faults_cpu. |
1383 | * | |
b5dd77c8 RR |
1384 | * Faults_cpu is used to decide whether memory should move |
1385 | * towards the CPU. As a consequence, these stats are weighted | |
1386 | * more by CPU use than by memory faults. | |
1387 | */ | |
04f5c362 | 1388 | unsigned long faults[]; |
b5dd77c8 RR |
1389 | }; |
1390 | ||
cb361d8c JH |
1391 | /* |
1392 | * For functions that can be called in multiple contexts that permit reading | |
1393 | * ->numa_group (see struct task_struct for locking rules). | |
1394 | */ | |
1395 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1396 | { | |
1397 | return rcu_dereference_check(p->numa_group, p == current || | |
9ef7e7e3 | 1398 | (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu))); |
cb361d8c JH |
1399 | } |
1400 | ||
1401 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1402 | { | |
1403 | return rcu_dereference_protected(p->numa_group, p == current); | |
1404 | } | |
1405 | ||
b5dd77c8 RR |
1406 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1407 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1408 | ||
598f0ec0 MG |
1409 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1410 | { | |
1411 | unsigned long rss = 0; | |
1412 | unsigned long nr_scan_pages; | |
1413 | ||
1414 | /* | |
1415 | * Calculations based on RSS as non-present and empty pages are skipped | |
1416 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1417 | * on resident pages | |
1418 | */ | |
1419 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1420 | rss = get_mm_rss(p->mm); | |
1421 | if (!rss) | |
1422 | rss = nr_scan_pages; | |
1423 | ||
1424 | rss = round_up(rss, nr_scan_pages); | |
1425 | return rss / nr_scan_pages; | |
1426 | } | |
1427 | ||
3b03706f | 1428 | /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ |
598f0ec0 MG |
1429 | #define MAX_SCAN_WINDOW 2560 |
1430 | ||
1431 | static unsigned int task_scan_min(struct task_struct *p) | |
1432 | { | |
316c1608 | 1433 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1434 | unsigned int scan, floor; |
1435 | unsigned int windows = 1; | |
1436 | ||
64192658 KT |
1437 | if (scan_size < MAX_SCAN_WINDOW) |
1438 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1439 | floor = 1000 / windows; |
1440 | ||
1441 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1442 | return max_t(unsigned int, floor, scan); | |
1443 | } | |
1444 | ||
b5dd77c8 RR |
1445 | static unsigned int task_scan_start(struct task_struct *p) |
1446 | { | |
1447 | unsigned long smin = task_scan_min(p); | |
1448 | unsigned long period = smin; | |
cb361d8c | 1449 | struct numa_group *ng; |
b5dd77c8 RR |
1450 | |
1451 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1452 | rcu_read_lock(); |
1453 | ng = rcu_dereference(p->numa_group); | |
1454 | if (ng) { | |
b5dd77c8 RR |
1455 | unsigned long shared = group_faults_shared(ng); |
1456 | unsigned long private = group_faults_priv(ng); | |
1457 | ||
c45a7795 | 1458 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1459 | period *= shared + 1; |
1460 | period /= private + shared + 1; | |
1461 | } | |
cb361d8c | 1462 | rcu_read_unlock(); |
b5dd77c8 RR |
1463 | |
1464 | return max(smin, period); | |
1465 | } | |
1466 | ||
598f0ec0 MG |
1467 | static unsigned int task_scan_max(struct task_struct *p) |
1468 | { | |
b5dd77c8 RR |
1469 | unsigned long smin = task_scan_min(p); |
1470 | unsigned long smax; | |
cb361d8c | 1471 | struct numa_group *ng; |
598f0ec0 MG |
1472 | |
1473 | /* Watch for min being lower than max due to floor calculations */ | |
1474 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1475 | |
1476 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1477 | ng = deref_curr_numa_group(p); |
1478 | if (ng) { | |
b5dd77c8 RR |
1479 | unsigned long shared = group_faults_shared(ng); |
1480 | unsigned long private = group_faults_priv(ng); | |
1481 | unsigned long period = smax; | |
1482 | ||
c45a7795 | 1483 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1484 | period *= shared + 1; |
1485 | period /= private + shared + 1; | |
1486 | ||
1487 | smax = max(smax, period); | |
1488 | } | |
1489 | ||
598f0ec0 MG |
1490 | return max(smin, smax); |
1491 | } | |
1492 | ||
0ec8aa00 PZ |
1493 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1494 | { | |
98fa15f3 | 1495 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1496 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1497 | } | |
1498 | ||
1499 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1500 | { | |
98fa15f3 | 1501 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1502 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1503 | } | |
1504 | ||
be1e4e76 RR |
1505 | /* Shared or private faults. */ |
1506 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1507 | ||
1508 | /* Memory and CPU locality */ | |
1509 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1510 | ||
1511 | /* Averaged statistics, and temporary buffers. */ | |
1512 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1513 | ||
e29cf08b MG |
1514 | pid_t task_numa_group_id(struct task_struct *p) |
1515 | { | |
cb361d8c JH |
1516 | struct numa_group *ng; |
1517 | pid_t gid = 0; | |
1518 | ||
1519 | rcu_read_lock(); | |
1520 | ng = rcu_dereference(p->numa_group); | |
1521 | if (ng) | |
1522 | gid = ng->gid; | |
1523 | rcu_read_unlock(); | |
1524 | ||
1525 | return gid; | |
e29cf08b MG |
1526 | } |
1527 | ||
44dba3d5 | 1528 | /* |
97fb7a0a | 1529 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1530 | * occupy the first half of the array. The second half of the |
1531 | * array is for current counters, which are averaged into the | |
1532 | * first set by task_numa_placement. | |
1533 | */ | |
1534 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1535 | { |
44dba3d5 | 1536 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1537 | } |
1538 | ||
1539 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1540 | { | |
44dba3d5 | 1541 | if (!p->numa_faults) |
ac8e895b MG |
1542 | return 0; |
1543 | ||
44dba3d5 IM |
1544 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1545 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1546 | } |
1547 | ||
83e1d2cd MG |
1548 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1549 | { | |
cb361d8c JH |
1550 | struct numa_group *ng = deref_task_numa_group(p); |
1551 | ||
1552 | if (!ng) | |
83e1d2cd MG |
1553 | return 0; |
1554 | ||
cb361d8c JH |
1555 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1556 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1557 | } |
1558 | ||
20e07dea RR |
1559 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1560 | { | |
5b763a14 BR |
1561 | return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] + |
1562 | group->faults[task_faults_idx(NUMA_CPU, nid, 1)]; | |
20e07dea RR |
1563 | } |
1564 | ||
b5dd77c8 RR |
1565 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1566 | { | |
1567 | unsigned long faults = 0; | |
1568 | int node; | |
1569 | ||
1570 | for_each_online_node(node) { | |
1571 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1572 | } | |
1573 | ||
1574 | return faults; | |
1575 | } | |
1576 | ||
1577 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1578 | { | |
1579 | unsigned long faults = 0; | |
1580 | int node; | |
1581 | ||
1582 | for_each_online_node(node) { | |
1583 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1584 | } | |
1585 | ||
1586 | return faults; | |
1587 | } | |
1588 | ||
4142c3eb RR |
1589 | /* |
1590 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1591 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1592 | * between these nodes are slowed down, to allow things to settle down. | |
1593 | */ | |
1594 | #define ACTIVE_NODE_FRACTION 3 | |
1595 | ||
1596 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1597 | { | |
1598 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1599 | } | |
1600 | ||
6c6b1193 RR |
1601 | /* Handle placement on systems where not all nodes are directly connected. */ |
1602 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
0fb3978b | 1603 | int lim_dist, bool task) |
6c6b1193 RR |
1604 | { |
1605 | unsigned long score = 0; | |
0fb3978b | 1606 | int node, max_dist; |
6c6b1193 RR |
1607 | |
1608 | /* | |
1609 | * All nodes are directly connected, and the same distance | |
1610 | * from each other. No need for fancy placement algorithms. | |
1611 | */ | |
1612 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1613 | return 0; | |
1614 | ||
0fb3978b HY |
1615 | /* sched_max_numa_distance may be changed in parallel. */ |
1616 | max_dist = READ_ONCE(sched_max_numa_distance); | |
6c6b1193 RR |
1617 | /* |
1618 | * This code is called for each node, introducing N^2 complexity, | |
b9e6e286 | 1619 | * which should be OK given the number of nodes rarely exceeds 8. |
6c6b1193 RR |
1620 | */ |
1621 | for_each_online_node(node) { | |
1622 | unsigned long faults; | |
1623 | int dist = node_distance(nid, node); | |
1624 | ||
1625 | /* | |
1626 | * The furthest away nodes in the system are not interesting | |
1627 | * for placement; nid was already counted. | |
1628 | */ | |
0fb3978b | 1629 | if (dist >= max_dist || node == nid) |
6c6b1193 RR |
1630 | continue; |
1631 | ||
1632 | /* | |
1633 | * On systems with a backplane NUMA topology, compare groups | |
1634 | * of nodes, and move tasks towards the group with the most | |
1635 | * memory accesses. When comparing two nodes at distance | |
1636 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1637 | * of each group. Skip other nodes. | |
1638 | */ | |
0fb3978b | 1639 | if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist) |
6c6b1193 RR |
1640 | continue; |
1641 | ||
1642 | /* Add up the faults from nearby nodes. */ | |
1643 | if (task) | |
1644 | faults = task_faults(p, node); | |
1645 | else | |
1646 | faults = group_faults(p, node); | |
1647 | ||
1648 | /* | |
1649 | * On systems with a glueless mesh NUMA topology, there are | |
1650 | * no fixed "groups of nodes". Instead, nodes that are not | |
1651 | * directly connected bounce traffic through intermediate | |
1652 | * nodes; a numa_group can occupy any set of nodes. | |
1653 | * The further away a node is, the less the faults count. | |
1654 | * This seems to result in good task placement. | |
1655 | */ | |
1656 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
0fb3978b HY |
1657 | faults *= (max_dist - dist); |
1658 | faults /= (max_dist - LOCAL_DISTANCE); | |
6c6b1193 RR |
1659 | } |
1660 | ||
1661 | score += faults; | |
1662 | } | |
1663 | ||
1664 | return score; | |
1665 | } | |
1666 | ||
83e1d2cd MG |
1667 | /* |
1668 | * These return the fraction of accesses done by a particular task, or | |
1669 | * task group, on a particular numa node. The group weight is given a | |
1670 | * larger multiplier, in order to group tasks together that are almost | |
1671 | * evenly spread out between numa nodes. | |
1672 | */ | |
7bd95320 RR |
1673 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1674 | int dist) | |
83e1d2cd | 1675 | { |
7bd95320 | 1676 | unsigned long faults, total_faults; |
83e1d2cd | 1677 | |
44dba3d5 | 1678 | if (!p->numa_faults) |
83e1d2cd MG |
1679 | return 0; |
1680 | ||
1681 | total_faults = p->total_numa_faults; | |
1682 | ||
1683 | if (!total_faults) | |
1684 | return 0; | |
1685 | ||
7bd95320 | 1686 | faults = task_faults(p, nid); |
6c6b1193 RR |
1687 | faults += score_nearby_nodes(p, nid, dist, true); |
1688 | ||
7bd95320 | 1689 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1690 | } |
1691 | ||
7bd95320 RR |
1692 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1693 | int dist) | |
83e1d2cd | 1694 | { |
cb361d8c | 1695 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1696 | unsigned long faults, total_faults; |
1697 | ||
cb361d8c | 1698 | if (!ng) |
7bd95320 RR |
1699 | return 0; |
1700 | ||
cb361d8c | 1701 | total_faults = ng->total_faults; |
7bd95320 RR |
1702 | |
1703 | if (!total_faults) | |
83e1d2cd MG |
1704 | return 0; |
1705 | ||
7bd95320 | 1706 | faults = group_faults(p, nid); |
6c6b1193 RR |
1707 | faults += score_nearby_nodes(p, nid, dist, false); |
1708 | ||
7bd95320 | 1709 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1710 | } |
1711 | ||
33024536 HY |
1712 | /* |
1713 | * If memory tiering mode is enabled, cpupid of slow memory page is | |
1714 | * used to record scan time instead of CPU and PID. When tiering mode | |
1715 | * is disabled at run time, the scan time (in cpupid) will be | |
1716 | * interpreted as CPU and PID. So CPU needs to be checked to avoid to | |
1717 | * access out of array bound. | |
1718 | */ | |
1719 | static inline bool cpupid_valid(int cpupid) | |
1720 | { | |
1721 | return cpupid_to_cpu(cpupid) < nr_cpu_ids; | |
1722 | } | |
1723 | ||
1724 | /* | |
1725 | * For memory tiering mode, if there are enough free pages (more than | |
1726 | * enough watermark defined here) in fast memory node, to take full | |
1727 | * advantage of fast memory capacity, all recently accessed slow | |
1728 | * memory pages will be migrated to fast memory node without | |
1729 | * considering hot threshold. | |
1730 | */ | |
1731 | static bool pgdat_free_space_enough(struct pglist_data *pgdat) | |
1732 | { | |
1733 | int z; | |
1734 | unsigned long enough_wmark; | |
1735 | ||
1736 | enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT, | |
1737 | pgdat->node_present_pages >> 4); | |
1738 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1739 | struct zone *zone = pgdat->node_zones + z; | |
1740 | ||
1741 | if (!populated_zone(zone)) | |
1742 | continue; | |
1743 | ||
1744 | if (zone_watermark_ok(zone, 0, | |
1745 | wmark_pages(zone, WMARK_PROMO) + enough_wmark, | |
1746 | ZONE_MOVABLE, 0)) | |
1747 | return true; | |
1748 | } | |
1749 | return false; | |
1750 | } | |
1751 | ||
1752 | /* | |
1753 | * For memory tiering mode, when page tables are scanned, the scan | |
1754 | * time will be recorded in struct page in addition to make page | |
1755 | * PROT_NONE for slow memory page. So when the page is accessed, in | |
1756 | * hint page fault handler, the hint page fault latency is calculated | |
1757 | * via, | |
1758 | * | |
1759 | * hint page fault latency = hint page fault time - scan time | |
1760 | * | |
1761 | * The smaller the hint page fault latency, the higher the possibility | |
1762 | * for the page to be hot. | |
1763 | */ | |
8c9ae56d | 1764 | static int numa_hint_fault_latency(struct folio *folio) |
33024536 HY |
1765 | { |
1766 | int last_time, time; | |
1767 | ||
1768 | time = jiffies_to_msecs(jiffies); | |
0b201c36 | 1769 | last_time = folio_xchg_access_time(folio, time); |
33024536 HY |
1770 | |
1771 | return (time - last_time) & PAGE_ACCESS_TIME_MASK; | |
1772 | } | |
1773 | ||
c6833e10 HY |
1774 | /* |
1775 | * For memory tiering mode, too high promotion/demotion throughput may | |
1776 | * hurt application latency. So we provide a mechanism to rate limit | |
1777 | * the number of pages that are tried to be promoted. | |
1778 | */ | |
1779 | static bool numa_promotion_rate_limit(struct pglist_data *pgdat, | |
1780 | unsigned long rate_limit, int nr) | |
1781 | { | |
1782 | unsigned long nr_cand; | |
1783 | unsigned int now, start; | |
1784 | ||
1785 | now = jiffies_to_msecs(jiffies); | |
1786 | mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr); | |
1787 | nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE); | |
1788 | start = pgdat->nbp_rl_start; | |
1789 | if (now - start > MSEC_PER_SEC && | |
1790 | cmpxchg(&pgdat->nbp_rl_start, start, now) == start) | |
1791 | pgdat->nbp_rl_nr_cand = nr_cand; | |
1792 | if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit) | |
1793 | return true; | |
1794 | return false; | |
1795 | } | |
1796 | ||
c959924b HY |
1797 | #define NUMA_MIGRATION_ADJUST_STEPS 16 |
1798 | ||
1799 | static void numa_promotion_adjust_threshold(struct pglist_data *pgdat, | |
1800 | unsigned long rate_limit, | |
1801 | unsigned int ref_th) | |
1802 | { | |
1803 | unsigned int now, start, th_period, unit_th, th; | |
1804 | unsigned long nr_cand, ref_cand, diff_cand; | |
1805 | ||
1806 | now = jiffies_to_msecs(jiffies); | |
1807 | th_period = sysctl_numa_balancing_scan_period_max; | |
1808 | start = pgdat->nbp_th_start; | |
1809 | if (now - start > th_period && | |
1810 | cmpxchg(&pgdat->nbp_th_start, start, now) == start) { | |
1811 | ref_cand = rate_limit * | |
1812 | sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC; | |
1813 | nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE); | |
1814 | diff_cand = nr_cand - pgdat->nbp_th_nr_cand; | |
1815 | unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS; | |
1816 | th = pgdat->nbp_threshold ? : ref_th; | |
1817 | if (diff_cand > ref_cand * 11 / 10) | |
1818 | th = max(th - unit_th, unit_th); | |
1819 | else if (diff_cand < ref_cand * 9 / 10) | |
1820 | th = min(th + unit_th, ref_th * 2); | |
1821 | pgdat->nbp_th_nr_cand = nr_cand; | |
1822 | pgdat->nbp_threshold = th; | |
1823 | } | |
1824 | } | |
1825 | ||
8c9ae56d | 1826 | bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio, |
10f39042 RR |
1827 | int src_nid, int dst_cpu) |
1828 | { | |
cb361d8c | 1829 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1830 | int dst_nid = cpu_to_node(dst_cpu); |
1831 | int last_cpupid, this_cpupid; | |
1832 | ||
3fb43636 BP |
1833 | /* |
1834 | * Cannot migrate to memoryless nodes. | |
1835 | */ | |
1836 | if (!node_state(dst_nid, N_MEMORY)) | |
1837 | return false; | |
1838 | ||
33024536 HY |
1839 | /* |
1840 | * The pages in slow memory node should be migrated according | |
1841 | * to hot/cold instead of private/shared. | |
1842 | */ | |
1843 | if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING && | |
1844 | !node_is_toptier(src_nid)) { | |
1845 | struct pglist_data *pgdat; | |
c959924b HY |
1846 | unsigned long rate_limit; |
1847 | unsigned int latency, th, def_th; | |
33024536 HY |
1848 | |
1849 | pgdat = NODE_DATA(dst_nid); | |
c959924b HY |
1850 | if (pgdat_free_space_enough(pgdat)) { |
1851 | /* workload changed, reset hot threshold */ | |
1852 | pgdat->nbp_threshold = 0; | |
33024536 | 1853 | return true; |
c959924b HY |
1854 | } |
1855 | ||
1856 | def_th = sysctl_numa_balancing_hot_threshold; | |
1857 | rate_limit = sysctl_numa_balancing_promote_rate_limit << \ | |
1858 | (20 - PAGE_SHIFT); | |
1859 | numa_promotion_adjust_threshold(pgdat, rate_limit, def_th); | |
33024536 | 1860 | |
c959924b | 1861 | th = pgdat->nbp_threshold ? : def_th; |
8c9ae56d | 1862 | latency = numa_hint_fault_latency(folio); |
33024536 HY |
1863 | if (latency >= th) |
1864 | return false; | |
1865 | ||
c6833e10 | 1866 | return !numa_promotion_rate_limit(pgdat, rate_limit, |
8c9ae56d | 1867 | folio_nr_pages(folio)); |
33024536 HY |
1868 | } |
1869 | ||
10f39042 | 1870 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); |
1b143cc7 | 1871 | last_cpupid = folio_xchg_last_cpupid(folio, this_cpupid); |
37355bdc | 1872 | |
33024536 HY |
1873 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) && |
1874 | !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid)) | |
1875 | return false; | |
1876 | ||
37355bdc MG |
1877 | /* |
1878 | * Allow first faults or private faults to migrate immediately early in | |
1879 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1880 | * two full passes of the "multi-stage node selection" test that is | |
1881 | * executed below. | |
1882 | */ | |
98fa15f3 | 1883 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1884 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1885 | return true; | |
10f39042 RR |
1886 | |
1887 | /* | |
1888 | * Multi-stage node selection is used in conjunction with a periodic | |
1889 | * migration fault to build a temporal task<->page relation. By using | |
1890 | * a two-stage filter we remove short/unlikely relations. | |
1891 | * | |
1892 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1893 | * a task's usage of a particular page (n_p) per total usage of this | |
1894 | * page (n_t) (in a given time-span) to a probability. | |
1895 | * | |
1896 | * Our periodic faults will sample this probability and getting the | |
1897 | * same result twice in a row, given these samples are fully | |
1898 | * independent, is then given by P(n)^2, provided our sample period | |
1899 | * is sufficiently short compared to the usage pattern. | |
1900 | * | |
1901 | * This quadric squishes small probabilities, making it less likely we | |
1902 | * act on an unlikely task<->page relation. | |
1903 | */ | |
10f39042 RR |
1904 | if (!cpupid_pid_unset(last_cpupid) && |
1905 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1906 | return false; | |
1907 | ||
1908 | /* Always allow migrate on private faults */ | |
1909 | if (cpupid_match_pid(p, last_cpupid)) | |
1910 | return true; | |
1911 | ||
1912 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1913 | if (!ng) | |
1914 | return true; | |
1915 | ||
1916 | /* | |
4142c3eb RR |
1917 | * Destination node is much more heavily used than the source |
1918 | * node? Allow migration. | |
10f39042 | 1919 | */ |
4142c3eb RR |
1920 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1921 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1922 | return true; |
1923 | ||
1924 | /* | |
4142c3eb RR |
1925 | * Distribute memory according to CPU & memory use on each node, |
1926 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1927 | * | |
1928 | * faults_cpu(dst) 3 faults_cpu(src) | |
1929 | * --------------- * - > --------------- | |
1930 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1931 | */ |
4142c3eb RR |
1932 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1933 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1934 | } |
1935 | ||
6499b1b2 VG |
1936 | /* |
1937 | * 'numa_type' describes the node at the moment of load balancing. | |
1938 | */ | |
1939 | enum numa_type { | |
1940 | /* The node has spare capacity that can be used to run more tasks. */ | |
1941 | node_has_spare = 0, | |
1942 | /* | |
1943 | * The node is fully used and the tasks don't compete for more CPU | |
1944 | * cycles. Nevertheless, some tasks might wait before running. | |
1945 | */ | |
1946 | node_fully_busy, | |
1947 | /* | |
1948 | * The node is overloaded and can't provide expected CPU cycles to all | |
1949 | * tasks. | |
1950 | */ | |
1951 | node_overloaded | |
1952 | }; | |
58d081b5 | 1953 | |
fb13c7ee | 1954 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1955 | struct numa_stats { |
1956 | unsigned long load; | |
8e0e0eda | 1957 | unsigned long runnable; |
6499b1b2 | 1958 | unsigned long util; |
fb13c7ee | 1959 | /* Total compute capacity of CPUs on a node */ |
5ef20ca1 | 1960 | unsigned long compute_capacity; |
6499b1b2 VG |
1961 | unsigned int nr_running; |
1962 | unsigned int weight; | |
1963 | enum numa_type node_type; | |
ff7db0bf | 1964 | int idle_cpu; |
58d081b5 | 1965 | }; |
e6628d5b | 1966 | |
58d081b5 MG |
1967 | struct task_numa_env { |
1968 | struct task_struct *p; | |
e6628d5b | 1969 | |
58d081b5 MG |
1970 | int src_cpu, src_nid; |
1971 | int dst_cpu, dst_nid; | |
e496132e | 1972 | int imb_numa_nr; |
e6628d5b | 1973 | |
58d081b5 | 1974 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1975 | |
40ea2b42 | 1976 | int imbalance_pct; |
7bd95320 | 1977 | int dist; |
fb13c7ee MG |
1978 | |
1979 | struct task_struct *best_task; | |
1980 | long best_imp; | |
58d081b5 MG |
1981 | int best_cpu; |
1982 | }; | |
1983 | ||
6499b1b2 | 1984 | static unsigned long cpu_load(struct rq *rq); |
8e0e0eda | 1985 | static unsigned long cpu_runnable(struct rq *rq); |
6499b1b2 VG |
1986 | |
1987 | static inline enum | |
1988 | numa_type numa_classify(unsigned int imbalance_pct, | |
1989 | struct numa_stats *ns) | |
1990 | { | |
1991 | if ((ns->nr_running > ns->weight) && | |
8e0e0eda VG |
1992 | (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || |
1993 | ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) | |
6499b1b2 VG |
1994 | return node_overloaded; |
1995 | ||
1996 | if ((ns->nr_running < ns->weight) || | |
8e0e0eda VG |
1997 | (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && |
1998 | ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) | |
6499b1b2 VG |
1999 | return node_has_spare; |
2000 | ||
2001 | return node_fully_busy; | |
2002 | } | |
2003 | ||
76c389ab VS |
2004 | #ifdef CONFIG_SCHED_SMT |
2005 | /* Forward declarations of select_idle_sibling helpers */ | |
398ba2b0 | 2006 | static inline bool test_idle_cores(int cpu); |
ff7db0bf MG |
2007 | static inline int numa_idle_core(int idle_core, int cpu) |
2008 | { | |
ff7db0bf | 2009 | if (!static_branch_likely(&sched_smt_present) || |
398ba2b0 | 2010 | idle_core >= 0 || !test_idle_cores(cpu)) |
ff7db0bf MG |
2011 | return idle_core; |
2012 | ||
2013 | /* | |
2014 | * Prefer cores instead of packing HT siblings | |
2015 | * and triggering future load balancing. | |
2016 | */ | |
2017 | if (is_core_idle(cpu)) | |
2018 | idle_core = cpu; | |
ff7db0bf MG |
2019 | |
2020 | return idle_core; | |
2021 | } | |
76c389ab VS |
2022 | #else |
2023 | static inline int numa_idle_core(int idle_core, int cpu) | |
2024 | { | |
2025 | return idle_core; | |
2026 | } | |
2027 | #endif | |
ff7db0bf | 2028 | |
6499b1b2 | 2029 | /* |
ff7db0bf MG |
2030 | * Gather all necessary information to make NUMA balancing placement |
2031 | * decisions that are compatible with standard load balancer. This | |
2032 | * borrows code and logic from update_sg_lb_stats but sharing a | |
2033 | * common implementation is impractical. | |
6499b1b2 VG |
2034 | */ |
2035 | static void update_numa_stats(struct task_numa_env *env, | |
ff7db0bf MG |
2036 | struct numa_stats *ns, int nid, |
2037 | bool find_idle) | |
6499b1b2 | 2038 | { |
ff7db0bf | 2039 | int cpu, idle_core = -1; |
6499b1b2 VG |
2040 | |
2041 | memset(ns, 0, sizeof(*ns)); | |
ff7db0bf MG |
2042 | ns->idle_cpu = -1; |
2043 | ||
0621df31 | 2044 | rcu_read_lock(); |
6499b1b2 VG |
2045 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
2046 | struct rq *rq = cpu_rq(cpu); | |
2047 | ||
2048 | ns->load += cpu_load(rq); | |
8e0e0eda | 2049 | ns->runnable += cpu_runnable(rq); |
82762d2a | 2050 | ns->util += cpu_util_cfs(cpu); |
6499b1b2 VG |
2051 | ns->nr_running += rq->cfs.h_nr_running; |
2052 | ns->compute_capacity += capacity_of(cpu); | |
ff7db0bf | 2053 | |
feaed763 | 2054 | if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) { |
ff7db0bf MG |
2055 | if (READ_ONCE(rq->numa_migrate_on) || |
2056 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) | |
2057 | continue; | |
2058 | ||
2059 | if (ns->idle_cpu == -1) | |
2060 | ns->idle_cpu = cpu; | |
2061 | ||
2062 | idle_core = numa_idle_core(idle_core, cpu); | |
2063 | } | |
6499b1b2 | 2064 | } |
0621df31 | 2065 | rcu_read_unlock(); |
6499b1b2 VG |
2066 | |
2067 | ns->weight = cpumask_weight(cpumask_of_node(nid)); | |
2068 | ||
2069 | ns->node_type = numa_classify(env->imbalance_pct, ns); | |
ff7db0bf MG |
2070 | |
2071 | if (idle_core >= 0) | |
2072 | ns->idle_cpu = idle_core; | |
6499b1b2 VG |
2073 | } |
2074 | ||
fb13c7ee MG |
2075 | static void task_numa_assign(struct task_numa_env *env, |
2076 | struct task_struct *p, long imp) | |
2077 | { | |
a4739eca SD |
2078 | struct rq *rq = cpu_rq(env->dst_cpu); |
2079 | ||
5fb52dd9 MG |
2080 | /* Check if run-queue part of active NUMA balance. */ |
2081 | if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { | |
2082 | int cpu; | |
2083 | int start = env->dst_cpu; | |
2084 | ||
2085 | /* Find alternative idle CPU. */ | |
8589018a | 2086 | for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) { |
5fb52dd9 MG |
2087 | if (cpu == env->best_cpu || !idle_cpu(cpu) || |
2088 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { | |
2089 | continue; | |
2090 | } | |
2091 | ||
2092 | env->dst_cpu = cpu; | |
2093 | rq = cpu_rq(env->dst_cpu); | |
2094 | if (!xchg(&rq->numa_migrate_on, 1)) | |
2095 | goto assign; | |
2096 | } | |
2097 | ||
2098 | /* Failed to find an alternative idle CPU */ | |
a4739eca | 2099 | return; |
5fb52dd9 | 2100 | } |
a4739eca | 2101 | |
5fb52dd9 | 2102 | assign: |
a4739eca SD |
2103 | /* |
2104 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
2105 | * found a better CPU to move/swap. | |
2106 | */ | |
5fb52dd9 | 2107 | if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { |
a4739eca SD |
2108 | rq = cpu_rq(env->best_cpu); |
2109 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
2110 | } | |
2111 | ||
fb13c7ee MG |
2112 | if (env->best_task) |
2113 | put_task_struct(env->best_task); | |
bac78573 ON |
2114 | if (p) |
2115 | get_task_struct(p); | |
fb13c7ee MG |
2116 | |
2117 | env->best_task = p; | |
2118 | env->best_imp = imp; | |
2119 | env->best_cpu = env->dst_cpu; | |
2120 | } | |
2121 | ||
28a21745 | 2122 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
2123 | struct task_numa_env *env) |
2124 | { | |
e4991b24 RR |
2125 | long imb, old_imb; |
2126 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
2127 | long src_capacity, dst_capacity; |
2128 | ||
2129 | /* | |
2130 | * The load is corrected for the CPU capacity available on each node. | |
2131 | * | |
2132 | * src_load dst_load | |
2133 | * ------------ vs --------- | |
2134 | * src_capacity dst_capacity | |
2135 | */ | |
2136 | src_capacity = env->src_stats.compute_capacity; | |
2137 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 2138 | |
5f95ba7a | 2139 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 2140 | |
28a21745 | 2141 | orig_src_load = env->src_stats.load; |
e4991b24 | 2142 | orig_dst_load = env->dst_stats.load; |
28a21745 | 2143 | |
5f95ba7a | 2144 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
2145 | |
2146 | /* Would this change make things worse? */ | |
2147 | return (imb > old_imb); | |
e63da036 RR |
2148 | } |
2149 | ||
6fd98e77 SD |
2150 | /* |
2151 | * Maximum NUMA importance can be 1998 (2*999); | |
2152 | * SMALLIMP @ 30 would be close to 1998/64. | |
2153 | * Used to deter task migration. | |
2154 | */ | |
2155 | #define SMALLIMP 30 | |
2156 | ||
fb13c7ee MG |
2157 | /* |
2158 | * This checks if the overall compute and NUMA accesses of the system would | |
2159 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
2160 | * into account that it might be best if task running on the dst_cpu should | |
2161 | * be exchanged with the source task | |
2162 | */ | |
a0f03b61 | 2163 | static bool task_numa_compare(struct task_numa_env *env, |
305c1fac | 2164 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 2165 | { |
cb361d8c | 2166 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 2167 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 2168 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 2169 | struct task_struct *cur; |
28a21745 | 2170 | long src_load, dst_load; |
7bd95320 | 2171 | int dist = env->dist; |
cb361d8c JH |
2172 | long moveimp = imp; |
2173 | long load; | |
a0f03b61 | 2174 | bool stopsearch = false; |
fb13c7ee | 2175 | |
a4739eca | 2176 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
a0f03b61 | 2177 | return false; |
a4739eca | 2178 | |
fb13c7ee | 2179 | rcu_read_lock(); |
154abafc | 2180 | cur = rcu_dereference(dst_rq->curr); |
bac78573 | 2181 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) |
fb13c7ee MG |
2182 | cur = NULL; |
2183 | ||
7af68335 PZ |
2184 | /* |
2185 | * Because we have preemption enabled we can get migrated around and | |
2186 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
2187 | */ | |
a0f03b61 MG |
2188 | if (cur == env->p) { |
2189 | stopsearch = true; | |
7af68335 | 2190 | goto unlock; |
a0f03b61 | 2191 | } |
7af68335 | 2192 | |
305c1fac | 2193 | if (!cur) { |
6fd98e77 | 2194 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
2195 | goto assign; |
2196 | else | |
2197 | goto unlock; | |
2198 | } | |
2199 | ||
88cca72c MG |
2200 | /* Skip this swap candidate if cannot move to the source cpu. */ |
2201 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) | |
2202 | goto unlock; | |
2203 | ||
2204 | /* | |
2205 | * Skip this swap candidate if it is not moving to its preferred | |
2206 | * node and the best task is. | |
2207 | */ | |
2208 | if (env->best_task && | |
2209 | env->best_task->numa_preferred_nid == env->src_nid && | |
2210 | cur->numa_preferred_nid != env->src_nid) { | |
2211 | goto unlock; | |
2212 | } | |
2213 | ||
fb13c7ee MG |
2214 | /* |
2215 | * "imp" is the fault differential for the source task between the | |
2216 | * source and destination node. Calculate the total differential for | |
2217 | * the source task and potential destination task. The more negative | |
305c1fac | 2218 | * the value is, the more remote accesses that would be expected to |
fb13c7ee | 2219 | * be incurred if the tasks were swapped. |
88cca72c | 2220 | * |
305c1fac SD |
2221 | * If dst and source tasks are in the same NUMA group, or not |
2222 | * in any group then look only at task weights. | |
2223 | */ | |
cb361d8c JH |
2224 | cur_ng = rcu_dereference(cur->numa_group); |
2225 | if (cur_ng == p_ng) { | |
13ede331 MG |
2226 | /* |
2227 | * Do not swap within a group or between tasks that have | |
2228 | * no group if there is spare capacity. Swapping does | |
2229 | * not address the load imbalance and helps one task at | |
2230 | * the cost of punishing another. | |
2231 | */ | |
2232 | if (env->dst_stats.node_type == node_has_spare) | |
2233 | goto unlock; | |
2234 | ||
305c1fac SD |
2235 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
2236 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 2237 | /* |
305c1fac SD |
2238 | * Add some hysteresis to prevent swapping the |
2239 | * tasks within a group over tiny differences. | |
887c290e | 2240 | */ |
cb361d8c | 2241 | if (cur_ng) |
305c1fac SD |
2242 | imp -= imp / 16; |
2243 | } else { | |
2244 | /* | |
2245 | * Compare the group weights. If a task is all by itself | |
2246 | * (not part of a group), use the task weight instead. | |
2247 | */ | |
cb361d8c | 2248 | if (cur_ng && p_ng) |
305c1fac SD |
2249 | imp += group_weight(cur, env->src_nid, dist) - |
2250 | group_weight(cur, env->dst_nid, dist); | |
2251 | else | |
2252 | imp += task_weight(cur, env->src_nid, dist) - | |
2253 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
2254 | } |
2255 | ||
88cca72c MG |
2256 | /* Discourage picking a task already on its preferred node */ |
2257 | if (cur->numa_preferred_nid == env->dst_nid) | |
2258 | imp -= imp / 16; | |
2259 | ||
2260 | /* | |
2261 | * Encourage picking a task that moves to its preferred node. | |
2262 | * This potentially makes imp larger than it's maximum of | |
2263 | * 1998 (see SMALLIMP and task_weight for why) but in this | |
2264 | * case, it does not matter. | |
2265 | */ | |
2266 | if (cur->numa_preferred_nid == env->src_nid) | |
2267 | imp += imp / 8; | |
2268 | ||
305c1fac | 2269 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 2270 | imp = moveimp; |
305c1fac | 2271 | cur = NULL; |
fb13c7ee | 2272 | goto assign; |
305c1fac | 2273 | } |
fb13c7ee | 2274 | |
88cca72c MG |
2275 | /* |
2276 | * Prefer swapping with a task moving to its preferred node over a | |
2277 | * task that is not. | |
2278 | */ | |
2279 | if (env->best_task && cur->numa_preferred_nid == env->src_nid && | |
2280 | env->best_task->numa_preferred_nid != env->src_nid) { | |
2281 | goto assign; | |
2282 | } | |
2283 | ||
6fd98e77 SD |
2284 | /* |
2285 | * If the NUMA importance is less than SMALLIMP, | |
2286 | * task migration might only result in ping pong | |
2287 | * of tasks and also hurt performance due to cache | |
2288 | * misses. | |
2289 | */ | |
2290 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
2291 | goto unlock; | |
2292 | ||
fb13c7ee MG |
2293 | /* |
2294 | * In the overloaded case, try and keep the load balanced. | |
2295 | */ | |
305c1fac SD |
2296 | load = task_h_load(env->p) - task_h_load(cur); |
2297 | if (!load) | |
2298 | goto assign; | |
2299 | ||
e720fff6 PZ |
2300 | dst_load = env->dst_stats.load + load; |
2301 | src_load = env->src_stats.load - load; | |
fb13c7ee | 2302 | |
28a21745 | 2303 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
2304 | goto unlock; |
2305 | ||
305c1fac | 2306 | assign: |
ff7db0bf | 2307 | /* Evaluate an idle CPU for a task numa move. */ |
10e2f1ac | 2308 | if (!cur) { |
ff7db0bf MG |
2309 | int cpu = env->dst_stats.idle_cpu; |
2310 | ||
2311 | /* Nothing cached so current CPU went idle since the search. */ | |
2312 | if (cpu < 0) | |
2313 | cpu = env->dst_cpu; | |
2314 | ||
10e2f1ac | 2315 | /* |
ff7db0bf MG |
2316 | * If the CPU is no longer truly idle and the previous best CPU |
2317 | * is, keep using it. | |
10e2f1ac | 2318 | */ |
ff7db0bf MG |
2319 | if (!idle_cpu(cpu) && env->best_cpu >= 0 && |
2320 | idle_cpu(env->best_cpu)) { | |
2321 | cpu = env->best_cpu; | |
2322 | } | |
2323 | ||
ff7db0bf | 2324 | env->dst_cpu = cpu; |
10e2f1ac | 2325 | } |
ba7e5a27 | 2326 | |
fb13c7ee | 2327 | task_numa_assign(env, cur, imp); |
a0f03b61 MG |
2328 | |
2329 | /* | |
2330 | * If a move to idle is allowed because there is capacity or load | |
2331 | * balance improves then stop the search. While a better swap | |
2332 | * candidate may exist, a search is not free. | |
2333 | */ | |
2334 | if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) | |
2335 | stopsearch = true; | |
2336 | ||
2337 | /* | |
2338 | * If a swap candidate must be identified and the current best task | |
2339 | * moves its preferred node then stop the search. | |
2340 | */ | |
2341 | if (!maymove && env->best_task && | |
2342 | env->best_task->numa_preferred_nid == env->src_nid) { | |
2343 | stopsearch = true; | |
2344 | } | |
fb13c7ee MG |
2345 | unlock: |
2346 | rcu_read_unlock(); | |
a0f03b61 MG |
2347 | |
2348 | return stopsearch; | |
fb13c7ee MG |
2349 | } |
2350 | ||
887c290e RR |
2351 | static void task_numa_find_cpu(struct task_numa_env *env, |
2352 | long taskimp, long groupimp) | |
2c8a50aa | 2353 | { |
305c1fac | 2354 | bool maymove = false; |
2c8a50aa MG |
2355 | int cpu; |
2356 | ||
305c1fac | 2357 | /* |
fb86f5b2 MG |
2358 | * If dst node has spare capacity, then check if there is an |
2359 | * imbalance that would be overruled by the load balancer. | |
305c1fac | 2360 | */ |
fb86f5b2 MG |
2361 | if (env->dst_stats.node_type == node_has_spare) { |
2362 | unsigned int imbalance; | |
2363 | int src_running, dst_running; | |
2364 | ||
2365 | /* | |
2366 | * Would movement cause an imbalance? Note that if src has | |
2367 | * more running tasks that the imbalance is ignored as the | |
2368 | * move improves the imbalance from the perspective of the | |
2369 | * CPU load balancer. | |
2370 | * */ | |
2371 | src_running = env->src_stats.nr_running - 1; | |
2372 | dst_running = env->dst_stats.nr_running + 1; | |
2373 | imbalance = max(0, dst_running - src_running); | |
7d2b5dd0 | 2374 | imbalance = adjust_numa_imbalance(imbalance, dst_running, |
e496132e | 2375 | env->imb_numa_nr); |
fb86f5b2 MG |
2376 | |
2377 | /* Use idle CPU if there is no imbalance */ | |
ff7db0bf | 2378 | if (!imbalance) { |
fb86f5b2 | 2379 | maymove = true; |
ff7db0bf MG |
2380 | if (env->dst_stats.idle_cpu >= 0) { |
2381 | env->dst_cpu = env->dst_stats.idle_cpu; | |
2382 | task_numa_assign(env, NULL, 0); | |
2383 | return; | |
2384 | } | |
2385 | } | |
fb86f5b2 MG |
2386 | } else { |
2387 | long src_load, dst_load, load; | |
2388 | /* | |
2389 | * If the improvement from just moving env->p direction is better | |
2390 | * than swapping tasks around, check if a move is possible. | |
2391 | */ | |
2392 | load = task_h_load(env->p); | |
2393 | dst_load = env->dst_stats.load + load; | |
2394 | src_load = env->src_stats.load - load; | |
2395 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
2396 | } | |
305c1fac | 2397 | |
2c8a50aa MG |
2398 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
2399 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 2400 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
2401 | continue; |
2402 | ||
2403 | env->dst_cpu = cpu; | |
a0f03b61 MG |
2404 | if (task_numa_compare(env, taskimp, groupimp, maymove)) |
2405 | break; | |
2c8a50aa MG |
2406 | } |
2407 | } | |
2408 | ||
58d081b5 MG |
2409 | static int task_numa_migrate(struct task_struct *p) |
2410 | { | |
58d081b5 MG |
2411 | struct task_numa_env env = { |
2412 | .p = p, | |
fb13c7ee | 2413 | |
58d081b5 | 2414 | .src_cpu = task_cpu(p), |
b32e86b4 | 2415 | .src_nid = task_node(p), |
fb13c7ee MG |
2416 | |
2417 | .imbalance_pct = 112, | |
2418 | ||
2419 | .best_task = NULL, | |
2420 | .best_imp = 0, | |
4142c3eb | 2421 | .best_cpu = -1, |
58d081b5 | 2422 | }; |
cb361d8c | 2423 | unsigned long taskweight, groupweight; |
58d081b5 | 2424 | struct sched_domain *sd; |
cb361d8c JH |
2425 | long taskimp, groupimp; |
2426 | struct numa_group *ng; | |
a4739eca | 2427 | struct rq *best_rq; |
7bd95320 | 2428 | int nid, ret, dist; |
e6628d5b | 2429 | |
58d081b5 | 2430 | /* |
fb13c7ee MG |
2431 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
2432 | * imbalance and would be the first to start moving tasks about. | |
2433 | * | |
2434 | * And we want to avoid any moving of tasks about, as that would create | |
2435 | * random movement of tasks -- counter the numa conditions we're trying | |
2436 | * to satisfy here. | |
58d081b5 MG |
2437 | */ |
2438 | rcu_read_lock(); | |
fb13c7ee | 2439 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
e496132e | 2440 | if (sd) { |
46a73e8a | 2441 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; |
e496132e MG |
2442 | env.imb_numa_nr = sd->imb_numa_nr; |
2443 | } | |
e6628d5b MG |
2444 | rcu_read_unlock(); |
2445 | ||
46a73e8a RR |
2446 | /* |
2447 | * Cpusets can break the scheduler domain tree into smaller | |
2448 | * balance domains, some of which do not cross NUMA boundaries. | |
2449 | * Tasks that are "trapped" in such domains cannot be migrated | |
2450 | * elsewhere, so there is no point in (re)trying. | |
2451 | */ | |
2452 | if (unlikely(!sd)) { | |
8cd45eee | 2453 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
2454 | return -EINVAL; |
2455 | } | |
2456 | ||
2c8a50aa | 2457 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
2458 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
2459 | taskweight = task_weight(p, env.src_nid, dist); | |
2460 | groupweight = group_weight(p, env.src_nid, dist); | |
ff7db0bf | 2461 | update_numa_stats(&env, &env.src_stats, env.src_nid, false); |
7bd95320 RR |
2462 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
2463 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
ff7db0bf | 2464 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
58d081b5 | 2465 | |
a43455a1 | 2466 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 2467 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 2468 | |
9de05d48 RR |
2469 | /* |
2470 | * Look at other nodes in these cases: | |
2471 | * - there is no space available on the preferred_nid | |
2472 | * - the task is part of a numa_group that is interleaved across | |
2473 | * multiple NUMA nodes; in order to better consolidate the group, | |
2474 | * we need to check other locations. | |
2475 | */ | |
cb361d8c JH |
2476 | ng = deref_curr_numa_group(p); |
2477 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
5c7b1aaf | 2478 | for_each_node_state(nid, N_CPU) { |
2c8a50aa MG |
2479 | if (nid == env.src_nid || nid == p->numa_preferred_nid) |
2480 | continue; | |
58d081b5 | 2481 | |
7bd95320 | 2482 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
2483 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
2484 | dist != env.dist) { | |
2485 | taskweight = task_weight(p, env.src_nid, dist); | |
2486 | groupweight = group_weight(p, env.src_nid, dist); | |
2487 | } | |
7bd95320 | 2488 | |
83e1d2cd | 2489 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
2490 | taskimp = task_weight(p, nid, dist) - taskweight; |
2491 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 2492 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
2493 | continue; |
2494 | ||
7bd95320 | 2495 | env.dist = dist; |
2c8a50aa | 2496 | env.dst_nid = nid; |
ff7db0bf | 2497 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
2d4056fa | 2498 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
2499 | } |
2500 | } | |
2501 | ||
68d1b02a RR |
2502 | /* |
2503 | * If the task is part of a workload that spans multiple NUMA nodes, | |
2504 | * and is migrating into one of the workload's active nodes, remember | |
2505 | * this node as the task's preferred numa node, so the workload can | |
2506 | * settle down. | |
2507 | * A task that migrated to a second choice node will be better off | |
2508 | * trying for a better one later. Do not set the preferred node here. | |
2509 | */ | |
cb361d8c | 2510 | if (ng) { |
db015dae RR |
2511 | if (env.best_cpu == -1) |
2512 | nid = env.src_nid; | |
2513 | else | |
8cd45eee | 2514 | nid = cpu_to_node(env.best_cpu); |
db015dae | 2515 | |
8cd45eee SD |
2516 | if (nid != p->numa_preferred_nid) |
2517 | sched_setnuma(p, nid); | |
db015dae RR |
2518 | } |
2519 | ||
2520 | /* No better CPU than the current one was found. */ | |
f22aef4a | 2521 | if (env.best_cpu == -1) { |
b2b2042b | 2522 | trace_sched_stick_numa(p, env.src_cpu, NULL, -1); |
db015dae | 2523 | return -EAGAIN; |
f22aef4a | 2524 | } |
0ec8aa00 | 2525 | |
a4739eca | 2526 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 2527 | if (env.best_task == NULL) { |
286549dc | 2528 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 2529 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc | 2530 | if (ret != 0) |
b2b2042b | 2531 | trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); |
fb13c7ee MG |
2532 | return ret; |
2533 | } | |
2534 | ||
0ad4e3df | 2535 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 2536 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 2537 | |
286549dc | 2538 | if (ret != 0) |
b2b2042b | 2539 | trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); |
fb13c7ee MG |
2540 | put_task_struct(env.best_task); |
2541 | return ret; | |
e6628d5b MG |
2542 | } |
2543 | ||
6b9a7460 MG |
2544 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
2545 | static void numa_migrate_preferred(struct task_struct *p) | |
2546 | { | |
5085e2a3 RR |
2547 | unsigned long interval = HZ; |
2548 | ||
2739d3ee | 2549 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 2550 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
2551 | return; |
2552 | ||
2739d3ee | 2553 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 2554 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 2555 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
2556 | |
2557 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 2558 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
2559 | return; |
2560 | ||
2561 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 2562 | task_numa_migrate(p); |
6b9a7460 MG |
2563 | } |
2564 | ||
20e07dea | 2565 | /* |
7d380f24 | 2566 | * Find out how many nodes the workload is actively running on. Do this by |
20e07dea RR |
2567 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
2568 | * be different from the set of nodes where the workload's memory is currently | |
2569 | * located. | |
20e07dea | 2570 | */ |
4142c3eb | 2571 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
2572 | { |
2573 | unsigned long faults, max_faults = 0; | |
4142c3eb | 2574 | int nid, active_nodes = 0; |
20e07dea | 2575 | |
5c7b1aaf | 2576 | for_each_node_state(nid, N_CPU) { |
20e07dea RR |
2577 | faults = group_faults_cpu(numa_group, nid); |
2578 | if (faults > max_faults) | |
2579 | max_faults = faults; | |
2580 | } | |
2581 | ||
5c7b1aaf | 2582 | for_each_node_state(nid, N_CPU) { |
20e07dea | 2583 | faults = group_faults_cpu(numa_group, nid); |
4142c3eb RR |
2584 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
2585 | active_nodes++; | |
20e07dea | 2586 | } |
4142c3eb RR |
2587 | |
2588 | numa_group->max_faults_cpu = max_faults; | |
2589 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
2590 | } |
2591 | ||
04bb2f94 RR |
2592 | /* |
2593 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
2594 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
2595 | * period will be for the next scan window. If local/(local+remote) ratio is |
2596 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
2597 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
2598 | */ |
2599 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 2600 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
2601 | |
2602 | /* | |
2603 | * Increase the scan period (slow down scanning) if the majority of | |
2604 | * our memory is already on our local node, or if the majority of | |
2605 | * the page accesses are shared with other processes. | |
2606 | * Otherwise, decrease the scan period. | |
2607 | */ | |
2608 | static void update_task_scan_period(struct task_struct *p, | |
2609 | unsigned long shared, unsigned long private) | |
2610 | { | |
2611 | unsigned int period_slot; | |
37ec97de | 2612 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
2613 | int diff; |
2614 | ||
2615 | unsigned long remote = p->numa_faults_locality[0]; | |
2616 | unsigned long local = p->numa_faults_locality[1]; | |
2617 | ||
2618 | /* | |
2619 | * If there were no record hinting faults then either the task is | |
7d380f24 | 2620 | * completely idle or all activity is in areas that are not of interest |
074c2381 MG |
2621 | * to automatic numa balancing. Related to that, if there were failed |
2622 | * migration then it implies we are migrating too quickly or the local | |
2623 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 2624 | */ |
074c2381 | 2625 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
2626 | p->numa_scan_period = min(p->numa_scan_period_max, |
2627 | p->numa_scan_period << 1); | |
2628 | ||
2629 | p->mm->numa_next_scan = jiffies + | |
2630 | msecs_to_jiffies(p->numa_scan_period); | |
2631 | ||
2632 | return; | |
2633 | } | |
2634 | ||
2635 | /* | |
2636 | * Prepare to scale scan period relative to the current period. | |
2637 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
2638 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
2639 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
2640 | */ | |
2641 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
2642 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
2643 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
2644 | ||
2645 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2646 | /* | |
2647 | * Most memory accesses are local. There is no need to | |
2648 | * do fast NUMA scanning, since memory is already local. | |
2649 | */ | |
2650 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
2651 | if (!slot) | |
2652 | slot = 1; | |
2653 | diff = slot * period_slot; | |
2654 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2655 | /* | |
2656 | * Most memory accesses are shared with other tasks. | |
2657 | * There is no point in continuing fast NUMA scanning, | |
2658 | * since other tasks may just move the memory elsewhere. | |
2659 | */ | |
2660 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
2661 | if (!slot) |
2662 | slot = 1; | |
2663 | diff = slot * period_slot; | |
2664 | } else { | |
04bb2f94 | 2665 | /* |
37ec97de RR |
2666 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
2667 | * yet they are not on the local NUMA node. Speed up | |
2668 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 2669 | */ |
37ec97de RR |
2670 | int ratio = max(lr_ratio, ps_ratio); |
2671 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
2672 | } |
2673 | ||
2674 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2675 | task_scan_min(p), task_scan_max(p)); | |
2676 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2677 | } | |
2678 | ||
7e2703e6 RR |
2679 | /* |
2680 | * Get the fraction of time the task has been running since the last | |
2681 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2682 | * decays those on a 32ms period, which is orders of magnitude off | |
2683 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2684 | * stats only if the task is so new there are no NUMA statistics yet. | |
2685 | */ | |
2686 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2687 | { | |
2688 | u64 runtime, delta, now; | |
2689 | /* Use the start of this time slice to avoid calculations. */ | |
2690 | now = p->se.exec_start; | |
2691 | runtime = p->se.sum_exec_runtime; | |
2692 | ||
2693 | if (p->last_task_numa_placement) { | |
2694 | delta = runtime - p->last_sum_exec_runtime; | |
2695 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2696 | |
2697 | /* Avoid time going backwards, prevent potential divide error: */ | |
2698 | if (unlikely((s64)*period < 0)) | |
2699 | *period = 0; | |
7e2703e6 | 2700 | } else { |
c7b50216 | 2701 | delta = p->se.avg.load_sum; |
9d89c257 | 2702 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2703 | } |
2704 | ||
2705 | p->last_sum_exec_runtime = runtime; | |
2706 | p->last_task_numa_placement = now; | |
2707 | ||
2708 | return delta; | |
2709 | } | |
2710 | ||
54009416 RR |
2711 | /* |
2712 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2713 | * be done in a way that produces consistent results with group_weight, | |
2714 | * otherwise workloads might not converge. | |
2715 | */ | |
2716 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2717 | { | |
2718 | nodemask_t nodes; | |
2719 | int dist; | |
2720 | ||
2721 | /* Direct connections between all NUMA nodes. */ | |
2722 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2723 | return nid; | |
2724 | ||
2725 | /* | |
2726 | * On a system with glueless mesh NUMA topology, group_weight | |
2727 | * scores nodes according to the number of NUMA hinting faults on | |
2728 | * both the node itself, and on nearby nodes. | |
2729 | */ | |
2730 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2731 | unsigned long score, max_score = 0; | |
2732 | int node, max_node = nid; | |
2733 | ||
2734 | dist = sched_max_numa_distance; | |
2735 | ||
5c7b1aaf | 2736 | for_each_node_state(node, N_CPU) { |
54009416 RR |
2737 | score = group_weight(p, node, dist); |
2738 | if (score > max_score) { | |
2739 | max_score = score; | |
2740 | max_node = node; | |
2741 | } | |
2742 | } | |
2743 | return max_node; | |
2744 | } | |
2745 | ||
2746 | /* | |
2747 | * Finding the preferred nid in a system with NUMA backplane | |
2748 | * interconnect topology is more involved. The goal is to locate | |
2749 | * tasks from numa_groups near each other in the system, and | |
2750 | * untangle workloads from different sides of the system. This requires | |
2751 | * searching down the hierarchy of node groups, recursively searching | |
2752 | * inside the highest scoring group of nodes. The nodemask tricks | |
2753 | * keep the complexity of the search down. | |
2754 | */ | |
5c7b1aaf | 2755 | nodes = node_states[N_CPU]; |
54009416 RR |
2756 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { |
2757 | unsigned long max_faults = 0; | |
81907478 | 2758 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2759 | int a, b; |
2760 | ||
2761 | /* Are there nodes at this distance from each other? */ | |
2762 | if (!find_numa_distance(dist)) | |
2763 | continue; | |
2764 | ||
2765 | for_each_node_mask(a, nodes) { | |
2766 | unsigned long faults = 0; | |
2767 | nodemask_t this_group; | |
2768 | nodes_clear(this_group); | |
2769 | ||
2770 | /* Sum group's NUMA faults; includes a==b case. */ | |
2771 | for_each_node_mask(b, nodes) { | |
2772 | if (node_distance(a, b) < dist) { | |
2773 | faults += group_faults(p, b); | |
2774 | node_set(b, this_group); | |
2775 | node_clear(b, nodes); | |
2776 | } | |
2777 | } | |
2778 | ||
2779 | /* Remember the top group. */ | |
2780 | if (faults > max_faults) { | |
2781 | max_faults = faults; | |
2782 | max_group = this_group; | |
2783 | /* | |
2784 | * subtle: at the smallest distance there is | |
2785 | * just one node left in each "group", the | |
2786 | * winner is the preferred nid. | |
2787 | */ | |
2788 | nid = a; | |
2789 | } | |
2790 | } | |
2791 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2792 | if (!max_faults) |
2793 | break; | |
54009416 RR |
2794 | nodes = max_group; |
2795 | } | |
2796 | return nid; | |
2797 | } | |
2798 | ||
cbee9f88 PZ |
2799 | static void task_numa_placement(struct task_struct *p) |
2800 | { | |
98fa15f3 | 2801 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2802 | unsigned long max_faults = 0; |
04bb2f94 | 2803 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2804 | unsigned long total_faults; |
2805 | u64 runtime, period; | |
7dbd13ed | 2806 | spinlock_t *group_lock = NULL; |
cb361d8c | 2807 | struct numa_group *ng; |
cbee9f88 | 2808 | |
7e5a2c17 JL |
2809 | /* |
2810 | * The p->mm->numa_scan_seq field gets updated without | |
2811 | * exclusive access. Use READ_ONCE() here to ensure | |
2812 | * that the field is read in a single access: | |
2813 | */ | |
316c1608 | 2814 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2815 | if (p->numa_scan_seq == seq) |
2816 | return; | |
2817 | p->numa_scan_seq = seq; | |
598f0ec0 | 2818 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2819 | |
7e2703e6 RR |
2820 | total_faults = p->numa_faults_locality[0] + |
2821 | p->numa_faults_locality[1]; | |
2822 | runtime = numa_get_avg_runtime(p, &period); | |
2823 | ||
7dbd13ed | 2824 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2825 | ng = deref_curr_numa_group(p); |
2826 | if (ng) { | |
2827 | group_lock = &ng->lock; | |
60e69eed | 2828 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2829 | } |
2830 | ||
688b7585 MG |
2831 | /* Find the node with the highest number of faults */ |
2832 | for_each_online_node(nid) { | |
44dba3d5 IM |
2833 | /* Keep track of the offsets in numa_faults array */ |
2834 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2835 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2836 | int priv; |
745d6147 | 2837 | |
be1e4e76 | 2838 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2839 | long diff, f_diff, f_weight; |
8c8a743c | 2840 | |
44dba3d5 IM |
2841 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2842 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2843 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2844 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2845 | |
ac8e895b | 2846 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2847 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2848 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2849 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2850 | |
7e2703e6 RR |
2851 | /* |
2852 | * Normalize the faults_from, so all tasks in a group | |
2853 | * count according to CPU use, instead of by the raw | |
2854 | * number of faults. Tasks with little runtime have | |
2855 | * little over-all impact on throughput, and thus their | |
2856 | * faults are less important. | |
2857 | */ | |
2858 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2859 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2860 | (total_faults + 1); |
44dba3d5 IM |
2861 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2862 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2863 | |
44dba3d5 IM |
2864 | p->numa_faults[mem_idx] += diff; |
2865 | p->numa_faults[cpu_idx] += f_diff; | |
2866 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2867 | p->total_numa_faults += diff; |
cb361d8c | 2868 | if (ng) { |
44dba3d5 IM |
2869 | /* |
2870 | * safe because we can only change our own group | |
2871 | * | |
2872 | * mem_idx represents the offset for a given | |
2873 | * nid and priv in a specific region because it | |
2874 | * is at the beginning of the numa_faults array. | |
2875 | */ | |
cb361d8c | 2876 | ng->faults[mem_idx] += diff; |
5b763a14 | 2877 | ng->faults[cpu_idx] += f_diff; |
cb361d8c JH |
2878 | ng->total_faults += diff; |
2879 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2880 | } |
ac8e895b MG |
2881 | } |
2882 | ||
cb361d8c | 2883 | if (!ng) { |
f03bb676 SD |
2884 | if (faults > max_faults) { |
2885 | max_faults = faults; | |
2886 | max_nid = nid; | |
2887 | } | |
2888 | } else if (group_faults > max_faults) { | |
2889 | max_faults = group_faults; | |
688b7585 MG |
2890 | max_nid = nid; |
2891 | } | |
83e1d2cd MG |
2892 | } |
2893 | ||
5c7b1aaf | 2894 | /* Cannot migrate task to CPU-less node */ |
d1db9fb4 | 2895 | max_nid = numa_nearest_node(max_nid, N_CPU); |
5c7b1aaf | 2896 | |
cb361d8c JH |
2897 | if (ng) { |
2898 | numa_group_count_active_nodes(ng); | |
60e69eed | 2899 | spin_unlock_irq(group_lock); |
f03bb676 | 2900 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2901 | } |
2902 | ||
bb97fc31 RR |
2903 | if (max_faults) { |
2904 | /* Set the new preferred node */ | |
2905 | if (max_nid != p->numa_preferred_nid) | |
2906 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2907 | } |
30619c89 SD |
2908 | |
2909 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2910 | } |
2911 | ||
8c8a743c PZ |
2912 | static inline int get_numa_group(struct numa_group *grp) |
2913 | { | |
c45a7795 | 2914 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2915 | } |
2916 | ||
2917 | static inline void put_numa_group(struct numa_group *grp) | |
2918 | { | |
c45a7795 | 2919 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2920 | kfree_rcu(grp, rcu); |
2921 | } | |
2922 | ||
3e6a9418 MG |
2923 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2924 | int *priv) | |
8c8a743c PZ |
2925 | { |
2926 | struct numa_group *grp, *my_grp; | |
2927 | struct task_struct *tsk; | |
2928 | bool join = false; | |
2929 | int cpu = cpupid_to_cpu(cpupid); | |
2930 | int i; | |
2931 | ||
cb361d8c | 2932 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2933 | unsigned int size = sizeof(struct numa_group) + |
7a2341fc BR |
2934 | NR_NUMA_HINT_FAULT_STATS * |
2935 | nr_node_ids * sizeof(unsigned long); | |
8c8a743c PZ |
2936 | |
2937 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2938 | if (!grp) | |
2939 | return; | |
2940 | ||
c45a7795 | 2941 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2942 | grp->active_nodes = 1; |
2943 | grp->max_faults_cpu = 0; | |
8c8a743c | 2944 | spin_lock_init(&grp->lock); |
e29cf08b | 2945 | grp->gid = p->pid; |
8c8a743c | 2946 | |
be1e4e76 | 2947 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2948 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2949 | |
989348b5 | 2950 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2951 | |
8c8a743c PZ |
2952 | grp->nr_tasks++; |
2953 | rcu_assign_pointer(p->numa_group, grp); | |
2954 | } | |
2955 | ||
2956 | rcu_read_lock(); | |
316c1608 | 2957 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2958 | |
2959 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2960 | goto no_join; |
8c8a743c PZ |
2961 | |
2962 | grp = rcu_dereference(tsk->numa_group); | |
2963 | if (!grp) | |
3354781a | 2964 | goto no_join; |
8c8a743c | 2965 | |
cb361d8c | 2966 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2967 | if (grp == my_grp) |
3354781a | 2968 | goto no_join; |
8c8a743c PZ |
2969 | |
2970 | /* | |
2971 | * Only join the other group if its bigger; if we're the bigger group, | |
2972 | * the other task will join us. | |
2973 | */ | |
2974 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2975 | goto no_join; |
8c8a743c PZ |
2976 | |
2977 | /* | |
2978 | * Tie-break on the grp address. | |
2979 | */ | |
2980 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2981 | goto no_join; |
8c8a743c | 2982 | |
dabe1d99 RR |
2983 | /* Always join threads in the same process. */ |
2984 | if (tsk->mm == current->mm) | |
2985 | join = true; | |
2986 | ||
2987 | /* Simple filter to avoid false positives due to PID collisions */ | |
2988 | if (flags & TNF_SHARED) | |
2989 | join = true; | |
8c8a743c | 2990 | |
3e6a9418 MG |
2991 | /* Update priv based on whether false sharing was detected */ |
2992 | *priv = !join; | |
2993 | ||
dabe1d99 | 2994 | if (join && !get_numa_group(grp)) |
3354781a | 2995 | goto no_join; |
8c8a743c | 2996 | |
8c8a743c PZ |
2997 | rcu_read_unlock(); |
2998 | ||
2999 | if (!join) | |
3000 | return; | |
3001 | ||
09348d75 | 3002 | WARN_ON_ONCE(irqs_disabled()); |
60e69eed | 3003 | double_lock_irq(&my_grp->lock, &grp->lock); |
989348b5 | 3004 | |
be1e4e76 | 3005 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
3006 | my_grp->faults[i] -= p->numa_faults[i]; |
3007 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 3008 | } |
989348b5 MG |
3009 | my_grp->total_faults -= p->total_numa_faults; |
3010 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 3011 | |
8c8a743c PZ |
3012 | my_grp->nr_tasks--; |
3013 | grp->nr_tasks++; | |
3014 | ||
3015 | spin_unlock(&my_grp->lock); | |
60e69eed | 3016 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
3017 | |
3018 | rcu_assign_pointer(p->numa_group, grp); | |
3019 | ||
3020 | put_numa_group(my_grp); | |
3354781a PZ |
3021 | return; |
3022 | ||
3023 | no_join: | |
3024 | rcu_read_unlock(); | |
3025 | return; | |
8c8a743c PZ |
3026 | } |
3027 | ||
16d51a59 | 3028 | /* |
3b03706f | 3029 | * Get rid of NUMA statistics associated with a task (either current or dead). |
16d51a59 JH |
3030 | * If @final is set, the task is dead and has reached refcount zero, so we can |
3031 | * safely free all relevant data structures. Otherwise, there might be | |
3032 | * concurrent reads from places like load balancing and procfs, and we should | |
3033 | * reset the data back to default state without freeing ->numa_faults. | |
3034 | */ | |
3035 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 3036 | { |
cb361d8c JH |
3037 | /* safe: p either is current or is being freed by current */ |
3038 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 3039 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
3040 | unsigned long flags; |
3041 | int i; | |
8c8a743c | 3042 | |
16d51a59 JH |
3043 | if (!numa_faults) |
3044 | return; | |
3045 | ||
8c8a743c | 3046 | if (grp) { |
e9dd685c | 3047 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 3048 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 3049 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 3050 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 3051 | |
8c8a743c | 3052 | grp->nr_tasks--; |
e9dd685c | 3053 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 3054 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
3055 | put_numa_group(grp); |
3056 | } | |
3057 | ||
16d51a59 JH |
3058 | if (final) { |
3059 | p->numa_faults = NULL; | |
3060 | kfree(numa_faults); | |
3061 | } else { | |
3062 | p->total_numa_faults = 0; | |
3063 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
3064 | numa_faults[i] = 0; | |
3065 | } | |
8c8a743c PZ |
3066 | } |
3067 | ||
cbee9f88 PZ |
3068 | /* |
3069 | * Got a PROT_NONE fault for a page on @node. | |
3070 | */ | |
58b46da3 | 3071 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
3072 | { |
3073 | struct task_struct *p = current; | |
6688cc05 | 3074 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 3075 | int cpu_node = task_node(current); |
792568ec | 3076 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 3077 | struct numa_group *ng; |
ac8e895b | 3078 | int priv; |
cbee9f88 | 3079 | |
2a595721 | 3080 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
3081 | return; |
3082 | ||
9ff1d9ff MG |
3083 | /* for example, ksmd faulting in a user's mm */ |
3084 | if (!p->mm) | |
3085 | return; | |
3086 | ||
33024536 HY |
3087 | /* |
3088 | * NUMA faults statistics are unnecessary for the slow memory | |
3089 | * node for memory tiering mode. | |
3090 | */ | |
3091 | if (!node_is_toptier(mem_node) && | |
3092 | (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING || | |
3093 | !cpupid_valid(last_cpupid))) | |
3094 | return; | |
3095 | ||
f809ca9a | 3096 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
3097 | if (unlikely(!p->numa_faults)) { |
3098 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 3099 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 3100 | |
44dba3d5 IM |
3101 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
3102 | if (!p->numa_faults) | |
f809ca9a | 3103 | return; |
745d6147 | 3104 | |
83e1d2cd | 3105 | p->total_numa_faults = 0; |
04bb2f94 | 3106 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 3107 | } |
cbee9f88 | 3108 | |
8c8a743c PZ |
3109 | /* |
3110 | * First accesses are treated as private, otherwise consider accesses | |
3111 | * to be private if the accessing pid has not changed | |
3112 | */ | |
3113 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
3114 | priv = 1; | |
3115 | } else { | |
3116 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 3117 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 3118 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
3119 | } |
3120 | ||
792568ec RR |
3121 | /* |
3122 | * If a workload spans multiple NUMA nodes, a shared fault that | |
3123 | * occurs wholly within the set of nodes that the workload is | |
3124 | * actively using should be counted as local. This allows the | |
3125 | * scan rate to slow down when a workload has settled down. | |
3126 | */ | |
cb361d8c | 3127 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
3128 | if (!priv && !local && ng && ng->active_nodes > 1 && |
3129 | numa_is_active_node(cpu_node, ng) && | |
3130 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
3131 | local = 1; |
3132 | ||
2739d3ee | 3133 | /* |
e1ff516a YW |
3134 | * Retry to migrate task to preferred node periodically, in case it |
3135 | * previously failed, or the scheduler moved us. | |
2739d3ee | 3136 | */ |
b6a60cf3 SD |
3137 | if (time_after(jiffies, p->numa_migrate_retry)) { |
3138 | task_numa_placement(p); | |
6b9a7460 | 3139 | numa_migrate_preferred(p); |
b6a60cf3 | 3140 | } |
6b9a7460 | 3141 | |
b32e86b4 IM |
3142 | if (migrated) |
3143 | p->numa_pages_migrated += pages; | |
074c2381 MG |
3144 | if (flags & TNF_MIGRATE_FAIL) |
3145 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 3146 | |
44dba3d5 IM |
3147 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
3148 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 3149 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
3150 | } |
3151 | ||
6e5fb223 PZ |
3152 | static void reset_ptenuma_scan(struct task_struct *p) |
3153 | { | |
7e5a2c17 JL |
3154 | /* |
3155 | * We only did a read acquisition of the mmap sem, so | |
3156 | * p->mm->numa_scan_seq is written to without exclusive access | |
3157 | * and the update is not guaranteed to be atomic. That's not | |
3158 | * much of an issue though, since this is just used for | |
3159 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
3160 | * expensive, to avoid any form of compiler optimizations: | |
3161 | */ | |
316c1608 | 3162 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
3163 | p->mm->numa_scan_offset = 0; |
3164 | } | |
3165 | ||
b7a5b537 | 3166 | static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma) |
fc137c0d | 3167 | { |
20f58648 | 3168 | unsigned long pids; |
fc137c0d R |
3169 | /* |
3170 | * Allow unconditional access first two times, so that all the (pages) | |
3171 | * of VMAs get prot_none fault introduced irrespective of accesses. | |
3172 | * This is also done to avoid any side effect of task scanning | |
3173 | * amplifying the unfairness of disjoint set of VMAs' access. | |
3174 | */ | |
84db47ca | 3175 | if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2) |
fc137c0d R |
3176 | return true; |
3177 | ||
f3a6c979 | 3178 | pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1]; |
b7a5b537 MG |
3179 | if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids)) |
3180 | return true; | |
3181 | ||
3182 | /* | |
3183 | * Complete a scan that has already started regardless of PID access, or | |
3184 | * some VMAs may never be scanned in multi-threaded applications: | |
3185 | */ | |
3186 | if (mm->numa_scan_offset > vma->vm_start) { | |
3187 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID); | |
3188 | return true; | |
3189 | } | |
3190 | ||
3191 | return false; | |
fc137c0d R |
3192 | } |
3193 | ||
20f58648 R |
3194 | #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay) |
3195 | ||
cbee9f88 PZ |
3196 | /* |
3197 | * The expensive part of numa migration is done from task_work context. | |
3198 | * Triggered from task_tick_numa(). | |
3199 | */ | |
9434f9f5 | 3200 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
3201 | { |
3202 | unsigned long migrate, next_scan, now = jiffies; | |
3203 | struct task_struct *p = current; | |
3204 | struct mm_struct *mm = p->mm; | |
51170840 | 3205 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 3206 | struct vm_area_struct *vma; |
9f40604c | 3207 | unsigned long start, end; |
598f0ec0 | 3208 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 3209 | long pages, virtpages; |
214dbc42 | 3210 | struct vma_iterator vmi; |
f169c62f MG |
3211 | bool vma_pids_skipped; |
3212 | bool vma_pids_forced = false; | |
cbee9f88 | 3213 | |
9148a3a1 | 3214 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 3215 | |
b34920d4 | 3216 | work->next = work; |
cbee9f88 PZ |
3217 | /* |
3218 | * Who cares about NUMA placement when they're dying. | |
3219 | * | |
3220 | * NOTE: make sure not to dereference p->mm before this check, | |
3221 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
3222 | * without p->mm even though we still had it when we enqueued this | |
3223 | * work. | |
3224 | */ | |
3225 | if (p->flags & PF_EXITING) | |
3226 | return; | |
3227 | ||
930aa174 | 3228 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
3229 | mm->numa_next_scan = now + |
3230 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
3231 | } |
3232 | ||
cbee9f88 PZ |
3233 | /* |
3234 | * Enforce maximal scan/migration frequency.. | |
3235 | */ | |
3236 | migrate = mm->numa_next_scan; | |
3237 | if (time_before(now, migrate)) | |
3238 | return; | |
3239 | ||
598f0ec0 MG |
3240 | if (p->numa_scan_period == 0) { |
3241 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 3242 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 3243 | } |
cbee9f88 | 3244 | |
fb003b80 | 3245 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
8baceabc | 3246 | if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan)) |
cbee9f88 PZ |
3247 | return; |
3248 | ||
19a78d11 PZ |
3249 | /* |
3250 | * Delay this task enough that another task of this mm will likely win | |
3251 | * the next time around. | |
3252 | */ | |
3253 | p->node_stamp += 2 * TICK_NSEC; | |
3254 | ||
9f40604c MG |
3255 | pages = sysctl_numa_balancing_scan_size; |
3256 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 3257 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
3258 | if (!pages) |
3259 | return; | |
cbee9f88 | 3260 | |
4620f8c1 | 3261 | |
d8ed45c5 | 3262 | if (!mmap_read_trylock(mm)) |
8655d549 | 3263 | return; |
f169c62f MG |
3264 | |
3265 | /* | |
3266 | * VMAs are skipped if the current PID has not trapped a fault within | |
3267 | * the VMA recently. Allow scanning to be forced if there is no | |
3268 | * suitable VMA remaining. | |
3269 | */ | |
3270 | vma_pids_skipped = false; | |
3271 | ||
3272 | retry_pids: | |
3273 | start = mm->numa_scan_offset; | |
214dbc42 LH |
3274 | vma_iter_init(&vmi, mm, start); |
3275 | vma = vma_next(&vmi); | |
6e5fb223 PZ |
3276 | if (!vma) { |
3277 | reset_ptenuma_scan(p); | |
9f40604c | 3278 | start = 0; |
214dbc42 LH |
3279 | vma_iter_set(&vmi, start); |
3280 | vma = vma_next(&vmi); | |
6e5fb223 | 3281 | } |
0cd4d02c | 3282 | |
214dbc42 | 3283 | do { |
6b79c57b | 3284 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 3285 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
ed2da8b7 | 3286 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE); |
6e5fb223 | 3287 | continue; |
6b79c57b | 3288 | } |
6e5fb223 | 3289 | |
4591ce4f MG |
3290 | /* |
3291 | * Shared library pages mapped by multiple processes are not | |
3292 | * migrated as it is expected they are cache replicated. Avoid | |
b9e6e286 | 3293 | * hinting faults in read-only file-backed mappings or the vDSO |
4591ce4f MG |
3294 | * as migrating the pages will be of marginal benefit. |
3295 | */ | |
3296 | if (!vma->vm_mm || | |
ed2da8b7 MG |
3297 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) { |
3298 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO); | |
4591ce4f | 3299 | continue; |
ed2da8b7 | 3300 | } |
4591ce4f | 3301 | |
3c67f474 MG |
3302 | /* |
3303 | * Skip inaccessible VMAs to avoid any confusion between | |
b9e6e286 | 3304 | * PROT_NONE and NUMA hinting PTEs |
3c67f474 | 3305 | */ |
ed2da8b7 MG |
3306 | if (!vma_is_accessible(vma)) { |
3307 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE); | |
3c67f474 | 3308 | continue; |
ed2da8b7 | 3309 | } |
4591ce4f | 3310 | |
ef6a22b7 MG |
3311 | /* Initialise new per-VMA NUMAB state. */ |
3312 | if (!vma->numab_state) { | |
3313 | vma->numab_state = kzalloc(sizeof(struct vma_numab_state), | |
3314 | GFP_KERNEL); | |
3315 | if (!vma->numab_state) | |
3316 | continue; | |
3317 | ||
84db47ca R |
3318 | vma->numab_state->start_scan_seq = mm->numa_scan_seq; |
3319 | ||
ef6a22b7 MG |
3320 | vma->numab_state->next_scan = now + |
3321 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
20f58648 R |
3322 | |
3323 | /* Reset happens after 4 times scan delay of scan start */ | |
f3a6c979 | 3324 | vma->numab_state->pids_active_reset = vma->numab_state->next_scan + |
20f58648 | 3325 | msecs_to_jiffies(VMA_PID_RESET_PERIOD); |
f169c62f MG |
3326 | |
3327 | /* | |
3328 | * Ensure prev_scan_seq does not match numa_scan_seq, | |
3329 | * to prevent VMAs being skipped prematurely on the | |
3330 | * first scan: | |
3331 | */ | |
3332 | vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1; | |
ef6a22b7 MG |
3333 | } |
3334 | ||
3335 | /* | |
b9e6e286 | 3336 | * Scanning the VMAs of short lived tasks add more overhead. So |
ef6a22b7 MG |
3337 | * delay the scan for new VMAs. |
3338 | */ | |
3339 | if (mm->numa_scan_seq && time_before(jiffies, | |
ed2da8b7 MG |
3340 | vma->numab_state->next_scan)) { |
3341 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY); | |
ef6a22b7 | 3342 | continue; |
ed2da8b7 | 3343 | } |
ef6a22b7 | 3344 | |
2e2675db | 3345 | /* RESET access PIDs regularly for old VMAs. */ |
20f58648 | 3346 | if (mm->numa_scan_seq && |
f3a6c979 MG |
3347 | time_after(jiffies, vma->numab_state->pids_active_reset)) { |
3348 | vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset + | |
20f58648 | 3349 | msecs_to_jiffies(VMA_PID_RESET_PERIOD); |
f3a6c979 MG |
3350 | vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]); |
3351 | vma->numab_state->pids_active[1] = 0; | |
20f58648 R |
3352 | } |
3353 | ||
f169c62f MG |
3354 | /* Do not rescan VMAs twice within the same sequence. */ |
3355 | if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) { | |
3356 | mm->numa_scan_offset = vma->vm_end; | |
3357 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED); | |
3358 | continue; | |
3359 | } | |
3360 | ||
3361 | /* | |
3362 | * Do not scan the VMA if task has not accessed it, unless no other | |
3363 | * VMA candidate exists. | |
3364 | */ | |
3365 | if (!vma_pids_forced && !vma_is_accessed(mm, vma)) { | |
3366 | vma_pids_skipped = true; | |
2e2675db R |
3367 | trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE); |
3368 | continue; | |
3369 | } | |
3370 | ||
9f40604c MG |
3371 | do { |
3372 | start = max(start, vma->vm_start); | |
3373 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
3374 | end = min(end, vma->vm_end); | |
4620f8c1 | 3375 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
3376 | |
3377 | /* | |
4620f8c1 RR |
3378 | * Try to scan sysctl_numa_balancing_size worth of |
3379 | * hpages that have at least one present PTE that | |
b9e6e286 | 3380 | * is not already PTE-numa. If the VMA contains |
4620f8c1 RR |
3381 | * areas that are unused or already full of prot_numa |
3382 | * PTEs, scan up to virtpages, to skip through those | |
3383 | * areas faster. | |
598f0ec0 MG |
3384 | */ |
3385 | if (nr_pte_updates) | |
3386 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 3387 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 3388 | |
9f40604c | 3389 | start = end; |
4620f8c1 | 3390 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 3391 | goto out; |
3cf1962c RR |
3392 | |
3393 | cond_resched(); | |
9f40604c | 3394 | } while (end != vma->vm_end); |
f169c62f MG |
3395 | |
3396 | /* VMA scan is complete, do not scan until next sequence. */ | |
3397 | vma->numab_state->prev_scan_seq = mm->numa_scan_seq; | |
3398 | ||
3399 | /* | |
3400 | * Only force scan within one VMA at a time, to limit the | |
3401 | * cost of scanning a potentially uninteresting VMA. | |
3402 | */ | |
3403 | if (vma_pids_forced) | |
3404 | break; | |
214dbc42 | 3405 | } for_each_vma(vmi, vma); |
6e5fb223 | 3406 | |
f169c62f MG |
3407 | /* |
3408 | * If no VMAs are remaining and VMAs were skipped due to the PID | |
3409 | * not accessing the VMA previously, then force a scan to ensure | |
3410 | * forward progress: | |
3411 | */ | |
3412 | if (!vma && !vma_pids_forced && vma_pids_skipped) { | |
3413 | vma_pids_forced = true; | |
3414 | goto retry_pids; | |
3415 | } | |
3416 | ||
9f40604c | 3417 | out: |
6e5fb223 | 3418 | /* |
c69307d5 PZ |
3419 | * It is possible to reach the end of the VMA list but the last few |
3420 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
3421 | * would find the !migratable VMA on the next scan but not reset the | |
3422 | * scanner to the start so check it now. | |
6e5fb223 PZ |
3423 | */ |
3424 | if (vma) | |
9f40604c | 3425 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
3426 | else |
3427 | reset_ptenuma_scan(p); | |
d8ed45c5 | 3428 | mmap_read_unlock(mm); |
51170840 RR |
3429 | |
3430 | /* | |
3431 | * Make sure tasks use at least 32x as much time to run other code | |
3432 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
3433 | * Usually update_task_scan_period slows down scanning enough; on an | |
3434 | * overloaded system we need to limit overhead on a per task basis. | |
3435 | */ | |
3436 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
3437 | u64 diff = p->se.sum_exec_runtime - runtime; | |
3438 | p->node_stamp += 32 * diff; | |
3439 | } | |
cbee9f88 PZ |
3440 | } |
3441 | ||
d35927a1 VS |
3442 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
3443 | { | |
3444 | int mm_users = 0; | |
3445 | struct mm_struct *mm = p->mm; | |
3446 | ||
3447 | if (mm) { | |
3448 | mm_users = atomic_read(&mm->mm_users); | |
3449 | if (mm_users == 1) { | |
3450 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
3451 | mm->numa_scan_seq = 0; | |
3452 | } | |
3453 | } | |
3454 | p->node_stamp = 0; | |
3455 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
3456 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
70ce3ea9 | 3457 | p->numa_migrate_retry = 0; |
b34920d4 | 3458 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
3459 | p->numa_work.next = &p->numa_work; |
3460 | p->numa_faults = NULL; | |
12bf8a7e HW |
3461 | p->numa_pages_migrated = 0; |
3462 | p->total_numa_faults = 0; | |
d35927a1 VS |
3463 | RCU_INIT_POINTER(p->numa_group, NULL); |
3464 | p->last_task_numa_placement = 0; | |
3465 | p->last_sum_exec_runtime = 0; | |
3466 | ||
b34920d4 VS |
3467 | init_task_work(&p->numa_work, task_numa_work); |
3468 | ||
d35927a1 VS |
3469 | /* New address space, reset the preferred nid */ |
3470 | if (!(clone_flags & CLONE_VM)) { | |
3471 | p->numa_preferred_nid = NUMA_NO_NODE; | |
3472 | return; | |
3473 | } | |
3474 | ||
3475 | /* | |
3476 | * New thread, keep existing numa_preferred_nid which should be copied | |
3477 | * already by arch_dup_task_struct but stagger when scans start. | |
3478 | */ | |
3479 | if (mm) { | |
3480 | unsigned int delay; | |
3481 | ||
3482 | delay = min_t(unsigned int, task_scan_max(current), | |
3483 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
3484 | delay += 2 * TICK_NSEC; | |
3485 | p->node_stamp = delay; | |
3486 | } | |
3487 | } | |
3488 | ||
cbee9f88 PZ |
3489 | /* |
3490 | * Drive the periodic memory faults.. | |
3491 | */ | |
b1546edc | 3492 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
3493 | { |
3494 | struct callback_head *work = &curr->numa_work; | |
3495 | u64 period, now; | |
3496 | ||
3497 | /* | |
3498 | * We don't care about NUMA placement if we don't have memory. | |
3499 | */ | |
b3f9916d | 3500 | if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) |
cbee9f88 PZ |
3501 | return; |
3502 | ||
3503 | /* | |
3504 | * Using runtime rather than walltime has the dual advantage that | |
3505 | * we (mostly) drive the selection from busy threads and that the | |
3506 | * task needs to have done some actual work before we bother with | |
3507 | * NUMA placement. | |
3508 | */ | |
3509 | now = curr->se.sum_exec_runtime; | |
3510 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
3511 | ||
25b3e5a3 | 3512 | if (now > curr->node_stamp + period) { |
4b96a29b | 3513 | if (!curr->node_stamp) |
b5dd77c8 | 3514 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 3515 | curr->node_stamp += period; |
cbee9f88 | 3516 | |
b34920d4 | 3517 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
91989c70 | 3518 | task_work_add(curr, work, TWA_RESUME); |
cbee9f88 PZ |
3519 | } |
3520 | } | |
3fed382b | 3521 | |
3f9672ba SD |
3522 | static void update_scan_period(struct task_struct *p, int new_cpu) |
3523 | { | |
3524 | int src_nid = cpu_to_node(task_cpu(p)); | |
3525 | int dst_nid = cpu_to_node(new_cpu); | |
3526 | ||
05cbdf4f MG |
3527 | if (!static_branch_likely(&sched_numa_balancing)) |
3528 | return; | |
3529 | ||
3f9672ba SD |
3530 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
3531 | return; | |
3532 | ||
05cbdf4f MG |
3533 | if (src_nid == dst_nid) |
3534 | return; | |
3535 | ||
3536 | /* | |
3537 | * Allow resets if faults have been trapped before one scan | |
3538 | * has completed. This is most likely due to a new task that | |
3539 | * is pulled cross-node due to wakeups or load balancing. | |
3540 | */ | |
3541 | if (p->numa_scan_seq) { | |
3542 | /* | |
3543 | * Avoid scan adjustments if moving to the preferred | |
3544 | * node or if the task was not previously running on | |
3545 | * the preferred node. | |
3546 | */ | |
3547 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
3548 | (p->numa_preferred_nid != NUMA_NO_NODE && |
3549 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
3550 | return; |
3551 | } | |
3552 | ||
3553 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
3554 | } |
3555 | ||
cbee9f88 PZ |
3556 | #else |
3557 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
3558 | { | |
3559 | } | |
0ec8aa00 PZ |
3560 | |
3561 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
3562 | { | |
3563 | } | |
3564 | ||
3565 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
3566 | { | |
3567 | } | |
3fed382b | 3568 | |
3f9672ba SD |
3569 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
3570 | { | |
3571 | } | |
3572 | ||
cbee9f88 PZ |
3573 | #endif /* CONFIG_NUMA_BALANCING */ |
3574 | ||
30cfdcfc DA |
3575 | static void |
3576 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3577 | { | |
3578 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 3579 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3580 | if (entity_is_task(se)) { |
3581 | struct rq *rq = rq_of(cfs_rq); | |
3582 | ||
3583 | account_numa_enqueue(rq, task_of(se)); | |
3584 | list_add(&se->group_node, &rq->cfs_tasks); | |
3585 | } | |
367456c7 | 3586 | #endif |
30cfdcfc | 3587 | cfs_rq->nr_running++; |
a480adde JD |
3588 | if (se_is_idle(se)) |
3589 | cfs_rq->idle_nr_running++; | |
30cfdcfc DA |
3590 | } |
3591 | ||
3592 | static void | |
3593 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3594 | { | |
3595 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 3596 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3597 | if (entity_is_task(se)) { |
3598 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 3599 | list_del_init(&se->group_node); |
0ec8aa00 | 3600 | } |
bfdb198c | 3601 | #endif |
30cfdcfc | 3602 | cfs_rq->nr_running--; |
a480adde JD |
3603 | if (se_is_idle(se)) |
3604 | cfs_rq->idle_nr_running--; | |
30cfdcfc DA |
3605 | } |
3606 | ||
8d5b9025 PZ |
3607 | /* |
3608 | * Signed add and clamp on underflow. | |
3609 | * | |
3610 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3611 | * memory. This allows lockless observations without ever seeing the negative | |
3612 | * values. | |
3613 | */ | |
3614 | #define add_positive(_ptr, _val) do { \ | |
3615 | typeof(_ptr) ptr = (_ptr); \ | |
3616 | typeof(_val) val = (_val); \ | |
3617 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3618 | \ | |
3619 | res = var + val; \ | |
3620 | \ | |
3621 | if (val < 0 && res > var) \ | |
3622 | res = 0; \ | |
3623 | \ | |
3624 | WRITE_ONCE(*ptr, res); \ | |
3625 | } while (0) | |
3626 | ||
3627 | /* | |
3628 | * Unsigned subtract and clamp on underflow. | |
3629 | * | |
3630 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3631 | * memory. This allows lockless observations without ever seeing the negative | |
3632 | * values. | |
3633 | */ | |
3634 | #define sub_positive(_ptr, _val) do { \ | |
3635 | typeof(_ptr) ptr = (_ptr); \ | |
3636 | typeof(*ptr) val = (_val); \ | |
3637 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3638 | res = var - val; \ | |
3639 | if (res > var) \ | |
3640 | res = 0; \ | |
3641 | WRITE_ONCE(*ptr, res); \ | |
3642 | } while (0) | |
3643 | ||
b5c0ce7b PB |
3644 | /* |
3645 | * Remove and clamp on negative, from a local variable. | |
3646 | * | |
3647 | * A variant of sub_positive(), which does not use explicit load-store | |
3648 | * and is thus optimized for local variable updates. | |
3649 | */ | |
3650 | #define lsub_positive(_ptr, _val) do { \ | |
3651 | typeof(_ptr) ptr = (_ptr); \ | |
3652 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
3653 | } while (0) | |
3654 | ||
8d5b9025 | 3655 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
3656 | static inline void |
3657 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3658 | { | |
3659 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3660 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
3661 | } | |
3662 | ||
3663 | static inline void | |
3664 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3665 | { | |
3666 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
2d02fa8c VG |
3667 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); |
3668 | /* See update_cfs_rq_load_avg() */ | |
3669 | cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, | |
3670 | cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); | |
8d5b9025 PZ |
3671 | } |
3672 | #else | |
3673 | static inline void | |
8d5b9025 PZ |
3674 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } |
3675 | static inline void | |
3676 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
3677 | #endif | |
3678 | ||
afae8002 | 3679 | static void reweight_eevdf(struct sched_entity *se, u64 avruntime, |
eab03c23 AW |
3680 | unsigned long weight) |
3681 | { | |
3682 | unsigned long old_weight = se->load.weight; | |
eab03c23 AW |
3683 | s64 vlag, vslice; |
3684 | ||
3685 | /* | |
3686 | * VRUNTIME | |
be8858db | 3687 | * -------- |
eab03c23 AW |
3688 | * |
3689 | * COROLLARY #1: The virtual runtime of the entity needs to be | |
3690 | * adjusted if re-weight at !0-lag point. | |
3691 | * | |
3692 | * Proof: For contradiction assume this is not true, so we can | |
3693 | * re-weight without changing vruntime at !0-lag point. | |
3694 | * | |
3695 | * Weight VRuntime Avg-VRuntime | |
3696 | * before w v V | |
3697 | * after w' v' V' | |
3698 | * | |
3699 | * Since lag needs to be preserved through re-weight: | |
3700 | * | |
3701 | * lag = (V - v)*w = (V'- v')*w', where v = v' | |
3702 | * ==> V' = (V - v)*w/w' + v (1) | |
3703 | * | |
3704 | * Let W be the total weight of the entities before reweight, | |
3705 | * since V' is the new weighted average of entities: | |
3706 | * | |
3707 | * V' = (WV + w'v - wv) / (W + w' - w) (2) | |
3708 | * | |
3709 | * by using (1) & (2) we obtain: | |
3710 | * | |
3711 | * (WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v | |
3712 | * ==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v | |
3713 | * ==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v | |
3714 | * ==> (V - v)*W/(W + w' - w) = (V - v)*w/w' (3) | |
3715 | * | |
3716 | * Since we are doing at !0-lag point which means V != v, we | |
3717 | * can simplify (3): | |
3718 | * | |
3719 | * ==> W / (W + w' - w) = w / w' | |
3720 | * ==> Ww' = Ww + ww' - ww | |
3721 | * ==> W * (w' - w) = w * (w' - w) | |
3722 | * ==> W = w (re-weight indicates w' != w) | |
3723 | * | |
3724 | * So the cfs_rq contains only one entity, hence vruntime of | |
3725 | * the entity @v should always equal to the cfs_rq's weighted | |
3726 | * average vruntime @V, which means we will always re-weight | |
3727 | * at 0-lag point, thus breach assumption. Proof completed. | |
3728 | * | |
3729 | * | |
3730 | * COROLLARY #2: Re-weight does NOT affect weighted average | |
3731 | * vruntime of all the entities. | |
3732 | * | |
3733 | * Proof: According to corollary #1, Eq. (1) should be: | |
3734 | * | |
3735 | * (V - v)*w = (V' - v')*w' | |
3736 | * ==> v' = V' - (V - v)*w/w' (4) | |
3737 | * | |
3738 | * According to the weighted average formula, we have: | |
3739 | * | |
3740 | * V' = (WV - wv + w'v') / (W - w + w') | |
3741 | * = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w') | |
3742 | * = (WV - wv + w'V' - Vw + wv) / (W - w + w') | |
3743 | * = (WV + w'V' - Vw) / (W - w + w') | |
3744 | * | |
3745 | * ==> V'*(W - w + w') = WV + w'V' - Vw | |
3746 | * ==> V' * (W - w) = (W - w) * V (5) | |
3747 | * | |
3748 | * If the entity is the only one in the cfs_rq, then reweight | |
3749 | * always occurs at 0-lag point, so V won't change. Or else | |
3750 | * there are other entities, hence W != w, then Eq. (5) turns | |
3751 | * into V' = V. So V won't change in either case, proof done. | |
3752 | * | |
3753 | * | |
3754 | * So according to corollary #1 & #2, the effect of re-weight | |
3755 | * on vruntime should be: | |
3756 | * | |
3757 | * v' = V' - (V - v) * w / w' (4) | |
3758 | * = V - (V - v) * w / w' | |
3759 | * = V - vl * w / w' | |
3760 | * = V - vl' | |
3761 | */ | |
3762 | if (avruntime != se->vruntime) { | |
1560d1f6 | 3763 | vlag = entity_lag(avruntime, se); |
eab03c23 AW |
3764 | vlag = div_s64(vlag * old_weight, weight); |
3765 | se->vruntime = avruntime - vlag; | |
3766 | } | |
3767 | ||
3768 | /* | |
3769 | * DEADLINE | |
be8858db | 3770 | * -------- |
eab03c23 AW |
3771 | * |
3772 | * When the weight changes, the virtual time slope changes and | |
3773 | * we should adjust the relative virtual deadline accordingly. | |
3774 | * | |
3775 | * d' = v' + (d - v)*w/w' | |
3776 | * = V' - (V - v)*w/w' + (d - v)*w/w' | |
3777 | * = V - (V - v)*w/w' + (d - v)*w/w' | |
3778 | * = V + (d - V)*w/w' | |
3779 | */ | |
3780 | vslice = (s64)(se->deadline - avruntime); | |
3781 | vslice = div_s64(vslice * old_weight, weight); | |
3782 | se->deadline = avruntime + vslice; | |
3783 | } | |
3784 | ||
9059393e | 3785 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
0dacee1b | 3786 | unsigned long weight) |
9059393e | 3787 | { |
eab03c23 | 3788 | bool curr = cfs_rq->curr == se; |
afae8002 | 3789 | u64 avruntime; |
86bfbb7c | 3790 | |
9059393e VG |
3791 | if (se->on_rq) { |
3792 | /* commit outstanding execution time */ | |
11b1b8bc | 3793 | update_curr(cfs_rq); |
afae8002 | 3794 | avruntime = avg_vruntime(cfs_rq); |
11b1b8bc | 3795 | if (!curr) |
eab03c23 | 3796 | __dequeue_entity(cfs_rq, se); |
1724b95b | 3797 | update_load_sub(&cfs_rq->load, se->load.weight); |
9059393e VG |
3798 | } |
3799 | dequeue_load_avg(cfs_rq, se); | |
3800 | ||
afae8002 TD |
3801 | if (se->on_rq) { |
3802 | reweight_eevdf(se, avruntime, weight); | |
3803 | } else { | |
86bfbb7c PZ |
3804 | /* |
3805 | * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i), | |
3806 | * we need to scale se->vlag when w_i changes. | |
3807 | */ | |
eab03c23 | 3808 | se->vlag = div_s64(se->vlag * se->load.weight, weight); |
86bfbb7c PZ |
3809 | } |
3810 | ||
eab03c23 AW |
3811 | update_load_set(&se->load, weight); |
3812 | ||
9059393e | 3813 | #ifdef CONFIG_SMP |
1ea6c46a | 3814 | do { |
87e867b4 | 3815 | u32 divider = get_pelt_divider(&se->avg); |
1ea6c46a PZ |
3816 | |
3817 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
1ea6c46a | 3818 | } while (0); |
9059393e VG |
3819 | #endif |
3820 | ||
3821 | enqueue_load_avg(cfs_rq, se); | |
af4cf404 | 3822 | if (se->on_rq) { |
1724b95b | 3823 | update_load_add(&cfs_rq->load, se->load.weight); |
5068d840 | 3824 | if (!curr) |
eab03c23 | 3825 | __enqueue_entity(cfs_rq, se); |
5068d840 YL |
3826 | |
3827 | /* | |
3828 | * The entity's vruntime has been adjusted, so let's check | |
3829 | * whether the rq-wide min_vruntime needs updated too. Since | |
3830 | * the calculations above require stable min_vruntime rather | |
3831 | * than up-to-date one, we do the update at the end of the | |
3832 | * reweight process. | |
3833 | */ | |
3834 | update_min_vruntime(cfs_rq); | |
af4cf404 | 3835 | } |
9059393e VG |
3836 | } |
3837 | ||
3838 | void reweight_task(struct task_struct *p, int prio) | |
3839 | { | |
3840 | struct sched_entity *se = &p->se; | |
3841 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3842 | struct load_weight *load = &se->load; | |
3843 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
3844 | ||
0dacee1b | 3845 | reweight_entity(cfs_rq, se, weight); |
9059393e VG |
3846 | load->inv_weight = sched_prio_to_wmult[prio]; |
3847 | } | |
3848 | ||
51bf903b CZ |
3849 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
3850 | ||
3ff6dcac | 3851 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 3852 | #ifdef CONFIG_SMP |
cef27403 PZ |
3853 | /* |
3854 | * All this does is approximate the hierarchical proportion which includes that | |
3855 | * global sum we all love to hate. | |
3856 | * | |
3857 | * That is, the weight of a group entity, is the proportional share of the | |
3858 | * group weight based on the group runqueue weights. That is: | |
3859 | * | |
3860 | * tg->weight * grq->load.weight | |
3861 | * ge->load.weight = ----------------------------- (1) | |
08f7c2f4 | 3862 | * \Sum grq->load.weight |
cef27403 PZ |
3863 | * |
3864 | * Now, because computing that sum is prohibitively expensive to compute (been | |
3865 | * there, done that) we approximate it with this average stuff. The average | |
3866 | * moves slower and therefore the approximation is cheaper and more stable. | |
3867 | * | |
3868 | * So instead of the above, we substitute: | |
3869 | * | |
3870 | * grq->load.weight -> grq->avg.load_avg (2) | |
3871 | * | |
3872 | * which yields the following: | |
3873 | * | |
3874 | * tg->weight * grq->avg.load_avg | |
3875 | * ge->load.weight = ------------------------------ (3) | |
08f7c2f4 | 3876 | * tg->load_avg |
cef27403 PZ |
3877 | * |
3878 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
3879 | * | |
3880 | * That is shares_avg, and it is right (given the approximation (2)). | |
3881 | * | |
3882 | * The problem with it is that because the average is slow -- it was designed | |
3883 | * to be exactly that of course -- this leads to transients in boundary | |
3884 | * conditions. In specific, the case where the group was idle and we start the | |
3885 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
3886 | * yielding bad latency etc.. | |
3887 | * | |
3888 | * Now, in that special case (1) reduces to: | |
3889 | * | |
3890 | * tg->weight * grq->load.weight | |
17de4ee0 | 3891 | * ge->load.weight = ----------------------------- = tg->weight (4) |
08f7c2f4 | 3892 | * grp->load.weight |
cef27403 PZ |
3893 | * |
3894 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
3895 | * | |
3896 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
3897 | * UP case, like: | |
3898 | * | |
3899 | * ge->load.weight = | |
3900 | * | |
3901 | * tg->weight * grq->load.weight | |
3902 | * --------------------------------------------------- (5) | |
3903 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
3904 | * | |
17de4ee0 PZ |
3905 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
3906 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
3907 | * | |
3908 | * | |
3909 | * tg->weight * grq->load.weight | |
3910 | * ge->load.weight = ----------------------------- (6) | |
08f7c2f4 | 3911 | * tg_load_avg' |
17de4ee0 PZ |
3912 | * |
3913 | * Where: | |
3914 | * | |
3915 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
3916 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
3917 | * |
3918 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
3919 | * (4) while in the normal case it approaches (3). It consistently | |
3920 | * overestimates the ge->load.weight and therefore: | |
3921 | * | |
3922 | * \Sum ge->load.weight >= tg->weight | |
3923 | * | |
3924 | * hence icky! | |
3925 | */ | |
2c8e4dce | 3926 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 3927 | { |
7c80cfc9 PZ |
3928 | long tg_weight, tg_shares, load, shares; |
3929 | struct task_group *tg = cfs_rq->tg; | |
3930 | ||
3931 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 3932 | |
3d4b60d3 | 3933 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 3934 | |
ea1dc6fc | 3935 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 3936 | |
ea1dc6fc PZ |
3937 | /* Ensure tg_weight >= load */ |
3938 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
3939 | tg_weight += load; | |
3ff6dcac | 3940 | |
7c80cfc9 | 3941 | shares = (tg_shares * load); |
cf5f0acf PZ |
3942 | if (tg_weight) |
3943 | shares /= tg_weight; | |
3ff6dcac | 3944 | |
b8fd8423 DE |
3945 | /* |
3946 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
3947 | * of a group with small tg->shares value. It is a floor value which is | |
3948 | * assigned as a minimum load.weight to the sched_entity representing | |
3949 | * the group on a CPU. | |
3950 | * | |
3951 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
3952 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
3953 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
3954 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
3955 | * instead of 0. | |
3956 | */ | |
7c80cfc9 | 3957 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 3958 | } |
387f77cc | 3959 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 3960 | |
1ea6c46a PZ |
3961 | /* |
3962 | * Recomputes the group entity based on the current state of its group | |
3963 | * runqueue. | |
3964 | */ | |
3965 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 3966 | { |
1ea6c46a | 3967 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
0dacee1b | 3968 | long shares; |
2069dd75 | 3969 | |
1ea6c46a | 3970 | if (!gcfs_rq) |
89ee048f VG |
3971 | return; |
3972 | ||
1ea6c46a | 3973 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3974 | return; |
89ee048f | 3975 | |
3ff6dcac | 3976 | #ifndef CONFIG_SMP |
0dacee1b | 3977 | shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 | 3978 | #else |
eab03c23 | 3979 | shares = calc_group_shares(gcfs_rq); |
3ff6dcac | 3980 | #endif |
eab03c23 AW |
3981 | if (unlikely(se->load.weight != shares)) |
3982 | reweight_entity(cfs_rq_of(se), se, shares); | |
2069dd75 | 3983 | } |
89ee048f | 3984 | |
2069dd75 | 3985 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3986 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3987 | { |
3988 | } | |
3989 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3990 | ||
ea14b57e | 3991 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3992 | { |
43964409 LT |
3993 | struct rq *rq = rq_of(cfs_rq); |
3994 | ||
a4f9a0e5 | 3995 | if (&rq->cfs == cfs_rq) { |
a030d738 VK |
3996 | /* |
3997 | * There are a few boundary cases this might miss but it should | |
3998 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3999 | * a real problem. |
a030d738 VK |
4000 | * |
4001 | * It will not get called when we go idle, because the idle | |
4002 | * thread is a different class (!fair), nor will the utilization | |
4003 | * number include things like RT tasks. | |
4004 | * | |
4005 | * As is, the util number is not freq-invariant (we'd have to | |
4006 | * implement arch_scale_freq_capacity() for that). | |
4007 | * | |
82762d2a | 4008 | * See cpu_util_cfs(). |
a030d738 | 4009 | */ |
ea14b57e | 4010 | cpufreq_update_util(rq, flags); |
a030d738 VK |
4011 | } |
4012 | } | |
4013 | ||
141965c7 | 4014 | #ifdef CONFIG_SMP |
e2f3e35f VD |
4015 | static inline bool load_avg_is_decayed(struct sched_avg *sa) |
4016 | { | |
4017 | if (sa->load_sum) | |
4018 | return false; | |
4019 | ||
4020 | if (sa->util_sum) | |
4021 | return false; | |
4022 | ||
4023 | if (sa->runnable_sum) | |
4024 | return false; | |
4025 | ||
4026 | /* | |
4027 | * _avg must be null when _sum are null because _avg = _sum / divider | |
4028 | * Make sure that rounding and/or propagation of PELT values never | |
4029 | * break this. | |
4030 | */ | |
4031 | SCHED_WARN_ON(sa->load_avg || | |
4032 | sa->util_avg || | |
4033 | sa->runnable_avg); | |
4034 | ||
4035 | return true; | |
4036 | } | |
4037 | ||
d05b4305 VD |
4038 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
4039 | { | |
4040 | return u64_u32_load_copy(cfs_rq->avg.last_update_time, | |
4041 | cfs_rq->last_update_time_copy); | |
4042 | } | |
c566e8e9 | 4043 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fdaba61e RR |
4044 | /* |
4045 | * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list | |
4046 | * immediately before a parent cfs_rq, and cfs_rqs are removed from the list | |
4047 | * bottom-up, we only have to test whether the cfs_rq before us on the list | |
4048 | * is our child. | |
4049 | * If cfs_rq is not on the list, test whether a child needs its to be added to | |
4050 | * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details). | |
4051 | */ | |
4052 | static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq) | |
4053 | { | |
4054 | struct cfs_rq *prev_cfs_rq; | |
4055 | struct list_head *prev; | |
4056 | ||
4057 | if (cfs_rq->on_list) { | |
4058 | prev = cfs_rq->leaf_cfs_rq_list.prev; | |
4059 | } else { | |
4060 | struct rq *rq = rq_of(cfs_rq); | |
4061 | ||
4062 | prev = rq->tmp_alone_branch; | |
4063 | } | |
4064 | ||
4065 | prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list); | |
4066 | ||
4067 | return (prev_cfs_rq->tg->parent == cfs_rq->tg); | |
4068 | } | |
a7b359fc OU |
4069 | |
4070 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) | |
4071 | { | |
4072 | if (cfs_rq->load.weight) | |
4073 | return false; | |
4074 | ||
e2f3e35f | 4075 | if (!load_avg_is_decayed(&cfs_rq->avg)) |
a7b359fc OU |
4076 | return false; |
4077 | ||
fdaba61e RR |
4078 | if (child_cfs_rq_on_list(cfs_rq)) |
4079 | return false; | |
4080 | ||
a7b359fc OU |
4081 | return true; |
4082 | } | |
4083 | ||
7c3edd2c PZ |
4084 | /** |
4085 | * update_tg_load_avg - update the tg's load avg | |
4086 | * @cfs_rq: the cfs_rq whose avg changed | |
7c3edd2c PZ |
4087 | * |
4088 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
4089 | * However, because tg->load_avg is a global value there are performance | |
4090 | * considerations. | |
4091 | * | |
4092 | * In order to avoid having to look at the other cfs_rq's, we use a | |
4093 | * differential update where we store the last value we propagated. This in | |
4094 | * turn allows skipping updates if the differential is 'small'. | |
4095 | * | |
815abf5a | 4096 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 4097 | */ |
fe749158 | 4098 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) |
bb17f655 | 4099 | { |
1528c661 AL |
4100 | long delta; |
4101 | u64 now; | |
bb17f655 | 4102 | |
aa0b7ae0 WL |
4103 | /* |
4104 | * No need to update load_avg for root_task_group as it is not used. | |
4105 | */ | |
4106 | if (cfs_rq->tg == &root_task_group) | |
4107 | return; | |
4108 | ||
f60a631a VG |
4109 | /* rq has been offline and doesn't contribute to the share anymore: */ |
4110 | if (!cpu_active(cpu_of(rq_of(cfs_rq)))) | |
4111 | return; | |
4112 | ||
1528c661 AL |
4113 | /* |
4114 | * For migration heavy workloads, access to tg->load_avg can be | |
4115 | * unbound. Limit the update rate to at most once per ms. | |
4116 | */ | |
4117 | now = sched_clock_cpu(cpu_of(rq_of(cfs_rq))); | |
4118 | if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC) | |
4119 | return; | |
4120 | ||
4121 | delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; | |
fe749158 | 4122 | if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
9d89c257 YD |
4123 | atomic_long_add(delta, &cfs_rq->tg->load_avg); |
4124 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
1528c661 | 4125 | cfs_rq->last_update_tg_load_avg = now; |
bb17f655 | 4126 | } |
8165e145 | 4127 | } |
f5f9739d | 4128 | |
f60a631a VG |
4129 | static inline void clear_tg_load_avg(struct cfs_rq *cfs_rq) |
4130 | { | |
4131 | long delta; | |
4132 | u64 now; | |
4133 | ||
4134 | /* | |
4135 | * No need to update load_avg for root_task_group, as it is not used. | |
4136 | */ | |
4137 | if (cfs_rq->tg == &root_task_group) | |
4138 | return; | |
4139 | ||
4140 | now = sched_clock_cpu(cpu_of(rq_of(cfs_rq))); | |
4141 | delta = 0 - cfs_rq->tg_load_avg_contrib; | |
4142 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
4143 | cfs_rq->tg_load_avg_contrib = 0; | |
4144 | cfs_rq->last_update_tg_load_avg = now; | |
4145 | } | |
4146 | ||
4147 | /* CPU offline callback: */ | |
4148 | static void __maybe_unused clear_tg_offline_cfs_rqs(struct rq *rq) | |
4149 | { | |
4150 | struct task_group *tg; | |
4151 | ||
4152 | lockdep_assert_rq_held(rq); | |
4153 | ||
4154 | /* | |
4155 | * The rq clock has already been updated in | |
4156 | * set_rq_offline(), so we should skip updating | |
4157 | * the rq clock again in unthrottle_cfs_rq(). | |
4158 | */ | |
4159 | rq_clock_start_loop_update(rq); | |
4160 | ||
4161 | rcu_read_lock(); | |
4162 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
4163 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4164 | ||
4165 | clear_tg_load_avg(cfs_rq); | |
4166 | } | |
4167 | rcu_read_unlock(); | |
4168 | ||
4169 | rq_clock_stop_loop_update(rq); | |
4170 | } | |
4171 | ||
ad936d86 | 4172 | /* |
97fb7a0a | 4173 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
4174 | * caller only guarantees p->pi_lock is held; no other assumptions, |
4175 | * including the state of rq->lock, should be made. | |
4176 | */ | |
4177 | void set_task_rq_fair(struct sched_entity *se, | |
4178 | struct cfs_rq *prev, struct cfs_rq *next) | |
4179 | { | |
0ccb977f PZ |
4180 | u64 p_last_update_time; |
4181 | u64 n_last_update_time; | |
4182 | ||
ad936d86 BP |
4183 | if (!sched_feat(ATTACH_AGE_LOAD)) |
4184 | return; | |
4185 | ||
4186 | /* | |
4187 | * We are supposed to update the task to "current" time, then its up to | |
4188 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
4189 | * getting what current time is, so simply throw away the out-of-date | |
4190 | * time. This will result in the wakee task is less decayed, but giving | |
4191 | * the wakee more load sounds not bad. | |
4192 | */ | |
0ccb977f PZ |
4193 | if (!(se->avg.last_update_time && prev)) |
4194 | return; | |
ad936d86 | 4195 | |
d05b4305 VD |
4196 | p_last_update_time = cfs_rq_last_update_time(prev); |
4197 | n_last_update_time = cfs_rq_last_update_time(next); | |
ad936d86 | 4198 | |
23127296 | 4199 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 4200 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 4201 | } |
09a43ace | 4202 | |
0e2d2aaa PZ |
4203 | /* |
4204 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
4205 | * propagate its contribution. The key to this propagation is the invariant | |
4206 | * that for each group: | |
4207 | * | |
4208 | * ge->avg == grq->avg (1) | |
4209 | * | |
4210 | * _IFF_ we look at the pure running and runnable sums. Because they | |
4211 | * represent the very same entity, just at different points in the hierarchy. | |
4212 | * | |
9f683953 VG |
4213 | * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial |
4214 | * and simply copies the running/runnable sum over (but still wrong, because | |
4215 | * the group entity and group rq do not have their PELT windows aligned). | |
0e2d2aaa | 4216 | * |
0dacee1b | 4217 | * However, update_tg_cfs_load() is more complex. So we have: |
0e2d2aaa PZ |
4218 | * |
4219 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
4220 | * | |
4221 | * And since, like util, the runnable part should be directly transferable, | |
4222 | * the following would _appear_ to be the straight forward approach: | |
4223 | * | |
a4c3c049 | 4224 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
4225 | * |
4226 | * And per (1) we have: | |
4227 | * | |
a4c3c049 | 4228 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
4229 | * |
4230 | * Which gives: | |
4231 | * | |
4232 | * ge->load.weight * grq->avg.load_avg | |
4233 | * ge->avg.load_avg = ----------------------------------- (4) | |
4234 | * grq->load.weight | |
4235 | * | |
4236 | * Except that is wrong! | |
4237 | * | |
4238 | * Because while for entities historical weight is not important and we | |
4239 | * really only care about our future and therefore can consider a pure | |
4240 | * runnable sum, runqueues can NOT do this. | |
4241 | * | |
4242 | * We specifically want runqueues to have a load_avg that includes | |
4243 | * historical weights. Those represent the blocked load, the load we expect | |
4244 | * to (shortly) return to us. This only works by keeping the weights as | |
4245 | * integral part of the sum. We therefore cannot decompose as per (3). | |
4246 | * | |
a4c3c049 VG |
4247 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
4248 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
4249 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
4250 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
4251 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
4252 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 4253 | * |
a4c3c049 | 4254 | * So we'll have to approximate.. :/ |
0e2d2aaa | 4255 | * |
a4c3c049 | 4256 | * Given the constraint: |
0e2d2aaa | 4257 | * |
a4c3c049 | 4258 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 4259 | * |
a4c3c049 VG |
4260 | * We can construct a rule that adds runnable to a rq by assuming minimal |
4261 | * overlap. | |
0e2d2aaa | 4262 | * |
a4c3c049 | 4263 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 4264 | * |
a4c3c049 | 4265 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 4266 | * |
a4c3c049 | 4267 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 4268 | * |
0e2d2aaa | 4269 | */ |
09a43ace | 4270 | static inline void |
0e2d2aaa | 4271 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 4272 | { |
7ceb7710 VG |
4273 | long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg; |
4274 | u32 new_sum, divider; | |
09a43ace VG |
4275 | |
4276 | /* Nothing to update */ | |
7ceb7710 | 4277 | if (!delta_avg) |
09a43ace VG |
4278 | return; |
4279 | ||
87e867b4 VG |
4280 | /* |
4281 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
4282 | * See ___update_load_avg() for details. | |
4283 | */ | |
4284 | divider = get_pelt_divider(&cfs_rq->avg); | |
4285 | ||
7ceb7710 | 4286 | |
09a43ace VG |
4287 | /* Set new sched_entity's utilization */ |
4288 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
7ceb7710 VG |
4289 | new_sum = se->avg.util_avg * divider; |
4290 | delta_sum = (long)new_sum - (long)se->avg.util_sum; | |
4291 | se->avg.util_sum = new_sum; | |
09a43ace VG |
4292 | |
4293 | /* Update parent cfs_rq utilization */ | |
7ceb7710 VG |
4294 | add_positive(&cfs_rq->avg.util_avg, delta_avg); |
4295 | add_positive(&cfs_rq->avg.util_sum, delta_sum); | |
4296 | ||
4297 | /* See update_cfs_rq_load_avg() */ | |
4298 | cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, | |
4299 | cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); | |
09a43ace VG |
4300 | } |
4301 | ||
9f683953 VG |
4302 | static inline void |
4303 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) | |
4304 | { | |
95246d1e VG |
4305 | long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; |
4306 | u32 new_sum, divider; | |
9f683953 VG |
4307 | |
4308 | /* Nothing to update */ | |
95246d1e | 4309 | if (!delta_avg) |
9f683953 VG |
4310 | return; |
4311 | ||
87e867b4 VG |
4312 | /* |
4313 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
4314 | * See ___update_load_avg() for details. | |
4315 | */ | |
4316 | divider = get_pelt_divider(&cfs_rq->avg); | |
4317 | ||
9f683953 VG |
4318 | /* Set new sched_entity's runnable */ |
4319 | se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; | |
95246d1e VG |
4320 | new_sum = se->avg.runnable_avg * divider; |
4321 | delta_sum = (long)new_sum - (long)se->avg.runnable_sum; | |
4322 | se->avg.runnable_sum = new_sum; | |
9f683953 VG |
4323 | |
4324 | /* Update parent cfs_rq runnable */ | |
95246d1e VG |
4325 | add_positive(&cfs_rq->avg.runnable_avg, delta_avg); |
4326 | add_positive(&cfs_rq->avg.runnable_sum, delta_sum); | |
4327 | /* See update_cfs_rq_load_avg() */ | |
4328 | cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, | |
4329 | cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); | |
9f683953 VG |
4330 | } |
4331 | ||
09a43ace | 4332 | static inline void |
0dacee1b | 4333 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 4334 | { |
2d02fa8c | 4335 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
0dacee1b VG |
4336 | unsigned long load_avg; |
4337 | u64 load_sum = 0; | |
2d02fa8c | 4338 | s64 delta_sum; |
95d68593 | 4339 | u32 divider; |
09a43ace | 4340 | |
0e2d2aaa PZ |
4341 | if (!runnable_sum) |
4342 | return; | |
09a43ace | 4343 | |
0e2d2aaa | 4344 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 4345 | |
95d68593 VG |
4346 | /* |
4347 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
4348 | * See ___update_load_avg() for details. | |
4349 | */ | |
87e867b4 | 4350 | divider = get_pelt_divider(&cfs_rq->avg); |
95d68593 | 4351 | |
a4c3c049 VG |
4352 | if (runnable_sum >= 0) { |
4353 | /* | |
4354 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
4355 | * the CPU is saturated running == runnable. | |
4356 | */ | |
4357 | runnable_sum += se->avg.load_sum; | |
95d68593 | 4358 | runnable_sum = min_t(long, runnable_sum, divider); |
a4c3c049 VG |
4359 | } else { |
4360 | /* | |
4361 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
4362 | * assuming all tasks are equally runnable. | |
4363 | */ | |
4364 | if (scale_load_down(gcfs_rq->load.weight)) { | |
2d02fa8c | 4365 | load_sum = div_u64(gcfs_rq->avg.load_sum, |
a4c3c049 VG |
4366 | scale_load_down(gcfs_rq->load.weight)); |
4367 | } | |
4368 | ||
4369 | /* But make sure to not inflate se's runnable */ | |
4370 | runnable_sum = min(se->avg.load_sum, load_sum); | |
4371 | } | |
4372 | ||
4373 | /* | |
4374 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
4375 | * Rescale running sum to be in the same range as runnable sum |
4376 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
4377 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 4378 | */ |
23127296 | 4379 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
4380 | runnable_sum = max(runnable_sum, running_sum); |
4381 | ||
2d02fa8c VG |
4382 | load_sum = se_weight(se) * runnable_sum; |
4383 | load_avg = div_u64(load_sum, divider); | |
83c5e9d5 | 4384 | |
2d02fa8c VG |
4385 | delta_avg = load_avg - se->avg.load_avg; |
4386 | if (!delta_avg) | |
83c5e9d5 | 4387 | return; |
09a43ace | 4388 | |
2d02fa8c | 4389 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
7c7ad626 | 4390 | |
2d02fa8c VG |
4391 | se->avg.load_sum = runnable_sum; |
4392 | se->avg.load_avg = load_avg; | |
4393 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
4394 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
4395 | /* See update_cfs_rq_load_avg() */ | |
4396 | cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, | |
4397 | cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); | |
09a43ace VG |
4398 | } |
4399 | ||
0e2d2aaa | 4400 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 4401 | { |
0e2d2aaa PZ |
4402 | cfs_rq->propagate = 1; |
4403 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
4404 | } |
4405 | ||
4406 | /* Update task and its cfs_rq load average */ | |
4407 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
4408 | { | |
0e2d2aaa | 4409 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
4410 | |
4411 | if (entity_is_task(se)) | |
4412 | return 0; | |
4413 | ||
0e2d2aaa PZ |
4414 | gcfs_rq = group_cfs_rq(se); |
4415 | if (!gcfs_rq->propagate) | |
09a43ace VG |
4416 | return 0; |
4417 | ||
0e2d2aaa PZ |
4418 | gcfs_rq->propagate = 0; |
4419 | ||
09a43ace VG |
4420 | cfs_rq = cfs_rq_of(se); |
4421 | ||
0e2d2aaa | 4422 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 4423 | |
0e2d2aaa | 4424 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
9f683953 | 4425 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); |
0dacee1b | 4426 | update_tg_cfs_load(cfs_rq, se, gcfs_rq); |
09a43ace | 4427 | |
ba19f51f | 4428 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 4429 | trace_pelt_se_tp(se); |
ba19f51f | 4430 | |
09a43ace VG |
4431 | return 1; |
4432 | } | |
4433 | ||
bc427898 VG |
4434 | /* |
4435 | * Check if we need to update the load and the utilization of a blocked | |
4436 | * group_entity: | |
4437 | */ | |
4438 | static inline bool skip_blocked_update(struct sched_entity *se) | |
4439 | { | |
4440 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
4441 | ||
4442 | /* | |
4443 | * If sched_entity still have not zero load or utilization, we have to | |
4444 | * decay it: | |
4445 | */ | |
4446 | if (se->avg.load_avg || se->avg.util_avg) | |
4447 | return false; | |
4448 | ||
4449 | /* | |
4450 | * If there is a pending propagation, we have to update the load and | |
4451 | * the utilization of the sched_entity: | |
4452 | */ | |
0e2d2aaa | 4453 | if (gcfs_rq->propagate) |
bc427898 VG |
4454 | return false; |
4455 | ||
4456 | /* | |
4457 | * Otherwise, the load and the utilization of the sched_entity is | |
4458 | * already zero and there is no pending propagation, so it will be a | |
4459 | * waste of time to try to decay it: | |
4460 | */ | |
4461 | return true; | |
4462 | } | |
4463 | ||
6e83125c | 4464 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 4465 | |
fe749158 | 4466 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} |
09a43ace | 4467 | |
f60a631a VG |
4468 | static inline void clear_tg_offline_cfs_rqs(struct rq *rq) {} |
4469 | ||
09a43ace VG |
4470 | static inline int propagate_entity_load_avg(struct sched_entity *se) |
4471 | { | |
4472 | return 0; | |
4473 | } | |
4474 | ||
0e2d2aaa | 4475 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 4476 | |
6e83125c | 4477 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 4478 | |
e2f3e35f VD |
4479 | #ifdef CONFIG_NO_HZ_COMMON |
4480 | static inline void migrate_se_pelt_lag(struct sched_entity *se) | |
4481 | { | |
4482 | u64 throttled = 0, now, lut; | |
4483 | struct cfs_rq *cfs_rq; | |
4484 | struct rq *rq; | |
4485 | bool is_idle; | |
4486 | ||
4487 | if (load_avg_is_decayed(&se->avg)) | |
4488 | return; | |
4489 | ||
4490 | cfs_rq = cfs_rq_of(se); | |
4491 | rq = rq_of(cfs_rq); | |
4492 | ||
4493 | rcu_read_lock(); | |
4494 | is_idle = is_idle_task(rcu_dereference(rq->curr)); | |
4495 | rcu_read_unlock(); | |
4496 | ||
4497 | /* | |
4498 | * The lag estimation comes with a cost we don't want to pay all the | |
4499 | * time. Hence, limiting to the case where the source CPU is idle and | |
4500 | * we know we are at the greatest risk to have an outdated clock. | |
4501 | */ | |
4502 | if (!is_idle) | |
4503 | return; | |
4504 | ||
4505 | /* | |
4506 | * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where: | |
4507 | * | |
4508 | * last_update_time (the cfs_rq's last_update_time) | |
4509 | * = cfs_rq_clock_pelt()@cfs_rq_idle | |
4510 | * = rq_clock_pelt()@cfs_rq_idle | |
4511 | * - cfs->throttled_clock_pelt_time@cfs_rq_idle | |
4512 | * | |
4513 | * cfs_idle_lag (delta between rq's update and cfs_rq's update) | |
4514 | * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle | |
4515 | * | |
4516 | * rq_idle_lag (delta between now and rq's update) | |
4517 | * = sched_clock_cpu() - rq_clock()@rq_idle | |
4518 | * | |
4519 | * We can then write: | |
4520 | * | |
4521 | * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time + | |
4522 | * sched_clock_cpu() - rq_clock()@rq_idle | |
4523 | * Where: | |
4524 | * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle | |
4525 | * rq_clock()@rq_idle is rq->clock_idle | |
4526 | * cfs->throttled_clock_pelt_time@cfs_rq_idle | |
4527 | * is cfs_rq->throttled_pelt_idle | |
4528 | */ | |
4529 | ||
4530 | #ifdef CONFIG_CFS_BANDWIDTH | |
4531 | throttled = u64_u32_load(cfs_rq->throttled_pelt_idle); | |
4532 | /* The clock has been stopped for throttling */ | |
4533 | if (throttled == U64_MAX) | |
4534 | return; | |
4535 | #endif | |
4536 | now = u64_u32_load(rq->clock_pelt_idle); | |
4537 | /* | |
4538 | * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case | |
4539 | * is observed the old clock_pelt_idle value and the new clock_idle, | |
4540 | * which lead to an underestimation. The opposite would lead to an | |
4541 | * overestimation. | |
4542 | */ | |
4543 | smp_rmb(); | |
4544 | lut = cfs_rq_last_update_time(cfs_rq); | |
4545 | ||
4546 | now -= throttled; | |
4547 | if (now < lut) | |
4548 | /* | |
4549 | * cfs_rq->avg.last_update_time is more recent than our | |
4550 | * estimation, let's use it. | |
4551 | */ | |
4552 | now = lut; | |
4553 | else | |
4554 | now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle); | |
4555 | ||
4556 | __update_load_avg_blocked_se(now, se); | |
4557 | } | |
4558 | #else | |
4559 | static void migrate_se_pelt_lag(struct sched_entity *se) {} | |
4560 | #endif | |
4561 | ||
3d30544f PZ |
4562 | /** |
4563 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 4564 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 4565 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
4566 | * |
4567 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
d6531ab6 | 4568 | * avg. The immediate corollary is that all (fair) tasks must be attached. |
3d30544f PZ |
4569 | * |
4570 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
4571 | * | |
a315da5e | 4572 | * Return: true if the load decayed or we removed load. |
7c3edd2c PZ |
4573 | * |
4574 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
4575 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 4576 | */ |
a2c6c91f | 4577 | static inline int |
3a123bbb | 4578 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 4579 | { |
9f683953 | 4580 | unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; |
9d89c257 | 4581 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 4582 | int decayed = 0; |
2dac754e | 4583 | |
2a2f5d4e PZ |
4584 | if (cfs_rq->removed.nr) { |
4585 | unsigned long r; | |
87e867b4 | 4586 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
2a2f5d4e PZ |
4587 | |
4588 | raw_spin_lock(&cfs_rq->removed.lock); | |
4589 | swap(cfs_rq->removed.util_avg, removed_util); | |
4590 | swap(cfs_rq->removed.load_avg, removed_load); | |
9f683953 | 4591 | swap(cfs_rq->removed.runnable_avg, removed_runnable); |
2a2f5d4e PZ |
4592 | cfs_rq->removed.nr = 0; |
4593 | raw_spin_unlock(&cfs_rq->removed.lock); | |
4594 | ||
2a2f5d4e | 4595 | r = removed_load; |
89741892 | 4596 | sub_positive(&sa->load_avg, r); |
2d02fa8c VG |
4597 | sub_positive(&sa->load_sum, r * divider); |
4598 | /* See sa->util_sum below */ | |
4599 | sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER); | |
2dac754e | 4600 | |
2a2f5d4e | 4601 | r = removed_util; |
89741892 | 4602 | sub_positive(&sa->util_avg, r); |
98b0d890 VG |
4603 | sub_positive(&sa->util_sum, r * divider); |
4604 | /* | |
4605 | * Because of rounding, se->util_sum might ends up being +1 more than | |
4606 | * cfs->util_sum. Although this is not a problem by itself, detaching | |
4607 | * a lot of tasks with the rounding problem between 2 updates of | |
4608 | * util_avg (~1ms) can make cfs->util_sum becoming null whereas | |
4609 | * cfs_util_avg is not. | |
4610 | * Check that util_sum is still above its lower bound for the new | |
4611 | * util_avg. Given that period_contrib might have moved since the last | |
4612 | * sync, we are only sure that util_sum must be above or equal to | |
4613 | * util_avg * minimum possible divider | |
4614 | */ | |
4615 | sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER); | |
2a2f5d4e | 4616 | |
9f683953 VG |
4617 | r = removed_runnable; |
4618 | sub_positive(&sa->runnable_avg, r); | |
95246d1e VG |
4619 | sub_positive(&sa->runnable_sum, r * divider); |
4620 | /* See sa->util_sum above */ | |
4621 | sa->runnable_sum = max_t(u32, sa->runnable_sum, | |
4622 | sa->runnable_avg * PELT_MIN_DIVIDER); | |
9f683953 VG |
4623 | |
4624 | /* | |
4625 | * removed_runnable is the unweighted version of removed_load so we | |
4626 | * can use it to estimate removed_load_sum. | |
4627 | */ | |
4628 | add_tg_cfs_propagate(cfs_rq, | |
4629 | -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); | |
2a2f5d4e PZ |
4630 | |
4631 | decayed = 1; | |
9d89c257 | 4632 | } |
36ee28e4 | 4633 | |
23127296 | 4634 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
d05b4305 VD |
4635 | u64_u32_store_copy(sa->last_update_time, |
4636 | cfs_rq->last_update_time_copy, | |
4637 | sa->last_update_time); | |
2a2f5d4e | 4638 | return decayed; |
21e96f88 SM |
4639 | } |
4640 | ||
3d30544f PZ |
4641 | /** |
4642 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
4643 | * @cfs_rq: cfs_rq to attach to | |
4644 | * @se: sched_entity to attach | |
4645 | * | |
4646 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
4647 | * cfs_rq->avg.last_update_time being current. | |
4648 | */ | |
a4f9a0e5 | 4649 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
a05e8c51 | 4650 | { |
95d68593 VG |
4651 | /* |
4652 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
4653 | * See ___update_load_avg() for details. | |
4654 | */ | |
87e867b4 | 4655 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
f207934f PZ |
4656 | |
4657 | /* | |
4658 | * When we attach the @se to the @cfs_rq, we must align the decay | |
4659 | * window because without that, really weird and wonderful things can | |
4660 | * happen. | |
4661 | * | |
4662 | * XXX illustrate | |
4663 | */ | |
a05e8c51 | 4664 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
4665 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
4666 | ||
4667 | /* | |
4668 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
4669 | * period_contrib. This isn't strictly correct, but since we're | |
4670 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
4671 | * _sum a little. | |
4672 | */ | |
4673 | se->avg.util_sum = se->avg.util_avg * divider; | |
4674 | ||
9f683953 VG |
4675 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
4676 | ||
40f5aa4c | 4677 | se->avg.load_sum = se->avg.load_avg * divider; |
4678 | if (se_weight(se) < se->avg.load_sum) | |
4679 | se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se)); | |
4680 | else | |
4681 | se->avg.load_sum = 1; | |
f207934f | 4682 | |
8d5b9025 | 4683 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
4684 | cfs_rq->avg.util_avg += se->avg.util_avg; |
4685 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
9f683953 VG |
4686 | cfs_rq->avg.runnable_avg += se->avg.runnable_avg; |
4687 | cfs_rq->avg.runnable_sum += se->avg.runnable_sum; | |
0e2d2aaa PZ |
4688 | |
4689 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 4690 | |
a4f9a0e5 | 4691 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
4692 | |
4693 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
4694 | } |
4695 | ||
3d30544f PZ |
4696 | /** |
4697 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
4698 | * @cfs_rq: cfs_rq to detach from | |
4699 | * @se: sched_entity to detach | |
4700 | * | |
4701 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
4702 | * cfs_rq->avg.last_update_time being current. | |
4703 | */ | |
a05e8c51 BP |
4704 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4705 | { | |
8d5b9025 | 4706 | dequeue_load_avg(cfs_rq, se); |
89741892 | 4707 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
7ceb7710 VG |
4708 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); |
4709 | /* See update_cfs_rq_load_avg() */ | |
4710 | cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, | |
4711 | cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); | |
4712 | ||
9f683953 | 4713 | sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); |
95246d1e VG |
4714 | sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); |
4715 | /* See update_cfs_rq_load_avg() */ | |
4716 | cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, | |
4717 | cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); | |
0e2d2aaa PZ |
4718 | |
4719 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 4720 | |
ea14b57e | 4721 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
4722 | |
4723 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
4724 | } |
4725 | ||
b382a531 PZ |
4726 | /* |
4727 | * Optional action to be done while updating the load average | |
4728 | */ | |
4729 | #define UPDATE_TG 0x1 | |
4730 | #define SKIP_AGE_LOAD 0x2 | |
4731 | #define DO_ATTACH 0x4 | |
e1f078f5 | 4732 | #define DO_DETACH 0x8 |
b382a531 PZ |
4733 | |
4734 | /* Update task and its cfs_rq load average */ | |
4735 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
4736 | { | |
23127296 | 4737 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
4738 | int decayed; |
4739 | ||
4740 | /* | |
4741 | * Track task load average for carrying it to new CPU after migrated, and | |
b9e6e286 | 4742 | * track group sched_entity load average for task_h_load calculation in migration |
b382a531 PZ |
4743 | */ |
4744 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 4745 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
4746 | |
4747 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
4748 | decayed |= propagate_entity_load_avg(se); | |
4749 | ||
4750 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
4751 | ||
ea14b57e PZ |
4752 | /* |
4753 | * DO_ATTACH means we're here from enqueue_entity(). | |
4754 | * !last_update_time means we've passed through | |
4755 | * migrate_task_rq_fair() indicating we migrated. | |
4756 | * | |
4757 | * IOW we're enqueueing a task on a new CPU. | |
4758 | */ | |
a4f9a0e5 | 4759 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 4760 | update_tg_load_avg(cfs_rq); |
b382a531 | 4761 | |
e1f078f5 CZ |
4762 | } else if (flags & DO_DETACH) { |
4763 | /* | |
4764 | * DO_DETACH means we're here from dequeue_entity() | |
4765 | * and we are migrating task out of the CPU. | |
4766 | */ | |
4767 | detach_entity_load_avg(cfs_rq, se); | |
4768 | update_tg_load_avg(cfs_rq); | |
bef69dd8 VG |
4769 | } else if (decayed) { |
4770 | cfs_rq_util_change(cfs_rq, 0); | |
4771 | ||
4772 | if (flags & UPDATE_TG) | |
fe749158 | 4773 | update_tg_load_avg(cfs_rq); |
bef69dd8 | 4774 | } |
b382a531 PZ |
4775 | } |
4776 | ||
104cb16d MR |
4777 | /* |
4778 | * Synchronize entity load avg of dequeued entity without locking | |
4779 | * the previous rq. | |
4780 | */ | |
71b47eaf | 4781 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
4782 | { |
4783 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4784 | u64 last_update_time; | |
4785 | ||
4786 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 4787 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
4788 | } |
4789 | ||
0905f04e YD |
4790 | /* |
4791 | * Task first catches up with cfs_rq, and then subtract | |
4792 | * itself from the cfs_rq (task must be off the queue now). | |
4793 | */ | |
71b47eaf | 4794 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
4795 | { |
4796 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 4797 | unsigned long flags; |
0905f04e YD |
4798 | |
4799 | /* | |
7dc603c9 | 4800 | * tasks cannot exit without having gone through wake_up_new_task() -> |
d6531ab6 CZ |
4801 | * enqueue_task_fair() which will have added things to the cfs_rq, |
4802 | * so we can remove unconditionally. | |
0905f04e | 4803 | */ |
0905f04e | 4804 | |
104cb16d | 4805 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
4806 | |
4807 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
4808 | ++cfs_rq->removed.nr; | |
4809 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
4810 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
9f683953 | 4811 | cfs_rq->removed.runnable_avg += se->avg.runnable_avg; |
2a2f5d4e | 4812 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 4813 | } |
642dbc39 | 4814 | |
9f683953 VG |
4815 | static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) |
4816 | { | |
4817 | return cfs_rq->avg.runnable_avg; | |
4818 | } | |
4819 | ||
7ea241af YD |
4820 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) |
4821 | { | |
4822 | return cfs_rq->avg.load_avg; | |
4823 | } | |
4824 | ||
7d058285 | 4825 | static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf); |
d91cecc1 | 4826 | |
7f65ea42 PB |
4827 | static inline unsigned long task_util(struct task_struct *p) |
4828 | { | |
4829 | return READ_ONCE(p->se.avg.util_avg); | |
4830 | } | |
4831 | ||
50181c0c VG |
4832 | static inline unsigned long task_runnable(struct task_struct *p) |
4833 | { | |
4834 | return READ_ONCE(p->se.avg.runnable_avg); | |
4835 | } | |
4836 | ||
7f65ea42 PB |
4837 | static inline unsigned long _task_util_est(struct task_struct *p) |
4838 | { | |
11137d38 | 4839 | return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED; |
7f65ea42 PB |
4840 | } |
4841 | ||
4842 | static inline unsigned long task_util_est(struct task_struct *p) | |
4843 | { | |
4844 | return max(task_util(p), _task_util_est(p)); | |
4845 | } | |
4846 | ||
4847 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, | |
4848 | struct task_struct *p) | |
4849 | { | |
4850 | unsigned int enqueued; | |
4851 | ||
4852 | if (!sched_feat(UTIL_EST)) | |
4853 | return; | |
4854 | ||
4855 | /* Update root cfs_rq's estimated utilization */ | |
11137d38 | 4856 | enqueued = cfs_rq->avg.util_est; |
92a801e5 | 4857 | enqueued += _task_util_est(p); |
11137d38 | 4858 | WRITE_ONCE(cfs_rq->avg.util_est, enqueued); |
4581bea8 VD |
4859 | |
4860 | trace_sched_util_est_cfs_tp(cfs_rq); | |
7f65ea42 PB |
4861 | } |
4862 | ||
8c1f560c XY |
4863 | static inline void util_est_dequeue(struct cfs_rq *cfs_rq, |
4864 | struct task_struct *p) | |
4865 | { | |
4866 | unsigned int enqueued; | |
4867 | ||
4868 | if (!sched_feat(UTIL_EST)) | |
4869 | return; | |
4870 | ||
4871 | /* Update root cfs_rq's estimated utilization */ | |
11137d38 | 4872 | enqueued = cfs_rq->avg.util_est; |
8c1f560c | 4873 | enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); |
11137d38 | 4874 | WRITE_ONCE(cfs_rq->avg.util_est, enqueued); |
8c1f560c XY |
4875 | |
4876 | trace_sched_util_est_cfs_tp(cfs_rq); | |
4877 | } | |
4878 | ||
b89997aa VD |
4879 | #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100) |
4880 | ||
8c1f560c XY |
4881 | static inline void util_est_update(struct cfs_rq *cfs_rq, |
4882 | struct task_struct *p, | |
4883 | bool task_sleep) | |
7f65ea42 | 4884 | { |
11137d38 | 4885 | unsigned int ewma, dequeued, last_ewma_diff; |
7f65ea42 PB |
4886 | |
4887 | if (!sched_feat(UTIL_EST)) | |
4888 | return; | |
4889 | ||
7f65ea42 PB |
4890 | /* |
4891 | * Skip update of task's estimated utilization when the task has not | |
4892 | * yet completed an activation, e.g. being migrated. | |
4893 | */ | |
4894 | if (!task_sleep) | |
4895 | return; | |
4896 | ||
11137d38 VG |
4897 | /* Get current estimate of utilization */ |
4898 | ewma = READ_ONCE(p->se.avg.util_est); | |
4899 | ||
d519329f PB |
4900 | /* |
4901 | * If the PELT values haven't changed since enqueue time, | |
4902 | * skip the util_est update. | |
4903 | */ | |
11137d38 | 4904 | if (ewma & UTIL_AVG_UNCHANGED) |
d519329f PB |
4905 | return; |
4906 | ||
11137d38 VG |
4907 | /* Get utilization at dequeue */ |
4908 | dequeued = task_util(p); | |
b89997aa | 4909 | |
b8c96361 PB |
4910 | /* |
4911 | * Reset EWMA on utilization increases, the moving average is used only | |
4912 | * to smooth utilization decreases. | |
4913 | */ | |
11137d38 VG |
4914 | if (ewma <= dequeued) { |
4915 | ewma = dequeued; | |
7736ae55 | 4916 | goto done; |
b8c96361 PB |
4917 | } |
4918 | ||
7f65ea42 | 4919 | /* |
b89997aa | 4920 | * Skip update of task's estimated utilization when its members are |
7f65ea42 PB |
4921 | * already ~1% close to its last activation value. |
4922 | */ | |
11137d38 VG |
4923 | last_ewma_diff = ewma - dequeued; |
4924 | if (last_ewma_diff < UTIL_EST_MARGIN) | |
4925 | goto done; | |
7f65ea42 | 4926 | |
10a35e68 VG |
4927 | /* |
4928 | * To avoid overestimation of actual task utilization, skip updates if | |
4929 | * we cannot grant there is idle time in this CPU. | |
4930 | */ | |
11137d38 | 4931 | if (dequeued > arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)))) |
10a35e68 VG |
4932 | return; |
4933 | ||
50181c0c VG |
4934 | /* |
4935 | * To avoid underestimate of task utilization, skip updates of EWMA if | |
4936 | * we cannot grant that thread got all CPU time it wanted. | |
4937 | */ | |
11137d38 | 4938 | if ((dequeued + UTIL_EST_MARGIN) < task_runnable(p)) |
50181c0c VG |
4939 | goto done; |
4940 | ||
4941 | ||
7f65ea42 PB |
4942 | /* |
4943 | * Update Task's estimated utilization | |
4944 | * | |
4945 | * When *p completes an activation we can consolidate another sample | |
11137d38 VG |
4946 | * of the task size. This is done by using this value to update the |
4947 | * Exponential Weighted Moving Average (EWMA): | |
7f65ea42 PB |
4948 | * |
4949 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
4950 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
4951 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
11137d38 VG |
4952 | * = w * ( -last_ewma_diff ) + ewma(t-1) |
4953 | * = w * (-last_ewma_diff + ewma(t-1) / w) | |
7f65ea42 PB |
4954 | * |
4955 | * Where 'w' is the weight of new samples, which is configured to be | |
4956 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
4957 | */ | |
11137d38 VG |
4958 | ewma <<= UTIL_EST_WEIGHT_SHIFT; |
4959 | ewma -= last_ewma_diff; | |
4960 | ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
b8c96361 | 4961 | done: |
11137d38 VG |
4962 | ewma |= UTIL_AVG_UNCHANGED; |
4963 | WRITE_ONCE(p->se.avg.util_est, ewma); | |
4581bea8 VD |
4964 | |
4965 | trace_sched_util_est_se_tp(&p->se); | |
7f65ea42 PB |
4966 | } |
4967 | ||
f1f8d0a2 VG |
4968 | static inline unsigned long get_actual_cpu_capacity(int cpu) |
4969 | { | |
4970 | unsigned long capacity = arch_scale_cpu_capacity(cpu); | |
4971 | ||
d4dbc991 | 4972 | capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu)); |
f1f8d0a2 VG |
4973 | |
4974 | return capacity; | |
4975 | } | |
4976 | ||
48d5e9da QY |
4977 | static inline int util_fits_cpu(unsigned long util, |
4978 | unsigned long uclamp_min, | |
4979 | unsigned long uclamp_max, | |
4980 | int cpu) | |
4981 | { | |
48d5e9da | 4982 | unsigned long capacity = capacity_of(cpu); |
f1f8d0a2 | 4983 | unsigned long capacity_orig; |
48d5e9da QY |
4984 | bool fits, uclamp_max_fits; |
4985 | ||
4986 | /* | |
4987 | * Check if the real util fits without any uclamp boost/cap applied. | |
4988 | */ | |
4989 | fits = fits_capacity(util, capacity); | |
4990 | ||
4991 | if (!uclamp_is_used()) | |
4992 | return fits; | |
4993 | ||
4994 | /* | |
7bc26384 | 4995 | * We must use arch_scale_cpu_capacity() for comparing against uclamp_min and |
48d5e9da QY |
4996 | * uclamp_max. We only care about capacity pressure (by using |
4997 | * capacity_of()) for comparing against the real util. | |
4998 | * | |
4999 | * If a task is boosted to 1024 for example, we don't want a tiny | |
5000 | * pressure to skew the check whether it fits a CPU or not. | |
5001 | * | |
7bc26384 | 5002 | * Similarly if a task is capped to arch_scale_cpu_capacity(little_cpu), it |
48d5e9da QY |
5003 | * should fit a little cpu even if there's some pressure. |
5004 | * | |
d4dbc991 | 5005 | * Only exception is for HW or cpufreq pressure since it has a direct impact |
48d5e9da QY |
5006 | * on available OPP of the system. |
5007 | * | |
5008 | * We honour it for uclamp_min only as a drop in performance level | |
5009 | * could result in not getting the requested minimum performance level. | |
5010 | * | |
5011 | * For uclamp_max, we can tolerate a drop in performance level as the | |
5012 | * goal is to cap the task. So it's okay if it's getting less. | |
48d5e9da | 5013 | */ |
7bc26384 | 5014 | capacity_orig = arch_scale_cpu_capacity(cpu); |
48d5e9da QY |
5015 | |
5016 | /* | |
5017 | * We want to force a task to fit a cpu as implied by uclamp_max. | |
5018 | * But we do have some corner cases to cater for.. | |
5019 | * | |
5020 | * | |
5021 | * C=z | |
5022 | * | ___ | |
5023 | * | C=y | | | |
5024 | * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max | |
5025 | * | C=x | | | | | |
5026 | * | ___ | | | | | |
5027 | * | | | | | | | (util somewhere in this region) | |
5028 | * | | | | | | | | |
5029 | * | | | | | | | | |
5030 | * +---------------------------------------- | |
b9e6e286 | 5031 | * CPU0 CPU1 CPU2 |
48d5e9da QY |
5032 | * |
5033 | * In the above example if a task is capped to a specific performance | |
5034 | * point, y, then when: | |
5035 | * | |
b9e6e286 IM |
5036 | * * util = 80% of x then it does not fit on CPU0 and should migrate |
5037 | * to CPU1 | |
5038 | * * util = 80% of y then it is forced to fit on CPU1 to honour | |
48d5e9da QY |
5039 | * uclamp_max request. |
5040 | * | |
5041 | * which is what we're enforcing here. A task always fits if | |
5042 | * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig, | |
5043 | * the normal upmigration rules should withhold still. | |
5044 | * | |
5045 | * Only exception is when we are on max capacity, then we need to be | |
5046 | * careful not to block overutilized state. This is so because: | |
5047 | * | |
5048 | * 1. There's no concept of capping at max_capacity! We can't go | |
5049 | * beyond this performance level anyway. | |
5050 | * 2. The system is being saturated when we're operating near | |
5051 | * max capacity, it doesn't make sense to block overutilized. | |
5052 | */ | |
5053 | uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE); | |
5054 | uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig); | |
5055 | fits = fits || uclamp_max_fits; | |
5056 | ||
5057 | /* | |
5058 | * | |
5059 | * C=z | |
5060 | * | ___ (region a, capped, util >= uclamp_max) | |
5061 | * | C=y | | | |
5062 | * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max | |
5063 | * | C=x | | | | | |
5064 | * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max) | |
5065 | * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min | |
5066 | * | | | | | | | | |
5067 | * | | | | | | | (region c, boosted, util < uclamp_min) | |
5068 | * +---------------------------------------- | |
b9e6e286 | 5069 | * CPU0 CPU1 CPU2 |
48d5e9da QY |
5070 | * |
5071 | * a) If util > uclamp_max, then we're capped, we don't care about | |
5072 | * actual fitness value here. We only care if uclamp_max fits | |
5073 | * capacity without taking margin/pressure into account. | |
5074 | * See comment above. | |
5075 | * | |
5076 | * b) If uclamp_min <= util <= uclamp_max, then the normal | |
5077 | * fits_capacity() rules apply. Except we need to ensure that we | |
5078 | * enforce we remain within uclamp_max, see comment above. | |
5079 | * | |
5080 | * c) If util < uclamp_min, then we are boosted. Same as (b) but we | |
5081 | * need to take into account the boosted value fits the CPU without | |
5082 | * taking margin/pressure into account. | |
5083 | * | |
5084 | * Cases (a) and (b) are handled in the 'fits' variable already. We | |
5085 | * just need to consider an extra check for case (c) after ensuring we | |
5086 | * handle the case uclamp_min > uclamp_max. | |
5087 | */ | |
5088 | uclamp_min = min(uclamp_min, uclamp_max); | |
f1f8d0a2 VG |
5089 | if (fits && (util < uclamp_min) && |
5090 | (uclamp_min > get_actual_cpu_capacity(cpu))) | |
e5ed0550 | 5091 | return -1; |
48d5e9da QY |
5092 | |
5093 | return fits; | |
5094 | } | |
5095 | ||
b48e16a6 | 5096 | static inline int task_fits_cpu(struct task_struct *p, int cpu) |
3b1baa64 | 5097 | { |
b48e16a6 QY |
5098 | unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN); |
5099 | unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX); | |
5100 | unsigned long util = task_util_est(p); | |
e5ed0550 VG |
5101 | /* |
5102 | * Return true only if the cpu fully fits the task requirements, which | |
5103 | * include the utilization but also the performance hints. | |
5104 | */ | |
5105 | return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0); | |
3b1baa64 MR |
5106 | } |
5107 | ||
5108 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
5109 | { | |
22d56074 QY |
5110 | int cpu = cpu_of(rq); |
5111 | ||
740cf8a7 | 5112 | if (!sched_asym_cpucap_active()) |
3b1baa64 MR |
5113 | return; |
5114 | ||
22d56074 QY |
5115 | /* |
5116 | * Affinity allows us to go somewhere higher? Or are we on biggest | |
5117 | * available CPU already? Or do we fit into this CPU ? | |
5118 | */ | |
5119 | if (!p || (p->nr_cpus_allowed == 1) || | |
5120 | (arch_scale_cpu_capacity(cpu) == p->max_allowed_capacity) || | |
5121 | task_fits_cpu(p, cpu)) { | |
3b1baa64 | 5122 | |
3b1baa64 MR |
5123 | rq->misfit_task_load = 0; |
5124 | return; | |
5125 | } | |
5126 | ||
01cfcde9 VG |
5127 | /* |
5128 | * Make sure that misfit_task_load will not be null even if | |
5129 | * task_h_load() returns 0. | |
5130 | */ | |
5131 | rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); | |
3b1baa64 MR |
5132 | } |
5133 | ||
38033c37 PZ |
5134 | #else /* CONFIG_SMP */ |
5135 | ||
a7b359fc OU |
5136 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
5137 | { | |
c0490bc9 | 5138 | return !cfs_rq->nr_running; |
a7b359fc OU |
5139 | } |
5140 | ||
d31b1a66 VG |
5141 | #define UPDATE_TG 0x0 |
5142 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 5143 | #define DO_ATTACH 0x0 |
e1f078f5 | 5144 | #define DO_DETACH 0x0 |
d31b1a66 | 5145 | |
88c0616e | 5146 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 5147 | { |
ea14b57e | 5148 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
5149 | } |
5150 | ||
9d89c257 | 5151 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 5152 | |
a05e8c51 | 5153 | static inline void |
a4f9a0e5 | 5154 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} |
a05e8c51 BP |
5155 | static inline void |
5156 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
5157 | ||
7d058285 | 5158 | static inline int sched_balance_newidle(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
5159 | { |
5160 | return 0; | |
5161 | } | |
5162 | ||
7f65ea42 PB |
5163 | static inline void |
5164 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
5165 | ||
5166 | static inline void | |
8c1f560c XY |
5167 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {} |
5168 | ||
5169 | static inline void | |
5170 | util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p, | |
5171 | bool task_sleep) {} | |
3b1baa64 | 5172 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 5173 | |
38033c37 | 5174 | #endif /* CONFIG_SMP */ |
9d85f21c | 5175 | |
aeb73b04 | 5176 | static void |
d07f09a1 | 5177 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
a53ce18c | 5178 | { |
2f2fc17b | 5179 | u64 vslice, vruntime = avg_vruntime(cfs_rq); |
86bfbb7c | 5180 | s64 lag = 0; |
a53ce18c | 5181 | |
2f2fc17b PZ |
5182 | se->slice = sysctl_sched_base_slice; |
5183 | vslice = calc_delta_fair(se->slice, se); | |
5184 | ||
86bfbb7c PZ |
5185 | /* |
5186 | * Due to how V is constructed as the weighted average of entities, | |
5187 | * adding tasks with positive lag, or removing tasks with negative lag | |
5188 | * will move 'time' backwards, this can screw around with the lag of | |
5189 | * other tasks. | |
5190 | * | |
5191 | * EEVDF: placement strategy #1 / #2 | |
5192 | */ | |
e8f331bc | 5193 | if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) { |
86bfbb7c PZ |
5194 | struct sched_entity *curr = cfs_rq->curr; |
5195 | unsigned long load; | |
a53ce18c | 5196 | |
86bfbb7c | 5197 | lag = se->vlag; |
a53ce18c | 5198 | |
a2e7a7eb | 5199 | /* |
86bfbb7c PZ |
5200 | * If we want to place a task and preserve lag, we have to |
5201 | * consider the effect of the new entity on the weighted | |
5202 | * average and compensate for this, otherwise lag can quickly | |
5203 | * evaporate. | |
5204 | * | |
5205 | * Lag is defined as: | |
5206 | * | |
5207 | * lag_i = S - s_i = w_i * (V - v_i) | |
5208 | * | |
5209 | * To avoid the 'w_i' term all over the place, we only track | |
5210 | * the virtual lag: | |
5211 | * | |
5212 | * vl_i = V - v_i <=> v_i = V - vl_i | |
5213 | * | |
5214 | * And we take V to be the weighted average of all v: | |
5215 | * | |
5216 | * V = (\Sum w_j*v_j) / W | |
5217 | * | |
5218 | * Where W is: \Sum w_j | |
5219 | * | |
5220 | * Then, the weighted average after adding an entity with lag | |
5221 | * vl_i is given by: | |
5222 | * | |
5223 | * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i) | |
5224 | * = (W*V + w_i*(V - vl_i)) / (W + w_i) | |
5225 | * = (W*V + w_i*V - w_i*vl_i) / (W + w_i) | |
5226 | * = (V*(W + w_i) - w_i*l) / (W + w_i) | |
5227 | * = V - w_i*vl_i / (W + w_i) | |
5228 | * | |
5229 | * And the actual lag after adding an entity with vl_i is: | |
5230 | * | |
5231 | * vl'_i = V' - v_i | |
5232 | * = V - w_i*vl_i / (W + w_i) - (V - vl_i) | |
5233 | * = vl_i - w_i*vl_i / (W + w_i) | |
5234 | * | |
5235 | * Which is strictly less than vl_i. So in order to preserve lag | |
5236 | * we should inflate the lag before placement such that the | |
5237 | * effective lag after placement comes out right. | |
5238 | * | |
5239 | * As such, invert the above relation for vl'_i to get the vl_i | |
5240 | * we need to use such that the lag after placement is the lag | |
5241 | * we computed before dequeue. | |
5242 | * | |
5243 | * vl'_i = vl_i - w_i*vl_i / (W + w_i) | |
5244 | * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i) | |
5245 | * | |
5246 | * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i | |
5247 | * = W*vl_i | |
5248 | * | |
5249 | * vl_i = (W + w_i)*vl'_i / W | |
a2e7a7eb | 5250 | */ |
86bfbb7c PZ |
5251 | load = cfs_rq->avg_load; |
5252 | if (curr && curr->on_rq) | |
147f3efa | 5253 | load += scale_load_down(curr->load.weight); |
a53ce18c | 5254 | |
147f3efa | 5255 | lag *= load + scale_load_down(se->load.weight); |
86bfbb7c PZ |
5256 | if (WARN_ON_ONCE(!load)) |
5257 | load = 1; | |
5258 | lag = div_s64(lag, load); | |
86bfbb7c | 5259 | } |
a53ce18c | 5260 | |
76cae9db | 5261 | se->vruntime = vruntime - lag; |
94dfb5e7 | 5262 | |
2cb8600e | 5263 | /* |
b9e6e286 | 5264 | * When joining the competition; the existing tasks will be, |
147f3efa PZ |
5265 | * on average, halfway through their slice, as such start tasks |
5266 | * off with half a slice to ease into the competition. | |
2cb8600e | 5267 | */ |
d07f09a1 | 5268 | if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL)) |
147f3efa | 5269 | vslice /= 2; |
2cae3948 | 5270 | |
147f3efa PZ |
5271 | /* |
5272 | * EEVDF: vd_i = ve_i + r_i/w_i | |
5273 | */ | |
5274 | se->deadline = se->vruntime + vslice; | |
aeb73b04 PZ |
5275 | } |
5276 | ||
d3d9dc33 | 5277 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
79462e8c | 5278 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq); |
d3d9dc33 | 5279 | |
fe61468b | 5280 | static inline bool cfs_bandwidth_used(void); |
b5179ac7 | 5281 | |
bf0f6f24 | 5282 | static void |
88ec22d3 | 5283 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 5284 | { |
2f950354 PZ |
5285 | bool curr = cfs_rq->curr == se; |
5286 | ||
88ec22d3 | 5287 | /* |
2f950354 PZ |
5288 | * If we're the current task, we must renormalise before calling |
5289 | * update_curr(). | |
88ec22d3 | 5290 | */ |
e8f331bc | 5291 | if (curr) |
d07f09a1 | 5292 | place_entity(cfs_rq, se, flags); |
88ec22d3 | 5293 | |
2f950354 PZ |
5294 | update_curr(cfs_rq); |
5295 | ||
89ee048f VG |
5296 | /* |
5297 | * When enqueuing a sched_entity, we must: | |
5298 | * - Update loads to have both entity and cfs_rq synced with now. | |
859f2062 CZ |
5299 | * - For group_entity, update its runnable_weight to reflect the new |
5300 | * h_nr_running of its group cfs_rq. | |
89ee048f VG |
5301 | * - For group_entity, update its weight to reflect the new share of |
5302 | * its group cfs_rq | |
5303 | * - Add its new weight to cfs_rq->load.weight | |
5304 | */ | |
b382a531 | 5305 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
9f683953 | 5306 | se_update_runnable(se); |
e8f331bc PZ |
5307 | /* |
5308 | * XXX update_load_avg() above will have attached us to the pelt sum; | |
5309 | * but update_cfs_group() here will re-adjust the weight and have to | |
5310 | * undo/redo all that. Seems wasteful. | |
5311 | */ | |
1ea6c46a | 5312 | update_cfs_group(se); |
bf0f6f24 | 5313 | |
e8f331bc PZ |
5314 | /* |
5315 | * XXX now that the entity has been re-weighted, and it's lag adjusted, | |
5316 | * we can place the entity. | |
5317 | */ | |
5318 | if (!curr) | |
d07f09a1 | 5319 | place_entity(cfs_rq, se, flags); |
e8f331bc | 5320 | |
17bc14b7 | 5321 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 5322 | |
a53ce18c VG |
5323 | /* Entity has migrated, no longer consider this task hot */ |
5324 | if (flags & ENQUEUE_MIGRATED) | |
5325 | se->exec_start = 0; | |
bf0f6f24 | 5326 | |
cb251765 | 5327 | check_schedstat_required(); |
60f2415e | 5328 | update_stats_enqueue_fair(cfs_rq, se, flags); |
2f950354 | 5329 | if (!curr) |
83b699ed | 5330 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 5331 | se->on_rq = 1; |
3d4b47b4 | 5332 | |
51bf903b | 5333 | if (cfs_rq->nr_running == 1) { |
d3d9dc33 | 5334 | check_enqueue_throttle(cfs_rq); |
79462e8c | 5335 | if (!throttled_hierarchy(cfs_rq)) { |
51bf903b | 5336 | list_add_leaf_cfs_rq(cfs_rq); |
79462e8c JD |
5337 | } else { |
5338 | #ifdef CONFIG_CFS_BANDWIDTH | |
677ea015 | 5339 | struct rq *rq = rq_of(cfs_rq); |
f1044799 | 5340 | |
79462e8c | 5341 | if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock) |
677ea015 JD |
5342 | cfs_rq->throttled_clock = rq_clock(rq); |
5343 | if (!cfs_rq->throttled_clock_self) | |
5344 | cfs_rq->throttled_clock_self = rq_clock(rq); | |
79462e8c JD |
5345 | #endif |
5346 | } | |
2c13c919 RR |
5347 | } |
5348 | } | |
2002c695 | 5349 | |
2c13c919 RR |
5350 | static void __clear_buddies_next(struct sched_entity *se) |
5351 | { | |
5352 | for_each_sched_entity(se) { | |
5353 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 5354 | if (cfs_rq->next != se) |
2c13c919 | 5355 | break; |
f1044799 PZ |
5356 | |
5357 | cfs_rq->next = NULL; | |
2c13c919 | 5358 | } |
2002c695 PZ |
5359 | } |
5360 | ||
a571bbea PZ |
5361 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
5362 | { | |
2c13c919 RR |
5363 | if (cfs_rq->next == se) |
5364 | __clear_buddies_next(se); | |
a571bbea PZ |
5365 | } |
5366 | ||
6c16a6dc | 5367 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 5368 | |
bf0f6f24 | 5369 | static void |
371fd7e7 | 5370 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 5371 | { |
e1f078f5 CZ |
5372 | int action = UPDATE_TG; |
5373 | ||
5374 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se))) | |
5375 | action |= DO_DETACH; | |
5376 | ||
a2a2d680 DA |
5377 | /* |
5378 | * Update run-time statistics of the 'current'. | |
5379 | */ | |
5380 | update_curr(cfs_rq); | |
89ee048f VG |
5381 | |
5382 | /* | |
5383 | * When dequeuing a sched_entity, we must: | |
5384 | * - Update loads to have both entity and cfs_rq synced with now. | |
859f2062 CZ |
5385 | * - For group_entity, update its runnable_weight to reflect the new |
5386 | * h_nr_running of its group cfs_rq. | |
dfcb245e | 5387 | * - Subtract its previous weight from cfs_rq->load.weight. |
89ee048f VG |
5388 | * - For group entity, update its weight to reflect the new share |
5389 | * of its group cfs_rq. | |
5390 | */ | |
e1f078f5 | 5391 | update_load_avg(cfs_rq, se, action); |
9f683953 | 5392 | se_update_runnable(se); |
a2a2d680 | 5393 | |
60f2415e | 5394 | update_stats_dequeue_fair(cfs_rq, se, flags); |
67e9fb2a | 5395 | |
2002c695 | 5396 | clear_buddies(cfs_rq, se); |
4793241b | 5397 | |
e8f331bc | 5398 | update_entity_lag(cfs_rq, se); |
83b699ed | 5399 | if (se != cfs_rq->curr) |
30cfdcfc | 5400 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 5401 | se->on_rq = 0; |
30cfdcfc | 5402 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 | 5403 | |
d8b4986d PT |
5404 | /* return excess runtime on last dequeue */ |
5405 | return_cfs_rq_runtime(cfs_rq); | |
5406 | ||
1ea6c46a | 5407 | update_cfs_group(se); |
b60205c7 PZ |
5408 | |
5409 | /* | |
5410 | * Now advance min_vruntime if @se was the entity holding it back, | |
5411 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
5412 | * put back on, and if we advance min_vruntime, we'll be placed back | |
b9e6e286 | 5413 | * further than we started -- i.e. we'll be penalized. |
b60205c7 | 5414 | */ |
9845c49c | 5415 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 5416 | update_min_vruntime(cfs_rq); |
e2f3e35f VD |
5417 | |
5418 | if (cfs_rq->nr_running == 0) | |
5419 | update_idle_cfs_rq_clock_pelt(cfs_rq); | |
bf0f6f24 IM |
5420 | } |
5421 | ||
83b699ed | 5422 | static void |
8494f412 | 5423 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 5424 | { |
21f56ffe PZ |
5425 | clear_buddies(cfs_rq, se); |
5426 | ||
83b699ed SV |
5427 | /* 'current' is not kept within the tree. */ |
5428 | if (se->on_rq) { | |
5429 | /* | |
5430 | * Any task has to be enqueued before it get to execute on | |
5431 | * a CPU. So account for the time it spent waiting on the | |
5432 | * runqueue. | |
5433 | */ | |
60f2415e | 5434 | update_stats_wait_end_fair(cfs_rq, se); |
83b699ed | 5435 | __dequeue_entity(cfs_rq, se); |
88c0616e | 5436 | update_load_avg(cfs_rq, se, UPDATE_TG); |
63304558 PZ |
5437 | /* |
5438 | * HACK, stash a copy of deadline at the point of pick in vlag, | |
5439 | * which isn't used until dequeue. | |
5440 | */ | |
5441 | se->vlag = se->deadline; | |
83b699ed SV |
5442 | } |
5443 | ||
79303e9e | 5444 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 5445 | cfs_rq->curr = se; |
4fa8d299 | 5446 | |
eba1ed4b IM |
5447 | /* |
5448 | * Track our maximum slice length, if the CPU's load is at | |
b9e6e286 | 5449 | * least twice that of our own weight (i.e. don't track it |
eba1ed4b IM |
5450 | * when there are only lesser-weight tasks around): |
5451 | */ | |
f2bedc47 DE |
5452 | if (schedstat_enabled() && |
5453 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
ceeadb83 YS |
5454 | struct sched_statistics *stats; |
5455 | ||
5456 | stats = __schedstats_from_se(se); | |
5457 | __schedstat_set(stats->slice_max, | |
5458 | max((u64)stats->slice_max, | |
a2dcb276 | 5459 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); |
eba1ed4b | 5460 | } |
4fa8d299 | 5461 | |
4a55b450 | 5462 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
5463 | } |
5464 | ||
ac53db59 RR |
5465 | /* |
5466 | * Pick the next process, keeping these things in mind, in this order: | |
5467 | * 1) keep things fair between processes/task groups | |
5468 | * 2) pick the "next" process, since someone really wants that to run | |
5469 | * 3) pick the "last" process, for cache locality | |
5470 | * 4) do not run the "skip" process, if something else is available | |
5471 | */ | |
678d5718 | 5472 | static struct sched_entity * |
4c456c9a | 5473 | pick_next_entity(struct cfs_rq *cfs_rq) |
aa2ac252 | 5474 | { |
ac53db59 | 5475 | /* |
5e963f2b | 5476 | * Enabling NEXT_BUDDY will affect latency but not fairness. |
ac53db59 | 5477 | */ |
5e963f2b PZ |
5478 | if (sched_feat(NEXT_BUDDY) && |
5479 | cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) | |
5480 | return cfs_rq->next; | |
ac53db59 | 5481 | |
5e963f2b | 5482 | return pick_eevdf(cfs_rq); |
aa2ac252 PZ |
5483 | } |
5484 | ||
678d5718 | 5485 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 5486 | |
ab6cde26 | 5487 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
5488 | { |
5489 | /* | |
5490 | * If still on the runqueue then deactivate_task() | |
5491 | * was not called and update_curr() has to be done: | |
5492 | */ | |
5493 | if (prev->on_rq) | |
b7cc0896 | 5494 | update_curr(cfs_rq); |
bf0f6f24 | 5495 | |
d3d9dc33 PT |
5496 | /* throttle cfs_rqs exceeding runtime */ |
5497 | check_cfs_rq_runtime(cfs_rq); | |
5498 | ||
30cfdcfc | 5499 | if (prev->on_rq) { |
60f2415e | 5500 | update_stats_wait_start_fair(cfs_rq, prev); |
30cfdcfc DA |
5501 | /* Put 'current' back into the tree. */ |
5502 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 5503 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 5504 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 5505 | } |
429d43bc | 5506 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
5507 | } |
5508 | ||
8f4d37ec PZ |
5509 | static void |
5510 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 5511 | { |
bf0f6f24 | 5512 | /* |
30cfdcfc | 5513 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 5514 | */ |
30cfdcfc | 5515 | update_curr(cfs_rq); |
bf0f6f24 | 5516 | |
9d85f21c PT |
5517 | /* |
5518 | * Ensure that runnable average is periodically updated. | |
5519 | */ | |
88c0616e | 5520 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 5521 | update_cfs_group(curr); |
9d85f21c | 5522 | |
8f4d37ec PZ |
5523 | #ifdef CONFIG_SCHED_HRTICK |
5524 | /* | |
5525 | * queued ticks are scheduled to match the slice, so don't bother | |
5526 | * validating it and just reschedule. | |
5527 | */ | |
983ed7a6 | 5528 | if (queued) { |
8875125e | 5529 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
5530 | return; |
5531 | } | |
8f4d37ec PZ |
5532 | /* |
5533 | * don't let the period tick interfere with the hrtick preemption | |
5534 | */ | |
5535 | if (!sched_feat(DOUBLE_TICK) && | |
5536 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
5537 | return; | |
5538 | #endif | |
bf0f6f24 IM |
5539 | } |
5540 | ||
ab84d31e PT |
5541 | |
5542 | /************************************************** | |
5543 | * CFS bandwidth control machinery | |
5544 | */ | |
5545 | ||
5546 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 5547 | |
e9666d10 | 5548 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 5549 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
5550 | |
5551 | static inline bool cfs_bandwidth_used(void) | |
5552 | { | |
c5905afb | 5553 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
5554 | } |
5555 | ||
1ee14e6c | 5556 | void cfs_bandwidth_usage_inc(void) |
029632fb | 5557 | { |
ce48c146 | 5558 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
5559 | } |
5560 | ||
5561 | void cfs_bandwidth_usage_dec(void) | |
5562 | { | |
ce48c146 | 5563 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 5564 | } |
e9666d10 | 5565 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
5566 | static bool cfs_bandwidth_used(void) |
5567 | { | |
5568 | return true; | |
5569 | } | |
5570 | ||
1ee14e6c BS |
5571 | void cfs_bandwidth_usage_inc(void) {} |
5572 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 5573 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 5574 | |
ab84d31e PT |
5575 | /* |
5576 | * default period for cfs group bandwidth. | |
5577 | * default: 0.1s, units: nanoseconds | |
5578 | */ | |
5579 | static inline u64 default_cfs_period(void) | |
5580 | { | |
5581 | return 100000000ULL; | |
5582 | } | |
ec12cb7f PT |
5583 | |
5584 | static inline u64 sched_cfs_bandwidth_slice(void) | |
5585 | { | |
5586 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
5587 | } | |
5588 | ||
a9cf55b2 | 5589 | /* |
763a9ec0 QC |
5590 | * Replenish runtime according to assigned quota. We use sched_clock_cpu |
5591 | * directly instead of rq->clock to avoid adding additional synchronization | |
5592 | * around rq->lock. | |
a9cf55b2 PT |
5593 | * |
5594 | * requires cfs_b->lock | |
5595 | */ | |
029632fb | 5596 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 | 5597 | { |
bcb1704a HC |
5598 | s64 runtime; |
5599 | ||
f4183717 HC |
5600 | if (unlikely(cfs_b->quota == RUNTIME_INF)) |
5601 | return; | |
5602 | ||
5603 | cfs_b->runtime += cfs_b->quota; | |
bcb1704a HC |
5604 | runtime = cfs_b->runtime_snap - cfs_b->runtime; |
5605 | if (runtime > 0) { | |
5606 | cfs_b->burst_time += runtime; | |
5607 | cfs_b->nr_burst++; | |
5608 | } | |
5609 | ||
f4183717 | 5610 | cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst); |
bcb1704a | 5611 | cfs_b->runtime_snap = cfs_b->runtime; |
a9cf55b2 PT |
5612 | } |
5613 | ||
029632fb PZ |
5614 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5615 | { | |
5616 | return &tg->cfs_bandwidth; | |
5617 | } | |
5618 | ||
85dac906 | 5619 | /* returns 0 on failure to allocate runtime */ |
e98fa02c PT |
5620 | static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, |
5621 | struct cfs_rq *cfs_rq, u64 target_runtime) | |
ec12cb7f | 5622 | { |
e98fa02c PT |
5623 | u64 min_amount, amount = 0; |
5624 | ||
5625 | lockdep_assert_held(&cfs_b->lock); | |
ec12cb7f PT |
5626 | |
5627 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
e98fa02c | 5628 | min_amount = target_runtime - cfs_rq->runtime_remaining; |
ec12cb7f | 5629 | |
ec12cb7f PT |
5630 | if (cfs_b->quota == RUNTIME_INF) |
5631 | amount = min_amount; | |
58088ad0 | 5632 | else { |
77a4d1a1 | 5633 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
5634 | |
5635 | if (cfs_b->runtime > 0) { | |
5636 | amount = min(cfs_b->runtime, min_amount); | |
5637 | cfs_b->runtime -= amount; | |
5638 | cfs_b->idle = 0; | |
5639 | } | |
ec12cb7f | 5640 | } |
ec12cb7f PT |
5641 | |
5642 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
5643 | |
5644 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
5645 | } |
5646 | ||
e98fa02c PT |
5647 | /* returns 0 on failure to allocate runtime */ |
5648 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5649 | { | |
5650 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5651 | int ret; | |
5652 | ||
5653 | raw_spin_lock(&cfs_b->lock); | |
5654 | ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); | |
5655 | raw_spin_unlock(&cfs_b->lock); | |
5656 | ||
5657 | return ret; | |
5658 | } | |
5659 | ||
9dbdb155 | 5660 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
5661 | { |
5662 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 5663 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
5664 | |
5665 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
5666 | return; |
5667 | ||
5e2d2cc2 L |
5668 | if (cfs_rq->throttled) |
5669 | return; | |
85dac906 PT |
5670 | /* |
5671 | * if we're unable to extend our runtime we resched so that the active | |
5672 | * hierarchy can be throttled | |
5673 | */ | |
5674 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 5675 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
5676 | } |
5677 | ||
6c16a6dc | 5678 | static __always_inline |
9dbdb155 | 5679 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 5680 | { |
56f570e5 | 5681 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
5682 | return; |
5683 | ||
5684 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
5685 | } | |
5686 | ||
85dac906 PT |
5687 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
5688 | { | |
56f570e5 | 5689 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
5690 | } |
5691 | ||
64660c86 PT |
5692 | /* check whether cfs_rq, or any parent, is throttled */ |
5693 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5694 | { | |
56f570e5 | 5695 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
5696 | } |
5697 | ||
5698 | /* | |
5699 | * Ensure that neither of the group entities corresponding to src_cpu or | |
5700 | * dest_cpu are members of a throttled hierarchy when performing group | |
5701 | * load-balance operations. | |
5702 | */ | |
5703 | static inline int throttled_lb_pair(struct task_group *tg, | |
5704 | int src_cpu, int dest_cpu) | |
5705 | { | |
5706 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
5707 | ||
5708 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
5709 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
5710 | ||
5711 | return throttled_hierarchy(src_cfs_rq) || | |
5712 | throttled_hierarchy(dest_cfs_rq); | |
5713 | } | |
5714 | ||
64660c86 PT |
5715 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
5716 | { | |
5717 | struct rq *rq = data; | |
5718 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
5719 | ||
5720 | cfs_rq->throttle_count--; | |
64660c86 | 5721 | if (!cfs_rq->throttle_count) { |
64eaf507 CZ |
5722 | cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) - |
5723 | cfs_rq->throttled_clock_pelt; | |
31bc6aea | 5724 | |
a7b359fc | 5725 | /* Add cfs_rq with load or one or more already running entities to the list */ |
0a00a354 | 5726 | if (!cfs_rq_is_decayed(cfs_rq)) |
31bc6aea | 5727 | list_add_leaf_cfs_rq(cfs_rq); |
677ea015 JD |
5728 | |
5729 | if (cfs_rq->throttled_clock_self) { | |
5730 | u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self; | |
5731 | ||
5732 | cfs_rq->throttled_clock_self = 0; | |
5733 | ||
5734 | if (SCHED_WARN_ON((s64)delta < 0)) | |
5735 | delta = 0; | |
5736 | ||
5737 | cfs_rq->throttled_clock_self_time += delta; | |
5738 | } | |
64660c86 | 5739 | } |
64660c86 PT |
5740 | |
5741 | return 0; | |
5742 | } | |
5743 | ||
5744 | static int tg_throttle_down(struct task_group *tg, void *data) | |
5745 | { | |
5746 | struct rq *rq = data; | |
5747 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
5748 | ||
82958366 | 5749 | /* group is entering throttled state, stop time */ |
31bc6aea | 5750 | if (!cfs_rq->throttle_count) { |
64eaf507 | 5751 | cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq); |
31bc6aea | 5752 | list_del_leaf_cfs_rq(cfs_rq); |
677ea015 JD |
5753 | |
5754 | SCHED_WARN_ON(cfs_rq->throttled_clock_self); | |
5755 | if (cfs_rq->nr_running) | |
5756 | cfs_rq->throttled_clock_self = rq_clock(rq); | |
31bc6aea | 5757 | } |
64660c86 PT |
5758 | cfs_rq->throttle_count++; |
5759 | ||
5760 | return 0; | |
5761 | } | |
5762 | ||
e98fa02c | 5763 | static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
5764 | { |
5765 | struct rq *rq = rq_of(cfs_rq); | |
5766 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5767 | struct sched_entity *se; | |
43e9f7f2 | 5768 | long task_delta, idle_task_delta, dequeue = 1; |
e98fa02c PT |
5769 | |
5770 | raw_spin_lock(&cfs_b->lock); | |
5771 | /* This will start the period timer if necessary */ | |
5772 | if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { | |
5773 | /* | |
5774 | * We have raced with bandwidth becoming available, and if we | |
5775 | * actually throttled the timer might not unthrottle us for an | |
5776 | * entire period. We additionally needed to make sure that any | |
5777 | * subsequent check_cfs_rq_runtime calls agree not to throttle | |
5778 | * us, as we may commit to do cfs put_prev+pick_next, so we ask | |
5779 | * for 1ns of runtime rather than just check cfs_b. | |
5780 | */ | |
5781 | dequeue = 0; | |
5782 | } else { | |
5783 | list_add_tail_rcu(&cfs_rq->throttled_list, | |
5784 | &cfs_b->throttled_cfs_rq); | |
5785 | } | |
5786 | raw_spin_unlock(&cfs_b->lock); | |
5787 | ||
5788 | if (!dequeue) | |
5789 | return false; /* Throttle no longer required. */ | |
85dac906 PT |
5790 | |
5791 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
5792 | ||
f1b17280 | 5793 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
5794 | rcu_read_lock(); |
5795 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
5796 | rcu_read_unlock(); | |
85dac906 PT |
5797 | |
5798 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 5799 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
5800 | for_each_sched_entity(se) { |
5801 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
5802 | /* throttled entity or throttle-on-deactivate */ | |
5803 | if (!se->on_rq) | |
b6d37a76 | 5804 | goto done; |
85dac906 | 5805 | |
b6d37a76 | 5806 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
6212437f | 5807 | |
30400039 JD |
5808 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5809 | idle_task_delta = cfs_rq->h_nr_running; | |
5810 | ||
85dac906 | 5811 | qcfs_rq->h_nr_running -= task_delta; |
43e9f7f2 | 5812 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 | 5813 | |
b6d37a76 PW |
5814 | if (qcfs_rq->load.weight) { |
5815 | /* Avoid re-evaluating load for this entity: */ | |
5816 | se = parent_entity(se); | |
5817 | break; | |
5818 | } | |
5819 | } | |
5820 | ||
5821 | for_each_sched_entity(se) { | |
5822 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
5823 | /* throttled entity or throttle-on-deactivate */ | |
5824 | if (!se->on_rq) | |
5825 | goto done; | |
5826 | ||
5827 | update_load_avg(qcfs_rq, se, 0); | |
5828 | se_update_runnable(se); | |
5829 | ||
30400039 JD |
5830 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5831 | idle_task_delta = cfs_rq->h_nr_running; | |
5832 | ||
b6d37a76 PW |
5833 | qcfs_rq->h_nr_running -= task_delta; |
5834 | qcfs_rq->idle_h_nr_running -= idle_task_delta; | |
85dac906 PT |
5835 | } |
5836 | ||
b6d37a76 PW |
5837 | /* At this point se is NULL and we are at root level*/ |
5838 | sub_nr_running(rq, task_delta); | |
85dac906 | 5839 | |
b6d37a76 | 5840 | done: |
c06f04c7 | 5841 | /* |
e98fa02c PT |
5842 | * Note: distribution will already see us throttled via the |
5843 | * throttled-list. rq->lock protects completion. | |
c06f04c7 | 5844 | */ |
e98fa02c | 5845 | cfs_rq->throttled = 1; |
79462e8c JD |
5846 | SCHED_WARN_ON(cfs_rq->throttled_clock); |
5847 | if (cfs_rq->nr_running) | |
5848 | cfs_rq->throttled_clock = rq_clock(rq); | |
e98fa02c | 5849 | return true; |
85dac906 PT |
5850 | } |
5851 | ||
029632fb | 5852 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
5853 | { |
5854 | struct rq *rq = rq_of(cfs_rq); | |
5855 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5856 | struct sched_entity *se; | |
43e9f7f2 | 5857 | long task_delta, idle_task_delta; |
671fd9da | 5858 | |
22b958d8 | 5859 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
5860 | |
5861 | cfs_rq->throttled = 0; | |
1a55af2e FW |
5862 | |
5863 | update_rq_clock(rq); | |
5864 | ||
671fd9da | 5865 | raw_spin_lock(&cfs_b->lock); |
79462e8c JD |
5866 | if (cfs_rq->throttled_clock) { |
5867 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; | |
5868 | cfs_rq->throttled_clock = 0; | |
5869 | } | |
671fd9da PT |
5870 | list_del_rcu(&cfs_rq->throttled_list); |
5871 | raw_spin_unlock(&cfs_b->lock); | |
5872 | ||
64660c86 PT |
5873 | /* update hierarchical throttle state */ |
5874 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
5875 | ||
2630cde2 | 5876 | if (!cfs_rq->load.weight) { |
51bf903b CZ |
5877 | if (!cfs_rq->on_list) |
5878 | return; | |
5879 | /* | |
5880 | * Nothing to run but something to decay (on_list)? | |
5881 | * Complete the branch. | |
5882 | */ | |
5883 | for_each_sched_entity(se) { | |
5884 | if (list_add_leaf_cfs_rq(cfs_rq_of(se))) | |
5885 | break; | |
5886 | } | |
5887 | goto unthrottle_throttle; | |
2630cde2 | 5888 | } |
671fd9da PT |
5889 | |
5890 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 5891 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da | 5892 | for_each_sched_entity(se) { |
30400039 JD |
5893 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
5894 | ||
671fd9da | 5895 | if (se->on_rq) |
39f23ce0 | 5896 | break; |
30400039 JD |
5897 | enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP); |
5898 | ||
5899 | if (cfs_rq_is_idle(group_cfs_rq(se))) | |
5900 | idle_task_delta = cfs_rq->h_nr_running; | |
39f23ce0 | 5901 | |
30400039 JD |
5902 | qcfs_rq->h_nr_running += task_delta; |
5903 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
5904 | |
5905 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 5906 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 VG |
5907 | goto unthrottle_throttle; |
5908 | } | |
671fd9da | 5909 | |
39f23ce0 | 5910 | for_each_sched_entity(se) { |
30400039 | 5911 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
39f23ce0 | 5912 | |
30400039 | 5913 | update_load_avg(qcfs_rq, se, UPDATE_TG); |
39f23ce0 | 5914 | se_update_runnable(se); |
6212437f | 5915 | |
30400039 JD |
5916 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5917 | idle_task_delta = cfs_rq->h_nr_running; | |
671fd9da | 5918 | |
30400039 JD |
5919 | qcfs_rq->h_nr_running += task_delta; |
5920 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
5921 | |
5922 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 5923 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 | 5924 | goto unthrottle_throttle; |
671fd9da PT |
5925 | } |
5926 | ||
39f23ce0 VG |
5927 | /* At this point se is NULL and we are at root level*/ |
5928 | add_nr_running(rq, task_delta); | |
671fd9da | 5929 | |
39f23ce0 | 5930 | unthrottle_throttle: |
fe61468b VG |
5931 | assert_list_leaf_cfs_rq(rq); |
5932 | ||
97fb7a0a | 5933 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 5934 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 5935 | resched_curr(rq); |
671fd9da PT |
5936 | } |
5937 | ||
8ad075c2 JD |
5938 | #ifdef CONFIG_SMP |
5939 | static void __cfsb_csd_unthrottle(void *arg) | |
671fd9da | 5940 | { |
8ad075c2 JD |
5941 | struct cfs_rq *cursor, *tmp; |
5942 | struct rq *rq = arg; | |
5943 | struct rq_flags rf; | |
5944 | ||
5945 | rq_lock(rq, &rf); | |
5946 | ||
ebb83d84 HJ |
5947 | /* |
5948 | * Iterating over the list can trigger several call to | |
5949 | * update_rq_clock() in unthrottle_cfs_rq(). | |
5950 | * Do it once and skip the potential next ones. | |
5951 | */ | |
5952 | update_rq_clock(rq); | |
5953 | rq_clock_start_loop_update(rq); | |
5954 | ||
8ad075c2 JD |
5955 | /* |
5956 | * Since we hold rq lock we're safe from concurrent manipulation of | |
5957 | * the CSD list. However, this RCU critical section annotates the | |
5958 | * fact that we pair with sched_free_group_rcu(), so that we cannot | |
5959 | * race with group being freed in the window between removing it | |
5960 | * from the list and advancing to the next entry in the list. | |
5961 | */ | |
5962 | rcu_read_lock(); | |
5963 | ||
5964 | list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list, | |
5965 | throttled_csd_list) { | |
5966 | list_del_init(&cursor->throttled_csd_list); | |
5967 | ||
5968 | if (cfs_rq_throttled(cursor)) | |
5969 | unthrottle_cfs_rq(cursor); | |
5970 | } | |
5971 | ||
5972 | rcu_read_unlock(); | |
5973 | ||
ebb83d84 | 5974 | rq_clock_stop_loop_update(rq); |
8ad075c2 JD |
5975 | rq_unlock(rq, &rf); |
5976 | } | |
5977 | ||
5978 | static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq) | |
5979 | { | |
5980 | struct rq *rq = rq_of(cfs_rq); | |
5981 | bool first; | |
5982 | ||
5983 | if (rq == this_rq()) { | |
5984 | unthrottle_cfs_rq(cfs_rq); | |
5985 | return; | |
5986 | } | |
5987 | ||
5988 | /* Already enqueued */ | |
5989 | if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list))) | |
5990 | return; | |
5991 | ||
5992 | first = list_empty(&rq->cfsb_csd_list); | |
5993 | list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list); | |
5994 | if (first) | |
5995 | smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd); | |
5996 | } | |
5997 | #else | |
5998 | static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq) | |
5999 | { | |
6000 | unthrottle_cfs_rq(cfs_rq); | |
6001 | } | |
6002 | #endif | |
6003 | ||
6004 | static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq) | |
6005 | { | |
6006 | lockdep_assert_rq_held(rq_of(cfs_rq)); | |
6007 | ||
6008 | if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) || | |
6009 | cfs_rq->runtime_remaining <= 0)) | |
6010 | return; | |
6011 | ||
6012 | __unthrottle_cfs_rq_async(cfs_rq); | |
6013 | } | |
6014 | ||
6015 | static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) | |
6016 | { | |
8ad075c2 | 6017 | int this_cpu = smp_processor_id(); |
26a8b127 | 6018 | u64 runtime, remaining = 1; |
8ad075c2 | 6019 | bool throttled = false; |
2f8c6229 | 6020 | struct cfs_rq *cfs_rq, *tmp; |
8ad075c2 JD |
6021 | struct rq_flags rf; |
6022 | struct rq *rq; | |
2f8c6229 | 6023 | LIST_HEAD(local_unthrottle); |
671fd9da PT |
6024 | |
6025 | rcu_read_lock(); | |
6026 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
6027 | throttled_list) { | |
8ad075c2 JD |
6028 | rq = rq_of(cfs_rq); |
6029 | ||
6030 | if (!remaining) { | |
6031 | throttled = true; | |
6032 | break; | |
6033 | } | |
671fd9da | 6034 | |
c0ad4aa4 | 6035 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
6036 | if (!cfs_rq_throttled(cfs_rq)) |
6037 | goto next; | |
6038 | ||
8ad075c2 JD |
6039 | /* Already queued for async unthrottle */ |
6040 | if (!list_empty(&cfs_rq->throttled_csd_list)) | |
6041 | goto next; | |
8ad075c2 JD |
6042 | |
6043 | /* By the above checks, this should never be true */ | |
5e2d2cc2 L |
6044 | SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); |
6045 | ||
26a8b127 | 6046 | raw_spin_lock(&cfs_b->lock); |
671fd9da | 6047 | runtime = -cfs_rq->runtime_remaining + 1; |
26a8b127 HC |
6048 | if (runtime > cfs_b->runtime) |
6049 | runtime = cfs_b->runtime; | |
6050 | cfs_b->runtime -= runtime; | |
6051 | remaining = cfs_b->runtime; | |
6052 | raw_spin_unlock(&cfs_b->lock); | |
671fd9da PT |
6053 | |
6054 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
6055 | |
6056 | /* we check whether we're throttled above */ | |
8ad075c2 | 6057 | if (cfs_rq->runtime_remaining > 0) { |
2f8c6229 | 6058 | if (cpu_of(rq) != this_cpu) { |
8ad075c2 | 6059 | unthrottle_cfs_rq_async(cfs_rq); |
2f8c6229 JD |
6060 | } else { |
6061 | /* | |
6062 | * We currently only expect to be unthrottling | |
6063 | * a single cfs_rq locally. | |
6064 | */ | |
6065 | SCHED_WARN_ON(!list_empty(&local_unthrottle)); | |
6066 | list_add_tail(&cfs_rq->throttled_csd_list, | |
6067 | &local_unthrottle); | |
6068 | } | |
8ad075c2 JD |
6069 | } else { |
6070 | throttled = true; | |
6071 | } | |
671fd9da PT |
6072 | |
6073 | next: | |
c0ad4aa4 | 6074 | rq_unlock_irqrestore(rq, &rf); |
671fd9da | 6075 | } |
8ad075c2 | 6076 | |
2f8c6229 JD |
6077 | list_for_each_entry_safe(cfs_rq, tmp, &local_unthrottle, |
6078 | throttled_csd_list) { | |
6079 | struct rq *rq = rq_of(cfs_rq); | |
6080 | ||
8ad075c2 | 6081 | rq_lock_irqsave(rq, &rf); |
2f8c6229 JD |
6082 | |
6083 | list_del_init(&cfs_rq->throttled_csd_list); | |
6084 | ||
6085 | if (cfs_rq_throttled(cfs_rq)) | |
6086 | unthrottle_cfs_rq(cfs_rq); | |
6087 | ||
8ad075c2 JD |
6088 | rq_unlock_irqrestore(rq, &rf); |
6089 | } | |
2f8c6229 JD |
6090 | SCHED_WARN_ON(!list_empty(&local_unthrottle)); |
6091 | ||
6092 | rcu_read_unlock(); | |
8ad075c2 JD |
6093 | |
6094 | return throttled; | |
671fd9da PT |
6095 | } |
6096 | ||
58088ad0 PT |
6097 | /* |
6098 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
6099 | * cfs_rqs as appropriate. If there has been no activity within the last | |
6100 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
6101 | * used to track this state. | |
6102 | */ | |
c0ad4aa4 | 6103 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 6104 | { |
51f2176d | 6105 | int throttled; |
58088ad0 | 6106 | |
58088ad0 PT |
6107 | /* no need to continue the timer with no bandwidth constraint */ |
6108 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 6109 | goto out_deactivate; |
58088ad0 | 6110 | |
671fd9da | 6111 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 6112 | cfs_b->nr_periods += overrun; |
671fd9da | 6113 | |
f4183717 HC |
6114 | /* Refill extra burst quota even if cfs_b->idle */ |
6115 | __refill_cfs_bandwidth_runtime(cfs_b); | |
6116 | ||
51f2176d BS |
6117 | /* |
6118 | * idle depends on !throttled (for the case of a large deficit), and if | |
6119 | * we're going inactive then everything else can be deferred | |
6120 | */ | |
6121 | if (cfs_b->idle && !throttled) | |
6122 | goto out_deactivate; | |
a9cf55b2 | 6123 | |
671fd9da PT |
6124 | if (!throttled) { |
6125 | /* mark as potentially idle for the upcoming period */ | |
6126 | cfs_b->idle = 1; | |
51f2176d | 6127 | return 0; |
671fd9da PT |
6128 | } |
6129 | ||
e8da1b18 NR |
6130 | /* account preceding periods in which throttling occurred */ |
6131 | cfs_b->nr_throttled += overrun; | |
6132 | ||
671fd9da | 6133 | /* |
26a8b127 | 6134 | * This check is repeated as we release cfs_b->lock while we unthrottle. |
671fd9da | 6135 | */ |
ab93a4bc | 6136 | while (throttled && cfs_b->runtime > 0) { |
c0ad4aa4 | 6137 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 6138 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
8ad075c2 | 6139 | throttled = distribute_cfs_runtime(cfs_b); |
c0ad4aa4 | 6140 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da | 6141 | } |
58088ad0 | 6142 | |
671fd9da PT |
6143 | /* |
6144 | * While we are ensured activity in the period following an | |
6145 | * unthrottle, this also covers the case in which the new bandwidth is | |
6146 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
6147 | * timer to remain active while there are any throttled entities.) | |
6148 | */ | |
6149 | cfs_b->idle = 0; | |
58088ad0 | 6150 | |
51f2176d BS |
6151 | return 0; |
6152 | ||
6153 | out_deactivate: | |
51f2176d | 6154 | return 1; |
58088ad0 | 6155 | } |
d3d9dc33 | 6156 | |
d8b4986d PT |
6157 | /* a cfs_rq won't donate quota below this amount */ |
6158 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
6159 | /* minimum remaining period time to redistribute slack quota */ | |
6160 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
6161 | /* how long we wait to gather additional slack before distributing */ | |
6162 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
6163 | ||
db06e78c BS |
6164 | /* |
6165 | * Are we near the end of the current quota period? | |
6166 | * | |
6167 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 6168 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
6169 | * migrate_hrtimers, base is never cleared, so we are fine. |
6170 | */ | |
d8b4986d PT |
6171 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
6172 | { | |
6173 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
72d0ad7c | 6174 | s64 remaining; |
d8b4986d PT |
6175 | |
6176 | /* if the call-back is running a quota refresh is already occurring */ | |
6177 | if (hrtimer_callback_running(refresh_timer)) | |
6178 | return 1; | |
6179 | ||
6180 | /* is a quota refresh about to occur? */ | |
6181 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
72d0ad7c | 6182 | if (remaining < (s64)min_expire) |
d8b4986d PT |
6183 | return 1; |
6184 | ||
6185 | return 0; | |
6186 | } | |
6187 | ||
6188 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
6189 | { | |
6190 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
6191 | ||
6192 | /* if there's a quota refresh soon don't bother with slack */ | |
6193 | if (runtime_refresh_within(cfs_b, min_left)) | |
6194 | return; | |
6195 | ||
66567fcb | 6196 | /* don't push forwards an existing deferred unthrottle */ |
6197 | if (cfs_b->slack_started) | |
6198 | return; | |
6199 | cfs_b->slack_started = true; | |
6200 | ||
4cfafd30 PZ |
6201 | hrtimer_start(&cfs_b->slack_timer, |
6202 | ns_to_ktime(cfs_bandwidth_slack_period), | |
6203 | HRTIMER_MODE_REL); | |
d8b4986d PT |
6204 | } |
6205 | ||
6206 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
6207 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
6208 | { | |
6209 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
6210 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
6211 | ||
6212 | if (slack_runtime <= 0) | |
6213 | return; | |
6214 | ||
6215 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 6216 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
6217 | cfs_b->runtime += slack_runtime; |
6218 | ||
6219 | /* we are under rq->lock, defer unthrottling using a timer */ | |
6220 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
6221 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
6222 | start_cfs_slack_bandwidth(cfs_b); | |
6223 | } | |
6224 | raw_spin_unlock(&cfs_b->lock); | |
6225 | ||
6226 | /* even if it's not valid for return we don't want to try again */ | |
6227 | cfs_rq->runtime_remaining -= slack_runtime; | |
6228 | } | |
6229 | ||
6230 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
6231 | { | |
56f570e5 PT |
6232 | if (!cfs_bandwidth_used()) |
6233 | return; | |
6234 | ||
fccfdc6f | 6235 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
6236 | return; |
6237 | ||
6238 | __return_cfs_rq_runtime(cfs_rq); | |
6239 | } | |
6240 | ||
6241 | /* | |
6242 | * This is done with a timer (instead of inline with bandwidth return) since | |
6243 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
6244 | */ | |
6245 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
6246 | { | |
6247 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 6248 | unsigned long flags; |
d8b4986d PT |
6249 | |
6250 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 6251 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 6252 | cfs_b->slack_started = false; |
baa9be4f | 6253 | |
db06e78c | 6254 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 6255 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 6256 | return; |
db06e78c | 6257 | } |
d8b4986d | 6258 | |
c06f04c7 | 6259 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 6260 | runtime = cfs_b->runtime; |
c06f04c7 | 6261 | |
c0ad4aa4 | 6262 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
6263 | |
6264 | if (!runtime) | |
6265 | return; | |
6266 | ||
26a8b127 | 6267 | distribute_cfs_runtime(cfs_b); |
d8b4986d PT |
6268 | } |
6269 | ||
d3d9dc33 PT |
6270 | /* |
6271 | * When a group wakes up we want to make sure that its quota is not already | |
6272 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
c034f48e | 6273 | * runtime as update_curr() throttling can not trigger until it's on-rq. |
d3d9dc33 PT |
6274 | */ |
6275 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
6276 | { | |
56f570e5 PT |
6277 | if (!cfs_bandwidth_used()) |
6278 | return; | |
6279 | ||
d3d9dc33 PT |
6280 | /* an active group must be handled by the update_curr()->put() path */ |
6281 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
6282 | return; | |
6283 | ||
6284 | /* ensure the group is not already throttled */ | |
6285 | if (cfs_rq_throttled(cfs_rq)) | |
6286 | return; | |
6287 | ||
6288 | /* update runtime allocation */ | |
6289 | account_cfs_rq_runtime(cfs_rq, 0); | |
6290 | if (cfs_rq->runtime_remaining <= 0) | |
6291 | throttle_cfs_rq(cfs_rq); | |
6292 | } | |
6293 | ||
55e16d30 PZ |
6294 | static void sync_throttle(struct task_group *tg, int cpu) |
6295 | { | |
6296 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
6297 | ||
6298 | if (!cfs_bandwidth_used()) | |
6299 | return; | |
6300 | ||
6301 | if (!tg->parent) | |
6302 | return; | |
6303 | ||
6304 | cfs_rq = tg->cfs_rq[cpu]; | |
6305 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
6306 | ||
6307 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
64eaf507 | 6308 | cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu)); |
55e16d30 PZ |
6309 | } |
6310 | ||
d3d9dc33 | 6311 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 6312 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 6313 | { |
56f570e5 | 6314 | if (!cfs_bandwidth_used()) |
678d5718 | 6315 | return false; |
56f570e5 | 6316 | |
d3d9dc33 | 6317 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 6318 | return false; |
d3d9dc33 PT |
6319 | |
6320 | /* | |
6321 | * it's possible for a throttled entity to be forced into a running | |
6322 | * state (e.g. set_curr_task), in this case we're finished. | |
6323 | */ | |
6324 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 6325 | return true; |
d3d9dc33 | 6326 | |
e98fa02c | 6327 | return throttle_cfs_rq(cfs_rq); |
d3d9dc33 | 6328 | } |
029632fb | 6329 | |
029632fb PZ |
6330 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
6331 | { | |
6332 | struct cfs_bandwidth *cfs_b = | |
6333 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 6334 | |
029632fb PZ |
6335 | do_sched_cfs_slack_timer(cfs_b); |
6336 | ||
6337 | return HRTIMER_NORESTART; | |
6338 | } | |
6339 | ||
2e8e1922 PA |
6340 | extern const u64 max_cfs_quota_period; |
6341 | ||
029632fb PZ |
6342 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
6343 | { | |
6344 | struct cfs_bandwidth *cfs_b = | |
6345 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 6346 | unsigned long flags; |
029632fb PZ |
6347 | int overrun; |
6348 | int idle = 0; | |
2e8e1922 | 6349 | int count = 0; |
029632fb | 6350 | |
c0ad4aa4 | 6351 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 6352 | for (;;) { |
77a4d1a1 | 6353 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
6354 | if (!overrun) |
6355 | break; | |
6356 | ||
5a6d6a6c HC |
6357 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
6358 | ||
2e8e1922 PA |
6359 | if (++count > 3) { |
6360 | u64 new, old = ktime_to_ns(cfs_b->period); | |
6361 | ||
4929a4e6 XZ |
6362 | /* |
6363 | * Grow period by a factor of 2 to avoid losing precision. | |
6364 | * Precision loss in the quota/period ratio can cause __cfs_schedulable | |
6365 | * to fail. | |
6366 | */ | |
6367 | new = old * 2; | |
6368 | if (new < max_cfs_quota_period) { | |
6369 | cfs_b->period = ns_to_ktime(new); | |
6370 | cfs_b->quota *= 2; | |
f4183717 | 6371 | cfs_b->burst *= 2; |
4929a4e6 XZ |
6372 | |
6373 | pr_warn_ratelimited( | |
6374 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
6375 | smp_processor_id(), | |
6376 | div_u64(new, NSEC_PER_USEC), | |
6377 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
6378 | } else { | |
6379 | pr_warn_ratelimited( | |
6380 | "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
6381 | smp_processor_id(), | |
6382 | div_u64(old, NSEC_PER_USEC), | |
6383 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
6384 | } | |
2e8e1922 PA |
6385 | |
6386 | /* reset count so we don't come right back in here */ | |
6387 | count = 0; | |
6388 | } | |
029632fb | 6389 | } |
4cfafd30 PZ |
6390 | if (idle) |
6391 | cfs_b->period_active = 0; | |
c0ad4aa4 | 6392 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
6393 | |
6394 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
6395 | } | |
6396 | ||
c98c1827 | 6397 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) |
029632fb PZ |
6398 | { |
6399 | raw_spin_lock_init(&cfs_b->lock); | |
6400 | cfs_b->runtime = 0; | |
6401 | cfs_b->quota = RUNTIME_INF; | |
6402 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
f4183717 | 6403 | cfs_b->burst = 0; |
c98c1827 | 6404 | cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF; |
029632fb PZ |
6405 | |
6406 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 6407 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb | 6408 | cfs_b->period_timer.function = sched_cfs_period_timer; |
41abdba9 SH |
6409 | |
6410 | /* Add a random offset so that timers interleave */ | |
6411 | hrtimer_set_expires(&cfs_b->period_timer, | |
6412 | get_random_u32_below(cfs_b->period)); | |
029632fb PZ |
6413 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
6414 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
66567fcb | 6415 | cfs_b->slack_started = false; |
029632fb PZ |
6416 | } |
6417 | ||
6418 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
6419 | { | |
6420 | cfs_rq->runtime_enabled = 0; | |
6421 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
8ad075c2 | 6422 | INIT_LIST_HEAD(&cfs_rq->throttled_csd_list); |
029632fb PZ |
6423 | } |
6424 | ||
77a4d1a1 | 6425 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 6426 | { |
4cfafd30 | 6427 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 6428 | |
f1d1be8a XP |
6429 | if (cfs_b->period_active) |
6430 | return; | |
6431 | ||
6432 | cfs_b->period_active = 1; | |
763a9ec0 | 6433 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); |
f1d1be8a | 6434 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
6435 | } |
6436 | ||
6437 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
6438 | { | |
8ad075c2 JD |
6439 | int __maybe_unused i; |
6440 | ||
7f1a169b TH |
6441 | /* init_cfs_bandwidth() was not called */ |
6442 | if (!cfs_b->throttled_cfs_rq.next) | |
6443 | return; | |
6444 | ||
029632fb PZ |
6445 | hrtimer_cancel(&cfs_b->period_timer); |
6446 | hrtimer_cancel(&cfs_b->slack_timer); | |
8ad075c2 JD |
6447 | |
6448 | /* | |
6449 | * It is possible that we still have some cfs_rq's pending on a CSD | |
6450 | * list, though this race is very rare. In order for this to occur, we | |
6451 | * must have raced with the last task leaving the group while there | |
6452 | * exist throttled cfs_rq(s), and the period_timer must have queued the | |
6453 | * CSD item but the remote cpu has not yet processed it. To handle this, | |
6454 | * we can simply flush all pending CSD work inline here. We're | |
6455 | * guaranteed at this point that no additional cfs_rq of this group can | |
6456 | * join a CSD list. | |
6457 | */ | |
6458 | #ifdef CONFIG_SMP | |
6459 | for_each_possible_cpu(i) { | |
6460 | struct rq *rq = cpu_rq(i); | |
6461 | unsigned long flags; | |
6462 | ||
6463 | if (list_empty(&rq->cfsb_csd_list)) | |
6464 | continue; | |
6465 | ||
6466 | local_irq_save(flags); | |
6467 | __cfsb_csd_unthrottle(rq); | |
6468 | local_irq_restore(flags); | |
6469 | } | |
6470 | #endif | |
029632fb PZ |
6471 | } |
6472 | ||
502ce005 | 6473 | /* |
97fb7a0a | 6474 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
6475 | * |
6476 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
6477 | * bits doesn't do much. | |
6478 | */ | |
6479 | ||
3b03706f | 6480 | /* cpu online callback */ |
0e59bdae KT |
6481 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
6482 | { | |
502ce005 | 6483 | struct task_group *tg; |
0e59bdae | 6484 | |
5cb9eaa3 | 6485 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
6486 | |
6487 | rcu_read_lock(); | |
6488 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
6489 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
6490 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
6491 | |
6492 | raw_spin_lock(&cfs_b->lock); | |
6493 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
6494 | raw_spin_unlock(&cfs_b->lock); | |
6495 | } | |
502ce005 | 6496 | rcu_read_unlock(); |
0e59bdae KT |
6497 | } |
6498 | ||
502ce005 | 6499 | /* cpu offline callback */ |
38dc3348 | 6500 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 6501 | { |
502ce005 PZ |
6502 | struct task_group *tg; |
6503 | ||
5cb9eaa3 | 6504 | lockdep_assert_rq_held(rq); |
502ce005 | 6505 | |
ebb83d84 HJ |
6506 | /* |
6507 | * The rq clock has already been updated in the | |
6508 | * set_rq_offline(), so we should skip updating | |
6509 | * the rq clock again in unthrottle_cfs_rq(). | |
6510 | */ | |
6511 | rq_clock_start_loop_update(rq); | |
6512 | ||
502ce005 PZ |
6513 | rcu_read_lock(); |
6514 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
6515 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 6516 | |
029632fb PZ |
6517 | if (!cfs_rq->runtime_enabled) |
6518 | continue; | |
6519 | ||
6520 | /* | |
6521 | * clock_task is not advancing so we just need to make sure | |
6522 | * there's some valid quota amount | |
6523 | */ | |
51f2176d | 6524 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 6525 | /* |
97fb7a0a | 6526 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
6527 | * in take_cpu_down(), so we prevent new cfs throttling here. |
6528 | */ | |
6529 | cfs_rq->runtime_enabled = 0; | |
6530 | ||
029632fb PZ |
6531 | if (cfs_rq_throttled(cfs_rq)) |
6532 | unthrottle_cfs_rq(cfs_rq); | |
6533 | } | |
502ce005 | 6534 | rcu_read_unlock(); |
ebb83d84 HJ |
6535 | |
6536 | rq_clock_stop_loop_update(rq); | |
029632fb PZ |
6537 | } |
6538 | ||
88c56cfe PA |
6539 | bool cfs_task_bw_constrained(struct task_struct *p) |
6540 | { | |
6541 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | |
6542 | ||
6543 | if (!cfs_bandwidth_used()) | |
6544 | return false; | |
6545 | ||
6546 | if (cfs_rq->runtime_enabled || | |
6547 | tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF) | |
6548 | return true; | |
6549 | ||
6550 | return false; | |
6551 | } | |
6552 | ||
6553 | #ifdef CONFIG_NO_HZ_FULL | |
6554 | /* called from pick_next_task_fair() */ | |
6555 | static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) | |
6556 | { | |
6557 | int cpu = cpu_of(rq); | |
6558 | ||
6559 | if (!sched_feat(HZ_BW) || !cfs_bandwidth_used()) | |
6560 | return; | |
6561 | ||
6562 | if (!tick_nohz_full_cpu(cpu)) | |
6563 | return; | |
6564 | ||
6565 | if (rq->nr_running != 1) | |
6566 | return; | |
6567 | ||
6568 | /* | |
6569 | * We know there is only one task runnable and we've just picked it. The | |
6570 | * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will | |
6571 | * be otherwise able to stop the tick. Just need to check if we are using | |
6572 | * bandwidth control. | |
6573 | */ | |
6574 | if (cfs_task_bw_constrained(p)) | |
6575 | tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED); | |
6576 | } | |
6577 | #endif | |
6578 | ||
029632fb | 6579 | #else /* CONFIG_CFS_BANDWIDTH */ |
f6783319 VG |
6580 | |
6581 | static inline bool cfs_bandwidth_used(void) | |
6582 | { | |
6583 | return false; | |
6584 | } | |
6585 | ||
9dbdb155 | 6586 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 6587 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 6588 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 6589 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 6590 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
6591 | |
6592 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
6593 | { | |
6594 | return 0; | |
6595 | } | |
64660c86 PT |
6596 | |
6597 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
6598 | { | |
6599 | return 0; | |
6600 | } | |
6601 | ||
6602 | static inline int throttled_lb_pair(struct task_group *tg, | |
6603 | int src_cpu, int dest_cpu) | |
6604 | { | |
6605 | return 0; | |
6606 | } | |
029632fb | 6607 | |
7aa55f2a | 6608 | #ifdef CONFIG_FAIR_GROUP_SCHED |
97efd283 | 6609 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {} |
029632fb | 6610 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
ab84d31e PT |
6611 | #endif |
6612 | ||
029632fb PZ |
6613 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
6614 | { | |
6615 | return NULL; | |
6616 | } | |
6617 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 6618 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 6619 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
88c56cfe PA |
6620 | #ifdef CONFIG_CGROUP_SCHED |
6621 | bool cfs_task_bw_constrained(struct task_struct *p) | |
6622 | { | |
6623 | return false; | |
6624 | } | |
6625 | #endif | |
029632fb PZ |
6626 | #endif /* CONFIG_CFS_BANDWIDTH */ |
6627 | ||
88c56cfe PA |
6628 | #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL) |
6629 | static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {} | |
6630 | #endif | |
6631 | ||
bf0f6f24 IM |
6632 | /************************************************** |
6633 | * CFS operations on tasks: | |
6634 | */ | |
6635 | ||
8f4d37ec PZ |
6636 | #ifdef CONFIG_SCHED_HRTICK |
6637 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
6638 | { | |
8f4d37ec | 6639 | struct sched_entity *se = &p->se; |
8f4d37ec | 6640 | |
9148a3a1 | 6641 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 6642 | |
8bf46a39 | 6643 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec | 6644 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; |
147f3efa | 6645 | u64 slice = se->slice; |
8f4d37ec PZ |
6646 | s64 delta = slice - ran; |
6647 | ||
6648 | if (delta < 0) { | |
65bcf072 | 6649 | if (task_current(rq, p)) |
8875125e | 6650 | resched_curr(rq); |
8f4d37ec PZ |
6651 | return; |
6652 | } | |
31656519 | 6653 | hrtick_start(rq, delta); |
8f4d37ec PZ |
6654 | } |
6655 | } | |
a4c2f00f PZ |
6656 | |
6657 | /* | |
6658 | * called from enqueue/dequeue and updates the hrtick when the | |
6659 | * current task is from our class and nr_running is low enough | |
6660 | * to matter. | |
6661 | */ | |
6662 | static void hrtick_update(struct rq *rq) | |
6663 | { | |
6664 | struct task_struct *curr = rq->curr; | |
6665 | ||
e0ee463c | 6666 | if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
6667 | return; |
6668 | ||
5e963f2b | 6669 | hrtick_start_fair(rq, curr); |
a4c2f00f | 6670 | } |
55e12e5e | 6671 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
6672 | static inline void |
6673 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
6674 | { | |
6675 | } | |
a4c2f00f PZ |
6676 | |
6677 | static inline void hrtick_update(struct rq *rq) | |
6678 | { | |
6679 | } | |
8f4d37ec PZ |
6680 | #endif |
6681 | ||
2802bf3c | 6682 | #ifdef CONFIG_SMP |
2802bf3c MR |
6683 | static inline bool cpu_overutilized(int cpu) |
6684 | { | |
be3a51e6 SH |
6685 | unsigned long rq_util_min, rq_util_max; |
6686 | ||
6687 | if (!sched_energy_enabled()) | |
6688 | return false; | |
6689 | ||
6690 | rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN); | |
6691 | rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX); | |
c56ab1b3 | 6692 | |
e5ed0550 | 6693 | /* Return true only if the utilization doesn't fit CPU's capacity */ |
c56ab1b3 | 6694 | return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu); |
2802bf3c MR |
6695 | } |
6696 | ||
d0f5d3ce | 6697 | /* |
902e786c | 6698 | * overutilized value make sense only if EAS is enabled |
d0f5d3ce | 6699 | */ |
4475cd8b | 6700 | static inline bool is_rd_overutilized(struct root_domain *rd) |
d0f5d3ce | 6701 | { |
902e786c | 6702 | return !sched_energy_enabled() || READ_ONCE(rd->overutilized); |
d0f5d3ce SH |
6703 | } |
6704 | ||
4475cd8b | 6705 | static inline void set_rd_overutilized(struct root_domain *rd, bool flag) |
2802bf3c | 6706 | { |
be3a51e6 SH |
6707 | if (!sched_energy_enabled()) |
6708 | return; | |
6709 | ||
4475cd8b IM |
6710 | WRITE_ONCE(rd->overutilized, flag); |
6711 | trace_sched_overutilized_tp(rd, flag); | |
be3a51e6 SH |
6712 | } |
6713 | ||
6714 | static inline void check_update_overutilized_status(struct rq *rq) | |
6715 | { | |
6716 | /* | |
6717 | * overutilized field is used for load balancing decisions only | |
6718 | * if energy aware scheduler is being used | |
6719 | */ | |
be3a51e6 | 6720 | |
d0f5d3ce | 6721 | if (!is_rd_overutilized(rq->rd) && cpu_overutilized(rq->cpu)) |
4475cd8b | 6722 | set_rd_overutilized(rq->rd, 1); |
2802bf3c MR |
6723 | } |
6724 | #else | |
be3a51e6 | 6725 | static inline void check_update_overutilized_status(struct rq *rq) { } |
2802bf3c MR |
6726 | #endif |
6727 | ||
323af6de VK |
6728 | /* Runqueue only has SCHED_IDLE tasks enqueued */ |
6729 | static int sched_idle_rq(struct rq *rq) | |
6730 | { | |
6731 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
6732 | rq->nr_running); | |
6733 | } | |
6734 | ||
afa70d94 | 6735 | #ifdef CONFIG_SMP |
323af6de VK |
6736 | static int sched_idle_cpu(int cpu) |
6737 | { | |
6738 | return sched_idle_rq(cpu_rq(cpu)); | |
6739 | } | |
afa70d94 | 6740 | #endif |
323af6de | 6741 | |
bf0f6f24 IM |
6742 | /* |
6743 | * The enqueue_task method is called before nr_running is | |
6744 | * increased. Here we update the fair scheduling stats and | |
6745 | * then put the task into the rbtree: | |
6746 | */ | |
ea87bb78 | 6747 | static void |
371fd7e7 | 6748 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
6749 | { |
6750 | struct cfs_rq *cfs_rq; | |
62fb1851 | 6751 | struct sched_entity *se = &p->se; |
43e9f7f2 | 6752 | int idle_h_nr_running = task_has_idle_policy(p); |
8e1ac429 | 6753 | int task_new = !(flags & ENQUEUE_WAKEUP); |
bf0f6f24 | 6754 | |
2539fc82 PB |
6755 | /* |
6756 | * The code below (indirectly) updates schedutil which looks at | |
6757 | * the cfs_rq utilization to select a frequency. | |
6758 | * Let's add the task's estimated utilization to the cfs_rq's | |
6759 | * estimated utilization, before we update schedutil. | |
6760 | */ | |
6761 | util_est_enqueue(&rq->cfs, p); | |
6762 | ||
8c34ab19 RW |
6763 | /* |
6764 | * If in_iowait is set, the code below may not trigger any cpufreq | |
6765 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
6766 | * passed. | |
6767 | */ | |
6768 | if (p->in_iowait) | |
674e7541 | 6769 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 6770 | |
bf0f6f24 | 6771 | for_each_sched_entity(se) { |
62fb1851 | 6772 | if (se->on_rq) |
bf0f6f24 IM |
6773 | break; |
6774 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 6775 | enqueue_entity(cfs_rq, se, flags); |
85dac906 | 6776 | |
953bfcd1 | 6777 | cfs_rq->h_nr_running++; |
43e9f7f2 | 6778 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 6779 | |
30400039 JD |
6780 | if (cfs_rq_is_idle(cfs_rq)) |
6781 | idle_h_nr_running = 1; | |
6782 | ||
6d4d2246 VG |
6783 | /* end evaluation on encountering a throttled cfs_rq */ |
6784 | if (cfs_rq_throttled(cfs_rq)) | |
6785 | goto enqueue_throttle; | |
6786 | ||
88ec22d3 | 6787 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 6788 | } |
8f4d37ec | 6789 | |
2069dd75 | 6790 | for_each_sched_entity(se) { |
0f317143 | 6791 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 6792 | |
88c0616e | 6793 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 6794 | se_update_runnable(se); |
1ea6c46a | 6795 | update_cfs_group(se); |
6d4d2246 VG |
6796 | |
6797 | cfs_rq->h_nr_running++; | |
6798 | cfs_rq->idle_h_nr_running += idle_h_nr_running; | |
5ab297ba | 6799 | |
30400039 JD |
6800 | if (cfs_rq_is_idle(cfs_rq)) |
6801 | idle_h_nr_running = 1; | |
6802 | ||
5ab297ba VG |
6803 | /* end evaluation on encountering a throttled cfs_rq */ |
6804 | if (cfs_rq_throttled(cfs_rq)) | |
6805 | goto enqueue_throttle; | |
2069dd75 PZ |
6806 | } |
6807 | ||
7d148be6 VG |
6808 | /* At this point se is NULL and we are at root level*/ |
6809 | add_nr_running(rq, 1); | |
2802bf3c | 6810 | |
7d148be6 VG |
6811 | /* |
6812 | * Since new tasks are assigned an initial util_avg equal to | |
6813 | * half of the spare capacity of their CPU, tiny tasks have the | |
6814 | * ability to cross the overutilized threshold, which will | |
6815 | * result in the load balancer ruining all the task placement | |
6816 | * done by EAS. As a way to mitigate that effect, do not account | |
6817 | * for the first enqueue operation of new tasks during the | |
6818 | * overutilized flag detection. | |
6819 | * | |
6820 | * A better way of solving this problem would be to wait for | |
6821 | * the PELT signals of tasks to converge before taking them | |
6822 | * into account, but that is not straightforward to implement, | |
6823 | * and the following generally works well enough in practice. | |
6824 | */ | |
8e1ac429 | 6825 | if (!task_new) |
be3a51e6 | 6826 | check_update_overutilized_status(rq); |
cd126afe | 6827 | |
7d148be6 | 6828 | enqueue_throttle: |
5d299eab PZ |
6829 | assert_list_leaf_cfs_rq(rq); |
6830 | ||
a4c2f00f | 6831 | hrtick_update(rq); |
bf0f6f24 IM |
6832 | } |
6833 | ||
2f36825b VP |
6834 | static void set_next_buddy(struct sched_entity *se); |
6835 | ||
bf0f6f24 IM |
6836 | /* |
6837 | * The dequeue_task method is called before nr_running is | |
6838 | * decreased. We remove the task from the rbtree and | |
6839 | * update the fair scheduling stats: | |
6840 | */ | |
371fd7e7 | 6841 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
6842 | { |
6843 | struct cfs_rq *cfs_rq; | |
62fb1851 | 6844 | struct sched_entity *se = &p->se; |
2f36825b | 6845 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 6846 | int idle_h_nr_running = task_has_idle_policy(p); |
323af6de | 6847 | bool was_sched_idle = sched_idle_rq(rq); |
bf0f6f24 | 6848 | |
8c1f560c XY |
6849 | util_est_dequeue(&rq->cfs, p); |
6850 | ||
bf0f6f24 IM |
6851 | for_each_sched_entity(se) { |
6852 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 6853 | dequeue_entity(cfs_rq, se, flags); |
85dac906 | 6854 | |
953bfcd1 | 6855 | cfs_rq->h_nr_running--; |
43e9f7f2 | 6856 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 6857 | |
30400039 JD |
6858 | if (cfs_rq_is_idle(cfs_rq)) |
6859 | idle_h_nr_running = 1; | |
6860 | ||
6d4d2246 VG |
6861 | /* end evaluation on encountering a throttled cfs_rq */ |
6862 | if (cfs_rq_throttled(cfs_rq)) | |
6863 | goto dequeue_throttle; | |
6864 | ||
bf0f6f24 | 6865 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 6866 | if (cfs_rq->load.weight) { |
754bd598 KK |
6867 | /* Avoid re-evaluating load for this entity: */ |
6868 | se = parent_entity(se); | |
2f36825b VP |
6869 | /* |
6870 | * Bias pick_next to pick a task from this cfs_rq, as | |
6871 | * p is sleeping when it is within its sched_slice. | |
6872 | */ | |
754bd598 KK |
6873 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
6874 | set_next_buddy(se); | |
bf0f6f24 | 6875 | break; |
2f36825b | 6876 | } |
371fd7e7 | 6877 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 6878 | } |
8f4d37ec | 6879 | |
2069dd75 | 6880 | for_each_sched_entity(se) { |
0f317143 | 6881 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 6882 | |
88c0616e | 6883 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 6884 | se_update_runnable(se); |
1ea6c46a | 6885 | update_cfs_group(se); |
6d4d2246 VG |
6886 | |
6887 | cfs_rq->h_nr_running--; | |
6888 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; | |
5ab297ba | 6889 | |
30400039 JD |
6890 | if (cfs_rq_is_idle(cfs_rq)) |
6891 | idle_h_nr_running = 1; | |
6892 | ||
5ab297ba VG |
6893 | /* end evaluation on encountering a throttled cfs_rq */ |
6894 | if (cfs_rq_throttled(cfs_rq)) | |
6895 | goto dequeue_throttle; | |
6896 | ||
2069dd75 PZ |
6897 | } |
6898 | ||
423d02e1 PW |
6899 | /* At this point se is NULL and we are at root level*/ |
6900 | sub_nr_running(rq, 1); | |
cd126afe | 6901 | |
323af6de VK |
6902 | /* balance early to pull high priority tasks */ |
6903 | if (unlikely(!was_sched_idle && sched_idle_rq(rq))) | |
6904 | rq->next_balance = jiffies; | |
6905 | ||
423d02e1 | 6906 | dequeue_throttle: |
8c1f560c | 6907 | util_est_update(&rq->cfs, p, task_sleep); |
a4c2f00f | 6908 | hrtick_update(rq); |
bf0f6f24 IM |
6909 | } |
6910 | ||
e7693a36 | 6911 | #ifdef CONFIG_SMP |
10e2f1ac | 6912 | |
d72cf624 | 6913 | /* Working cpumask for: sched_balance_rq(), sched_balance_newidle(). */ |
18c31c97 BH |
6914 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
6915 | static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask); | |
f8858d96 | 6916 | static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask); |
10e2f1ac | 6917 | |
9fd81dd5 | 6918 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
6919 | |
6920 | static struct { | |
6921 | cpumask_var_t idle_cpus_mask; | |
6922 | atomic_t nr_cpus; | |
f643ea22 | 6923 | int has_blocked; /* Idle CPUS has blocked load */ |
7fd7a9e0 | 6924 | int needs_update; /* Newly idle CPUs need their next_balance collated */ |
e022e0d3 | 6925 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 6926 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
6927 | } nohz ____cacheline_aligned; |
6928 | ||
9fd81dd5 | 6929 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 6930 | |
b0fb1eb4 VG |
6931 | static unsigned long cpu_load(struct rq *rq) |
6932 | { | |
6933 | return cfs_rq_load_avg(&rq->cfs); | |
6934 | } | |
6935 | ||
3318544b VG |
6936 | /* |
6937 | * cpu_load_without - compute CPU load without any contributions from *p | |
6938 | * @cpu: the CPU which load is requested | |
6939 | * @p: the task which load should be discounted | |
6940 | * | |
6941 | * The load of a CPU is defined by the load of tasks currently enqueued on that | |
6942 | * CPU as well as tasks which are currently sleeping after an execution on that | |
6943 | * CPU. | |
6944 | * | |
6945 | * This method returns the load of the specified CPU by discounting the load of | |
6946 | * the specified task, whenever the task is currently contributing to the CPU | |
6947 | * load. | |
6948 | */ | |
6949 | static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) | |
6950 | { | |
6951 | struct cfs_rq *cfs_rq; | |
6952 | unsigned int load; | |
6953 | ||
6954 | /* Task has no contribution or is new */ | |
6955 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
6956 | return cpu_load(rq); | |
6957 | ||
6958 | cfs_rq = &rq->cfs; | |
6959 | load = READ_ONCE(cfs_rq->avg.load_avg); | |
6960 | ||
6961 | /* Discount task's util from CPU's util */ | |
6962 | lsub_positive(&load, task_h_load(p)); | |
6963 | ||
6964 | return load; | |
6965 | } | |
6966 | ||
9f683953 VG |
6967 | static unsigned long cpu_runnable(struct rq *rq) |
6968 | { | |
6969 | return cfs_rq_runnable_avg(&rq->cfs); | |
6970 | } | |
6971 | ||
070f5e86 VG |
6972 | static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) |
6973 | { | |
6974 | struct cfs_rq *cfs_rq; | |
6975 | unsigned int runnable; | |
6976 | ||
6977 | /* Task has no contribution or is new */ | |
6978 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
6979 | return cpu_runnable(rq); | |
6980 | ||
6981 | cfs_rq = &rq->cfs; | |
6982 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
6983 | ||
6984 | /* Discount task's runnable from CPU's runnable */ | |
6985 | lsub_positive(&runnable, p->se.avg.runnable_avg); | |
6986 | ||
6987 | return runnable; | |
6988 | } | |
6989 | ||
ced549fa | 6990 | static unsigned long capacity_of(int cpu) |
029632fb | 6991 | { |
ced549fa | 6992 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
6993 | } |
6994 | ||
c58d25f3 PZ |
6995 | static void record_wakee(struct task_struct *p) |
6996 | { | |
6997 | /* | |
6998 | * Only decay a single time; tasks that have less then 1 wakeup per | |
6999 | * jiffy will not have built up many flips. | |
7000 | */ | |
7001 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
7002 | current->wakee_flips >>= 1; | |
7003 | current->wakee_flip_decay_ts = jiffies; | |
7004 | } | |
7005 | ||
7006 | if (current->last_wakee != p) { | |
7007 | current->last_wakee = p; | |
7008 | current->wakee_flips++; | |
7009 | } | |
7010 | } | |
7011 | ||
63b0e9ed MG |
7012 | /* |
7013 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 7014 | * |
63b0e9ed | 7015 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
7016 | * at a frequency roughly N times higher than one of its wakees. |
7017 | * | |
7018 | * In order to determine whether we should let the load spread vs consolidating | |
7019 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
7020 | * partner, and a factor of lls_size higher frequency in the other. | |
7021 | * | |
7022 | * With both conditions met, we can be relatively sure that the relationship is | |
7023 | * non-monogamous, with partner count exceeding socket size. | |
7024 | * | |
7025 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
7026 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
7027 | * socket size. | |
63b0e9ed | 7028 | */ |
62470419 MW |
7029 | static int wake_wide(struct task_struct *p) |
7030 | { | |
63b0e9ed MG |
7031 | unsigned int master = current->wakee_flips; |
7032 | unsigned int slave = p->wakee_flips; | |
17c891ab | 7033 | int factor = __this_cpu_read(sd_llc_size); |
62470419 | 7034 | |
63b0e9ed MG |
7035 | if (master < slave) |
7036 | swap(master, slave); | |
7037 | if (slave < factor || master < slave * factor) | |
7038 | return 0; | |
7039 | return 1; | |
62470419 MW |
7040 | } |
7041 | ||
90001d67 | 7042 | /* |
d153b153 PZ |
7043 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
7044 | * soonest. For the purpose of speed we only consider the waking and previous | |
7045 | * CPU. | |
90001d67 | 7046 | * |
7332dec0 MG |
7047 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
7048 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
7049 | * |
7050 | * wake_affine_weight() - considers the weight to reflect the average | |
7051 | * scheduling latency of the CPUs. This seems to work | |
7052 | * for the overloaded case. | |
90001d67 | 7053 | */ |
3b76c4a3 | 7054 | static int |
89a55f56 | 7055 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 7056 | { |
7332dec0 MG |
7057 | /* |
7058 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
7059 | * context. Only allow the move if cache is shared. Otherwise an | |
7060 | * interrupt intensive workload could force all tasks onto one | |
7061 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
7062 | * |
7063 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
7064 | * There is no guarantee that the cache hot data from an interrupt | |
7065 | * is more important than cache hot data on the prev_cpu and from | |
7066 | * a cpufreq perspective, it's better to have higher utilisation | |
7067 | * on one CPU. | |
7332dec0 | 7068 | */ |
943d355d RJ |
7069 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
7070 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 7071 | |
d153b153 | 7072 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 7073 | return this_cpu; |
90001d67 | 7074 | |
d8fcb81f JL |
7075 | if (available_idle_cpu(prev_cpu)) |
7076 | return prev_cpu; | |
7077 | ||
3b76c4a3 | 7078 | return nr_cpumask_bits; |
90001d67 PZ |
7079 | } |
7080 | ||
3b76c4a3 | 7081 | static int |
f2cdd9cc PZ |
7082 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
7083 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 7084 | { |
90001d67 PZ |
7085 | s64 this_eff_load, prev_eff_load; |
7086 | unsigned long task_load; | |
7087 | ||
11f10e54 | 7088 | this_eff_load = cpu_load(cpu_rq(this_cpu)); |
90001d67 | 7089 | |
90001d67 PZ |
7090 | if (sync) { |
7091 | unsigned long current_load = task_h_load(current); | |
7092 | ||
f2cdd9cc | 7093 | if (current_load > this_eff_load) |
3b76c4a3 | 7094 | return this_cpu; |
90001d67 | 7095 | |
f2cdd9cc | 7096 | this_eff_load -= current_load; |
90001d67 PZ |
7097 | } |
7098 | ||
90001d67 PZ |
7099 | task_load = task_h_load(p); |
7100 | ||
f2cdd9cc PZ |
7101 | this_eff_load += task_load; |
7102 | if (sched_feat(WA_BIAS)) | |
7103 | this_eff_load *= 100; | |
7104 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 7105 | |
11f10e54 | 7106 | prev_eff_load = cpu_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
7107 | prev_eff_load -= task_load; |
7108 | if (sched_feat(WA_BIAS)) | |
7109 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
7110 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 7111 | |
082f764a MG |
7112 | /* |
7113 | * If sync, adjust the weight of prev_eff_load such that if | |
7114 | * prev_eff == this_eff that select_idle_sibling() will consider | |
7115 | * stacking the wakee on top of the waker if no other CPU is | |
7116 | * idle. | |
7117 | */ | |
7118 | if (sync) | |
7119 | prev_eff_load += 1; | |
7120 | ||
7121 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
7122 | } |
7123 | ||
772bd008 | 7124 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 7125 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 7126 | { |
3b76c4a3 | 7127 | int target = nr_cpumask_bits; |
098fb9db | 7128 | |
89a55f56 | 7129 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 7130 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 7131 | |
3b76c4a3 MG |
7132 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
7133 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 7134 | |
ceeadb83 | 7135 | schedstat_inc(p->stats.nr_wakeups_affine_attempts); |
39afe5d6 | 7136 | if (target != this_cpu) |
3b76c4a3 | 7137 | return prev_cpu; |
098fb9db | 7138 | |
3b76c4a3 | 7139 | schedstat_inc(sd->ttwu_move_affine); |
ceeadb83 | 7140 | schedstat_inc(p->stats.nr_wakeups_affine); |
3b76c4a3 | 7141 | return target; |
098fb9db IM |
7142 | } |
7143 | ||
aaee1203 | 7144 | static struct sched_group * |
a88b1708 | 7145 | sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); |
aaee1203 PZ |
7146 | |
7147 | /* | |
646ebaf5 | 7148 | * sched_balance_find_dst_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
7149 | */ |
7150 | static int | |
646ebaf5 | 7151 | sched_balance_find_dst_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
7152 | { |
7153 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
7154 | unsigned int min_exit_latency = UINT_MAX; |
7155 | u64 latest_idle_timestamp = 0; | |
7156 | int least_loaded_cpu = this_cpu; | |
17346452 | 7157 | int shallowest_idle_cpu = -1; |
aaee1203 PZ |
7158 | int i; |
7159 | ||
eaecf41f MR |
7160 | /* Check if we have any choice: */ |
7161 | if (group->group_weight == 1) | |
ae4df9d6 | 7162 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 7163 | |
aaee1203 | 7164 | /* Traverse only the allowed CPUs */ |
3bd37062 | 7165 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
97886d9d AL |
7166 | struct rq *rq = cpu_rq(i); |
7167 | ||
7168 | if (!sched_core_cookie_match(rq, p)) | |
7169 | continue; | |
7170 | ||
17346452 VK |
7171 | if (sched_idle_cpu(i)) |
7172 | return i; | |
7173 | ||
943d355d | 7174 | if (available_idle_cpu(i)) { |
83a0a96a NP |
7175 | struct cpuidle_state *idle = idle_get_state(rq); |
7176 | if (idle && idle->exit_latency < min_exit_latency) { | |
7177 | /* | |
7178 | * We give priority to a CPU whose idle state | |
7179 | * has the smallest exit latency irrespective | |
7180 | * of any idle timestamp. | |
7181 | */ | |
7182 | min_exit_latency = idle->exit_latency; | |
7183 | latest_idle_timestamp = rq->idle_stamp; | |
7184 | shallowest_idle_cpu = i; | |
7185 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
7186 | rq->idle_stamp > latest_idle_timestamp) { | |
7187 | /* | |
7188 | * If equal or no active idle state, then | |
7189 | * the most recently idled CPU might have | |
7190 | * a warmer cache. | |
7191 | */ | |
7192 | latest_idle_timestamp = rq->idle_stamp; | |
7193 | shallowest_idle_cpu = i; | |
7194 | } | |
17346452 | 7195 | } else if (shallowest_idle_cpu == -1) { |
11f10e54 | 7196 | load = cpu_load(cpu_rq(i)); |
18cec7e0 | 7197 | if (load < min_load) { |
83a0a96a NP |
7198 | min_load = load; |
7199 | least_loaded_cpu = i; | |
7200 | } | |
e7693a36 GH |
7201 | } |
7202 | } | |
7203 | ||
17346452 | 7204 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 7205 | } |
e7693a36 | 7206 | |
686d148c | 7207 | static inline int sched_balance_find_dst_cpu(struct sched_domain *sd, struct task_struct *p, |
18bd1b4b BJ |
7208 | int cpu, int prev_cpu, int sd_flag) |
7209 | { | |
93f50f90 | 7210 | int new_cpu = cpu; |
18bd1b4b | 7211 | |
3bd37062 | 7212 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
7213 | return prev_cpu; |
7214 | ||
c976a862 | 7215 | /* |
57abff06 | 7216 | * We need task's util for cpu_util_without, sync it up to |
c469933e | 7217 | * prev_cpu's last_update_time. |
c976a862 VK |
7218 | */ |
7219 | if (!(sd_flag & SD_BALANCE_FORK)) | |
7220 | sync_entity_load_avg(&p->se); | |
7221 | ||
18bd1b4b BJ |
7222 | while (sd) { |
7223 | struct sched_group *group; | |
7224 | struct sched_domain *tmp; | |
7225 | int weight; | |
7226 | ||
7227 | if (!(sd->flags & sd_flag)) { | |
7228 | sd = sd->child; | |
7229 | continue; | |
7230 | } | |
7231 | ||
a88b1708 | 7232 | group = sched_balance_find_dst_group(sd, p, cpu); |
18bd1b4b BJ |
7233 | if (!group) { |
7234 | sd = sd->child; | |
7235 | continue; | |
7236 | } | |
7237 | ||
646ebaf5 | 7238 | new_cpu = sched_balance_find_dst_group_cpu(group, p, cpu); |
e90381ea | 7239 | if (new_cpu == cpu) { |
97fb7a0a | 7240 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
7241 | sd = sd->child; |
7242 | continue; | |
7243 | } | |
7244 | ||
97fb7a0a | 7245 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
7246 | cpu = new_cpu; |
7247 | weight = sd->span_weight; | |
7248 | sd = NULL; | |
7249 | for_each_domain(cpu, tmp) { | |
7250 | if (weight <= tmp->span_weight) | |
7251 | break; | |
7252 | if (tmp->flags & sd_flag) | |
7253 | sd = tmp; | |
7254 | } | |
18bd1b4b BJ |
7255 | } |
7256 | ||
7257 | return new_cpu; | |
7258 | } | |
7259 | ||
97886d9d | 7260 | static inline int __select_idle_cpu(int cpu, struct task_struct *p) |
9fe1f127 | 7261 | { |
97886d9d AL |
7262 | if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) && |
7263 | sched_cpu_cookie_match(cpu_rq(cpu), p)) | |
9fe1f127 MG |
7264 | return cpu; |
7265 | ||
7266 | return -1; | |
7267 | } | |
7268 | ||
10e2f1ac | 7269 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 7270 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 7271 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
7272 | |
7273 | static inline void set_idle_cores(int cpu, int val) | |
7274 | { | |
7275 | struct sched_domain_shared *sds; | |
7276 | ||
7277 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
7278 | if (sds) | |
7279 | WRITE_ONCE(sds->has_idle_cores, val); | |
7280 | } | |
7281 | ||
398ba2b0 | 7282 | static inline bool test_idle_cores(int cpu) |
10e2f1ac PZ |
7283 | { |
7284 | struct sched_domain_shared *sds; | |
7285 | ||
c722f35b RR |
7286 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
7287 | if (sds) | |
7288 | return READ_ONCE(sds->has_idle_cores); | |
10e2f1ac | 7289 | |
398ba2b0 | 7290 | return false; |
10e2f1ac PZ |
7291 | } |
7292 | ||
7293 | /* | |
7294 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
7295 | * information in sd_llc_shared->has_idle_cores. | |
7296 | * | |
7297 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
7298 | * state should be fairly cheap. | |
7299 | */ | |
1b568f0a | 7300 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
7301 | { |
7302 | int core = cpu_of(rq); | |
7303 | int cpu; | |
7304 | ||
7305 | rcu_read_lock(); | |
398ba2b0 | 7306 | if (test_idle_cores(core)) |
10e2f1ac PZ |
7307 | goto unlock; |
7308 | ||
7309 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
7310 | if (cpu == core) | |
7311 | continue; | |
7312 | ||
943d355d | 7313 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
7314 | goto unlock; |
7315 | } | |
7316 | ||
7317 | set_idle_cores(core, 1); | |
7318 | unlock: | |
7319 | rcu_read_unlock(); | |
7320 | } | |
7321 | ||
7322 | /* | |
7323 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
7324 | * there are no idle cores left in the system; tracked through | |
7325 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
7326 | */ | |
9fe1f127 | 7327 | static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) |
10e2f1ac | 7328 | { |
9fe1f127 MG |
7329 | bool idle = true; |
7330 | int cpu; | |
10e2f1ac | 7331 | |
9fe1f127 MG |
7332 | for_each_cpu(cpu, cpu_smt_mask(core)) { |
7333 | if (!available_idle_cpu(cpu)) { | |
7334 | idle = false; | |
7335 | if (*idle_cpu == -1) { | |
23d04d8c | 7336 | if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) { |
9fe1f127 MG |
7337 | *idle_cpu = cpu; |
7338 | break; | |
7339 | } | |
7340 | continue; | |
bec2860a | 7341 | } |
9fe1f127 | 7342 | break; |
10e2f1ac | 7343 | } |
23d04d8c | 7344 | if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus)) |
9fe1f127 | 7345 | *idle_cpu = cpu; |
10e2f1ac PZ |
7346 | } |
7347 | ||
9fe1f127 MG |
7348 | if (idle) |
7349 | return core; | |
10e2f1ac | 7350 | |
9fe1f127 | 7351 | cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); |
10e2f1ac PZ |
7352 | return -1; |
7353 | } | |
7354 | ||
c722f35b RR |
7355 | /* |
7356 | * Scan the local SMT mask for idle CPUs. | |
7357 | */ | |
8aeaffef | 7358 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) |
c722f35b RR |
7359 | { |
7360 | int cpu; | |
7361 | ||
3e6efe87 | 7362 | for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) { |
b9bae704 AW |
7363 | if (cpu == target) |
7364 | continue; | |
8aeaffef KN |
7365 | /* |
7366 | * Check if the CPU is in the LLC scheduling domain of @target. | |
7367 | * Due to isolcpus, there is no guarantee that all the siblings are in the domain. | |
7368 | */ | |
7369 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
7370 | continue; | |
c722f35b RR |
7371 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) |
7372 | return cpu; | |
7373 | } | |
7374 | ||
7375 | return -1; | |
7376 | } | |
7377 | ||
10e2f1ac PZ |
7378 | #else /* CONFIG_SCHED_SMT */ |
7379 | ||
9fe1f127 | 7380 | static inline void set_idle_cores(int cpu, int val) |
10e2f1ac | 7381 | { |
9fe1f127 MG |
7382 | } |
7383 | ||
398ba2b0 | 7384 | static inline bool test_idle_cores(int cpu) |
9fe1f127 | 7385 | { |
398ba2b0 | 7386 | return false; |
9fe1f127 MG |
7387 | } |
7388 | ||
7389 | static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) | |
7390 | { | |
97886d9d | 7391 | return __select_idle_cpu(core, p); |
10e2f1ac PZ |
7392 | } |
7393 | ||
8aeaffef | 7394 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) |
c722f35b RR |
7395 | { |
7396 | return -1; | |
7397 | } | |
7398 | ||
10e2f1ac PZ |
7399 | #endif /* CONFIG_SCHED_SMT */ |
7400 | ||
7401 | /* | |
7402 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
7403 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
7404 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 7405 | */ |
c722f35b | 7406 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target) |
10e2f1ac | 7407 | { |
ec4fc801 | 7408 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
9fe1f127 | 7409 | int i, cpu, idle_cpu = -1, nr = INT_MAX; |
70fb5ccf | 7410 | struct sched_domain_shared *sd_share; |
10e2f1ac | 7411 | |
bae4ec13 MG |
7412 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
7413 | ||
70fb5ccf CY |
7414 | if (sched_feat(SIS_UTIL)) { |
7415 | sd_share = rcu_dereference(per_cpu(sd_llc_shared, target)); | |
7416 | if (sd_share) { | |
7417 | /* because !--nr is the condition to stop scan */ | |
7418 | nr = READ_ONCE(sd_share->nr_idle_scan) + 1; | |
7419 | /* overloaded LLC is unlikely to have idle cpu/core */ | |
7420 | if (nr == 1) | |
7421 | return -1; | |
7422 | } | |
7423 | } | |
7424 | ||
8881e163 BS |
7425 | if (static_branch_unlikely(&sched_cluster_active)) { |
7426 | struct sched_group *sg = sd->groups; | |
7427 | ||
7428 | if (sg->flags & SD_CLUSTER) { | |
7429 | for_each_cpu_wrap(cpu, sched_group_span(sg), target + 1) { | |
7430 | if (!cpumask_test_cpu(cpu, cpus)) | |
7431 | continue; | |
7432 | ||
7433 | if (has_idle_core) { | |
7434 | i = select_idle_core(p, cpu, cpus, &idle_cpu); | |
7435 | if ((unsigned int)i < nr_cpumask_bits) | |
7436 | return i; | |
7437 | } else { | |
7438 | if (--nr <= 0) | |
7439 | return -1; | |
7440 | idle_cpu = __select_idle_cpu(cpu, p); | |
7441 | if ((unsigned int)idle_cpu < nr_cpumask_bits) | |
7442 | return idle_cpu; | |
7443 | } | |
7444 | } | |
7445 | cpumask_andnot(cpus, cpus, sched_group_span(sg)); | |
7446 | } | |
7447 | } | |
7448 | ||
56498cfb | 7449 | for_each_cpu_wrap(cpu, cpus, target + 1) { |
c722f35b | 7450 | if (has_idle_core) { |
9fe1f127 MG |
7451 | i = select_idle_core(p, cpu, cpus, &idle_cpu); |
7452 | if ((unsigned int)i < nr_cpumask_bits) | |
7453 | return i; | |
7454 | ||
7455 | } else { | |
8881e163 | 7456 | if (--nr <= 0) |
9fe1f127 | 7457 | return -1; |
97886d9d | 7458 | idle_cpu = __select_idle_cpu(cpu, p); |
9fe1f127 MG |
7459 | if ((unsigned int)idle_cpu < nr_cpumask_bits) |
7460 | break; | |
7461 | } | |
10e2f1ac PZ |
7462 | } |
7463 | ||
c722f35b | 7464 | if (has_idle_core) |
02dbb724 | 7465 | set_idle_cores(target, false); |
9fe1f127 | 7466 | |
9fe1f127 | 7467 | return idle_cpu; |
10e2f1ac PZ |
7468 | } |
7469 | ||
b7a33161 MR |
7470 | /* |
7471 | * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which | |
7472 | * the task fits. If no CPU is big enough, but there are idle ones, try to | |
7473 | * maximize capacity. | |
7474 | */ | |
7475 | static int | |
7476 | select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) | |
7477 | { | |
b759caa1 | 7478 | unsigned long task_util, util_min, util_max, best_cap = 0; |
e5ed0550 | 7479 | int fits, best_fits = 0; |
b7a33161 MR |
7480 | int cpu, best_cpu = -1; |
7481 | struct cpumask *cpus; | |
7482 | ||
ec4fc801 | 7483 | cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
b7a33161 MR |
7484 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
7485 | ||
b759caa1 QY |
7486 | task_util = task_util_est(p); |
7487 | util_min = uclamp_eff_value(p, UCLAMP_MIN); | |
7488 | util_max = uclamp_eff_value(p, UCLAMP_MAX); | |
b4c9c9f1 | 7489 | |
7ee7642c | 7490 | for_each_cpu_wrap(cpu, cpus, target) { |
b7a33161 MR |
7491 | unsigned long cpu_cap = capacity_of(cpu); |
7492 | ||
7493 | if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) | |
7494 | continue; | |
e5ed0550 VG |
7495 | |
7496 | fits = util_fits_cpu(task_util, util_min, util_max, cpu); | |
7497 | ||
7498 | /* This CPU fits with all requirements */ | |
7499 | if (fits > 0) | |
b7a33161 | 7500 | return cpu; |
e5ed0550 VG |
7501 | /* |
7502 | * Only the min performance hint (i.e. uclamp_min) doesn't fit. | |
7503 | * Look for the CPU with best capacity. | |
7504 | */ | |
7505 | else if (fits < 0) | |
f1f8d0a2 | 7506 | cpu_cap = get_actual_cpu_capacity(cpu); |
b7a33161 | 7507 | |
e5ed0550 VG |
7508 | /* |
7509 | * First, select CPU which fits better (-1 being better than 0). | |
7510 | * Then, select the one with best capacity at same level. | |
7511 | */ | |
7512 | if ((fits < best_fits) || | |
7513 | ((fits == best_fits) && (cpu_cap > best_cap))) { | |
b7a33161 MR |
7514 | best_cap = cpu_cap; |
7515 | best_cpu = cpu; | |
e5ed0550 | 7516 | best_fits = fits; |
b7a33161 MR |
7517 | } |
7518 | } | |
7519 | ||
7520 | return best_cpu; | |
7521 | } | |
7522 | ||
a2e7f03e QY |
7523 | static inline bool asym_fits_cpu(unsigned long util, |
7524 | unsigned long util_min, | |
7525 | unsigned long util_max, | |
7526 | int cpu) | |
b4c9c9f1 | 7527 | { |
740cf8a7 | 7528 | if (sched_asym_cpucap_active()) |
e5ed0550 VG |
7529 | /* |
7530 | * Return true only if the cpu fully fits the task requirements | |
7531 | * which include the utilization and the performance hints. | |
7532 | */ | |
7533 | return (util_fits_cpu(util, util_min, util_max, cpu) > 0); | |
b4c9c9f1 VG |
7534 | |
7535 | return true; | |
7536 | } | |
7537 | ||
10e2f1ac PZ |
7538 | /* |
7539 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 7540 | */ |
772bd008 | 7541 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 7542 | { |
c722f35b | 7543 | bool has_idle_core = false; |
99bd5e2f | 7544 | struct sched_domain *sd; |
a2e7f03e | 7545 | unsigned long task_util, util_min, util_max; |
22165f61 | 7546 | int i, recent_used_cpu, prev_aff = -1; |
a50bde51 | 7547 | |
b7a33161 | 7548 | /* |
b4c9c9f1 | 7549 | * On asymmetric system, update task utilization because we will check |
b9e6e286 | 7550 | * that the task fits with CPU's capacity. |
b7a33161 | 7551 | */ |
740cf8a7 | 7552 | if (sched_asym_cpucap_active()) { |
b4c9c9f1 | 7553 | sync_entity_load_avg(&p->se); |
a2e7f03e QY |
7554 | task_util = task_util_est(p); |
7555 | util_min = uclamp_eff_value(p, UCLAMP_MIN); | |
7556 | util_max = uclamp_eff_value(p, UCLAMP_MAX); | |
b7a33161 MR |
7557 | } |
7558 | ||
9099a147 | 7559 | /* |
ec4fc801 | 7560 | * per-cpu select_rq_mask usage |
9099a147 PZ |
7561 | */ |
7562 | lockdep_assert_irqs_disabled(); | |
7563 | ||
b4c9c9f1 | 7564 | if ((available_idle_cpu(target) || sched_idle_cpu(target)) && |
a2e7f03e | 7565 | asym_fits_cpu(task_util, util_min, util_max, target)) |
e0a79f52 | 7566 | return target; |
99bd5e2f SS |
7567 | |
7568 | /* | |
97fb7a0a | 7569 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 7570 | */ |
3c29e651 | 7571 | if (prev != target && cpus_share_cache(prev, target) && |
b4c9c9f1 | 7572 | (available_idle_cpu(prev) || sched_idle_cpu(prev)) && |
8881e163 BS |
7573 | asym_fits_cpu(task_util, util_min, util_max, prev)) { |
7574 | ||
7575 | if (!static_branch_unlikely(&sched_cluster_active) || | |
7576 | cpus_share_resources(prev, target)) | |
7577 | return prev; | |
22165f61 YY |
7578 | |
7579 | prev_aff = prev; | |
8881e163 | 7580 | } |
a50bde51 | 7581 | |
52262ee5 MG |
7582 | /* |
7583 | * Allow a per-cpu kthread to stack with the wakee if the | |
7584 | * kworker thread and the tasks previous CPUs are the same. | |
7585 | * The assumption is that the wakee queued work for the | |
7586 | * per-cpu kthread that is now complete and the wakeup is | |
7587 | * essentially a sync wakeup. An obvious example of this | |
7588 | * pattern is IO completions. | |
7589 | */ | |
7590 | if (is_per_cpu_kthread(current) && | |
8b4e74cc | 7591 | in_task() && |
52262ee5 | 7592 | prev == smp_processor_id() && |
014ba44e | 7593 | this_rq()->nr_running <= 1 && |
a2e7f03e | 7594 | asym_fits_cpu(task_util, util_min, util_max, prev)) { |
52262ee5 MG |
7595 | return prev; |
7596 | } | |
7597 | ||
97fb7a0a | 7598 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd | 7599 | recent_used_cpu = p->recent_used_cpu; |
89aafd67 | 7600 | p->recent_used_cpu = prev; |
32e839dd MG |
7601 | if (recent_used_cpu != prev && |
7602 | recent_used_cpu != target && | |
7603 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 7604 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
ae2ad293 | 7605 | cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) && |
a2e7f03e | 7606 | asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) { |
8881e163 BS |
7607 | |
7608 | if (!static_branch_unlikely(&sched_cluster_active) || | |
7609 | cpus_share_resources(recent_used_cpu, target)) | |
7610 | return recent_used_cpu; | |
7611 | ||
22165f61 YY |
7612 | } else { |
7613 | recent_used_cpu = -1; | |
32e839dd MG |
7614 | } |
7615 | ||
b4c9c9f1 VG |
7616 | /* |
7617 | * For asymmetric CPU capacity systems, our domain of interest is | |
7618 | * sd_asym_cpucapacity rather than sd_llc. | |
7619 | */ | |
740cf8a7 | 7620 | if (sched_asym_cpucap_active()) { |
b4c9c9f1 VG |
7621 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); |
7622 | /* | |
7623 | * On an asymmetric CPU capacity system where an exclusive | |
7624 | * cpuset defines a symmetric island (i.e. one unique | |
7625 | * capacity_orig value through the cpuset), the key will be set | |
7626 | * but the CPUs within that cpuset will not have a domain with | |
7627 | * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric | |
7628 | * capacity path. | |
7629 | */ | |
7630 | if (sd) { | |
7631 | i = select_idle_capacity(p, sd, target); | |
7632 | return ((unsigned)i < nr_cpumask_bits) ? i : target; | |
7633 | } | |
7634 | } | |
7635 | ||
518cd623 | 7636 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
7637 | if (!sd) |
7638 | return target; | |
772bd008 | 7639 | |
c722f35b | 7640 | if (sched_smt_active()) { |
398ba2b0 | 7641 | has_idle_core = test_idle_cores(target); |
c722f35b RR |
7642 | |
7643 | if (!has_idle_core && cpus_share_cache(prev, target)) { | |
8aeaffef | 7644 | i = select_idle_smt(p, sd, prev); |
c722f35b RR |
7645 | if ((unsigned int)i < nr_cpumask_bits) |
7646 | return i; | |
7647 | } | |
7648 | } | |
7649 | ||
7650 | i = select_idle_cpu(p, sd, has_idle_core, target); | |
10e2f1ac PZ |
7651 | if ((unsigned)i < nr_cpumask_bits) |
7652 | return i; | |
7653 | ||
22165f61 YY |
7654 | /* |
7655 | * For cluster machines which have lower sharing cache like L2 or | |
7656 | * LLC Tag, we tend to find an idle CPU in the target's cluster | |
7657 | * first. But prev_cpu or recent_used_cpu may also be a good candidate, | |
7658 | * use them if possible when no idle CPU found in select_idle_cpu(). | |
7659 | */ | |
7660 | if ((unsigned int)prev_aff < nr_cpumask_bits) | |
7661 | return prev_aff; | |
7662 | if ((unsigned int)recent_used_cpu < nr_cpumask_bits) | |
7663 | return recent_used_cpu; | |
7664 | ||
a50bde51 PZ |
7665 | return target; |
7666 | } | |
231678b7 | 7667 | |
3eb6d6ec DE |
7668 | /** |
7669 | * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks. | |
7670 | * @cpu: the CPU to get the utilization for | |
7671 | * @p: task for which the CPU utilization should be predicted or NULL | |
7672 | * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL | |
7d0583cf | 7673 | * @boost: 1 to enable boosting, otherwise 0 |
3eb6d6ec DE |
7674 | * |
7675 | * The unit of the return value must be the same as the one of CPU capacity | |
7676 | * so that CPU utilization can be compared with CPU capacity. | |
7677 | * | |
7678 | * CPU utilization is the sum of running time of runnable tasks plus the | |
7679 | * recent utilization of currently non-runnable tasks on that CPU. | |
7680 | * It represents the amount of CPU capacity currently used by CFS tasks in | |
7681 | * the range [0..max CPU capacity] with max CPU capacity being the CPU | |
7682 | * capacity at f_max. | |
7683 | * | |
7684 | * The estimated CPU utilization is defined as the maximum between CPU | |
7685 | * utilization and sum of the estimated utilization of the currently | |
7686 | * runnable tasks on that CPU. It preserves a utilization "snapshot" of | |
7687 | * previously-executed tasks, which helps better deduce how busy a CPU will | |
7688 | * be when a long-sleeping task wakes up. The contribution to CPU utilization | |
7689 | * of such a task would be significantly decayed at this point of time. | |
7690 | * | |
7d0583cf DE |
7691 | * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization). |
7692 | * CPU contention for CFS tasks can be detected by CPU runnable > CPU | |
7693 | * utilization. Boosting is implemented in cpu_util() so that internal | |
7694 | * users (e.g. EAS) can use it next to external users (e.g. schedutil), | |
7695 | * latter via cpu_util_cfs_boost(). | |
7696 | * | |
3eb6d6ec DE |
7697 | * CPU utilization can be higher than the current CPU capacity |
7698 | * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because | |
7699 | * of rounding errors as well as task migrations or wakeups of new tasks. | |
7700 | * CPU utilization has to be capped to fit into the [0..max CPU capacity] | |
7701 | * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%) | |
7702 | * could be seen as over-utilized even though CPU1 has 20% of spare CPU | |
7703 | * capacity. CPU utilization is allowed to overshoot current CPU capacity | |
7704 | * though since this is useful for predicting the CPU capacity required | |
7705 | * after task migrations (scheduler-driven DVFS). | |
7706 | * | |
7d0583cf | 7707 | * Return: (Boosted) (estimated) utilization for the specified CPU. |
390031e4 | 7708 | */ |
7d0583cf DE |
7709 | static unsigned long |
7710 | cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost) | |
390031e4 QP |
7711 | { |
7712 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
4e3c7d33 | 7713 | unsigned long util = READ_ONCE(cfs_rq->avg.util_avg); |
7d0583cf DE |
7714 | unsigned long runnable; |
7715 | ||
7716 | if (boost) { | |
7717 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
7718 | util = max(util, runnable); | |
7719 | } | |
390031e4 QP |
7720 | |
7721 | /* | |
4e3c7d33 DE |
7722 | * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its |
7723 | * contribution. If @p migrates from another CPU to @cpu add its | |
7724 | * contribution. In all the other cases @cpu is not impacted by the | |
7725 | * migration so its util_avg is already correct. | |
390031e4 | 7726 | */ |
3eb6d6ec | 7727 | if (p && task_cpu(p) == cpu && dst_cpu != cpu) |
736cc6b3 | 7728 | lsub_positive(&util, task_util(p)); |
3eb6d6ec | 7729 | else if (p && task_cpu(p) != cpu && dst_cpu == cpu) |
390031e4 QP |
7730 | util += task_util(p); |
7731 | ||
7732 | if (sched_feat(UTIL_EST)) { | |
4e3c7d33 DE |
7733 | unsigned long util_est; |
7734 | ||
11137d38 | 7735 | util_est = READ_ONCE(cfs_rq->avg.util_est); |
390031e4 QP |
7736 | |
7737 | /* | |
4e3c7d33 | 7738 | * During wake-up @p isn't enqueued yet and doesn't contribute |
11137d38 | 7739 | * to any cpu_rq(cpu)->cfs.avg.util_est. |
4e3c7d33 DE |
7740 | * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p |
7741 | * has been enqueued. | |
7742 | * | |
7743 | * During exec (@dst_cpu = -1) @p is enqueued and does | |
11137d38 | 7744 | * contribute to cpu_rq(cpu)->cfs.util_est. |
4e3c7d33 DE |
7745 | * Remove it to "simulate" cpu_util without @p's contribution. |
7746 | * | |
7747 | * Despite the task_on_rq_queued(@p) check there is still a | |
7748 | * small window for a possible race when an exec | |
7749 | * select_task_rq_fair() races with LB's detach_task(). | |
7750 | * | |
7751 | * detach_task() | |
7752 | * deactivate_task() | |
7753 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
7754 | * -------------------------------- A | |
7755 | * dequeue_task() \ | |
7756 | * dequeue_task_fair() + Race Time | |
7757 | * util_est_dequeue() / | |
7758 | * -------------------------------- B | |
7759 | * | |
7760 | * The additional check "current == p" is required to further | |
7761 | * reduce the race window. | |
390031e4 QP |
7762 | */ |
7763 | if (dst_cpu == cpu) | |
7764 | util_est += _task_util_est(p); | |
3eb6d6ec | 7765 | else if (p && unlikely(task_on_rq_queued(p) || current == p)) |
4e3c7d33 | 7766 | lsub_positive(&util_est, _task_util_est(p)); |
390031e4 QP |
7767 | |
7768 | util = max(util, util_est); | |
7769 | } | |
7770 | ||
7bc26384 | 7771 | return min(util, arch_scale_cpu_capacity(cpu)); |
390031e4 QP |
7772 | } |
7773 | ||
3eb6d6ec DE |
7774 | unsigned long cpu_util_cfs(int cpu) |
7775 | { | |
7d0583cf DE |
7776 | return cpu_util(cpu, NULL, -1, 0); |
7777 | } | |
7778 | ||
7779 | unsigned long cpu_util_cfs_boost(int cpu) | |
7780 | { | |
7781 | return cpu_util(cpu, NULL, -1, 1); | |
3eb6d6ec DE |
7782 | } |
7783 | ||
4e3c7d33 DE |
7784 | /* |
7785 | * cpu_util_without: compute cpu utilization without any contributions from *p | |
7786 | * @cpu: the CPU which utilization is requested | |
7787 | * @p: the task which utilization should be discounted | |
7788 | * | |
7789 | * The utilization of a CPU is defined by the utilization of tasks currently | |
7790 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
7791 | * execution on that CPU. | |
7792 | * | |
7793 | * This method returns the utilization of the specified CPU by discounting the | |
7794 | * utilization of the specified task, whenever the task is currently | |
7795 | * contributing to the CPU utilization. | |
7796 | */ | |
7797 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) | |
7798 | { | |
7799 | /* Task has no contribution or is new */ | |
7800 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
3eb6d6ec | 7801 | p = NULL; |
4e3c7d33 | 7802 | |
7d0583cf | 7803 | return cpu_util(cpu, p, -1, 0); |
4e3c7d33 DE |
7804 | } |
7805 | ||
390031e4 | 7806 | /* |
3e8c6c9a VD |
7807 | * energy_env - Utilization landscape for energy estimation. |
7808 | * @task_busy_time: Utilization contribution by the task for which we test the | |
7809 | * placement. Given by eenv_task_busy_time(). | |
7810 | * @pd_busy_time: Utilization of the whole perf domain without the task | |
7811 | * contribution. Given by eenv_pd_busy_time(). | |
7812 | * @cpu_cap: Maximum CPU capacity for the perf domain. | |
7813 | * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap). | |
390031e4 | 7814 | */ |
3e8c6c9a VD |
7815 | struct energy_env { |
7816 | unsigned long task_busy_time; | |
7817 | unsigned long pd_busy_time; | |
7818 | unsigned long cpu_cap; | |
7819 | unsigned long pd_cap; | |
7820 | }; | |
7821 | ||
7822 | /* | |
7823 | * Compute the task busy time for compute_energy(). This time cannot be | |
7824 | * injected directly into effective_cpu_util() because of the IRQ scaling. | |
7825 | * The latter only makes sense with the most recent CPUs where the task has | |
7826 | * run. | |
7827 | */ | |
7828 | static inline void eenv_task_busy_time(struct energy_env *eenv, | |
7829 | struct task_struct *p, int prev_cpu) | |
390031e4 | 7830 | { |
3e8c6c9a VD |
7831 | unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu); |
7832 | unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu)); | |
7833 | ||
7834 | if (unlikely(irq >= max_cap)) | |
7835 | busy_time = max_cap; | |
7836 | else | |
7837 | busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap); | |
7838 | ||
7839 | eenv->task_busy_time = busy_time; | |
7840 | } | |
7841 | ||
7842 | /* | |
7843 | * Compute the perf_domain (PD) busy time for compute_energy(). Based on the | |
7844 | * utilization for each @pd_cpus, it however doesn't take into account | |
7845 | * clamping since the ratio (utilization / cpu_capacity) is already enough to | |
7846 | * scale the EM reported power consumption at the (eventually clamped) | |
7847 | * cpu_capacity. | |
7848 | * | |
7849 | * The contribution of the task @p for which we want to estimate the | |
3eb6d6ec | 7850 | * energy cost is removed (by cpu_util()) and must be calculated |
3e8c6c9a VD |
7851 | * separately (see eenv_task_busy_time). This ensures: |
7852 | * | |
7853 | * - A stable PD utilization, no matter which CPU of that PD we want to place | |
7854 | * the task on. | |
7855 | * | |
7856 | * - A fair comparison between CPUs as the task contribution (task_util()) | |
7857 | * will always be the same no matter which CPU utilization we rely on | |
7858 | * (util_avg or util_est). | |
7859 | * | |
7860 | * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't | |
7861 | * exceed @eenv->pd_cap. | |
7862 | */ | |
7863 | static inline void eenv_pd_busy_time(struct energy_env *eenv, | |
7864 | struct cpumask *pd_cpus, | |
7865 | struct task_struct *p) | |
7866 | { | |
7867 | unsigned long busy_time = 0; | |
390031e4 QP |
7868 | int cpu; |
7869 | ||
3e8c6c9a | 7870 | for_each_cpu(cpu, pd_cpus) { |
7d0583cf | 7871 | unsigned long util = cpu_util(cpu, p, -1, 0); |
489f1645 | 7872 | |
9c0b4bb7 | 7873 | busy_time += effective_cpu_util(cpu, util, NULL, NULL); |
3e8c6c9a | 7874 | } |
0372e1cf | 7875 | |
3e8c6c9a VD |
7876 | eenv->pd_busy_time = min(eenv->pd_cap, busy_time); |
7877 | } | |
af24bde8 | 7878 | |
3e8c6c9a VD |
7879 | /* |
7880 | * Compute the maximum utilization for compute_energy() when the task @p | |
7881 | * is placed on the cpu @dst_cpu. | |
7882 | * | |
7883 | * Returns the maximum utilization among @eenv->cpus. This utilization can't | |
7884 | * exceed @eenv->cpu_cap. | |
7885 | */ | |
7886 | static inline unsigned long | |
7887 | eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus, | |
7888 | struct task_struct *p, int dst_cpu) | |
7889 | { | |
7890 | unsigned long max_util = 0; | |
7891 | int cpu; | |
489f1645 | 7892 | |
3e8c6c9a VD |
7893 | for_each_cpu(cpu, pd_cpus) { |
7894 | struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL; | |
7d0583cf | 7895 | unsigned long util = cpu_util(cpu, p, dst_cpu, 1); |
9c0b4bb7 | 7896 | unsigned long eff_util, min, max; |
af24bde8 | 7897 | |
390031e4 | 7898 | /* |
eb92692b QP |
7899 | * Performance domain frequency: utilization clamping |
7900 | * must be considered since it affects the selection | |
7901 | * of the performance domain frequency. | |
7cb7fb5b CL |
7902 | * NOTE: in case RT tasks are running, by default the min |
7903 | * utilization can be max OPP. | |
390031e4 | 7904 | */ |
9c0b4bb7 VG |
7905 | eff_util = effective_cpu_util(cpu, util, &min, &max); |
7906 | ||
7907 | /* Task's uclamp can modify min and max value */ | |
7908 | if (tsk && uclamp_is_used()) { | |
7909 | min = max(min, uclamp_eff_value(p, UCLAMP_MIN)); | |
7910 | ||
7911 | /* | |
7912 | * If there is no active max uclamp constraint, | |
7913 | * directly use task's one, otherwise keep max. | |
7914 | */ | |
7915 | if (uclamp_rq_is_idle(cpu_rq(cpu))) | |
7916 | max = uclamp_eff_value(p, UCLAMP_MAX); | |
7917 | else | |
7918 | max = max(max, uclamp_eff_value(p, UCLAMP_MAX)); | |
7919 | } | |
7920 | ||
7921 | eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max); | |
a707df30 | 7922 | max_util = max(max_util, eff_util); |
390031e4 QP |
7923 | } |
7924 | ||
3e8c6c9a VD |
7925 | return min(max_util, eenv->cpu_cap); |
7926 | } | |
7927 | ||
7928 | /* | |
7929 | * compute_energy(): Use the Energy Model to estimate the energy that @pd would | |
7930 | * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task | |
7931 | * contribution is ignored. | |
7932 | */ | |
7933 | static inline unsigned long | |
7934 | compute_energy(struct energy_env *eenv, struct perf_domain *pd, | |
7935 | struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu) | |
7936 | { | |
7937 | unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu); | |
7938 | unsigned long busy_time = eenv->pd_busy_time; | |
15874a3d | 7939 | unsigned long energy; |
3e8c6c9a VD |
7940 | |
7941 | if (dst_cpu >= 0) | |
7942 | busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time); | |
7943 | ||
15874a3d QY |
7944 | energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap); |
7945 | ||
7946 | trace_sched_compute_energy_tp(p, dst_cpu, energy, max_util, busy_time); | |
7947 | ||
7948 | return energy; | |
390031e4 QP |
7949 | } |
7950 | ||
732cd75b QP |
7951 | /* |
7952 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
7953 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
7954 | * spare capacity in each performance domain and uses it as a potential | |
7955 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
7956 | * out which of the CPU candidates is the most energy-efficient. | |
7957 | * | |
7958 | * The rationale for this heuristic is as follows. In a performance domain, | |
7959 | * all the most energy efficient CPU candidates (according to the Energy | |
7960 | * Model) are those for which we'll request a low frequency. When there are | |
7961 | * several CPUs for which the frequency request will be the same, we don't | |
7962 | * have enough data to break the tie between them, because the Energy Model | |
7963 | * only includes active power costs. With this model, if we assume that | |
7964 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
7965 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
7966 | * the best candidates of the performance domain. | |
7967 | * | |
7968 | * In practice, it could be preferable from an energy standpoint to pack | |
7969 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
7970 | * but that could also hurt our chances to go cluster idle, and we have no | |
7971 | * ways to tell with the current Energy Model if this is actually a good | |
7972 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
7973 | * cluster-packing, and spreading inside a cluster. That should at least be | |
7974 | * a good thing for latency, and this is consistent with the idea that most | |
7975 | * of the energy savings of EAS come from the asymmetry of the system, and | |
7976 | * not so much from breaking the tie between identical CPUs. That's also the | |
7977 | * reason why EAS is enabled in the topology code only for systems where | |
7978 | * SD_ASYM_CPUCAPACITY is set. | |
7979 | * | |
7980 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
7981 | * they don't have any useful utilization data yet and it's not possible to | |
7982 | * forecast their impact on energy consumption. Consequently, they will be | |
686d148c | 7983 | * placed by sched_balance_find_dst_cpu() on the least loaded CPU, which might turn out |
732cd75b QP |
7984 | * to be energy-inefficient in some use-cases. The alternative would be to |
7985 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
7986 | * their util_avg from the parent task, but those heuristics could hurt | |
7987 | * other use-cases too. So, until someone finds a better way to solve this, | |
7988 | * let's keep things simple by re-using the existing slow path. | |
7989 | */ | |
732cd75b QP |
7990 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) |
7991 | { | |
9b340131 | 7992 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
eb92692b | 7993 | unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; |
24422603 QY |
7994 | unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0; |
7995 | unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024; | |
3e8c6c9a | 7996 | struct root_domain *rd = this_rq()->rd; |
b812fc97 | 7997 | int cpu, best_energy_cpu, target = -1; |
e5ed0550 | 7998 | int prev_fits = -1, best_fits = -1; |
f1f8d0a2 VG |
7999 | unsigned long best_actual_cap = 0; |
8000 | unsigned long prev_actual_cap = 0; | |
732cd75b | 8001 | struct sched_domain *sd; |
eb92692b | 8002 | struct perf_domain *pd; |
3e8c6c9a | 8003 | struct energy_env eenv; |
732cd75b QP |
8004 | |
8005 | rcu_read_lock(); | |
8006 | pd = rcu_dereference(rd->pd); | |
902e786c | 8007 | if (!pd) |
619e090c | 8008 | goto unlock; |
732cd75b QP |
8009 | |
8010 | /* | |
8011 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
8012 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
8013 | */ | |
8014 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
8015 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
8016 | sd = sd->parent; | |
8017 | if (!sd) | |
619e090c PG |
8018 | goto unlock; |
8019 | ||
8020 | target = prev_cpu; | |
732cd75b QP |
8021 | |
8022 | sync_entity_load_avg(&p->se); | |
23c9519d | 8023 | if (!task_util_est(p) && p_util_min == 0) |
732cd75b QP |
8024 | goto unlock; |
8025 | ||
3e8c6c9a VD |
8026 | eenv_task_busy_time(&eenv, p, prev_cpu); |
8027 | ||
732cd75b | 8028 | for (; pd; pd = pd->next) { |
e26fd28d | 8029 | unsigned long util_min = p_util_min, util_max = p_util_max; |
f1f8d0a2 | 8030 | unsigned long cpu_cap, cpu_actual_cap, util; |
6b00a401 | 8031 | long prev_spare_cap = -1, max_spare_cap = -1; |
24422603 | 8032 | unsigned long rq_util_min, rq_util_max; |
6b00a401 | 8033 | unsigned long cur_delta, base_energy; |
732cd75b | 8034 | int max_spare_cap_cpu = -1; |
e5ed0550 | 8035 | int fits, max_fits = -1; |
732cd75b | 8036 | |
9b340131 DE |
8037 | cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask); |
8038 | ||
3e8c6c9a VD |
8039 | if (cpumask_empty(cpus)) |
8040 | continue; | |
8041 | ||
f1f8d0a2 | 8042 | /* Account external pressure for the energy estimation */ |
3e8c6c9a | 8043 | cpu = cpumask_first(cpus); |
f1f8d0a2 | 8044 | cpu_actual_cap = get_actual_cpu_capacity(cpu); |
3e8c6c9a | 8045 | |
f1f8d0a2 | 8046 | eenv.cpu_cap = cpu_actual_cap; |
3e8c6c9a VD |
8047 | eenv.pd_cap = 0; |
8048 | ||
8049 | for_each_cpu(cpu, cpus) { | |
e26fd28d QY |
8050 | struct rq *rq = cpu_rq(cpu); |
8051 | ||
f1f8d0a2 | 8052 | eenv.pd_cap += cpu_actual_cap; |
3e8c6c9a VD |
8053 | |
8054 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
8055 | continue; | |
8056 | ||
3bd37062 | 8057 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
8058 | continue; |
8059 | ||
7d0583cf | 8060 | util = cpu_util(cpu, p, cpu, 0); |
732cd75b | 8061 | cpu_cap = capacity_of(cpu); |
1d42509e VS |
8062 | |
8063 | /* | |
8064 | * Skip CPUs that cannot satisfy the capacity request. | |
8065 | * IOW, placing the task there would make the CPU | |
8066 | * overutilized. Take uclamp into account to see how | |
8067 | * much capacity we can get out of the CPU; this is | |
a5418be9 | 8068 | * aligned with sched_cpu_util(). |
1d42509e | 8069 | */ |
e26fd28d QY |
8070 | if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) { |
8071 | /* | |
8072 | * Open code uclamp_rq_util_with() except for | |
b9e6e286 | 8073 | * the clamp() part. I.e.: apply max aggregation |
e26fd28d QY |
8074 | * only. util_fits_cpu() logic requires to |
8075 | * operate on non clamped util but must use the | |
8076 | * max-aggregated uclamp_{min, max}. | |
8077 | */ | |
8078 | rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN); | |
8079 | rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX); | |
8080 | ||
8081 | util_min = max(rq_util_min, p_util_min); | |
8082 | util_max = max(rq_util_max, p_util_max); | |
24422603 | 8083 | } |
e5ed0550 VG |
8084 | |
8085 | fits = util_fits_cpu(util, util_min, util_max, cpu); | |
8086 | if (!fits) | |
732cd75b QP |
8087 | continue; |
8088 | ||
3e8c6c9a VD |
8089 | lsub_positive(&cpu_cap, util); |
8090 | ||
732cd75b | 8091 | if (cpu == prev_cpu) { |
8d4c97c1 | 8092 | /* Always use prev_cpu as a candidate. */ |
ad841e56 | 8093 | prev_spare_cap = cpu_cap; |
e5ed0550 VG |
8094 | prev_fits = fits; |
8095 | } else if ((fits > max_fits) || | |
6b00a401 | 8096 | ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) { |
8d4c97c1 PG |
8097 | /* |
8098 | * Find the CPU with the maximum spare capacity | |
ad841e56 PG |
8099 | * among the remaining CPUs in the performance |
8100 | * domain. | |
8d4c97c1 | 8101 | */ |
3e8c6c9a | 8102 | max_spare_cap = cpu_cap; |
732cd75b | 8103 | max_spare_cap_cpu = cpu; |
e5ed0550 | 8104 | max_fits = fits; |
732cd75b QP |
8105 | } |
8106 | } | |
8107 | ||
6b00a401 | 8108 | if (max_spare_cap_cpu < 0 && prev_spare_cap < 0) |
8d4c97c1 PG |
8109 | continue; |
8110 | ||
3e8c6c9a | 8111 | eenv_pd_busy_time(&eenv, cpus, p); |
8d4c97c1 | 8112 | /* Compute the 'base' energy of the pd, without @p */ |
b812fc97 | 8113 | base_energy = compute_energy(&eenv, pd, cpus, p, -1); |
8d4c97c1 PG |
8114 | |
8115 | /* Evaluate the energy impact of using prev_cpu. */ | |
6b00a401 | 8116 | if (prev_spare_cap > -1) { |
3e8c6c9a VD |
8117 | prev_delta = compute_energy(&eenv, pd, cpus, p, |
8118 | prev_cpu); | |
8119 | /* CPU utilization has changed */ | |
b812fc97 | 8120 | if (prev_delta < base_energy) |
619e090c | 8121 | goto unlock; |
b812fc97 | 8122 | prev_delta -= base_energy; |
f1f8d0a2 | 8123 | prev_actual_cap = cpu_actual_cap; |
8d4c97c1 PG |
8124 | best_delta = min(best_delta, prev_delta); |
8125 | } | |
8126 | ||
8127 | /* Evaluate the energy impact of using max_spare_cap_cpu. */ | |
ad841e56 | 8128 | if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) { |
e5ed0550 VG |
8129 | /* Current best energy cpu fits better */ |
8130 | if (max_fits < best_fits) | |
8131 | continue; | |
8132 | ||
8133 | /* | |
8134 | * Both don't fit performance hint (i.e. uclamp_min) | |
8135 | * but best energy cpu has better capacity. | |
8136 | */ | |
8137 | if ((max_fits < 0) && | |
f1f8d0a2 | 8138 | (cpu_actual_cap <= best_actual_cap)) |
e5ed0550 VG |
8139 | continue; |
8140 | ||
3e8c6c9a VD |
8141 | cur_delta = compute_energy(&eenv, pd, cpus, p, |
8142 | max_spare_cap_cpu); | |
8143 | /* CPU utilization has changed */ | |
b812fc97 | 8144 | if (cur_delta < base_energy) |
619e090c | 8145 | goto unlock; |
b812fc97 | 8146 | cur_delta -= base_energy; |
e5ed0550 VG |
8147 | |
8148 | /* | |
8149 | * Both fit for the task but best energy cpu has lower | |
8150 | * energy impact. | |
8151 | */ | |
8152 | if ((max_fits > 0) && (best_fits > 0) && | |
8153 | (cur_delta >= best_delta)) | |
8154 | continue; | |
8155 | ||
8156 | best_delta = cur_delta; | |
8157 | best_energy_cpu = max_spare_cap_cpu; | |
8158 | best_fits = max_fits; | |
f1f8d0a2 | 8159 | best_actual_cap = cpu_actual_cap; |
732cd75b QP |
8160 | } |
8161 | } | |
732cd75b QP |
8162 | rcu_read_unlock(); |
8163 | ||
e5ed0550 VG |
8164 | if ((best_fits > prev_fits) || |
8165 | ((best_fits > 0) && (best_delta < prev_delta)) || | |
f1f8d0a2 | 8166 | ((best_fits < 0) && (best_actual_cap > prev_actual_cap))) |
619e090c | 8167 | target = best_energy_cpu; |
732cd75b | 8168 | |
619e090c | 8169 | return target; |
732cd75b | 8170 | |
619e090c | 8171 | unlock: |
732cd75b QP |
8172 | rcu_read_unlock(); |
8173 | ||
619e090c | 8174 | return target; |
732cd75b QP |
8175 | } |
8176 | ||
aaee1203 | 8177 | /* |
de91b9cb | 8178 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
3aef1551 | 8179 | * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE, |
de91b9cb | 8180 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. |
aaee1203 | 8181 | * |
97fb7a0a IM |
8182 | * Balances load by selecting the idlest CPU in the idlest group, or under |
8183 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 8184 | * |
97fb7a0a | 8185 | * Returns the target CPU number. |
aaee1203 | 8186 | */ |
0017d735 | 8187 | static int |
3aef1551 | 8188 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags) |
aaee1203 | 8189 | { |
3aef1551 | 8190 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
f1d88b44 | 8191 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 8192 | int cpu = smp_processor_id(); |
63b0e9ed | 8193 | int new_cpu = prev_cpu; |
99bd5e2f | 8194 | int want_affine = 0; |
3aef1551 VS |
8195 | /* SD_flags and WF_flags share the first nibble */ |
8196 | int sd_flag = wake_flags & 0xF; | |
c88d5910 | 8197 | |
9099a147 PZ |
8198 | /* |
8199 | * required for stable ->cpus_allowed | |
8200 | */ | |
8201 | lockdep_assert_held(&p->pi_lock); | |
dc824eb8 | 8202 | if (wake_flags & WF_TTWU) { |
c58d25f3 | 8203 | record_wakee(p); |
732cd75b | 8204 | |
ab83f455 PO |
8205 | if ((wake_flags & WF_CURRENT_CPU) && |
8206 | cpumask_test_cpu(cpu, p->cpus_ptr)) | |
8207 | return cpu; | |
8208 | ||
902e786c | 8209 | if (!is_rd_overutilized(this_rq()->rd)) { |
732cd75b QP |
8210 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
8211 | if (new_cpu >= 0) | |
8212 | return new_cpu; | |
8213 | new_cpu = prev_cpu; | |
8214 | } | |
8215 | ||
00061968 | 8216 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 8217 | } |
aaee1203 | 8218 | |
dce840a0 | 8219 | rcu_read_lock(); |
aaee1203 | 8220 | for_each_domain(cpu, tmp) { |
fe3bcfe1 | 8221 | /* |
97fb7a0a | 8222 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 8223 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 8224 | */ |
99bd5e2f SS |
8225 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
8226 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
8227 | if (cpu != prev_cpu) |
8228 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
8229 | ||
8230 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 8231 | break; |
f03542a7 | 8232 | } |
29cd8bae | 8233 | |
2917406c BS |
8234 | /* |
8235 | * Usually only true for WF_EXEC and WF_FORK, as sched_domains | |
8236 | * usually do not have SD_BALANCE_WAKE set. That means wakeup | |
8237 | * will usually go to the fast path. | |
8238 | */ | |
f03542a7 | 8239 | if (tmp->flags & sd_flag) |
29cd8bae | 8240 | sd = tmp; |
63b0e9ed MG |
8241 | else if (!want_affine) |
8242 | break; | |
29cd8bae PZ |
8243 | } |
8244 | ||
f1d88b44 VK |
8245 | if (unlikely(sd)) { |
8246 | /* Slow path */ | |
686d148c | 8247 | new_cpu = sched_balance_find_dst_cpu(sd, p, cpu, prev_cpu, sd_flag); |
dc824eb8 | 8248 | } else if (wake_flags & WF_TTWU) { /* XXX always ? */ |
f1d88b44 | 8249 | /* Fast path */ |
f1d88b44 | 8250 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
e7693a36 | 8251 | } |
dce840a0 | 8252 | rcu_read_unlock(); |
e7693a36 | 8253 | |
c88d5910 | 8254 | return new_cpu; |
e7693a36 | 8255 | } |
0a74bef8 PT |
8256 | |
8257 | /* | |
97fb7a0a | 8258 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 8259 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 8260 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 8261 | */ |
3f9672ba | 8262 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 8263 | { |
e2f3e35f VD |
8264 | struct sched_entity *se = &p->se; |
8265 | ||
e1f078f5 | 8266 | if (!task_on_rq_migrating(p)) { |
e2f3e35f VD |
8267 | remove_entity_load_avg(se); |
8268 | ||
144d8487 | 8269 | /* |
e2f3e35f VD |
8270 | * Here, the task's PELT values have been updated according to |
8271 | * the current rq's clock. But if that clock hasn't been | |
8272 | * updated in a while, a substantial idle time will be missed, | |
8273 | * leading to an inflation after wake-up on the new rq. | |
8274 | * | |
8275 | * Estimate the missing time from the cfs_rq last_update_time | |
8276 | * and update sched_avg to improve the PELT continuity after | |
8277 | * migration. | |
144d8487 | 8278 | */ |
e2f3e35f | 8279 | migrate_se_pelt_lag(se); |
144d8487 | 8280 | } |
9d89c257 YD |
8281 | |
8282 | /* Tell new CPU we are migrated */ | |
e2f3e35f | 8283 | se->avg.last_update_time = 0; |
3944a927 | 8284 | |
3f9672ba | 8285 | update_scan_period(p, new_cpu); |
0a74bef8 | 8286 | } |
12695578 YD |
8287 | |
8288 | static void task_dead_fair(struct task_struct *p) | |
8289 | { | |
8290 | remove_entity_load_avg(&p->se); | |
8291 | } | |
6e2df058 | 8292 | |
22d56074 QY |
8293 | /* |
8294 | * Set the max capacity the task is allowed to run at for misfit detection. | |
8295 | */ | |
8296 | static void set_task_max_allowed_capacity(struct task_struct *p) | |
8297 | { | |
8298 | struct asym_cap_data *entry; | |
8299 | ||
8300 | if (!sched_asym_cpucap_active()) | |
8301 | return; | |
8302 | ||
8303 | rcu_read_lock(); | |
8304 | list_for_each_entry_rcu(entry, &asym_cap_list, link) { | |
8305 | cpumask_t *cpumask; | |
8306 | ||
8307 | cpumask = cpu_capacity_span(entry); | |
8308 | if (!cpumask_intersects(p->cpus_ptr, cpumask)) | |
8309 | continue; | |
8310 | ||
8311 | p->max_allowed_capacity = entry->capacity; | |
8312 | break; | |
8313 | } | |
8314 | rcu_read_unlock(); | |
8315 | } | |
8316 | ||
8317 | static void set_cpus_allowed_fair(struct task_struct *p, struct affinity_context *ctx) | |
8318 | { | |
8319 | set_cpus_allowed_common(p, ctx); | |
8320 | set_task_max_allowed_capacity(p); | |
8321 | } | |
8322 | ||
6e2df058 PZ |
8323 | static int |
8324 | balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
8325 | { | |
8326 | if (rq->nr_running) | |
8327 | return 1; | |
8328 | ||
7d058285 | 8329 | return sched_balance_newidle(rq, rf) != 0; |
6e2df058 | 8330 | } |
22d56074 QY |
8331 | #else |
8332 | static inline void set_task_max_allowed_capacity(struct task_struct *p) {} | |
e7693a36 GH |
8333 | #endif /* CONFIG_SMP */ |
8334 | ||
02479099 PZ |
8335 | static void set_next_buddy(struct sched_entity *se) |
8336 | { | |
c5ae366e DA |
8337 | for_each_sched_entity(se) { |
8338 | if (SCHED_WARN_ON(!se->on_rq)) | |
8339 | return; | |
30400039 JD |
8340 | if (se_is_idle(se)) |
8341 | return; | |
69c80f3e | 8342 | cfs_rq_of(se)->next = se; |
c5ae366e | 8343 | } |
02479099 PZ |
8344 | } |
8345 | ||
bf0f6f24 IM |
8346 | /* |
8347 | * Preempt the current task with a newly woken task if needed: | |
8348 | */ | |
82845683 | 8349 | static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
8350 | { |
8351 | struct task_struct *curr = rq->curr; | |
8651a86c | 8352 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 8353 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
30400039 | 8354 | int cse_is_idle, pse_is_idle; |
bf0f6f24 | 8355 | |
4ae7d5ce IM |
8356 | if (unlikely(se == pse)) |
8357 | return; | |
8358 | ||
5238cdd3 | 8359 | /* |
163122b7 | 8360 | * This is possible from callers such as attach_tasks(), in which we |
e23edc86 | 8361 | * unconditionally wakeup_preempt() after an enqueue (which may have |
5238cdd3 PT |
8362 | * lead to a throttle). This both saves work and prevents false |
8363 | * next-buddy nomination below. | |
8364 | */ | |
8365 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
8366 | return; | |
8367 | ||
5e963f2b | 8368 | if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) { |
3cb63d52 | 8369 | set_next_buddy(pse); |
2f36825b | 8370 | } |
57fdc26d | 8371 | |
aec0a514 BR |
8372 | /* |
8373 | * We can come here with TIF_NEED_RESCHED already set from new task | |
8374 | * wake up path. | |
5238cdd3 PT |
8375 | * |
8376 | * Note: this also catches the edge-case of curr being in a throttled | |
8377 | * group (e.g. via set_curr_task), since update_curr() (in the | |
8378 | * enqueue of curr) will have resulted in resched being set. This | |
8379 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
8380 | * below. | |
aec0a514 BR |
8381 | */ |
8382 | if (test_tsk_need_resched(curr)) | |
8383 | return; | |
8384 | ||
a2f5c9ab | 8385 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
8386 | if (unlikely(task_has_idle_policy(curr)) && |
8387 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
8388 | goto preempt; |
8389 | ||
91c234b4 | 8390 | /* |
a2f5c9ab DH |
8391 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
8392 | * is driven by the tick): | |
91c234b4 | 8393 | */ |
8ed92e51 | 8394 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 8395 | return; |
bf0f6f24 | 8396 | |
464b7527 | 8397 | find_matching_se(&se, &pse); |
09348d75 | 8398 | WARN_ON_ONCE(!pse); |
30400039 JD |
8399 | |
8400 | cse_is_idle = se_is_idle(se); | |
8401 | pse_is_idle = se_is_idle(pse); | |
8402 | ||
8403 | /* | |
8404 | * Preempt an idle group in favor of a non-idle group (and don't preempt | |
8405 | * in the inverse case). | |
8406 | */ | |
8407 | if (cse_is_idle && !pse_is_idle) | |
8408 | goto preempt; | |
8409 | if (cse_is_idle != pse_is_idle) | |
8410 | return; | |
8411 | ||
147f3efa PZ |
8412 | cfs_rq = cfs_rq_of(se); |
8413 | update_curr(cfs_rq); | |
8414 | ||
5e963f2b PZ |
8415 | /* |
8416 | * XXX pick_eevdf(cfs_rq) != se ? | |
8417 | */ | |
8418 | if (pick_eevdf(cfs_rq) == pse) | |
3a7e73a2 | 8419 | goto preempt; |
464b7527 | 8420 | |
3a7e73a2 | 8421 | return; |
a65ac745 | 8422 | |
3a7e73a2 | 8423 | preempt: |
8875125e | 8424 | resched_curr(rq); |
bf0f6f24 IM |
8425 | } |
8426 | ||
21f56ffe PZ |
8427 | #ifdef CONFIG_SMP |
8428 | static struct task_struct *pick_task_fair(struct rq *rq) | |
8429 | { | |
8430 | struct sched_entity *se; | |
8431 | struct cfs_rq *cfs_rq; | |
8432 | ||
8433 | again: | |
8434 | cfs_rq = &rq->cfs; | |
8435 | if (!cfs_rq->nr_running) | |
8436 | return NULL; | |
8437 | ||
8438 | do { | |
8439 | struct sched_entity *curr = cfs_rq->curr; | |
8440 | ||
8441 | /* When we pick for a remote RQ, we'll not have done put_prev_entity() */ | |
8442 | if (curr) { | |
8443 | if (curr->on_rq) | |
8444 | update_curr(cfs_rq); | |
8445 | else | |
8446 | curr = NULL; | |
8447 | ||
8448 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
8449 | goto again; | |
8450 | } | |
8451 | ||
4c456c9a | 8452 | se = pick_next_entity(cfs_rq); |
21f56ffe PZ |
8453 | cfs_rq = group_cfs_rq(se); |
8454 | } while (cfs_rq); | |
8455 | ||
8456 | return task_of(se); | |
8457 | } | |
8458 | #endif | |
8459 | ||
5d7d6056 | 8460 | struct task_struct * |
d8ac8971 | 8461 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
8462 | { |
8463 | struct cfs_rq *cfs_rq = &rq->cfs; | |
8464 | struct sched_entity *se; | |
678d5718 | 8465 | struct task_struct *p; |
37e117c0 | 8466 | int new_tasks; |
678d5718 | 8467 | |
6e83125c | 8468 | again: |
6e2df058 | 8469 | if (!sched_fair_runnable(rq)) |
38033c37 | 8470 | goto idle; |
678d5718 | 8471 | |
9674f5ca | 8472 | #ifdef CONFIG_FAIR_GROUP_SCHED |
67692435 | 8473 | if (!prev || prev->sched_class != &fair_sched_class) |
678d5718 PZ |
8474 | goto simple; |
8475 | ||
8476 | /* | |
8477 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
8478 | * likely that a next task is from the same cgroup as the current. | |
8479 | * | |
8480 | * Therefore attempt to avoid putting and setting the entire cgroup | |
8481 | * hierarchy, only change the part that actually changes. | |
8482 | */ | |
8483 | ||
8484 | do { | |
8485 | struct sched_entity *curr = cfs_rq->curr; | |
8486 | ||
8487 | /* | |
8488 | * Since we got here without doing put_prev_entity() we also | |
8489 | * have to consider cfs_rq->curr. If it is still a runnable | |
8490 | * entity, update_curr() will update its vruntime, otherwise | |
8491 | * forget we've ever seen it. | |
8492 | */ | |
54d27365 BS |
8493 | if (curr) { |
8494 | if (curr->on_rq) | |
8495 | update_curr(cfs_rq); | |
8496 | else | |
8497 | curr = NULL; | |
678d5718 | 8498 | |
54d27365 BS |
8499 | /* |
8500 | * This call to check_cfs_rq_runtime() will do the | |
8501 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 8502 | * Therefore the nr_running test will indeed |
54d27365 BS |
8503 | * be correct. |
8504 | */ | |
9674f5ca VK |
8505 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
8506 | cfs_rq = &rq->cfs; | |
8507 | ||
8508 | if (!cfs_rq->nr_running) | |
8509 | goto idle; | |
8510 | ||
54d27365 | 8511 | goto simple; |
9674f5ca | 8512 | } |
54d27365 | 8513 | } |
678d5718 | 8514 | |
4c456c9a | 8515 | se = pick_next_entity(cfs_rq); |
678d5718 PZ |
8516 | cfs_rq = group_cfs_rq(se); |
8517 | } while (cfs_rq); | |
8518 | ||
8519 | p = task_of(se); | |
8520 | ||
8521 | /* | |
8522 | * Since we haven't yet done put_prev_entity and if the selected task | |
8523 | * is a different task than we started out with, try and touch the | |
8524 | * least amount of cfs_rqs. | |
8525 | */ | |
8526 | if (prev != p) { | |
8527 | struct sched_entity *pse = &prev->se; | |
8528 | ||
8529 | while (!(cfs_rq = is_same_group(se, pse))) { | |
8530 | int se_depth = se->depth; | |
8531 | int pse_depth = pse->depth; | |
8532 | ||
8533 | if (se_depth <= pse_depth) { | |
8534 | put_prev_entity(cfs_rq_of(pse), pse); | |
8535 | pse = parent_entity(pse); | |
8536 | } | |
8537 | if (se_depth >= pse_depth) { | |
8538 | set_next_entity(cfs_rq_of(se), se); | |
8539 | se = parent_entity(se); | |
8540 | } | |
8541 | } | |
8542 | ||
8543 | put_prev_entity(cfs_rq, pse); | |
8544 | set_next_entity(cfs_rq, se); | |
8545 | } | |
8546 | ||
93824900 | 8547 | goto done; |
678d5718 | 8548 | simple: |
678d5718 | 8549 | #endif |
67692435 PZ |
8550 | if (prev) |
8551 | put_prev_task(rq, prev); | |
606dba2e | 8552 | |
bf0f6f24 | 8553 | do { |
4c456c9a | 8554 | se = pick_next_entity(cfs_rq); |
f4b6755f | 8555 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
8556 | cfs_rq = group_cfs_rq(se); |
8557 | } while (cfs_rq); | |
8558 | ||
8f4d37ec | 8559 | p = task_of(se); |
678d5718 | 8560 | |
13a453c2 | 8561 | done: __maybe_unused; |
93824900 UR |
8562 | #ifdef CONFIG_SMP |
8563 | /* | |
8564 | * Move the next running task to the front of | |
8565 | * the list, so our cfs_tasks list becomes MRU | |
8566 | * one. | |
8567 | */ | |
8568 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
8569 | #endif | |
8570 | ||
e0ee463c | 8571 | if (hrtick_enabled_fair(rq)) |
b39e66ea | 8572 | hrtick_start_fair(rq, p); |
8f4d37ec | 8573 | |
3b1baa64 | 8574 | update_misfit_status(p, rq); |
88c56cfe | 8575 | sched_fair_update_stop_tick(rq, p); |
3b1baa64 | 8576 | |
8f4d37ec | 8577 | return p; |
38033c37 PZ |
8578 | |
8579 | idle: | |
67692435 PZ |
8580 | if (!rf) |
8581 | return NULL; | |
8582 | ||
7d058285 | 8583 | new_tasks = sched_balance_newidle(rq, rf); |
46f69fa3 | 8584 | |
37e117c0 | 8585 | /* |
7d058285 | 8586 | * Because sched_balance_newidle() releases (and re-acquires) rq->lock, it is |
37e117c0 PZ |
8587 | * possible for any higher priority task to appear. In that case we |
8588 | * must re-start the pick_next_entity() loop. | |
8589 | */ | |
e4aa358b | 8590 | if (new_tasks < 0) |
37e117c0 PZ |
8591 | return RETRY_TASK; |
8592 | ||
e4aa358b | 8593 | if (new_tasks > 0) |
38033c37 | 8594 | goto again; |
38033c37 | 8595 | |
23127296 VG |
8596 | /* |
8597 | * rq is about to be idle, check if we need to update the | |
8598 | * lost_idle_time of clock_pelt | |
8599 | */ | |
8600 | update_idle_rq_clock_pelt(rq); | |
8601 | ||
38033c37 | 8602 | return NULL; |
bf0f6f24 IM |
8603 | } |
8604 | ||
98c2f700 PZ |
8605 | static struct task_struct *__pick_next_task_fair(struct rq *rq) |
8606 | { | |
8607 | return pick_next_task_fair(rq, NULL, NULL); | |
8608 | } | |
8609 | ||
bf0f6f24 IM |
8610 | /* |
8611 | * Account for a descheduled task: | |
8612 | */ | |
6e2df058 | 8613 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
8614 | { |
8615 | struct sched_entity *se = &prev->se; | |
8616 | struct cfs_rq *cfs_rq; | |
8617 | ||
8618 | for_each_sched_entity(se) { | |
8619 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 8620 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
8621 | } |
8622 | } | |
8623 | ||
ac53db59 RR |
8624 | /* |
8625 | * sched_yield() is very simple | |
ac53db59 RR |
8626 | */ |
8627 | static void yield_task_fair(struct rq *rq) | |
8628 | { | |
8629 | struct task_struct *curr = rq->curr; | |
8630 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
8631 | struct sched_entity *se = &curr->se; | |
8632 | ||
8633 | /* | |
8634 | * Are we the only task in the tree? | |
8635 | */ | |
8636 | if (unlikely(rq->nr_running == 1)) | |
8637 | return; | |
8638 | ||
8639 | clear_buddies(cfs_rq, se); | |
8640 | ||
5e963f2b PZ |
8641 | update_rq_clock(rq); |
8642 | /* | |
8643 | * Update run-time statistics of the 'current'. | |
8644 | */ | |
8645 | update_curr(cfs_rq); | |
8646 | /* | |
8647 | * Tell update_rq_clock() that we've just updated, | |
8648 | * so we don't do microscopic update in schedule() | |
8649 | * and double the fastpath cost. | |
8650 | */ | |
8651 | rq_clock_skip_update(rq); | |
ac53db59 | 8652 | |
5e963f2b | 8653 | se->deadline += calc_delta_fair(se->slice, se); |
ac53db59 RR |
8654 | } |
8655 | ||
0900acf2 | 8656 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) |
d95f4122 MG |
8657 | { |
8658 | struct sched_entity *se = &p->se; | |
8659 | ||
5238cdd3 PT |
8660 | /* throttled hierarchies are not runnable */ |
8661 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
8662 | return false; |
8663 | ||
b9e6e286 | 8664 | /* Tell the scheduler that we'd really like se to run next. */ |
d95f4122 MG |
8665 | set_next_buddy(se); |
8666 | ||
d95f4122 MG |
8667 | yield_task_fair(rq); |
8668 | ||
8669 | return true; | |
8670 | } | |
8671 | ||
681f3e68 | 8672 | #ifdef CONFIG_SMP |
bf0f6f24 | 8673 | /************************************************** |
e9c84cb8 PZ |
8674 | * Fair scheduling class load-balancing methods. |
8675 | * | |
8676 | * BASICS | |
8677 | * | |
8678 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 8679 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
8680 | * time to each task. This is expressed in the following equation: |
8681 | * | |
8682 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
8683 | * | |
97fb7a0a | 8684 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
8685 | * W_i,0 is defined as: |
8686 | * | |
8687 | * W_i,0 = \Sum_j w_i,j (2) | |
8688 | * | |
97fb7a0a | 8689 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 8690 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
8691 | * |
8692 | * The weight average is an exponential decay average of the instantaneous | |
8693 | * weight: | |
8694 | * | |
8695 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
8696 | * | |
97fb7a0a | 8697 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
8698 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
8699 | * can also include other factors [XXX]. | |
8700 | * | |
8701 | * To achieve this balance we define a measure of imbalance which follows | |
8702 | * directly from (1): | |
8703 | * | |
ced549fa | 8704 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
8705 | * |
8706 | * We them move tasks around to minimize the imbalance. In the continuous | |
8707 | * function space it is obvious this converges, in the discrete case we get | |
8708 | * a few fun cases generally called infeasible weight scenarios. | |
8709 | * | |
8710 | * [XXX expand on: | |
8711 | * - infeasible weights; | |
8712 | * - local vs global optima in the discrete case. ] | |
8713 | * | |
8714 | * | |
8715 | * SCHED DOMAINS | |
8716 | * | |
8717 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 8718 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 8719 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 8720 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 8721 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 8722 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
8723 | * the groups. |
8724 | * | |
8725 | * This yields: | |
8726 | * | |
8727 | * log_2 n 1 n | |
8728 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
8729 | * i = 0 2^i 2^i | |
8730 | * `- size of each group | |
97fb7a0a | 8731 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
8732 | * | `- freq |
8733 | * `- sum over all levels | |
8734 | * | |
8735 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
8736 | * this makes (5) the runtime complexity of the balancer. | |
8737 | * | |
8738 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 8739 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
8740 | * |
8741 | * The adjacency matrix of the resulting graph is given by: | |
8742 | * | |
97a7142f | 8743 | * log_2 n |
e9c84cb8 PZ |
8744 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
8745 | * k = 0 | |
8746 | * | |
8747 | * And you'll find that: | |
8748 | * | |
8749 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
8750 | * | |
97fb7a0a | 8751 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
8752 | * The task movement gives a factor of O(m), giving a convergence complexity |
8753 | * of: | |
8754 | * | |
8755 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
8756 | * | |
8757 | * | |
8758 | * WORK CONSERVING | |
8759 | * | |
8760 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 8761 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
8762 | * tree itself instead of relying on other CPUs to bring it work. |
8763 | * | |
8764 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
8765 | * time. | |
8766 | * | |
8767 | * [XXX more?] | |
8768 | * | |
8769 | * | |
8770 | * CGROUPS | |
8771 | * | |
8772 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
8773 | * | |
8774 | * s_k,i | |
8775 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
8776 | * S_k | |
8777 | * | |
8778 | * Where | |
8779 | * | |
8780 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
8781 | * | |
97fb7a0a | 8782 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
8783 | * |
8784 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
8785 | * property. | |
8786 | * | |
8787 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
8788 | * rewrite all of this once again.] | |
97a7142f | 8789 | */ |
bf0f6f24 | 8790 | |
ed387b78 HS |
8791 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
8792 | ||
0ec8aa00 PZ |
8793 | enum fbq_type { regular, remote, all }; |
8794 | ||
0b0695f2 | 8795 | /* |
a9723389 VG |
8796 | * 'group_type' describes the group of CPUs at the moment of load balancing. |
8797 | * | |
0b0695f2 | 8798 | * The enum is ordered by pulling priority, with the group with lowest priority |
a9723389 VG |
8799 | * first so the group_type can simply be compared when selecting the busiest |
8800 | * group. See update_sd_pick_busiest(). | |
0b0695f2 | 8801 | */ |
3b1baa64 | 8802 | enum group_type { |
a9723389 | 8803 | /* The group has spare capacity that can be used to run more tasks. */ |
0b0695f2 | 8804 | group_has_spare = 0, |
a9723389 VG |
8805 | /* |
8806 | * The group is fully used and the tasks don't compete for more CPU | |
8807 | * cycles. Nevertheless, some tasks might wait before running. | |
8808 | */ | |
0b0695f2 | 8809 | group_fully_busy, |
a9723389 | 8810 | /* |
c82a6962 VG |
8811 | * One task doesn't fit with CPU's capacity and must be migrated to a |
8812 | * more powerful CPU. | |
a9723389 | 8813 | */ |
3b1baa64 | 8814 | group_misfit_task, |
fee1759e TC |
8815 | /* |
8816 | * Balance SMT group that's fully busy. Can benefit from migration | |
8817 | * a task on SMT with busy sibling to another CPU on idle core. | |
8818 | */ | |
8819 | group_smt_balance, | |
a9723389 VG |
8820 | /* |
8821 | * SD_ASYM_PACKING only: One local CPU with higher capacity is available, | |
8822 | * and the task should be migrated to it instead of running on the | |
8823 | * current CPU. | |
8824 | */ | |
0b0695f2 | 8825 | group_asym_packing, |
a9723389 VG |
8826 | /* |
8827 | * The tasks' affinity constraints previously prevented the scheduler | |
8828 | * from balancing the load across the system. | |
8829 | */ | |
3b1baa64 | 8830 | group_imbalanced, |
a9723389 VG |
8831 | /* |
8832 | * The CPU is overloaded and can't provide expected CPU cycles to all | |
8833 | * tasks. | |
8834 | */ | |
0b0695f2 VG |
8835 | group_overloaded |
8836 | }; | |
8837 | ||
8838 | enum migration_type { | |
8839 | migrate_load = 0, | |
8840 | migrate_util, | |
8841 | migrate_task, | |
8842 | migrate_misfit | |
3b1baa64 MR |
8843 | }; |
8844 | ||
ddcdf6e7 | 8845 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 8846 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
8847 | #define LBF_DST_PINNED 0x04 |
8848 | #define LBF_SOME_PINNED 0x08 | |
23fb06d9 | 8849 | #define LBF_ACTIVE_LB 0x10 |
ddcdf6e7 PZ |
8850 | |
8851 | struct lb_env { | |
8852 | struct sched_domain *sd; | |
8853 | ||
ddcdf6e7 | 8854 | struct rq *src_rq; |
85c1e7da | 8855 | int src_cpu; |
ddcdf6e7 PZ |
8856 | |
8857 | int dst_cpu; | |
8858 | struct rq *dst_rq; | |
8859 | ||
88b8dac0 SV |
8860 | struct cpumask *dst_grpmask; |
8861 | int new_dst_cpu; | |
ddcdf6e7 | 8862 | enum cpu_idle_type idle; |
bd939f45 | 8863 | long imbalance; |
b9403130 MW |
8864 | /* The set of CPUs under consideration for load-balancing */ |
8865 | struct cpumask *cpus; | |
8866 | ||
ddcdf6e7 | 8867 | unsigned int flags; |
367456c7 PZ |
8868 | |
8869 | unsigned int loop; | |
8870 | unsigned int loop_break; | |
8871 | unsigned int loop_max; | |
0ec8aa00 PZ |
8872 | |
8873 | enum fbq_type fbq_type; | |
0b0695f2 | 8874 | enum migration_type migration_type; |
163122b7 | 8875 | struct list_head tasks; |
ddcdf6e7 PZ |
8876 | }; |
8877 | ||
029632fb PZ |
8878 | /* |
8879 | * Is this task likely cache-hot: | |
8880 | */ | |
5d5e2b1b | 8881 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
8882 | { |
8883 | s64 delta; | |
8884 | ||
5cb9eaa3 | 8885 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 8886 | |
029632fb PZ |
8887 | if (p->sched_class != &fair_sched_class) |
8888 | return 0; | |
8889 | ||
1da1843f | 8890 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
8891 | return 0; |
8892 | ||
ec73240b JD |
8893 | /* SMT siblings share cache */ |
8894 | if (env->sd->flags & SD_SHARE_CPUCAPACITY) | |
8895 | return 0; | |
8896 | ||
029632fb PZ |
8897 | /* |
8898 | * Buddy candidates are cache hot: | |
8899 | */ | |
5d5e2b1b | 8900 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
5e963f2b | 8901 | (&p->se == cfs_rq_of(&p->se)->next)) |
029632fb PZ |
8902 | return 1; |
8903 | ||
8904 | if (sysctl_sched_migration_cost == -1) | |
8905 | return 1; | |
97886d9d AL |
8906 | |
8907 | /* | |
8908 | * Don't migrate task if the task's cookie does not match | |
8909 | * with the destination CPU's core cookie. | |
8910 | */ | |
8911 | if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p)) | |
8912 | return 1; | |
8913 | ||
029632fb PZ |
8914 | if (sysctl_sched_migration_cost == 0) |
8915 | return 0; | |
8916 | ||
5d5e2b1b | 8917 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
8918 | |
8919 | return delta < (s64)sysctl_sched_migration_cost; | |
8920 | } | |
8921 | ||
3a7053b3 | 8922 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 8923 | /* |
2a1ed24c SD |
8924 | * Returns 1, if task migration degrades locality |
8925 | * Returns 0, if task migration improves locality i.e migration preferred. | |
8926 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 8927 | */ |
2a1ed24c | 8928 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 8929 | { |
b1ad065e | 8930 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
8931 | unsigned long src_weight, dst_weight; |
8932 | int src_nid, dst_nid, dist; | |
3a7053b3 | 8933 | |
2a595721 | 8934 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
8935 | return -1; |
8936 | ||
c3b9bc5b | 8937 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 8938 | return -1; |
7a0f3083 MG |
8939 | |
8940 | src_nid = cpu_to_node(env->src_cpu); | |
8941 | dst_nid = cpu_to_node(env->dst_cpu); | |
8942 | ||
83e1d2cd | 8943 | if (src_nid == dst_nid) |
2a1ed24c | 8944 | return -1; |
7a0f3083 | 8945 | |
2a1ed24c SD |
8946 | /* Migrating away from the preferred node is always bad. */ |
8947 | if (src_nid == p->numa_preferred_nid) { | |
8948 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
8949 | return 1; | |
8950 | else | |
8951 | return -1; | |
8952 | } | |
b1ad065e | 8953 | |
c1ceac62 RR |
8954 | /* Encourage migration to the preferred node. */ |
8955 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 8956 | return 0; |
b1ad065e | 8957 | |
739294fb | 8958 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 8959 | if (env->idle == CPU_IDLE) |
739294fb RR |
8960 | return -1; |
8961 | ||
f35678b6 | 8962 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 8963 | if (numa_group) { |
f35678b6 SD |
8964 | src_weight = group_weight(p, src_nid, dist); |
8965 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 8966 | } else { |
f35678b6 SD |
8967 | src_weight = task_weight(p, src_nid, dist); |
8968 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
8969 | } |
8970 | ||
f35678b6 | 8971 | return dst_weight < src_weight; |
7a0f3083 MG |
8972 | } |
8973 | ||
3a7053b3 | 8974 | #else |
2a1ed24c | 8975 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
8976 | struct lb_env *env) |
8977 | { | |
2a1ed24c | 8978 | return -1; |
7a0f3083 | 8979 | } |
3a7053b3 MG |
8980 | #endif |
8981 | ||
1e3c88bd PZ |
8982 | /* |
8983 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
8984 | */ | |
8985 | static | |
8e45cb54 | 8986 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 8987 | { |
2a1ed24c | 8988 | int tsk_cache_hot; |
e5673f28 | 8989 | |
5cb9eaa3 | 8990 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 8991 | |
1e3c88bd PZ |
8992 | /* |
8993 | * We do not migrate tasks that are: | |
d3198084 | 8994 | * 1) throttled_lb_pair, or |
3bd37062 | 8995 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
8996 | * 3) running (obviously), or |
8997 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 8998 | */ |
d3198084 JK |
8999 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
9000 | return 0; | |
9001 | ||
b9e6e286 | 9002 | /* Disregard percpu kthreads; they are where they need to be. */ |
3a7956e2 | 9003 | if (kthread_is_per_cpu(p)) |
9bcb959d LC |
9004 | return 0; |
9005 | ||
3bd37062 | 9006 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 9007 | int cpu; |
88b8dac0 | 9008 | |
ceeadb83 | 9009 | schedstat_inc(p->stats.nr_failed_migrations_affine); |
88b8dac0 | 9010 | |
6263322c PZ |
9011 | env->flags |= LBF_SOME_PINNED; |
9012 | ||
88b8dac0 | 9013 | /* |
97fb7a0a | 9014 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
9015 | * our sched_group. We may want to revisit it if we couldn't |
9016 | * meet load balance goals by pulling other tasks on src_cpu. | |
9017 | * | |
23fb06d9 VS |
9018 | * Avoid computing new_dst_cpu |
9019 | * - for NEWLY_IDLE | |
9020 | * - if we have already computed one in current iteration | |
9021 | * - if it's an active balance | |
88b8dac0 | 9022 | */ |
23fb06d9 VS |
9023 | if (env->idle == CPU_NEWLY_IDLE || |
9024 | env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) | |
88b8dac0 SV |
9025 | return 0; |
9026 | ||
97fb7a0a | 9027 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 9028 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 9029 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 9030 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
9031 | env->new_dst_cpu = cpu; |
9032 | break; | |
9033 | } | |
88b8dac0 | 9034 | } |
e02e60c1 | 9035 | |
1e3c88bd PZ |
9036 | return 0; |
9037 | } | |
88b8dac0 | 9038 | |
3b03706f | 9039 | /* Record that we found at least one task that could run on dst_cpu */ |
8e45cb54 | 9040 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 9041 | |
0b9d46fc | 9042 | if (task_on_cpu(env->src_rq, p)) { |
ceeadb83 | 9043 | schedstat_inc(p->stats.nr_failed_migrations_running); |
1e3c88bd PZ |
9044 | return 0; |
9045 | } | |
9046 | ||
9047 | /* | |
9048 | * Aggressive migration if: | |
23fb06d9 VS |
9049 | * 1) active balance |
9050 | * 2) destination numa is preferred | |
9051 | * 3) task is cache cold, or | |
9052 | * 4) too many balance attempts have failed. | |
1e3c88bd | 9053 | */ |
23fb06d9 VS |
9054 | if (env->flags & LBF_ACTIVE_LB) |
9055 | return 1; | |
9056 | ||
2a1ed24c SD |
9057 | tsk_cache_hot = migrate_degrades_locality(p, env); |
9058 | if (tsk_cache_hot == -1) | |
9059 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 9060 | |
2a1ed24c | 9061 | if (tsk_cache_hot <= 0 || |
7a96c231 | 9062 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 9063 | if (tsk_cache_hot == 1) { |
ae92882e | 9064 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
ceeadb83 | 9065 | schedstat_inc(p->stats.nr_forced_migrations); |
3a7053b3 | 9066 | } |
1e3c88bd PZ |
9067 | return 1; |
9068 | } | |
9069 | ||
ceeadb83 | 9070 | schedstat_inc(p->stats.nr_failed_migrations_hot); |
4e2dcb73 | 9071 | return 0; |
1e3c88bd PZ |
9072 | } |
9073 | ||
897c395f | 9074 | /* |
163122b7 KT |
9075 | * detach_task() -- detach the task for the migration specified in env |
9076 | */ | |
9077 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
9078 | { | |
5cb9eaa3 | 9079 | lockdep_assert_rq_held(env->src_rq); |
163122b7 | 9080 | |
5704ac0a | 9081 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
9082 | set_task_cpu(p, env->dst_cpu); |
9083 | } | |
9084 | ||
897c395f | 9085 | /* |
e5673f28 | 9086 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 9087 | * part of active balancing operations within "domain". |
897c395f | 9088 | * |
e5673f28 | 9089 | * Returns a task if successful and NULL otherwise. |
897c395f | 9090 | */ |
e5673f28 | 9091 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 9092 | { |
93824900 | 9093 | struct task_struct *p; |
897c395f | 9094 | |
5cb9eaa3 | 9095 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 9096 | |
93824900 UR |
9097 | list_for_each_entry_reverse(p, |
9098 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
9099 | if (!can_migrate_task(p, env)) |
9100 | continue; | |
897c395f | 9101 | |
163122b7 | 9102 | detach_task(p, env); |
e5673f28 | 9103 | |
367456c7 | 9104 | /* |
e5673f28 | 9105 | * Right now, this is only the second place where |
163122b7 | 9106 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 9107 | * so we can safely collect stats here rather than |
163122b7 | 9108 | * inside detach_tasks(). |
367456c7 | 9109 | */ |
ae92882e | 9110 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 9111 | return p; |
897c395f | 9112 | } |
e5673f28 | 9113 | return NULL; |
897c395f PZ |
9114 | } |
9115 | ||
5d6523eb | 9116 | /* |
0b0695f2 | 9117 | * detach_tasks() -- tries to detach up to imbalance load/util/tasks from |
163122b7 | 9118 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 9119 | * |
163122b7 | 9120 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 9121 | */ |
163122b7 | 9122 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 9123 | { |
5d6523eb | 9124 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
0b0695f2 | 9125 | unsigned long util, load; |
5d6523eb | 9126 | struct task_struct *p; |
163122b7 KT |
9127 | int detached = 0; |
9128 | ||
5cb9eaa3 | 9129 | lockdep_assert_rq_held(env->src_rq); |
1e3c88bd | 9130 | |
acb4decc AL |
9131 | /* |
9132 | * Source run queue has been emptied by another CPU, clear | |
9133 | * LBF_ALL_PINNED flag as we will not test any task. | |
9134 | */ | |
9135 | if (env->src_rq->nr_running <= 1) { | |
9136 | env->flags &= ~LBF_ALL_PINNED; | |
9137 | return 0; | |
9138 | } | |
9139 | ||
bd939f45 | 9140 | if (env->imbalance <= 0) |
5d6523eb | 9141 | return 0; |
1e3c88bd | 9142 | |
5d6523eb | 9143 | while (!list_empty(tasks)) { |
985d3a4c YD |
9144 | /* |
9145 | * We don't want to steal all, otherwise we may be treated likewise, | |
9146 | * which could at worst lead to a livelock crash. | |
9147 | */ | |
38d707c5 | 9148 | if (env->idle && env->src_rq->nr_running <= 1) |
985d3a4c YD |
9149 | break; |
9150 | ||
367456c7 | 9151 | env->loop++; |
b0defa7a VG |
9152 | /* |
9153 | * We've more or less seen every task there is, call it quits | |
9154 | * unless we haven't found any movable task yet. | |
9155 | */ | |
9156 | if (env->loop > env->loop_max && | |
9157 | !(env->flags & LBF_ALL_PINNED)) | |
367456c7 | 9158 | break; |
5d6523eb PZ |
9159 | |
9160 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 9161 | if (env->loop > env->loop_break) { |
c59862f8 | 9162 | env->loop_break += SCHED_NR_MIGRATE_BREAK; |
8e45cb54 | 9163 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 9164 | break; |
a195f004 | 9165 | } |
1e3c88bd | 9166 | |
7e9518ba VG |
9167 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
9168 | ||
d3198084 | 9169 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
9170 | goto next; |
9171 | ||
0b0695f2 VG |
9172 | switch (env->migration_type) { |
9173 | case migrate_load: | |
01cfcde9 VG |
9174 | /* |
9175 | * Depending of the number of CPUs and tasks and the | |
9176 | * cgroup hierarchy, task_h_load() can return a null | |
9177 | * value. Make sure that env->imbalance decreases | |
9178 | * otherwise detach_tasks() will stop only after | |
9179 | * detaching up to loop_max tasks. | |
9180 | */ | |
9181 | load = max_t(unsigned long, task_h_load(p), 1); | |
5d6523eb | 9182 | |
0b0695f2 VG |
9183 | if (sched_feat(LB_MIN) && |
9184 | load < 16 && !env->sd->nr_balance_failed) | |
9185 | goto next; | |
367456c7 | 9186 | |
6cf82d55 VG |
9187 | /* |
9188 | * Make sure that we don't migrate too much load. | |
9189 | * Nevertheless, let relax the constraint if | |
9190 | * scheduler fails to find a good waiting task to | |
9191 | * migrate. | |
9192 | */ | |
39a2a6eb | 9193 | if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) |
0b0695f2 VG |
9194 | goto next; |
9195 | ||
9196 | env->imbalance -= load; | |
9197 | break; | |
9198 | ||
9199 | case migrate_util: | |
9200 | util = task_util_est(p); | |
9201 | ||
3af7524b | 9202 | if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance) |
0b0695f2 VG |
9203 | goto next; |
9204 | ||
9205 | env->imbalance -= util; | |
9206 | break; | |
9207 | ||
9208 | case migrate_task: | |
9209 | env->imbalance--; | |
9210 | break; | |
9211 | ||
9212 | case migrate_misfit: | |
c63be7be | 9213 | /* This is not a misfit task */ |
b48e16a6 | 9214 | if (task_fits_cpu(p, env->src_cpu)) |
0b0695f2 VG |
9215 | goto next; |
9216 | ||
9217 | env->imbalance = 0; | |
9218 | break; | |
9219 | } | |
1e3c88bd | 9220 | |
163122b7 KT |
9221 | detach_task(p, env); |
9222 | list_add(&p->se.group_node, &env->tasks); | |
9223 | ||
9224 | detached++; | |
1e3c88bd | 9225 | |
c1a280b6 | 9226 | #ifdef CONFIG_PREEMPTION |
ee00e66f PZ |
9227 | /* |
9228 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 9229 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
9230 | * the critical section. |
9231 | */ | |
5d6523eb | 9232 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 9233 | break; |
1e3c88bd PZ |
9234 | #endif |
9235 | ||
ee00e66f PZ |
9236 | /* |
9237 | * We only want to steal up to the prescribed amount of | |
0b0695f2 | 9238 | * load/util/tasks. |
ee00e66f | 9239 | */ |
bd939f45 | 9240 | if (env->imbalance <= 0) |
ee00e66f | 9241 | break; |
367456c7 PZ |
9242 | |
9243 | continue; | |
9244 | next: | |
93824900 | 9245 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 9246 | } |
5d6523eb | 9247 | |
1e3c88bd | 9248 | /* |
163122b7 KT |
9249 | * Right now, this is one of only two places we collect this stat |
9250 | * so we can safely collect detach_one_task() stats here rather | |
9251 | * than inside detach_one_task(). | |
1e3c88bd | 9252 | */ |
ae92882e | 9253 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 9254 | |
163122b7 KT |
9255 | return detached; |
9256 | } | |
9257 | ||
9258 | /* | |
9259 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
9260 | */ | |
9261 | static void attach_task(struct rq *rq, struct task_struct *p) | |
9262 | { | |
5cb9eaa3 | 9263 | lockdep_assert_rq_held(rq); |
163122b7 | 9264 | |
09348d75 | 9265 | WARN_ON_ONCE(task_rq(p) != rq); |
5704ac0a | 9266 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
e23edc86 | 9267 | wakeup_preempt(rq, p, 0); |
163122b7 KT |
9268 | } |
9269 | ||
9270 | /* | |
9271 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
9272 | * its new rq. | |
9273 | */ | |
9274 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
9275 | { | |
8a8c69c3 PZ |
9276 | struct rq_flags rf; |
9277 | ||
9278 | rq_lock(rq, &rf); | |
5704ac0a | 9279 | update_rq_clock(rq); |
163122b7 | 9280 | attach_task(rq, p); |
8a8c69c3 | 9281 | rq_unlock(rq, &rf); |
163122b7 KT |
9282 | } |
9283 | ||
9284 | /* | |
9285 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
9286 | * new rq. | |
9287 | */ | |
9288 | static void attach_tasks(struct lb_env *env) | |
9289 | { | |
9290 | struct list_head *tasks = &env->tasks; | |
9291 | struct task_struct *p; | |
8a8c69c3 | 9292 | struct rq_flags rf; |
163122b7 | 9293 | |
8a8c69c3 | 9294 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 9295 | update_rq_clock(env->dst_rq); |
163122b7 KT |
9296 | |
9297 | while (!list_empty(tasks)) { | |
9298 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
9299 | list_del_init(&p->se.group_node); | |
1e3c88bd | 9300 | |
163122b7 KT |
9301 | attach_task(env->dst_rq, p); |
9302 | } | |
9303 | ||
8a8c69c3 | 9304 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
9305 | } |
9306 | ||
b0c79224 | 9307 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
9308 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
9309 | { | |
9310 | if (cfs_rq->avg.load_avg) | |
9311 | return true; | |
9312 | ||
9313 | if (cfs_rq->avg.util_avg) | |
9314 | return true; | |
9315 | ||
9316 | return false; | |
9317 | } | |
9318 | ||
91c27493 | 9319 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 | 9320 | { |
8b936fc1 | 9321 | if (cpu_util_rt(rq)) |
371bf427 VG |
9322 | return true; |
9323 | ||
8b936fc1 | 9324 | if (cpu_util_dl(rq)) |
3727e0e1 VG |
9325 | return true; |
9326 | ||
d4dbc991 | 9327 | if (hw_load_avg(rq)) |
b4eccf5f TG |
9328 | return true; |
9329 | ||
a6965b31 | 9330 | if (cpu_util_irq(rq)) |
91c27493 | 9331 | return true; |
91c27493 | 9332 | |
371bf427 VG |
9333 | return false; |
9334 | } | |
9335 | ||
39b6a429 | 9336 | static inline void update_blocked_load_tick(struct rq *rq) |
b0c79224 | 9337 | { |
39b6a429 VG |
9338 | WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); |
9339 | } | |
b0c79224 | 9340 | |
39b6a429 VG |
9341 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
9342 | { | |
b0c79224 VS |
9343 | if (!has_blocked) |
9344 | rq->has_blocked_load = 0; | |
9345 | } | |
9346 | #else | |
9347 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
9348 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
39b6a429 | 9349 | static inline void update_blocked_load_tick(struct rq *rq) {} |
b0c79224 VS |
9350 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} |
9351 | #endif | |
9352 | ||
bef69dd8 VG |
9353 | static bool __update_blocked_others(struct rq *rq, bool *done) |
9354 | { | |
9355 | const struct sched_class *curr_class; | |
9356 | u64 now = rq_clock_pelt(rq); | |
d4dbc991 | 9357 | unsigned long hw_pressure; |
bef69dd8 VG |
9358 | bool decayed; |
9359 | ||
9360 | /* | |
9361 | * update_load_avg() can call cpufreq_update_util(). Make sure that RT, | |
9362 | * DL and IRQ signals have been updated before updating CFS. | |
9363 | */ | |
9364 | curr_class = rq->curr->sched_class; | |
9365 | ||
d4dbc991 | 9366 | hw_pressure = arch_scale_hw_pressure(cpu_of(rq)); |
b4eccf5f | 9367 | |
bef69dd8 VG |
9368 | decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | |
9369 | update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | | |
97450eb9 | 9370 | update_hw_load_avg(now, rq, hw_pressure) | |
bef69dd8 VG |
9371 | update_irq_load_avg(rq, 0); |
9372 | ||
9373 | if (others_have_blocked(rq)) | |
9374 | *done = false; | |
9375 | ||
9376 | return decayed; | |
9377 | } | |
9378 | ||
1936c53c VG |
9379 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9380 | ||
bef69dd8 | 9381 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 9382 | { |
039ae8bc | 9383 | struct cfs_rq *cfs_rq, *pos; |
bef69dd8 VG |
9384 | bool decayed = false; |
9385 | int cpu = cpu_of(rq); | |
b90f7c9d | 9386 | |
9763b67f PZ |
9387 | /* |
9388 | * Iterates the task_group tree in a bottom up fashion, see | |
9389 | * list_add_leaf_cfs_rq() for details. | |
9390 | */ | |
039ae8bc | 9391 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
9392 | struct sched_entity *se; |
9393 | ||
bef69dd8 | 9394 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { |
fe749158 | 9395 | update_tg_load_avg(cfs_rq); |
4e516076 | 9396 | |
e2f3e35f VD |
9397 | if (cfs_rq->nr_running == 0) |
9398 | update_idle_cfs_rq_clock_pelt(cfs_rq); | |
9399 | ||
bef69dd8 VG |
9400 | if (cfs_rq == &rq->cfs) |
9401 | decayed = true; | |
9402 | } | |
9403 | ||
bc427898 VG |
9404 | /* Propagate pending load changes to the parent, if any: */ |
9405 | se = cfs_rq->tg->se[cpu]; | |
9406 | if (se && !skip_blocked_update(se)) | |
02da26ad | 9407 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
a9e7f654 | 9408 | |
039ae8bc VG |
9409 | /* |
9410 | * There can be a lot of idle CPU cgroups. Don't let fully | |
9411 | * decayed cfs_rqs linger on the list. | |
9412 | */ | |
9413 | if (cfs_rq_is_decayed(cfs_rq)) | |
9414 | list_del_leaf_cfs_rq(cfs_rq); | |
9415 | ||
1936c53c VG |
9416 | /* Don't need periodic decay once load/util_avg are null */ |
9417 | if (cfs_rq_has_blocked(cfs_rq)) | |
bef69dd8 | 9418 | *done = false; |
9d89c257 | 9419 | } |
12b04875 | 9420 | |
bef69dd8 | 9421 | return decayed; |
9e3081ca PZ |
9422 | } |
9423 | ||
9763b67f | 9424 | /* |
68520796 | 9425 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
9426 | * This needs to be done in a top-down fashion because the load of a child |
9427 | * group is a fraction of its parents load. | |
9428 | */ | |
68520796 | 9429 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 9430 | { |
68520796 VD |
9431 | struct rq *rq = rq_of(cfs_rq); |
9432 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 9433 | unsigned long now = jiffies; |
68520796 | 9434 | unsigned long load; |
a35b6466 | 9435 | |
68520796 | 9436 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
9437 | return; |
9438 | ||
0e9f0245 | 9439 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
9440 | for_each_sched_entity(se) { |
9441 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 9442 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
9443 | if (cfs_rq->last_h_load_update == now) |
9444 | break; | |
9445 | } | |
a35b6466 | 9446 | |
68520796 | 9447 | if (!se) { |
7ea241af | 9448 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
9449 | cfs_rq->last_h_load_update = now; |
9450 | } | |
9451 | ||
0e9f0245 | 9452 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 9453 | load = cfs_rq->h_load; |
7ea241af YD |
9454 | load = div64_ul(load * se->avg.load_avg, |
9455 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
9456 | cfs_rq = group_cfs_rq(se); |
9457 | cfs_rq->h_load = load; | |
9458 | cfs_rq->last_h_load_update = now; | |
9459 | } | |
9763b67f PZ |
9460 | } |
9461 | ||
367456c7 | 9462 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 9463 | { |
367456c7 | 9464 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 9465 | |
68520796 | 9466 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 9467 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 9468 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
9469 | } |
9470 | #else | |
bef69dd8 | 9471 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 9472 | { |
6c1d47c0 | 9473 | struct cfs_rq *cfs_rq = &rq->cfs; |
bef69dd8 | 9474 | bool decayed; |
b90f7c9d | 9475 | |
bef69dd8 VG |
9476 | decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
9477 | if (cfs_rq_has_blocked(cfs_rq)) | |
9478 | *done = false; | |
b90f7c9d | 9479 | |
bef69dd8 | 9480 | return decayed; |
9e3081ca PZ |
9481 | } |
9482 | ||
367456c7 | 9483 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 9484 | { |
9d89c257 | 9485 | return p->se.avg.load_avg; |
1e3c88bd | 9486 | } |
230059de | 9487 | #endif |
1e3c88bd | 9488 | |
391b7a53 | 9489 | static void sched_balance_update_blocked_averages(int cpu) |
bef69dd8 VG |
9490 | { |
9491 | bool decayed = false, done = true; | |
9492 | struct rq *rq = cpu_rq(cpu); | |
9493 | struct rq_flags rf; | |
9494 | ||
9495 | rq_lock_irqsave(rq, &rf); | |
39b6a429 | 9496 | update_blocked_load_tick(rq); |
bef69dd8 VG |
9497 | update_rq_clock(rq); |
9498 | ||
9499 | decayed |= __update_blocked_others(rq, &done); | |
9500 | decayed |= __update_blocked_fair(rq, &done); | |
9501 | ||
9502 | update_blocked_load_status(rq, !done); | |
9503 | if (decayed) | |
9504 | cpufreq_update_util(rq, 0); | |
9505 | rq_unlock_irqrestore(rq, &rf); | |
9506 | } | |
9507 | ||
82cf9214 | 9508 | /********** Helpers for sched_balance_find_src_group ************************/ |
caeb178c | 9509 | |
1e3c88bd | 9510 | /* |
33928ed8 | 9511 | * sg_lb_stats - stats of a sched_group required for load-balancing: |
1e3c88bd PZ |
9512 | */ |
9513 | struct sg_lb_stats { | |
33928ed8 IM |
9514 | unsigned long avg_load; /* Avg load over the CPUs of the group */ |
9515 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
9516 | unsigned long group_capacity; /* Capacity over the CPUs of the group */ | |
9517 | unsigned long group_util; /* Total utilization over the CPUs of the group */ | |
e492e1b0 | 9518 | unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ |
33928ed8 | 9519 | unsigned int sum_nr_running; /* Nr of all tasks running in the group */ |
e492e1b0 | 9520 | unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ |
33928ed8 | 9521 | unsigned int idle_cpus; /* Nr of idle CPUs in the group */ |
147c5fc2 | 9522 | unsigned int group_weight; |
caeb178c | 9523 | enum group_type group_type; |
e492e1b0 IM |
9524 | unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ |
9525 | unsigned int group_smt_balance; /* Task on busy SMT be moved */ | |
9526 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ | |
0ec8aa00 PZ |
9527 | #ifdef CONFIG_NUMA_BALANCING |
9528 | unsigned int nr_numa_running; | |
9529 | unsigned int nr_preferred_running; | |
9530 | #endif | |
1e3c88bd PZ |
9531 | }; |
9532 | ||
56cf515b | 9533 | /* |
33928ed8 | 9534 | * sd_lb_stats - stats of a sched_domain required for load-balancing: |
56cf515b JK |
9535 | */ |
9536 | struct sd_lb_stats { | |
e492e1b0 IM |
9537 | struct sched_group *busiest; /* Busiest group in this sd */ |
9538 | struct sched_group *local; /* Local group in this sd */ | |
9539 | unsigned long total_load; /* Total load of all groups in sd */ | |
9540 | unsigned long total_capacity; /* Total capacity of all groups in sd */ | |
9541 | unsigned long avg_load; /* Average load across all groups in sd */ | |
33928ed8 | 9542 | unsigned int prefer_sibling; /* Tasks should go to sibling first */ |
e492e1b0 IM |
9543 | |
9544 | struct sg_lb_stats busiest_stat; /* Statistics of the busiest group */ | |
9545 | struct sg_lb_stats local_stat; /* Statistics of the local group */ | |
56cf515b JK |
9546 | }; |
9547 | ||
147c5fc2 PZ |
9548 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
9549 | { | |
9550 | /* | |
9551 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
9552 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
0b0695f2 VG |
9553 | * We must however set busiest_stat::group_type and |
9554 | * busiest_stat::idle_cpus to the worst busiest group because | |
9555 | * update_sd_pick_busiest() reads these before assignment. | |
147c5fc2 PZ |
9556 | */ |
9557 | *sds = (struct sd_lb_stats){ | |
9558 | .busiest = NULL, | |
9559 | .local = NULL, | |
9560 | .total_load = 0UL, | |
63b2ca30 | 9561 | .total_capacity = 0UL, |
147c5fc2 | 9562 | .busiest_stat = { |
0b0695f2 VG |
9563 | .idle_cpus = UINT_MAX, |
9564 | .group_type = group_has_spare, | |
147c5fc2 PZ |
9565 | }, |
9566 | }; | |
9567 | } | |
9568 | ||
1ca2034e | 9569 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd | 9570 | { |
f1f8d0a2 | 9571 | unsigned long max = get_actual_cpu_capacity(cpu); |
1e3c88bd | 9572 | struct rq *rq = cpu_rq(cpu); |
523e979d | 9573 | unsigned long used, free; |
523e979d | 9574 | unsigned long irq; |
b654f7de | 9575 | |
2e62c474 | 9576 | irq = cpu_util_irq(rq); |
cadefd3d | 9577 | |
523e979d VG |
9578 | if (unlikely(irq >= max)) |
9579 | return 1; | |
aa483808 | 9580 | |
467b7d01 TG |
9581 | /* |
9582 | * avg_rt.util_avg and avg_dl.util_avg track binary signals | |
9583 | * (running and not running) with weights 0 and 1024 respectively. | |
467b7d01 | 9584 | */ |
8b936fc1 SH |
9585 | used = cpu_util_rt(rq); |
9586 | used += cpu_util_dl(rq); | |
1e3c88bd | 9587 | |
523e979d VG |
9588 | if (unlikely(used >= max)) |
9589 | return 1; | |
1e3c88bd | 9590 | |
523e979d | 9591 | free = max - used; |
2e62c474 VG |
9592 | |
9593 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
9594 | } |
9595 | ||
ced549fa | 9596 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 9597 | { |
1ca2034e | 9598 | unsigned long capacity = scale_rt_capacity(cpu); |
1e3c88bd PZ |
9599 | struct sched_group *sdg = sd->groups; |
9600 | ||
ced549fa NP |
9601 | if (!capacity) |
9602 | capacity = 1; | |
1e3c88bd | 9603 | |
a2e90611 VG |
9604 | cpu_rq(cpu)->cpu_capacity = capacity; |
9605 | trace_sched_cpu_capacity_tp(cpu_rq(cpu)); | |
51cf18c9 | 9606 | |
ced549fa | 9607 | sdg->sgc->capacity = capacity; |
bf475ce0 | 9608 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 9609 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
9610 | } |
9611 | ||
63b2ca30 | 9612 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
9613 | { |
9614 | struct sched_domain *child = sd->child; | |
9615 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 9616 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
9617 | unsigned long interval; |
9618 | ||
9619 | interval = msecs_to_jiffies(sd->balance_interval); | |
9620 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 9621 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
9622 | |
9623 | if (!child) { | |
ced549fa | 9624 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
9625 | return; |
9626 | } | |
9627 | ||
dc7ff76e | 9628 | capacity = 0; |
bf475ce0 | 9629 | min_capacity = ULONG_MAX; |
e3d6d0cb | 9630 | max_capacity = 0; |
1e3c88bd | 9631 | |
74a5ce20 PZ |
9632 | if (child->flags & SD_OVERLAP) { |
9633 | /* | |
9634 | * SD_OVERLAP domains cannot assume that child groups | |
9635 | * span the current group. | |
9636 | */ | |
9637 | ||
ae4df9d6 | 9638 | for_each_cpu(cpu, sched_group_span(sdg)) { |
4c58f57f | 9639 | unsigned long cpu_cap = capacity_of(cpu); |
863bffc8 | 9640 | |
4c58f57f PL |
9641 | capacity += cpu_cap; |
9642 | min_capacity = min(cpu_cap, min_capacity); | |
9643 | max_capacity = max(cpu_cap, max_capacity); | |
863bffc8 | 9644 | } |
74a5ce20 PZ |
9645 | } else { |
9646 | /* | |
9647 | * !SD_OVERLAP domains can assume that child groups | |
9648 | * span the current group. | |
97a7142f | 9649 | */ |
74a5ce20 PZ |
9650 | |
9651 | group = child->groups; | |
9652 | do { | |
bf475ce0 MR |
9653 | struct sched_group_capacity *sgc = group->sgc; |
9654 | ||
9655 | capacity += sgc->capacity; | |
9656 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 9657 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
9658 | group = group->next; |
9659 | } while (group != child->groups); | |
9660 | } | |
1e3c88bd | 9661 | |
63b2ca30 | 9662 | sdg->sgc->capacity = capacity; |
bf475ce0 | 9663 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 9664 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
9665 | } |
9666 | ||
9d5efe05 | 9667 | /* |
ea67821b VG |
9668 | * Check whether the capacity of the rq has been noticeably reduced by side |
9669 | * activity. The imbalance_pct is used for the threshold. | |
9670 | * Return true is the capacity is reduced | |
9d5efe05 SV |
9671 | */ |
9672 | static inline int | |
ea67821b | 9673 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 9674 | { |
ea67821b | 9675 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
7bc26384 | 9676 | (arch_scale_cpu_capacity(cpu_of(rq)) * 100)); |
9d5efe05 SV |
9677 | } |
9678 | ||
22d56074 QY |
9679 | /* Check if the rq has a misfit task */ |
9680 | static inline bool check_misfit_status(struct rq *rq) | |
a0fe2cf0 | 9681 | { |
22d56074 | 9682 | return rq->misfit_task_load; |
a0fe2cf0 VS |
9683 | } |
9684 | ||
30ce5dab PZ |
9685 | /* |
9686 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 9687 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 9688 | * |
97fb7a0a IM |
9689 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
9690 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
9691 | * Something like: |
9692 | * | |
2b4d5b25 IM |
9693 | * { 0 1 2 3 } { 4 5 6 7 } |
9694 | * * * * * | |
30ce5dab PZ |
9695 | * |
9696 | * If we were to balance group-wise we'd place two tasks in the first group and | |
9697 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 9698 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
9699 | * |
9700 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
9701 | * by noticing the lower domain failed to reach balance and had difficulty |
9702 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
9703 | * |
9704 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 9705 | * update_sd_pick_busiest(). And calculate_imbalance() and |
82cf9214 | 9706 | * sched_balance_find_src_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
9707 | * to create an effective group imbalance. |
9708 | * | |
9709 | * This is a somewhat tricky proposition since the next run might not find the | |
9710 | * group imbalance and decide the groups need to be balanced again. A most | |
9711 | * subtle and fragile situation. | |
9712 | */ | |
9713 | ||
6263322c | 9714 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 9715 | { |
63b2ca30 | 9716 | return group->sgc->imbalance; |
30ce5dab PZ |
9717 | } |
9718 | ||
b37d9316 | 9719 | /* |
ea67821b VG |
9720 | * group_has_capacity returns true if the group has spare capacity that could |
9721 | * be used by some tasks. | |
fb95a5a0 | 9722 | * We consider that a group has spare capacity if the number of task is |
9e91d61d DE |
9723 | * smaller than the number of CPUs or if the utilization is lower than the |
9724 | * available capacity for CFS tasks. | |
ea67821b VG |
9725 | * For the latter, we use a threshold to stabilize the state, to take into |
9726 | * account the variance of the tasks' load and to return true if the available | |
9727 | * capacity in meaningful for the load balancer. | |
9728 | * As an example, an available capacity of 1% can appear but it doesn't make | |
9729 | * any benefit for the load balance. | |
b37d9316 | 9730 | */ |
ea67821b | 9731 | static inline bool |
57abff06 | 9732 | group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
b37d9316 | 9733 | { |
5e23e474 | 9734 | if (sgs->sum_nr_running < sgs->group_weight) |
ea67821b | 9735 | return true; |
c61037e9 | 9736 | |
070f5e86 VG |
9737 | if ((sgs->group_capacity * imbalance_pct) < |
9738 | (sgs->group_runnable * 100)) | |
9739 | return false; | |
9740 | ||
ea67821b | 9741 | if ((sgs->group_capacity * 100) > |
57abff06 | 9742 | (sgs->group_util * imbalance_pct)) |
ea67821b | 9743 | return true; |
b37d9316 | 9744 | |
ea67821b VG |
9745 | return false; |
9746 | } | |
9747 | ||
9748 | /* | |
9749 | * group_is_overloaded returns true if the group has more tasks than it can | |
9750 | * handle. | |
9751 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
9752 | * with the exact right number of tasks, has no more spare capacity but is not | |
9753 | * overloaded so both group_has_capacity and group_is_overloaded return | |
9754 | * false. | |
9755 | */ | |
9756 | static inline bool | |
57abff06 | 9757 | group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
ea67821b | 9758 | { |
5e23e474 | 9759 | if (sgs->sum_nr_running <= sgs->group_weight) |
ea67821b | 9760 | return false; |
b37d9316 | 9761 | |
ea67821b | 9762 | if ((sgs->group_capacity * 100) < |
57abff06 | 9763 | (sgs->group_util * imbalance_pct)) |
ea67821b | 9764 | return true; |
b37d9316 | 9765 | |
070f5e86 VG |
9766 | if ((sgs->group_capacity * imbalance_pct) < |
9767 | (sgs->group_runnable * 100)) | |
9768 | return true; | |
9769 | ||
ea67821b | 9770 | return false; |
b37d9316 PZ |
9771 | } |
9772 | ||
79a89f92 | 9773 | static inline enum |
57abff06 | 9774 | group_type group_classify(unsigned int imbalance_pct, |
0b0695f2 | 9775 | struct sched_group *group, |
79a89f92 | 9776 | struct sg_lb_stats *sgs) |
caeb178c | 9777 | { |
57abff06 | 9778 | if (group_is_overloaded(imbalance_pct, sgs)) |
caeb178c RR |
9779 | return group_overloaded; |
9780 | ||
9781 | if (sg_imbalanced(group)) | |
9782 | return group_imbalanced; | |
9783 | ||
0b0695f2 VG |
9784 | if (sgs->group_asym_packing) |
9785 | return group_asym_packing; | |
9786 | ||
fee1759e TC |
9787 | if (sgs->group_smt_balance) |
9788 | return group_smt_balance; | |
9789 | ||
3b1baa64 MR |
9790 | if (sgs->group_misfit_task_load) |
9791 | return group_misfit_task; | |
9792 | ||
57abff06 | 9793 | if (!group_has_capacity(imbalance_pct, sgs)) |
0b0695f2 VG |
9794 | return group_fully_busy; |
9795 | ||
9796 | return group_has_spare; | |
caeb178c RR |
9797 | } |
9798 | ||
eefefa71 RN |
9799 | /** |
9800 | * sched_use_asym_prio - Check whether asym_packing priority must be used | |
9801 | * @sd: The scheduling domain of the load balancing | |
9802 | * @cpu: A CPU | |
9803 | * | |
9804 | * Always use CPU priority when balancing load between SMT siblings. When | |
9805 | * balancing load between cores, it is not sufficient that @cpu is idle. Only | |
9806 | * use CPU priority if the whole core is idle. | |
9807 | * | |
9808 | * Returns: True if the priority of @cpu must be followed. False otherwise. | |
9809 | */ | |
9810 | static bool sched_use_asym_prio(struct sched_domain *sd, int cpu) | |
9811 | { | |
fbc44986 AS |
9812 | if (!(sd->flags & SD_ASYM_PACKING)) |
9813 | return false; | |
9814 | ||
eefefa71 RN |
9815 | if (!sched_smt_active()) |
9816 | return true; | |
9817 | ||
9818 | return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu); | |
9819 | } | |
9820 | ||
45de2062 AS |
9821 | static inline bool sched_asym(struct sched_domain *sd, int dst_cpu, int src_cpu) |
9822 | { | |
9823 | /* | |
9824 | * First check if @dst_cpu can do asym_packing load balance. Only do it | |
9825 | * if it has higher priority than @src_cpu. | |
9826 | */ | |
9827 | return sched_use_asym_prio(sd, dst_cpu) && | |
9828 | sched_asym_prefer(dst_cpu, src_cpu); | |
9829 | } | |
9830 | ||
4006a72b | 9831 | /** |
45de2062 | 9832 | * sched_group_asym - Check if the destination CPU can do asym_packing balance |
c9ca0788 | 9833 | * @env: The load balancing environment |
4006a72b | 9834 | * @sgs: Load-balancing statistics of the candidate busiest group |
c9ca0788 | 9835 | * @group: The candidate busiest group |
4006a72b | 9836 | * |
c9ca0788 RN |
9837 | * @env::dst_cpu can do asym_packing if it has higher priority than the |
9838 | * preferred CPU of @group. | |
4006a72b | 9839 | * |
c9ca0788 RN |
9840 | * Return: true if @env::dst_cpu can do with asym_packing load balance. False |
9841 | * otherwise. | |
4006a72b | 9842 | */ |
aafc917a | 9843 | static inline bool |
45de2062 | 9844 | sched_group_asym(struct lb_env *env, struct sg_lb_stats *sgs, struct sched_group *group) |
aafc917a | 9845 | { |
c9ca0788 | 9846 | /* |
45de2062 | 9847 | * CPU priorities do not make sense for SMT cores with more than one |
c9ca0788 RN |
9848 | * busy sibling. |
9849 | */ | |
45de2062 AS |
9850 | if ((group->flags & SD_SHARE_CPUCAPACITY) && |
9851 | (sgs->group_weight - sgs->idle_cpus != 1)) | |
9852 | return false; | |
4006a72b | 9853 | |
45de2062 | 9854 | return sched_asym(env->sd, env->dst_cpu, group->asym_prefer_cpu); |
aafc917a RN |
9855 | } |
9856 | ||
fee1759e TC |
9857 | /* One group has more than one SMT CPU while the other group does not */ |
9858 | static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1, | |
9859 | struct sched_group *sg2) | |
9860 | { | |
9861 | if (!sg1 || !sg2) | |
9862 | return false; | |
9863 | ||
9864 | return (sg1->flags & SD_SHARE_CPUCAPACITY) != | |
9865 | (sg2->flags & SD_SHARE_CPUCAPACITY); | |
9866 | } | |
9867 | ||
9868 | static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs, | |
9869 | struct sched_group *group) | |
9870 | { | |
38d707c5 | 9871 | if (!env->idle) |
fee1759e TC |
9872 | return false; |
9873 | ||
9874 | /* | |
9875 | * For SMT source group, it is better to move a task | |
9876 | * to a CPU that doesn't have multiple tasks sharing its CPU capacity. | |
9877 | * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY | |
9878 | * will not be on. | |
9879 | */ | |
9880 | if (group->flags & SD_SHARE_CPUCAPACITY && | |
9881 | sgs->sum_h_nr_running > 1) | |
9882 | return true; | |
9883 | ||
9884 | return false; | |
9885 | } | |
9886 | ||
7ff16932 TC |
9887 | static inline long sibling_imbalance(struct lb_env *env, |
9888 | struct sd_lb_stats *sds, | |
9889 | struct sg_lb_stats *busiest, | |
9890 | struct sg_lb_stats *local) | |
9891 | { | |
9892 | int ncores_busiest, ncores_local; | |
9893 | long imbalance; | |
9894 | ||
38d707c5 | 9895 | if (!env->idle || !busiest->sum_nr_running) |
7ff16932 TC |
9896 | return 0; |
9897 | ||
9898 | ncores_busiest = sds->busiest->cores; | |
9899 | ncores_local = sds->local->cores; | |
9900 | ||
9901 | if (ncores_busiest == ncores_local) { | |
9902 | imbalance = busiest->sum_nr_running; | |
9903 | lsub_positive(&imbalance, local->sum_nr_running); | |
9904 | return imbalance; | |
9905 | } | |
9906 | ||
9907 | /* Balance such that nr_running/ncores ratio are same on both groups */ | |
9908 | imbalance = ncores_local * busiest->sum_nr_running; | |
9909 | lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running); | |
9910 | /* Normalize imbalance and do rounding on normalization */ | |
9911 | imbalance = 2 * imbalance + ncores_local + ncores_busiest; | |
9912 | imbalance /= ncores_local + ncores_busiest; | |
9913 | ||
9914 | /* Take advantage of resource in an empty sched group */ | |
450e7497 | 9915 | if (imbalance <= 1 && local->sum_nr_running == 0 && |
7ff16932 TC |
9916 | busiest->sum_nr_running > 1) |
9917 | imbalance = 2; | |
9918 | ||
9919 | return imbalance; | |
9920 | } | |
9921 | ||
c82a6962 VG |
9922 | static inline bool |
9923 | sched_reduced_capacity(struct rq *rq, struct sched_domain *sd) | |
9924 | { | |
9925 | /* | |
9926 | * When there is more than 1 task, the group_overloaded case already | |
9927 | * takes care of cpu with reduced capacity | |
9928 | */ | |
9929 | if (rq->cfs.h_nr_running != 1) | |
9930 | return false; | |
9931 | ||
9932 | return check_cpu_capacity(rq, sd); | |
9933 | } | |
9934 | ||
1e3c88bd PZ |
9935 | /** |
9936 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 9937 | * @env: The load balancing environment. |
a315da5e | 9938 | * @sds: Load-balancing data with statistics of the local group. |
1e3c88bd | 9939 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 9940 | * @sgs: variable to hold the statistics for this group. |
4475cd8b IM |
9941 | * @sg_overloaded: sched_group is overloaded |
9942 | * @sg_overutilized: sched_group is overutilized | |
1e3c88bd | 9943 | */ |
bd939f45 | 9944 | static inline void update_sg_lb_stats(struct lb_env *env, |
c0d14b57 | 9945 | struct sd_lb_stats *sds, |
630246a0 QP |
9946 | struct sched_group *group, |
9947 | struct sg_lb_stats *sgs, | |
4475cd8b IM |
9948 | bool *sg_overloaded, |
9949 | bool *sg_overutilized) | |
1e3c88bd | 9950 | { |
0b0695f2 | 9951 | int i, nr_running, local_group; |
1e3c88bd | 9952 | |
b72ff13c PZ |
9953 | memset(sgs, 0, sizeof(*sgs)); |
9954 | ||
c0d14b57 | 9955 | local_group = group == sds->local; |
0b0695f2 | 9956 | |
ae4df9d6 | 9957 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd | 9958 | struct rq *rq = cpu_rq(i); |
c82a6962 | 9959 | unsigned long load = cpu_load(rq); |
1e3c88bd | 9960 | |
c82a6962 | 9961 | sgs->group_load += load; |
82762d2a | 9962 | sgs->group_util += cpu_util_cfs(i); |
070f5e86 | 9963 | sgs->group_runnable += cpu_runnable(rq); |
a3498347 | 9964 | sgs->sum_h_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 9965 | |
a426f99c | 9966 | nr_running = rq->nr_running; |
5e23e474 VG |
9967 | sgs->sum_nr_running += nr_running; |
9968 | ||
a426f99c | 9969 | if (nr_running > 1) |
4475cd8b | 9970 | *sg_overloaded = 1; |
4486edd1 | 9971 | |
2802bf3c | 9972 | if (cpu_overutilized(i)) |
4475cd8b | 9973 | *sg_overutilized = 1; |
4486edd1 | 9974 | |
0ec8aa00 PZ |
9975 | #ifdef CONFIG_NUMA_BALANCING |
9976 | sgs->nr_numa_running += rq->nr_numa_running; | |
9977 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
9978 | #endif | |
a426f99c WL |
9979 | /* |
9980 | * No need to call idle_cpu() if nr_running is not 0 | |
9981 | */ | |
0b0695f2 | 9982 | if (!nr_running && idle_cpu(i)) { |
aae6d3dd | 9983 | sgs->idle_cpus++; |
0b0695f2 VG |
9984 | /* Idle cpu can't have misfit task */ |
9985 | continue; | |
9986 | } | |
9987 | ||
9988 | if (local_group) | |
9989 | continue; | |
3b1baa64 | 9990 | |
c82a6962 VG |
9991 | if (env->sd->flags & SD_ASYM_CPUCAPACITY) { |
9992 | /* Check for a misfit task on the cpu */ | |
9993 | if (sgs->group_misfit_task_load < rq->misfit_task_load) { | |
9994 | sgs->group_misfit_task_load = rq->misfit_task_load; | |
4475cd8b | 9995 | *sg_overloaded = 1; |
c82a6962 | 9996 | } |
38d707c5 | 9997 | } else if (env->idle && sched_reduced_capacity(rq, env->sd)) { |
c82a6962 VG |
9998 | /* Check for a task running on a CPU with reduced capacity */ |
9999 | if (sgs->group_misfit_task_load < load) | |
10000 | sgs->group_misfit_task_load = load; | |
757ffdd7 | 10001 | } |
1e3c88bd PZ |
10002 | } |
10003 | ||
aafc917a RN |
10004 | sgs->group_capacity = group->sgc->capacity; |
10005 | ||
10006 | sgs->group_weight = group->group_weight; | |
10007 | ||
0b0695f2 | 10008 | /* Check if dst CPU is idle and preferred to this group */ |
38d707c5 | 10009 | if (!local_group && env->idle && sgs->sum_h_nr_running && |
fbc44986 | 10010 | sched_group_asym(env, sgs, group)) |
0b0695f2 | 10011 | sgs->group_asym_packing = 1; |
0b0695f2 | 10012 | |
fee1759e TC |
10013 | /* Check for loaded SMT group to be balanced to dst CPU */ |
10014 | if (!local_group && smt_balance(env, sgs, group)) | |
10015 | sgs->group_smt_balance = 1; | |
10016 | ||
57abff06 | 10017 | sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); |
0b0695f2 VG |
10018 | |
10019 | /* Computing avg_load makes sense only when group is overloaded */ | |
10020 | if (sgs->group_type == group_overloaded) | |
10021 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / | |
10022 | sgs->group_capacity; | |
1e3c88bd PZ |
10023 | } |
10024 | ||
532cb4c4 MN |
10025 | /** |
10026 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 10027 | * @env: The load balancing environment. |
532cb4c4 MN |
10028 | * @sds: sched_domain statistics |
10029 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 10030 | * @sgs: sched_group statistics |
532cb4c4 MN |
10031 | * |
10032 | * Determine if @sg is a busier group than the previously selected | |
10033 | * busiest group. | |
e69f6186 YB |
10034 | * |
10035 | * Return: %true if @sg is a busier group than the previously selected | |
10036 | * busiest group. %false otherwise. | |
532cb4c4 | 10037 | */ |
bd939f45 | 10038 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
10039 | struct sd_lb_stats *sds, |
10040 | struct sched_group *sg, | |
bd939f45 | 10041 | struct sg_lb_stats *sgs) |
532cb4c4 | 10042 | { |
caeb178c | 10043 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 10044 | |
0b0695f2 VG |
10045 | /* Make sure that there is at least one task to pull */ |
10046 | if (!sgs->sum_h_nr_running) | |
10047 | return false; | |
10048 | ||
cad68e55 MR |
10049 | /* |
10050 | * Don't try to pull misfit tasks we can't help. | |
10051 | * We can use max_capacity here as reduction in capacity on some | |
10052 | * CPUs in the group should either be possible to resolve | |
10053 | * internally or be covered by avg_load imbalance (eventually). | |
10054 | */ | |
c82a6962 VG |
10055 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && |
10056 | (sgs->group_type == group_misfit_task) && | |
4aed8aa4 | 10057 | (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || |
0b0695f2 | 10058 | sds->local_stat.group_type != group_has_spare)) |
cad68e55 MR |
10059 | return false; |
10060 | ||
caeb178c | 10061 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
10062 | return true; |
10063 | ||
caeb178c RR |
10064 | if (sgs->group_type < busiest->group_type) |
10065 | return false; | |
10066 | ||
9e0994c0 | 10067 | /* |
0b0695f2 VG |
10068 | * The candidate and the current busiest group are the same type of |
10069 | * group. Let check which one is the busiest according to the type. | |
9e0994c0 | 10070 | */ |
9e0994c0 | 10071 | |
0b0695f2 VG |
10072 | switch (sgs->group_type) { |
10073 | case group_overloaded: | |
10074 | /* Select the overloaded group with highest avg_load. */ | |
7e9f7d17 | 10075 | return sgs->avg_load > busiest->avg_load; |
0b0695f2 VG |
10076 | |
10077 | case group_imbalanced: | |
10078 | /* | |
10079 | * Select the 1st imbalanced group as we don't have any way to | |
10080 | * choose one more than another. | |
10081 | */ | |
9e0994c0 MR |
10082 | return false; |
10083 | ||
0b0695f2 VG |
10084 | case group_asym_packing: |
10085 | /* Prefer to move from lowest priority CPU's work */ | |
7e9f7d17 | 10086 | return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu); |
532cb4c4 | 10087 | |
0b0695f2 VG |
10088 | case group_misfit_task: |
10089 | /* | |
10090 | * If we have more than one misfit sg go with the biggest | |
10091 | * misfit. | |
10092 | */ | |
7e9f7d17 | 10093 | return sgs->group_misfit_task_load > busiest->group_misfit_task_load; |
532cb4c4 | 10094 | |
fee1759e | 10095 | case group_smt_balance: |
450e7497 TC |
10096 | /* |
10097 | * Check if we have spare CPUs on either SMT group to | |
10098 | * choose has spare or fully busy handling. | |
10099 | */ | |
10100 | if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0) | |
10101 | goto has_spare; | |
10102 | ||
10103 | fallthrough; | |
10104 | ||
0b0695f2 VG |
10105 | case group_fully_busy: |
10106 | /* | |
10107 | * Select the fully busy group with highest avg_load. In | |
10108 | * theory, there is no need to pull task from such kind of | |
10109 | * group because tasks have all compute capacity that they need | |
10110 | * but we can still improve the overall throughput by reducing | |
10111 | * contention when accessing shared HW resources. | |
10112 | * | |
10113 | * XXX for now avg_load is not computed and always 0 so we | |
5fd6d7f4 RN |
10114 | * select the 1st one, except if @sg is composed of SMT |
10115 | * siblings. | |
0b0695f2 | 10116 | */ |
5fd6d7f4 RN |
10117 | |
10118 | if (sgs->avg_load < busiest->avg_load) | |
0b0695f2 | 10119 | return false; |
5fd6d7f4 RN |
10120 | |
10121 | if (sgs->avg_load == busiest->avg_load) { | |
10122 | /* | |
10123 | * SMT sched groups need more help than non-SMT groups. | |
10124 | * If @sg happens to also be SMT, either choice is good. | |
10125 | */ | |
10126 | if (sds->busiest->flags & SD_SHARE_CPUCAPACITY) | |
10127 | return false; | |
10128 | } | |
10129 | ||
0b0695f2 VG |
10130 | break; |
10131 | ||
10132 | case group_has_spare: | |
fee1759e TC |
10133 | /* |
10134 | * Do not pick sg with SMT CPUs over sg with pure CPUs, | |
10135 | * as we do not want to pull task off SMT core with one task | |
10136 | * and make the core idle. | |
10137 | */ | |
10138 | if (smt_vs_nonsmt_groups(sds->busiest, sg)) { | |
10139 | if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1) | |
10140 | return false; | |
10141 | else | |
10142 | return true; | |
10143 | } | |
450e7497 | 10144 | has_spare: |
fee1759e | 10145 | |
0b0695f2 | 10146 | /* |
b9e6e286 | 10147 | * Select not overloaded group with lowest number of idle CPUs |
5f68eb19 VG |
10148 | * and highest number of running tasks. We could also compare |
10149 | * the spare capacity which is more stable but it can end up | |
10150 | * that the group has less spare capacity but finally more idle | |
0b0695f2 VG |
10151 | * CPUs which means less opportunity to pull tasks. |
10152 | */ | |
5f68eb19 | 10153 | if (sgs->idle_cpus > busiest->idle_cpus) |
0b0695f2 | 10154 | return false; |
5f68eb19 VG |
10155 | else if ((sgs->idle_cpus == busiest->idle_cpus) && |
10156 | (sgs->sum_nr_running <= busiest->sum_nr_running)) | |
10157 | return false; | |
10158 | ||
0b0695f2 | 10159 | break; |
532cb4c4 MN |
10160 | } |
10161 | ||
0b0695f2 VG |
10162 | /* |
10163 | * Candidate sg has no more than one task per CPU and has higher | |
10164 | * per-CPU capacity. Migrating tasks to less capable CPUs may harm | |
10165 | * throughput. Maximize throughput, power/energy consequences are not | |
10166 | * considered. | |
10167 | */ | |
10168 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && | |
10169 | (sgs->group_type <= group_fully_busy) && | |
4aed8aa4 | 10170 | (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) |
0b0695f2 VG |
10171 | return false; |
10172 | ||
10173 | return true; | |
532cb4c4 MN |
10174 | } |
10175 | ||
0ec8aa00 PZ |
10176 | #ifdef CONFIG_NUMA_BALANCING |
10177 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
10178 | { | |
a3498347 | 10179 | if (sgs->sum_h_nr_running > sgs->nr_numa_running) |
0ec8aa00 | 10180 | return regular; |
a3498347 | 10181 | if (sgs->sum_h_nr_running > sgs->nr_preferred_running) |
0ec8aa00 PZ |
10182 | return remote; |
10183 | return all; | |
10184 | } | |
10185 | ||
10186 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
10187 | { | |
10188 | if (rq->nr_running > rq->nr_numa_running) | |
10189 | return regular; | |
10190 | if (rq->nr_running > rq->nr_preferred_running) | |
10191 | return remote; | |
10192 | return all; | |
10193 | } | |
10194 | #else | |
10195 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
10196 | { | |
10197 | return all; | |
10198 | } | |
10199 | ||
10200 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
10201 | { | |
10202 | return regular; | |
10203 | } | |
10204 | #endif /* CONFIG_NUMA_BALANCING */ | |
10205 | ||
57abff06 VG |
10206 | |
10207 | struct sg_lb_stats; | |
10208 | ||
3318544b VG |
10209 | /* |
10210 | * task_running_on_cpu - return 1 if @p is running on @cpu. | |
10211 | */ | |
10212 | ||
10213 | static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) | |
10214 | { | |
10215 | /* Task has no contribution or is new */ | |
10216 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
10217 | return 0; | |
10218 | ||
10219 | if (task_on_rq_queued(p)) | |
10220 | return 1; | |
10221 | ||
10222 | return 0; | |
10223 | } | |
10224 | ||
10225 | /** | |
10226 | * idle_cpu_without - would a given CPU be idle without p ? | |
10227 | * @cpu: the processor on which idleness is tested. | |
10228 | * @p: task which should be ignored. | |
10229 | * | |
10230 | * Return: 1 if the CPU would be idle. 0 otherwise. | |
10231 | */ | |
10232 | static int idle_cpu_without(int cpu, struct task_struct *p) | |
10233 | { | |
10234 | struct rq *rq = cpu_rq(cpu); | |
10235 | ||
10236 | if (rq->curr != rq->idle && rq->curr != p) | |
10237 | return 0; | |
10238 | ||
10239 | /* | |
10240 | * rq->nr_running can't be used but an updated version without the | |
10241 | * impact of p on cpu must be used instead. The updated nr_running | |
10242 | * be computed and tested before calling idle_cpu_without(). | |
10243 | */ | |
10244 | ||
126c2092 | 10245 | if (rq->ttwu_pending) |
3318544b | 10246 | return 0; |
3318544b VG |
10247 | |
10248 | return 1; | |
10249 | } | |
10250 | ||
57abff06 VG |
10251 | /* |
10252 | * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. | |
3318544b | 10253 | * @sd: The sched_domain level to look for idlest group. |
57abff06 VG |
10254 | * @group: sched_group whose statistics are to be updated. |
10255 | * @sgs: variable to hold the statistics for this group. | |
3318544b | 10256 | * @p: The task for which we look for the idlest group/CPU. |
57abff06 VG |
10257 | */ |
10258 | static inline void update_sg_wakeup_stats(struct sched_domain *sd, | |
10259 | struct sched_group *group, | |
10260 | struct sg_lb_stats *sgs, | |
10261 | struct task_struct *p) | |
10262 | { | |
10263 | int i, nr_running; | |
10264 | ||
10265 | memset(sgs, 0, sizeof(*sgs)); | |
10266 | ||
b48e16a6 QY |
10267 | /* Assume that task can't fit any CPU of the group */ |
10268 | if (sd->flags & SD_ASYM_CPUCAPACITY) | |
10269 | sgs->group_misfit_task_load = 1; | |
10270 | ||
57abff06 VG |
10271 | for_each_cpu(i, sched_group_span(group)) { |
10272 | struct rq *rq = cpu_rq(i); | |
3318544b | 10273 | unsigned int local; |
57abff06 | 10274 | |
3318544b | 10275 | sgs->group_load += cpu_load_without(rq, p); |
57abff06 | 10276 | sgs->group_util += cpu_util_without(i, p); |
070f5e86 | 10277 | sgs->group_runnable += cpu_runnable_without(rq, p); |
3318544b VG |
10278 | local = task_running_on_cpu(i, p); |
10279 | sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; | |
57abff06 | 10280 | |
3318544b | 10281 | nr_running = rq->nr_running - local; |
57abff06 VG |
10282 | sgs->sum_nr_running += nr_running; |
10283 | ||
10284 | /* | |
3318544b | 10285 | * No need to call idle_cpu_without() if nr_running is not 0 |
57abff06 | 10286 | */ |
3318544b | 10287 | if (!nr_running && idle_cpu_without(i, p)) |
57abff06 VG |
10288 | sgs->idle_cpus++; |
10289 | ||
b48e16a6 QY |
10290 | /* Check if task fits in the CPU */ |
10291 | if (sd->flags & SD_ASYM_CPUCAPACITY && | |
10292 | sgs->group_misfit_task_load && | |
10293 | task_fits_cpu(p, i)) | |
10294 | sgs->group_misfit_task_load = 0; | |
57abff06 | 10295 | |
57abff06 VG |
10296 | } |
10297 | ||
10298 | sgs->group_capacity = group->sgc->capacity; | |
10299 | ||
289de359 VG |
10300 | sgs->group_weight = group->group_weight; |
10301 | ||
57abff06 VG |
10302 | sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); |
10303 | ||
10304 | /* | |
10305 | * Computing avg_load makes sense only when group is fully busy or | |
10306 | * overloaded | |
10307 | */ | |
6c8116c9 TZ |
10308 | if (sgs->group_type == group_fully_busy || |
10309 | sgs->group_type == group_overloaded) | |
57abff06 VG |
10310 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / |
10311 | sgs->group_capacity; | |
10312 | } | |
10313 | ||
10314 | static bool update_pick_idlest(struct sched_group *idlest, | |
10315 | struct sg_lb_stats *idlest_sgs, | |
10316 | struct sched_group *group, | |
10317 | struct sg_lb_stats *sgs) | |
10318 | { | |
10319 | if (sgs->group_type < idlest_sgs->group_type) | |
10320 | return true; | |
10321 | ||
10322 | if (sgs->group_type > idlest_sgs->group_type) | |
10323 | return false; | |
10324 | ||
10325 | /* | |
10326 | * The candidate and the current idlest group are the same type of | |
10327 | * group. Let check which one is the idlest according to the type. | |
10328 | */ | |
10329 | ||
10330 | switch (sgs->group_type) { | |
10331 | case group_overloaded: | |
10332 | case group_fully_busy: | |
10333 | /* Select the group with lowest avg_load. */ | |
10334 | if (idlest_sgs->avg_load <= sgs->avg_load) | |
10335 | return false; | |
10336 | break; | |
10337 | ||
10338 | case group_imbalanced: | |
10339 | case group_asym_packing: | |
fee1759e | 10340 | case group_smt_balance: |
57abff06 VG |
10341 | /* Those types are not used in the slow wakeup path */ |
10342 | return false; | |
10343 | ||
10344 | case group_misfit_task: | |
10345 | /* Select group with the highest max capacity */ | |
10346 | if (idlest->sgc->max_capacity >= group->sgc->max_capacity) | |
10347 | return false; | |
10348 | break; | |
10349 | ||
10350 | case group_has_spare: | |
10351 | /* Select group with most idle CPUs */ | |
3edecfef | 10352 | if (idlest_sgs->idle_cpus > sgs->idle_cpus) |
57abff06 | 10353 | return false; |
3edecfef PP |
10354 | |
10355 | /* Select group with lowest group_util */ | |
10356 | if (idlest_sgs->idle_cpus == sgs->idle_cpus && | |
10357 | idlest_sgs->group_util <= sgs->group_util) | |
10358 | return false; | |
10359 | ||
57abff06 VG |
10360 | break; |
10361 | } | |
10362 | ||
10363 | return true; | |
10364 | } | |
10365 | ||
10366 | /* | |
a88b1708 | 10367 | * sched_balance_find_dst_group() finds and returns the least busy CPU group within the |
57abff06 VG |
10368 | * domain. |
10369 | * | |
10370 | * Assumes p is allowed on at least one CPU in sd. | |
10371 | */ | |
10372 | static struct sched_group * | |
a88b1708 | 10373 | sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
57abff06 VG |
10374 | { |
10375 | struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; | |
10376 | struct sg_lb_stats local_sgs, tmp_sgs; | |
10377 | struct sg_lb_stats *sgs; | |
10378 | unsigned long imbalance; | |
10379 | struct sg_lb_stats idlest_sgs = { | |
10380 | .avg_load = UINT_MAX, | |
10381 | .group_type = group_overloaded, | |
10382 | }; | |
10383 | ||
57abff06 VG |
10384 | do { |
10385 | int local_group; | |
10386 | ||
10387 | /* Skip over this group if it has no CPUs allowed */ | |
10388 | if (!cpumask_intersects(sched_group_span(group), | |
10389 | p->cpus_ptr)) | |
10390 | continue; | |
10391 | ||
97886d9d AL |
10392 | /* Skip over this group if no cookie matched */ |
10393 | if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group)) | |
10394 | continue; | |
10395 | ||
57abff06 VG |
10396 | local_group = cpumask_test_cpu(this_cpu, |
10397 | sched_group_span(group)); | |
10398 | ||
10399 | if (local_group) { | |
10400 | sgs = &local_sgs; | |
10401 | local = group; | |
10402 | } else { | |
10403 | sgs = &tmp_sgs; | |
10404 | } | |
10405 | ||
10406 | update_sg_wakeup_stats(sd, group, sgs, p); | |
10407 | ||
10408 | if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { | |
10409 | idlest = group; | |
10410 | idlest_sgs = *sgs; | |
10411 | } | |
10412 | ||
10413 | } while (group = group->next, group != sd->groups); | |
10414 | ||
10415 | ||
10416 | /* There is no idlest group to push tasks to */ | |
10417 | if (!idlest) | |
10418 | return NULL; | |
10419 | ||
7ed735c3 VG |
10420 | /* The local group has been skipped because of CPU affinity */ |
10421 | if (!local) | |
10422 | return idlest; | |
10423 | ||
57abff06 VG |
10424 | /* |
10425 | * If the local group is idler than the selected idlest group | |
10426 | * don't try and push the task. | |
10427 | */ | |
10428 | if (local_sgs.group_type < idlest_sgs.group_type) | |
10429 | return NULL; | |
10430 | ||
10431 | /* | |
10432 | * If the local group is busier than the selected idlest group | |
10433 | * try and push the task. | |
10434 | */ | |
10435 | if (local_sgs.group_type > idlest_sgs.group_type) | |
10436 | return idlest; | |
10437 | ||
10438 | switch (local_sgs.group_type) { | |
10439 | case group_overloaded: | |
10440 | case group_fully_busy: | |
5c339005 MG |
10441 | |
10442 | /* Calculate allowed imbalance based on load */ | |
10443 | imbalance = scale_load_down(NICE_0_LOAD) * | |
10444 | (sd->imbalance_pct-100) / 100; | |
10445 | ||
57abff06 VG |
10446 | /* |
10447 | * When comparing groups across NUMA domains, it's possible for | |
10448 | * the local domain to be very lightly loaded relative to the | |
10449 | * remote domains but "imbalance" skews the comparison making | |
10450 | * remote CPUs look much more favourable. When considering | |
10451 | * cross-domain, add imbalance to the load on the remote node | |
10452 | * and consider staying local. | |
10453 | */ | |
10454 | ||
10455 | if ((sd->flags & SD_NUMA) && | |
10456 | ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) | |
10457 | return NULL; | |
10458 | ||
10459 | /* | |
10460 | * If the local group is less loaded than the selected | |
10461 | * idlest group don't try and push any tasks. | |
10462 | */ | |
10463 | if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) | |
10464 | return NULL; | |
10465 | ||
10466 | if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) | |
10467 | return NULL; | |
10468 | break; | |
10469 | ||
10470 | case group_imbalanced: | |
10471 | case group_asym_packing: | |
fee1759e | 10472 | case group_smt_balance: |
57abff06 VG |
10473 | /* Those type are not used in the slow wakeup path */ |
10474 | return NULL; | |
10475 | ||
10476 | case group_misfit_task: | |
10477 | /* Select group with the highest max capacity */ | |
10478 | if (local->sgc->max_capacity >= idlest->sgc->max_capacity) | |
10479 | return NULL; | |
10480 | break; | |
10481 | ||
10482 | case group_has_spare: | |
cb29a5c1 | 10483 | #ifdef CONFIG_NUMA |
57abff06 | 10484 | if (sd->flags & SD_NUMA) { |
f5b2eeb4 | 10485 | int imb_numa_nr = sd->imb_numa_nr; |
57abff06 VG |
10486 | #ifdef CONFIG_NUMA_BALANCING |
10487 | int idlest_cpu; | |
10488 | /* | |
10489 | * If there is spare capacity at NUMA, try to select | |
10490 | * the preferred node | |
10491 | */ | |
10492 | if (cpu_to_node(this_cpu) == p->numa_preferred_nid) | |
10493 | return NULL; | |
10494 | ||
10495 | idlest_cpu = cpumask_first(sched_group_span(idlest)); | |
10496 | if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) | |
10497 | return idlest; | |
cb29a5c1 | 10498 | #endif /* CONFIG_NUMA_BALANCING */ |
57abff06 | 10499 | /* |
2cfb7a1b MG |
10500 | * Otherwise, keep the task close to the wakeup source |
10501 | * and improve locality if the number of running tasks | |
10502 | * would remain below threshold where an imbalance is | |
f5b2eeb4 PN |
10503 | * allowed while accounting for the possibility the |
10504 | * task is pinned to a subset of CPUs. If there is a | |
10505 | * real need of migration, periodic load balance will | |
10506 | * take care of it. | |
57abff06 | 10507 | */ |
f5b2eeb4 | 10508 | if (p->nr_cpus_allowed != NR_CPUS) { |
ec4fc801 | 10509 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
f5b2eeb4 PN |
10510 | |
10511 | cpumask_and(cpus, sched_group_span(local), p->cpus_ptr); | |
10512 | imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr); | |
10513 | } | |
10514 | ||
cb29a5c1 MG |
10515 | imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus); |
10516 | if (!adjust_numa_imbalance(imbalance, | |
10517 | local_sgs.sum_nr_running + 1, | |
f5b2eeb4 | 10518 | imb_numa_nr)) { |
57abff06 | 10519 | return NULL; |
cb29a5c1 | 10520 | } |
57abff06 | 10521 | } |
cb29a5c1 | 10522 | #endif /* CONFIG_NUMA */ |
57abff06 VG |
10523 | |
10524 | /* | |
10525 | * Select group with highest number of idle CPUs. We could also | |
10526 | * compare the utilization which is more stable but it can end | |
10527 | * up that the group has less spare capacity but finally more | |
10528 | * idle CPUs which means more opportunity to run task. | |
10529 | */ | |
10530 | if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) | |
10531 | return NULL; | |
10532 | break; | |
10533 | } | |
10534 | ||
10535 | return idlest; | |
10536 | } | |
10537 | ||
70fb5ccf CY |
10538 | static void update_idle_cpu_scan(struct lb_env *env, |
10539 | unsigned long sum_util) | |
10540 | { | |
10541 | struct sched_domain_shared *sd_share; | |
10542 | int llc_weight, pct; | |
10543 | u64 x, y, tmp; | |
10544 | /* | |
10545 | * Update the number of CPUs to scan in LLC domain, which could | |
10546 | * be used as a hint in select_idle_cpu(). The update of sd_share | |
10547 | * could be expensive because it is within a shared cache line. | |
10548 | * So the write of this hint only occurs during periodic load | |
10549 | * balancing, rather than CPU_NEWLY_IDLE, because the latter | |
10550 | * can fire way more frequently than the former. | |
10551 | */ | |
10552 | if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE) | |
10553 | return; | |
10554 | ||
10555 | llc_weight = per_cpu(sd_llc_size, env->dst_cpu); | |
10556 | if (env->sd->span_weight != llc_weight) | |
10557 | return; | |
10558 | ||
10559 | sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu)); | |
10560 | if (!sd_share) | |
10561 | return; | |
10562 | ||
10563 | /* | |
10564 | * The number of CPUs to search drops as sum_util increases, when | |
10565 | * sum_util hits 85% or above, the scan stops. | |
10566 | * The reason to choose 85% as the threshold is because this is the | |
10567 | * imbalance_pct(117) when a LLC sched group is overloaded. | |
10568 | * | |
10569 | * let y = SCHED_CAPACITY_SCALE - p * x^2 [1] | |
10570 | * and y'= y / SCHED_CAPACITY_SCALE | |
10571 | * | |
10572 | * x is the ratio of sum_util compared to the CPU capacity: | |
10573 | * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE) | |
10574 | * y' is the ratio of CPUs to be scanned in the LLC domain, | |
10575 | * and the number of CPUs to scan is calculated by: | |
10576 | * | |
10577 | * nr_scan = llc_weight * y' [2] | |
10578 | * | |
10579 | * When x hits the threshold of overloaded, AKA, when | |
10580 | * x = 100 / pct, y drops to 0. According to [1], | |
10581 | * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000 | |
10582 | * | |
10583 | * Scale x by SCHED_CAPACITY_SCALE: | |
10584 | * x' = sum_util / llc_weight; [3] | |
10585 | * | |
10586 | * and finally [1] becomes: | |
10587 | * y = SCHED_CAPACITY_SCALE - | |
10588 | * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4] | |
10589 | * | |
10590 | */ | |
10591 | /* equation [3] */ | |
10592 | x = sum_util; | |
10593 | do_div(x, llc_weight); | |
10594 | ||
10595 | /* equation [4] */ | |
10596 | pct = env->sd->imbalance_pct; | |
10597 | tmp = x * x * pct * pct; | |
10598 | do_div(tmp, 10000 * SCHED_CAPACITY_SCALE); | |
10599 | tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE); | |
10600 | y = SCHED_CAPACITY_SCALE - tmp; | |
10601 | ||
10602 | /* equation [2] */ | |
10603 | y *= llc_weight; | |
10604 | do_div(y, SCHED_CAPACITY_SCALE); | |
10605 | if ((int)y != sd_share->nr_idle_scan) | |
10606 | WRITE_ONCE(sd_share->nr_idle_scan, (int)y); | |
10607 | } | |
10608 | ||
1e3c88bd | 10609 | /** |
461819ac | 10610 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 10611 | * @env: The load balancing environment. |
1e3c88bd PZ |
10612 | * @sds: variable to hold the statistics for this sched_domain. |
10613 | */ | |
0b0695f2 | 10614 | |
0ec8aa00 | 10615 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 10616 | { |
bd939f45 | 10617 | struct sched_group *sg = env->sd->groups; |
05b40e05 | 10618 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 10619 | struct sg_lb_stats tmp_sgs; |
70fb5ccf | 10620 | unsigned long sum_util = 0; |
4475cd8b | 10621 | bool sg_overloaded = 0, sg_overutilized = 0; |
1e3c88bd | 10622 | |
1e3c88bd | 10623 | do { |
56cf515b | 10624 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
10625 | int local_group; |
10626 | ||
ae4df9d6 | 10627 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
10628 | if (local_group) { |
10629 | sds->local = sg; | |
05b40e05 | 10630 | sgs = local; |
b72ff13c PZ |
10631 | |
10632 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
10633 | time_after_eq(jiffies, sg->sgc->next_update)) |
10634 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 10635 | } |
1e3c88bd | 10636 | |
4475cd8b | 10637 | update_sg_lb_stats(env, sds, sg, sgs, &sg_overloaded, &sg_overutilized); |
1e3c88bd | 10638 | |
9dfbc26d | 10639 | if (!local_group && update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 10640 | sds->busiest = sg; |
56cf515b | 10641 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
10642 | } |
10643 | ||
b72ff13c PZ |
10644 | /* Now, start updating sd_lb_stats */ |
10645 | sds->total_load += sgs->group_load; | |
63b2ca30 | 10646 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 10647 | |
70fb5ccf | 10648 | sum_util += sgs->group_util; |
532cb4c4 | 10649 | sg = sg->next; |
bd939f45 | 10650 | } while (sg != env->sd->groups); |
0ec8aa00 | 10651 | |
43726bde RN |
10652 | /* |
10653 | * Indicate that the child domain of the busiest group prefers tasks | |
10654 | * go to a child's sibling domains first. NB the flags of a sched group | |
10655 | * are those of the child domain. | |
10656 | */ | |
10657 | if (sds->busiest) | |
10658 | sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING); | |
0b0695f2 | 10659 | |
f643ea22 | 10660 | |
0ec8aa00 PZ |
10661 | if (env->sd->flags & SD_NUMA) |
10662 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
10663 | |
10664 | if (!env->sd->parent) { | |
10665 | /* update overload indicator if we are at root domain */ | |
4475cd8b | 10666 | set_rd_overloaded(env->dst_rq->rd, sg_overloaded); |
2802bf3c MR |
10667 | |
10668 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
cd18bec6 | 10669 | set_rd_overutilized(env->dst_rq->rd, sg_overutilized); |
4475cd8b IM |
10670 | } else if (sg_overutilized) { |
10671 | set_rd_overutilized(env->dst_rq->rd, sg_overutilized); | |
4486edd1 | 10672 | } |
70fb5ccf CY |
10673 | |
10674 | update_idle_cpu_scan(env, sum_util); | |
532cb4c4 MN |
10675 | } |
10676 | ||
1e3c88bd PZ |
10677 | /** |
10678 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
10679 | * groups of a given sched_domain during load balance. | |
bd939f45 | 10680 | * @env: load balance environment |
1e3c88bd | 10681 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 10682 | */ |
bd939f45 | 10683 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 10684 | { |
56cf515b JK |
10685 | struct sg_lb_stats *local, *busiest; |
10686 | ||
10687 | local = &sds->local_stat; | |
56cf515b | 10688 | busiest = &sds->busiest_stat; |
dd5feea1 | 10689 | |
0b0695f2 | 10690 | if (busiest->group_type == group_misfit_task) { |
c82a6962 VG |
10691 | if (env->sd->flags & SD_ASYM_CPUCAPACITY) { |
10692 | /* Set imbalance to allow misfit tasks to be balanced. */ | |
10693 | env->migration_type = migrate_misfit; | |
10694 | env->imbalance = 1; | |
10695 | } else { | |
10696 | /* | |
10697 | * Set load imbalance to allow moving task from cpu | |
10698 | * with reduced capacity. | |
10699 | */ | |
10700 | env->migration_type = migrate_load; | |
10701 | env->imbalance = busiest->group_misfit_task_load; | |
10702 | } | |
0b0695f2 VG |
10703 | return; |
10704 | } | |
10705 | ||
10706 | if (busiest->group_type == group_asym_packing) { | |
10707 | /* | |
10708 | * In case of asym capacity, we will try to migrate all load to | |
10709 | * the preferred CPU. | |
10710 | */ | |
10711 | env->migration_type = migrate_task; | |
10712 | env->imbalance = busiest->sum_h_nr_running; | |
10713 | return; | |
10714 | } | |
10715 | ||
fee1759e TC |
10716 | if (busiest->group_type == group_smt_balance) { |
10717 | /* Reduce number of tasks sharing CPU capacity */ | |
10718 | env->migration_type = migrate_task; | |
10719 | env->imbalance = 1; | |
10720 | return; | |
10721 | } | |
10722 | ||
0b0695f2 VG |
10723 | if (busiest->group_type == group_imbalanced) { |
10724 | /* | |
10725 | * In the group_imb case we cannot rely on group-wide averages | |
10726 | * to ensure CPU-load equilibrium, try to move any task to fix | |
10727 | * the imbalance. The next load balance will take care of | |
10728 | * balancing back the system. | |
10729 | */ | |
10730 | env->migration_type = migrate_task; | |
10731 | env->imbalance = 1; | |
490ba971 VG |
10732 | return; |
10733 | } | |
10734 | ||
1e3c88bd | 10735 | /* |
0b0695f2 | 10736 | * Try to use spare capacity of local group without overloading it or |
a9723389 | 10737 | * emptying busiest. |
1e3c88bd | 10738 | */ |
0b0695f2 | 10739 | if (local->group_type == group_has_spare) { |
16b0a7a1 | 10740 | if ((busiest->group_type > group_fully_busy) && |
54de4427 | 10741 | !(env->sd->flags & SD_SHARE_LLC)) { |
0b0695f2 VG |
10742 | /* |
10743 | * If busiest is overloaded, try to fill spare | |
10744 | * capacity. This might end up creating spare capacity | |
10745 | * in busiest or busiest still being overloaded but | |
10746 | * there is no simple way to directly compute the | |
10747 | * amount of load to migrate in order to balance the | |
10748 | * system. | |
10749 | */ | |
10750 | env->migration_type = migrate_util; | |
10751 | env->imbalance = max(local->group_capacity, local->group_util) - | |
10752 | local->group_util; | |
10753 | ||
10754 | /* | |
10755 | * In some cases, the group's utilization is max or even | |
10756 | * higher than capacity because of migrations but the | |
10757 | * local CPU is (newly) idle. There is at least one | |
10758 | * waiting task in this overloaded busiest group. Let's | |
10759 | * try to pull it. | |
10760 | */ | |
38d707c5 | 10761 | if (env->idle && env->imbalance == 0) { |
0b0695f2 VG |
10762 | env->migration_type = migrate_task; |
10763 | env->imbalance = 1; | |
10764 | } | |
10765 | ||
10766 | return; | |
10767 | } | |
10768 | ||
10769 | if (busiest->group_weight == 1 || sds->prefer_sibling) { | |
0b0695f2 VG |
10770 | /* |
10771 | * When prefer sibling, evenly spread running tasks on | |
10772 | * groups. | |
10773 | */ | |
10774 | env->migration_type = migrate_task; | |
7ff16932 | 10775 | env->imbalance = sibling_imbalance(env, sds, busiest, local); |
b396f523 | 10776 | } else { |
0b0695f2 | 10777 | |
b396f523 MG |
10778 | /* |
10779 | * If there is no overload, we just want to even the number of | |
b9e6e286 | 10780 | * idle CPUs. |
b396f523 MG |
10781 | */ |
10782 | env->migration_type = migrate_task; | |
cb29a5c1 MG |
10783 | env->imbalance = max_t(long, 0, |
10784 | (local->idle_cpus - busiest->idle_cpus)); | |
b396f523 MG |
10785 | } |
10786 | ||
cb29a5c1 | 10787 | #ifdef CONFIG_NUMA |
b396f523 | 10788 | /* Consider allowing a small imbalance between NUMA groups */ |
7d2b5dd0 | 10789 | if (env->sd->flags & SD_NUMA) { |
fb86f5b2 | 10790 | env->imbalance = adjust_numa_imbalance(env->imbalance, |
cb29a5c1 MG |
10791 | local->sum_nr_running + 1, |
10792 | env->sd->imb_numa_nr); | |
7d2b5dd0 | 10793 | } |
cb29a5c1 MG |
10794 | #endif |
10795 | ||
10796 | /* Number of tasks to move to restore balance */ | |
10797 | env->imbalance >>= 1; | |
b396f523 | 10798 | |
fcf0553d | 10799 | return; |
1e3c88bd PZ |
10800 | } |
10801 | ||
9a5d9ba6 | 10802 | /* |
0b0695f2 VG |
10803 | * Local is fully busy but has to take more load to relieve the |
10804 | * busiest group | |
9a5d9ba6 | 10805 | */ |
0b0695f2 VG |
10806 | if (local->group_type < group_overloaded) { |
10807 | /* | |
10808 | * Local will become overloaded so the avg_load metrics are | |
10809 | * finally needed. | |
10810 | */ | |
10811 | ||
10812 | local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / | |
10813 | local->group_capacity; | |
10814 | ||
111688ca AL |
10815 | /* |
10816 | * If the local group is more loaded than the selected | |
10817 | * busiest group don't try to pull any tasks. | |
10818 | */ | |
10819 | if (local->avg_load >= busiest->avg_load) { | |
10820 | env->imbalance = 0; | |
10821 | return; | |
10822 | } | |
06354900 | 10823 | |
10824 | sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / | |
10825 | sds->total_capacity; | |
91dcf1e8 VG |
10826 | |
10827 | /* | |
10828 | * If the local group is more loaded than the average system | |
10829 | * load, don't try to pull any tasks. | |
10830 | */ | |
10831 | if (local->avg_load >= sds->avg_load) { | |
10832 | env->imbalance = 0; | |
10833 | return; | |
10834 | } | |
10835 | ||
dd5feea1 SS |
10836 | } |
10837 | ||
10838 | /* | |
0b0695f2 VG |
10839 | * Both group are or will become overloaded and we're trying to get all |
10840 | * the CPUs to the average_load, so we don't want to push ourselves | |
10841 | * above the average load, nor do we wish to reduce the max loaded CPU | |
10842 | * below the average load. At the same time, we also don't want to | |
10843 | * reduce the group load below the group capacity. Thus we look for | |
10844 | * the minimum possible imbalance. | |
dd5feea1 | 10845 | */ |
0b0695f2 | 10846 | env->migration_type = migrate_load; |
56cf515b | 10847 | env->imbalance = min( |
0b0695f2 | 10848 | (busiest->avg_load - sds->avg_load) * busiest->group_capacity, |
63b2ca30 | 10849 | (sds->avg_load - local->avg_load) * local->group_capacity |
ca8ce3d0 | 10850 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 10851 | } |
fab47622 | 10852 | |
82cf9214 | 10853 | /******* sched_balance_find_src_group() helpers end here *********************/ |
1e3c88bd | 10854 | |
0b0695f2 VG |
10855 | /* |
10856 | * Decision matrix according to the local and busiest group type: | |
10857 | * | |
10858 | * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded | |
10859 | * has_spare nr_idle balanced N/A N/A balanced balanced | |
10860 | * fully_busy nr_idle nr_idle N/A N/A balanced balanced | |
a6583531 | 10861 | * misfit_task force N/A N/A N/A N/A N/A |
0b0695f2 VG |
10862 | * asym_packing force force N/A N/A force force |
10863 | * imbalanced force force N/A N/A force force | |
10864 | * overloaded force force N/A N/A force avg_load | |
10865 | * | |
10866 | * N/A : Not Applicable because already filtered while updating | |
10867 | * statistics. | |
10868 | * balanced : The system is balanced for these 2 groups. | |
10869 | * force : Calculate the imbalance as load migration is probably needed. | |
10870 | * avg_load : Only if imbalance is significant enough. | |
10871 | * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite | |
10872 | * different in groups. | |
10873 | */ | |
10874 | ||
1e3c88bd | 10875 | /** |
82cf9214 | 10876 | * sched_balance_find_src_group - Returns the busiest group within the sched_domain |
0a9b23ce | 10877 | * if there is an imbalance. |
a315da5e | 10878 | * @env: The load balancing environment. |
1e3c88bd | 10879 | * |
a3df0679 | 10880 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
10881 | * to restore balance. |
10882 | * | |
e69f6186 | 10883 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 10884 | */ |
82cf9214 | 10885 | static struct sched_group *sched_balance_find_src_group(struct lb_env *env) |
1e3c88bd | 10886 | { |
56cf515b | 10887 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
10888 | struct sd_lb_stats sds; |
10889 | ||
147c5fc2 | 10890 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
10891 | |
10892 | /* | |
b0fb1eb4 | 10893 | * Compute the various statistics relevant for load balancing at |
1e3c88bd PZ |
10894 | * this level. |
10895 | */ | |
23f0d209 | 10896 | update_sd_lb_stats(env, &sds); |
2802bf3c | 10897 | |
cc57aa8f | 10898 | /* There is no busy sibling group to pull tasks from */ |
0b0695f2 | 10899 | if (!sds.busiest) |
1e3c88bd PZ |
10900 | goto out_balanced; |
10901 | ||
e5ed0550 VG |
10902 | busiest = &sds.busiest_stat; |
10903 | ||
0b0695f2 VG |
10904 | /* Misfit tasks should be dealt with regardless of the avg load */ |
10905 | if (busiest->group_type == group_misfit_task) | |
10906 | goto force_balance; | |
10907 | ||
902e786c SH |
10908 | if (!is_rd_overutilized(env->dst_rq->rd) && |
10909 | rcu_dereference(env->dst_rq->rd->pd)) | |
10910 | goto out_balanced; | |
e5ed0550 | 10911 | |
0b0695f2 VG |
10912 | /* ASYM feature bypasses nice load balance check */ |
10913 | if (busiest->group_type == group_asym_packing) | |
10914 | goto force_balance; | |
b0432d8f | 10915 | |
866ab43e PZ |
10916 | /* |
10917 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 10918 | * work because they assume all things are equal, which typically |
3bd37062 | 10919 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 10920 | */ |
caeb178c | 10921 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
10922 | goto force_balance; |
10923 | ||
e5ed0550 | 10924 | local = &sds.local_stat; |
cc57aa8f | 10925 | /* |
9c58c79a | 10926 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
10927 | * don't try and pull any tasks. |
10928 | */ | |
0b0695f2 | 10929 | if (local->group_type > busiest->group_type) |
1e3c88bd PZ |
10930 | goto out_balanced; |
10931 | ||
cc57aa8f | 10932 | /* |
0b0695f2 VG |
10933 | * When groups are overloaded, use the avg_load to ensure fairness |
10934 | * between tasks. | |
cc57aa8f | 10935 | */ |
0b0695f2 VG |
10936 | if (local->group_type == group_overloaded) { |
10937 | /* | |
10938 | * If the local group is more loaded than the selected | |
10939 | * busiest group don't try to pull any tasks. | |
10940 | */ | |
10941 | if (local->avg_load >= busiest->avg_load) | |
10942 | goto out_balanced; | |
10943 | ||
10944 | /* XXX broken for overlapping NUMA groups */ | |
10945 | sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / | |
10946 | sds.total_capacity; | |
1e3c88bd | 10947 | |
aae6d3dd | 10948 | /* |
0b0695f2 VG |
10949 | * Don't pull any tasks if this group is already above the |
10950 | * domain average load. | |
aae6d3dd | 10951 | */ |
0b0695f2 | 10952 | if (local->avg_load >= sds.avg_load) |
aae6d3dd | 10953 | goto out_balanced; |
0b0695f2 | 10954 | |
c186fafe | 10955 | /* |
0b0695f2 VG |
10956 | * If the busiest group is more loaded, use imbalance_pct to be |
10957 | * conservative. | |
c186fafe | 10958 | */ |
56cf515b JK |
10959 | if (100 * busiest->avg_load <= |
10960 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 10961 | goto out_balanced; |
aae6d3dd | 10962 | } |
1e3c88bd | 10963 | |
43726bde RN |
10964 | /* |
10965 | * Try to move all excess tasks to a sibling domain of the busiest | |
10966 | * group's child domain. | |
10967 | */ | |
0b0695f2 | 10968 | if (sds.prefer_sibling && local->group_type == group_has_spare && |
7ff16932 | 10969 | sibling_imbalance(env, &sds, busiest, local) > 1) |
0b0695f2 VG |
10970 | goto force_balance; |
10971 | ||
2ab4092f | 10972 | if (busiest->group_type != group_overloaded) { |
38d707c5 | 10973 | if (!env->idle) { |
2ab4092f VG |
10974 | /* |
10975 | * If the busiest group is not overloaded (and as a | |
10976 | * result the local one too) but this CPU is already | |
10977 | * busy, let another idle CPU try to pull task. | |
10978 | */ | |
10979 | goto out_balanced; | |
fee1759e TC |
10980 | } |
10981 | ||
10982 | if (busiest->group_type == group_smt_balance && | |
10983 | smt_vs_nonsmt_groups(sds.local, sds.busiest)) { | |
10984 | /* Let non SMT CPU pull from SMT CPU sharing with sibling */ | |
10985 | goto force_balance; | |
10986 | } | |
2ab4092f VG |
10987 | |
10988 | if (busiest->group_weight > 1 && | |
fee1759e | 10989 | local->idle_cpus <= (busiest->idle_cpus + 1)) { |
2ab4092f VG |
10990 | /* |
10991 | * If the busiest group is not overloaded | |
10992 | * and there is no imbalance between this and busiest | |
10993 | * group wrt idle CPUs, it is balanced. The imbalance | |
10994 | * becomes significant if the diff is greater than 1 | |
10995 | * otherwise we might end up to just move the imbalance | |
10996 | * on another group. Of course this applies only if | |
10997 | * there is more than 1 CPU per group. | |
10998 | */ | |
10999 | goto out_balanced; | |
fee1759e | 11000 | } |
2ab4092f | 11001 | |
fee1759e | 11002 | if (busiest->sum_h_nr_running == 1) { |
2ab4092f VG |
11003 | /* |
11004 | * busiest doesn't have any tasks waiting to run | |
11005 | */ | |
11006 | goto out_balanced; | |
fee1759e | 11007 | } |
2ab4092f | 11008 | } |
0b0695f2 | 11009 | |
fab47622 | 11010 | force_balance: |
1e3c88bd | 11011 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 11012 | calculate_imbalance(env, &sds); |
bb3485c8 | 11013 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
11014 | |
11015 | out_balanced: | |
bd939f45 | 11016 | env->imbalance = 0; |
1e3c88bd PZ |
11017 | return NULL; |
11018 | } | |
11019 | ||
11020 | /* | |
f1cd2e2e | 11021 | * sched_balance_find_src_rq - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 11022 | */ |
f1cd2e2e | 11023 | static struct rq *sched_balance_find_src_rq(struct lb_env *env, |
b9403130 | 11024 | struct sched_group *group) |
1e3c88bd PZ |
11025 | { |
11026 | struct rq *busiest = NULL, *rq; | |
0b0695f2 VG |
11027 | unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; |
11028 | unsigned int busiest_nr = 0; | |
1e3c88bd PZ |
11029 | int i; |
11030 | ||
ae4df9d6 | 11031 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
0b0695f2 VG |
11032 | unsigned long capacity, load, util; |
11033 | unsigned int nr_running; | |
0ec8aa00 PZ |
11034 | enum fbq_type rt; |
11035 | ||
11036 | rq = cpu_rq(i); | |
11037 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 11038 | |
0ec8aa00 PZ |
11039 | /* |
11040 | * We classify groups/runqueues into three groups: | |
11041 | * - regular: there are !numa tasks | |
11042 | * - remote: there are numa tasks that run on the 'wrong' node | |
11043 | * - all: there is no distinction | |
11044 | * | |
11045 | * In order to avoid migrating ideally placed numa tasks, | |
11046 | * ignore those when there's better options. | |
11047 | * | |
11048 | * If we ignore the actual busiest queue to migrate another | |
11049 | * task, the next balance pass can still reduce the busiest | |
11050 | * queue by moving tasks around inside the node. | |
11051 | * | |
11052 | * If we cannot move enough load due to this classification | |
11053 | * the next pass will adjust the group classification and | |
11054 | * allow migration of more tasks. | |
11055 | * | |
11056 | * Both cases only affect the total convergence complexity. | |
11057 | */ | |
11058 | if (rt > env->fbq_type) | |
11059 | continue; | |
11060 | ||
0b0695f2 | 11061 | nr_running = rq->cfs.h_nr_running; |
fc488ffd VG |
11062 | if (!nr_running) |
11063 | continue; | |
11064 | ||
11065 | capacity = capacity_of(i); | |
9d5efe05 | 11066 | |
4ad3831a CR |
11067 | /* |
11068 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
11069 | * eventually lead to active_balancing high->low capacity. | |
11070 | * Higher per-CPU capacity is considered better than balancing | |
11071 | * average load. | |
11072 | */ | |
11073 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
4aed8aa4 | 11074 | !capacity_greater(capacity_of(env->dst_cpu), capacity) && |
0b0695f2 | 11075 | nr_running == 1) |
4ad3831a CR |
11076 | continue; |
11077 | ||
18ad3453 RN |
11078 | /* |
11079 | * Make sure we only pull tasks from a CPU of lower priority | |
11080 | * when balancing between SMT siblings. | |
11081 | * | |
11082 | * If balancing between cores, let lower priority CPUs help | |
11083 | * SMT cores with more than one busy sibling. | |
11084 | */ | |
fbc44986 | 11085 | if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1) |
4006a72b RN |
11086 | continue; |
11087 | ||
0b0695f2 VG |
11088 | switch (env->migration_type) { |
11089 | case migrate_load: | |
11090 | /* | |
b0fb1eb4 VG |
11091 | * When comparing with load imbalance, use cpu_load() |
11092 | * which is not scaled with the CPU capacity. | |
0b0695f2 | 11093 | */ |
b0fb1eb4 | 11094 | load = cpu_load(rq); |
1e3c88bd | 11095 | |
0b0695f2 VG |
11096 | if (nr_running == 1 && load > env->imbalance && |
11097 | !check_cpu_capacity(rq, env->sd)) | |
11098 | break; | |
ea67821b | 11099 | |
0b0695f2 VG |
11100 | /* |
11101 | * For the load comparisons with the other CPUs, | |
b0fb1eb4 VG |
11102 | * consider the cpu_load() scaled with the CPU |
11103 | * capacity, so that the load can be moved away | |
11104 | * from the CPU that is potentially running at a | |
11105 | * lower capacity. | |
0b0695f2 VG |
11106 | * |
11107 | * Thus we're looking for max(load_i / capacity_i), | |
11108 | * crosswise multiplication to rid ourselves of the | |
11109 | * division works out to: | |
11110 | * load_i * capacity_j > load_j * capacity_i; | |
11111 | * where j is our previous maximum. | |
11112 | */ | |
11113 | if (load * busiest_capacity > busiest_load * capacity) { | |
11114 | busiest_load = load; | |
11115 | busiest_capacity = capacity; | |
11116 | busiest = rq; | |
11117 | } | |
11118 | break; | |
11119 | ||
11120 | case migrate_util: | |
7d0583cf | 11121 | util = cpu_util_cfs_boost(i); |
0b0695f2 | 11122 | |
c32b4308 VG |
11123 | /* |
11124 | * Don't try to pull utilization from a CPU with one | |
11125 | * running task. Whatever its utilization, we will fail | |
11126 | * detach the task. | |
11127 | */ | |
11128 | if (nr_running <= 1) | |
11129 | continue; | |
11130 | ||
0b0695f2 VG |
11131 | if (busiest_util < util) { |
11132 | busiest_util = util; | |
11133 | busiest = rq; | |
11134 | } | |
11135 | break; | |
11136 | ||
11137 | case migrate_task: | |
11138 | if (busiest_nr < nr_running) { | |
11139 | busiest_nr = nr_running; | |
11140 | busiest = rq; | |
11141 | } | |
11142 | break; | |
11143 | ||
11144 | case migrate_misfit: | |
11145 | /* | |
11146 | * For ASYM_CPUCAPACITY domains with misfit tasks we | |
11147 | * simply seek the "biggest" misfit task. | |
11148 | */ | |
11149 | if (rq->misfit_task_load > busiest_load) { | |
11150 | busiest_load = rq->misfit_task_load; | |
11151 | busiest = rq; | |
11152 | } | |
11153 | ||
11154 | break; | |
1e3c88bd | 11155 | |
1e3c88bd PZ |
11156 | } |
11157 | } | |
11158 | ||
11159 | return busiest; | |
11160 | } | |
11161 | ||
11162 | /* | |
11163 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
11164 | * so long as it is large enough. | |
11165 | */ | |
11166 | #define MAX_PINNED_INTERVAL 512 | |
11167 | ||
46a745d9 VG |
11168 | static inline bool |
11169 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 11170 | { |
46a745d9 | 11171 | /* |
eefefa71 RN |
11172 | * ASYM_PACKING needs to force migrate tasks from busy but lower |
11173 | * priority CPUs in order to pack all tasks in the highest priority | |
11174 | * CPUs. When done between cores, do it only if the whole core if the | |
11175 | * whole core is idle. | |
18ad3453 RN |
11176 | * |
11177 | * If @env::src_cpu is an SMT core with busy siblings, let | |
11178 | * the lower priority @env::dst_cpu help it. Do not follow | |
11179 | * CPU priority. | |
46a745d9 | 11180 | */ |
38d707c5 | 11181 | return env->idle && sched_use_asym_prio(env->sd, env->dst_cpu) && |
18ad3453 RN |
11182 | (sched_asym_prefer(env->dst_cpu, env->src_cpu) || |
11183 | !sched_use_asym_prio(env->sd, env->src_cpu)); | |
46a745d9 | 11184 | } |
bd939f45 | 11185 | |
46a745d9 | 11186 | static inline bool |
e9b9734b VG |
11187 | imbalanced_active_balance(struct lb_env *env) |
11188 | { | |
11189 | struct sched_domain *sd = env->sd; | |
11190 | ||
11191 | /* | |
11192 | * The imbalanced case includes the case of pinned tasks preventing a fair | |
11193 | * distribution of the load on the system but also the even distribution of the | |
11194 | * threads on a system with spare capacity | |
11195 | */ | |
11196 | if ((env->migration_type == migrate_task) && | |
11197 | (sd->nr_balance_failed > sd->cache_nice_tries+2)) | |
11198 | return 1; | |
11199 | ||
11200 | return 0; | |
11201 | } | |
11202 | ||
11203 | static int need_active_balance(struct lb_env *env) | |
46a745d9 VG |
11204 | { |
11205 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 11206 | |
46a745d9 VG |
11207 | if (asym_active_balance(env)) |
11208 | return 1; | |
1af3ed3d | 11209 | |
e9b9734b VG |
11210 | if (imbalanced_active_balance(env)) |
11211 | return 1; | |
11212 | ||
1aaf90a4 VG |
11213 | /* |
11214 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
11215 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
11216 | * because of other sched_class or IRQs if more capacity stays | |
11217 | * available on dst_cpu. | |
11218 | */ | |
38d707c5 | 11219 | if (env->idle && |
1aaf90a4 VG |
11220 | (env->src_rq->cfs.h_nr_running == 1)) { |
11221 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
11222 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
11223 | return 1; | |
11224 | } | |
11225 | ||
0b0695f2 | 11226 | if (env->migration_type == migrate_misfit) |
cad68e55 MR |
11227 | return 1; |
11228 | ||
46a745d9 VG |
11229 | return 0; |
11230 | } | |
11231 | ||
969c7921 TH |
11232 | static int active_load_balance_cpu_stop(void *data); |
11233 | ||
23f0d209 JK |
11234 | static int should_we_balance(struct lb_env *env) |
11235 | { | |
f8858d96 | 11236 | struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask); |
23f0d209 | 11237 | struct sched_group *sg = env->sd->groups; |
b1bfeab9 | 11238 | int cpu, idle_smt = -1; |
23f0d209 | 11239 | |
024c9d2f PZ |
11240 | /* |
11241 | * Ensure the balancing environment is consistent; can happen | |
11242 | * when the softirq triggers 'during' hotplug. | |
11243 | */ | |
11244 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
11245 | return 0; | |
11246 | ||
23f0d209 | 11247 | /* |
97fb7a0a | 11248 | * In the newly idle case, we will allow all the CPUs |
23f0d209 | 11249 | * to do the newly idle load balance. |
792b9f65 JD |
11250 | * |
11251 | * However, we bail out if we already have tasks or a wakeup pending, | |
11252 | * to optimize wakeup latency. | |
23f0d209 | 11253 | */ |
792b9f65 JD |
11254 | if (env->idle == CPU_NEWLY_IDLE) { |
11255 | if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending) | |
11256 | return 0; | |
23f0d209 | 11257 | return 1; |
792b9f65 | 11258 | } |
23f0d209 | 11259 | |
f8858d96 | 11260 | cpumask_copy(swb_cpus, group_balance_mask(sg)); |
97fb7a0a | 11261 | /* Try to find first idle CPU */ |
f8858d96 | 11262 | for_each_cpu_and(cpu, swb_cpus, env->cpus) { |
af218122 | 11263 | if (!idle_cpu(cpu)) |
23f0d209 JK |
11264 | continue; |
11265 | ||
b1bfeab9 RN |
11266 | /* |
11267 | * Don't balance to idle SMT in busy core right away when | |
11268 | * balancing cores, but remember the first idle SMT CPU for | |
11269 | * later consideration. Find CPU on an idle core first. | |
11270 | */ | |
11271 | if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) { | |
11272 | if (idle_smt == -1) | |
11273 | idle_smt = cpu; | |
f8858d96 SH |
11274 | /* |
11275 | * If the core is not idle, and first SMT sibling which is | |
11276 | * idle has been found, then its not needed to check other | |
11277 | * SMT siblings for idleness: | |
11278 | */ | |
11279 | #ifdef CONFIG_SCHED_SMT | |
11280 | cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu)); | |
11281 | #endif | |
b1bfeab9 RN |
11282 | continue; |
11283 | } | |
11284 | ||
6d7e4782 KN |
11285 | /* |
11286 | * Are we the first idle core in a non-SMT domain or higher, | |
11287 | * or the first idle CPU in a SMT domain? | |
11288 | */ | |
64297f2b | 11289 | return cpu == env->dst_cpu; |
23f0d209 JK |
11290 | } |
11291 | ||
6d7e4782 KN |
11292 | /* Are we the first idle CPU with busy siblings? */ |
11293 | if (idle_smt != -1) | |
11294 | return idle_smt == env->dst_cpu; | |
b1bfeab9 | 11295 | |
64297f2b PW |
11296 | /* Are we the first CPU of this group ? */ |
11297 | return group_balance_cpu(sg) == env->dst_cpu; | |
23f0d209 JK |
11298 | } |
11299 | ||
1e3c88bd PZ |
11300 | /* |
11301 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
11302 | * tasks if there is an imbalance. | |
11303 | */ | |
4c3e509e | 11304 | static int sched_balance_rq(int this_cpu, struct rq *this_rq, |
1e3c88bd | 11305 | struct sched_domain *sd, enum cpu_idle_type idle, |
23f0d209 | 11306 | int *continue_balancing) |
1e3c88bd | 11307 | { |
88b8dac0 | 11308 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 11309 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 11310 | struct sched_group *group; |
1e3c88bd | 11311 | struct rq *busiest; |
8a8c69c3 | 11312 | struct rq_flags rf; |
4ba29684 | 11313 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
8e45cb54 PZ |
11314 | struct lb_env env = { |
11315 | .sd = sd, | |
ddcdf6e7 PZ |
11316 | .dst_cpu = this_cpu, |
11317 | .dst_rq = this_rq, | |
0dd37d6d | 11318 | .dst_grpmask = group_balance_mask(sd->groups), |
8e45cb54 | 11319 | .idle = idle, |
c59862f8 | 11320 | .loop_break = SCHED_NR_MIGRATE_BREAK, |
b9403130 | 11321 | .cpus = cpus, |
0ec8aa00 | 11322 | .fbq_type = all, |
163122b7 | 11323 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
11324 | }; |
11325 | ||
65a4433a | 11326 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 11327 | |
ae92882e | 11328 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
11329 | |
11330 | redo: | |
23f0d209 JK |
11331 | if (!should_we_balance(&env)) { |
11332 | *continue_balancing = 0; | |
1e3c88bd | 11333 | goto out_balanced; |
23f0d209 | 11334 | } |
1e3c88bd | 11335 | |
82cf9214 | 11336 | group = sched_balance_find_src_group(&env); |
1e3c88bd | 11337 | if (!group) { |
ae92882e | 11338 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
11339 | goto out_balanced; |
11340 | } | |
11341 | ||
f1cd2e2e | 11342 | busiest = sched_balance_find_src_rq(&env, group); |
1e3c88bd | 11343 | if (!busiest) { |
ae92882e | 11344 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
11345 | goto out_balanced; |
11346 | } | |
11347 | ||
09348d75 | 11348 | WARN_ON_ONCE(busiest == env.dst_rq); |
1e3c88bd | 11349 | |
ae92882e | 11350 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 11351 | |
1aaf90a4 VG |
11352 | env.src_cpu = busiest->cpu; |
11353 | env.src_rq = busiest; | |
11354 | ||
1e3c88bd | 11355 | ld_moved = 0; |
8a41dfcd VG |
11356 | /* Clear this flag as soon as we find a pullable task */ |
11357 | env.flags |= LBF_ALL_PINNED; | |
1e3c88bd PZ |
11358 | if (busiest->nr_running > 1) { |
11359 | /* | |
82cf9214 | 11360 | * Attempt to move tasks. If sched_balance_find_src_group has found |
1e3c88bd PZ |
11361 | * an imbalance but busiest->nr_running <= 1, the group is |
11362 | * still unbalanced. ld_moved simply stays zero, so it is | |
11363 | * correctly treated as an imbalance. | |
11364 | */ | |
c82513e5 | 11365 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 11366 | |
5d6523eb | 11367 | more_balance: |
8a8c69c3 | 11368 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 11369 | update_rq_clock(busiest); |
88b8dac0 SV |
11370 | |
11371 | /* | |
11372 | * cur_ld_moved - load moved in current iteration | |
11373 | * ld_moved - cumulative load moved across iterations | |
11374 | */ | |
163122b7 | 11375 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
11376 | |
11377 | /* | |
163122b7 KT |
11378 | * We've detached some tasks from busiest_rq. Every |
11379 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
11380 | * unlock busiest->lock, and we are able to be sure | |
11381 | * that nobody can manipulate the tasks in parallel. | |
11382 | * See task_rq_lock() family for the details. | |
1e3c88bd | 11383 | */ |
163122b7 | 11384 | |
8a8c69c3 | 11385 | rq_unlock(busiest, &rf); |
163122b7 KT |
11386 | |
11387 | if (cur_ld_moved) { | |
11388 | attach_tasks(&env); | |
11389 | ld_moved += cur_ld_moved; | |
11390 | } | |
11391 | ||
8a8c69c3 | 11392 | local_irq_restore(rf.flags); |
88b8dac0 | 11393 | |
f1cd0858 JK |
11394 | if (env.flags & LBF_NEED_BREAK) { |
11395 | env.flags &= ~LBF_NEED_BREAK; | |
b0defa7a VG |
11396 | /* Stop if we tried all running tasks */ |
11397 | if (env.loop < busiest->nr_running) | |
11398 | goto more_balance; | |
f1cd0858 JK |
11399 | } |
11400 | ||
88b8dac0 SV |
11401 | /* |
11402 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
11403 | * us and move them to an alternate dst_cpu in our sched_group | |
11404 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 11405 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
11406 | * sched_group. |
11407 | * | |
11408 | * This changes load balance semantics a bit on who can move | |
11409 | * load to a given_cpu. In addition to the given_cpu itself | |
11410 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
11411 | * nohz-idle), we now have balance_cpu in a position to move | |
11412 | * load to given_cpu. In rare situations, this may cause | |
11413 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
11414 | * _independently_ and at _same_ time to move some load to | |
3b03706f | 11415 | * given_cpu) causing excess load to be moved to given_cpu. |
88b8dac0 SV |
11416 | * This however should not happen so much in practice and |
11417 | * moreover subsequent load balance cycles should correct the | |
11418 | * excess load moved. | |
11419 | */ | |
6263322c | 11420 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 11421 | |
97fb7a0a | 11422 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 11423 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 11424 | |
78feefc5 | 11425 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 11426 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 11427 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 | 11428 | env.loop = 0; |
c59862f8 | 11429 | env.loop_break = SCHED_NR_MIGRATE_BREAK; |
e02e60c1 | 11430 | |
88b8dac0 SV |
11431 | /* |
11432 | * Go back to "more_balance" rather than "redo" since we | |
11433 | * need to continue with same src_cpu. | |
11434 | */ | |
11435 | goto more_balance; | |
11436 | } | |
1e3c88bd | 11437 | |
6263322c PZ |
11438 | /* |
11439 | * We failed to reach balance because of affinity. | |
11440 | */ | |
11441 | if (sd_parent) { | |
63b2ca30 | 11442 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 11443 | |
afdeee05 | 11444 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 11445 | *group_imbalance = 1; |
6263322c PZ |
11446 | } |
11447 | ||
1e3c88bd | 11448 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 11449 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 11450 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
11451 | /* |
11452 | * Attempting to continue load balancing at the current | |
11453 | * sched_domain level only makes sense if there are | |
11454 | * active CPUs remaining as possible busiest CPUs to | |
11455 | * pull load from which are not contained within the | |
11456 | * destination group that is receiving any migrated | |
11457 | * load. | |
11458 | */ | |
11459 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 | 11460 | env.loop = 0; |
c59862f8 | 11461 | env.loop_break = SCHED_NR_MIGRATE_BREAK; |
1e3c88bd | 11462 | goto redo; |
bbf18b19 | 11463 | } |
afdeee05 | 11464 | goto out_all_pinned; |
1e3c88bd PZ |
11465 | } |
11466 | } | |
11467 | ||
11468 | if (!ld_moved) { | |
ae92882e | 11469 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
11470 | /* |
11471 | * Increment the failure counter only on periodic balance. | |
11472 | * We do not want newidle balance, which can be very | |
11473 | * frequent, pollute the failure counter causing | |
11474 | * excessive cache_hot migrations and active balances. | |
58eeb2d7 QY |
11475 | * |
11476 | * Similarly for migration_misfit which is not related to | |
11477 | * load/util migration, don't pollute nr_balance_failed. | |
58b26c4c | 11478 | */ |
58eeb2d7 QY |
11479 | if (idle != CPU_NEWLY_IDLE && |
11480 | env.migration_type != migrate_misfit) | |
58b26c4c | 11481 | sd->nr_balance_failed++; |
1e3c88bd | 11482 | |
bd939f45 | 11483 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
11484 | unsigned long flags; |
11485 | ||
5cb9eaa3 | 11486 | raw_spin_rq_lock_irqsave(busiest, flags); |
1e3c88bd | 11487 | |
97fb7a0a IM |
11488 | /* |
11489 | * Don't kick the active_load_balance_cpu_stop, | |
11490 | * if the curr task on busiest CPU can't be | |
11491 | * moved to this_cpu: | |
1e3c88bd | 11492 | */ |
3bd37062 | 11493 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
5cb9eaa3 | 11494 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
1e3c88bd PZ |
11495 | goto out_one_pinned; |
11496 | } | |
11497 | ||
8a41dfcd VG |
11498 | /* Record that we found at least one task that could run on this_cpu */ |
11499 | env.flags &= ~LBF_ALL_PINNED; | |
11500 | ||
969c7921 TH |
11501 | /* |
11502 | * ->active_balance synchronizes accesses to | |
11503 | * ->active_balance_work. Once set, it's cleared | |
11504 | * only after active load balance is finished. | |
11505 | */ | |
1e3c88bd PZ |
11506 | if (!busiest->active_balance) { |
11507 | busiest->active_balance = 1; | |
11508 | busiest->push_cpu = this_cpu; | |
11509 | active_balance = 1; | |
11510 | } | |
969c7921 | 11511 | |
f0498d2a PZ |
11512 | preempt_disable(); |
11513 | raw_spin_rq_unlock_irqrestore(busiest, flags); | |
bd939f45 | 11514 | if (active_balance) { |
969c7921 TH |
11515 | stop_one_cpu_nowait(cpu_of(busiest), |
11516 | active_load_balance_cpu_stop, busiest, | |
11517 | &busiest->active_balance_work); | |
bd939f45 | 11518 | } |
f0498d2a | 11519 | preempt_enable(); |
1e3c88bd | 11520 | } |
e9b9734b | 11521 | } else { |
1e3c88bd | 11522 | sd->nr_balance_failed = 0; |
e9b9734b | 11523 | } |
1e3c88bd | 11524 | |
e9b9734b | 11525 | if (likely(!active_balance) || need_active_balance(&env)) { |
1e3c88bd PZ |
11526 | /* We were unbalanced, so reset the balancing interval */ |
11527 | sd->balance_interval = sd->min_interval; | |
1e3c88bd PZ |
11528 | } |
11529 | ||
1e3c88bd PZ |
11530 | goto out; |
11531 | ||
11532 | out_balanced: | |
afdeee05 VG |
11533 | /* |
11534 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
11535 | * constraints. Clear the imbalance flag only if other tasks got |
11536 | * a chance to move and fix the imbalance. | |
afdeee05 | 11537 | */ |
f6cad8df | 11538 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
11539 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
11540 | ||
11541 | if (*group_imbalance) | |
11542 | *group_imbalance = 0; | |
11543 | } | |
11544 | ||
11545 | out_all_pinned: | |
11546 | /* | |
11547 | * We reach balance because all tasks are pinned at this level so | |
11548 | * we can't migrate them. Let the imbalance flag set so parent level | |
11549 | * can try to migrate them. | |
11550 | */ | |
ae92882e | 11551 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
11552 | |
11553 | sd->nr_balance_failed = 0; | |
11554 | ||
11555 | out_one_pinned: | |
3f130a37 VS |
11556 | ld_moved = 0; |
11557 | ||
11558 | /* | |
7d058285 | 11559 | * sched_balance_newidle() disregards balance intervals, so we could |
5ba553ef | 11560 | * repeatedly reach this code, which would lead to balance_interval |
3b03706f | 11561 | * skyrocketing in a short amount of time. Skip the balance_interval |
5ba553ef | 11562 | * increase logic to avoid that. |
58eeb2d7 QY |
11563 | * |
11564 | * Similarly misfit migration which is not necessarily an indication of | |
11565 | * the system being busy and requires lb to backoff to let it settle | |
11566 | * down. | |
3f130a37 | 11567 | */ |
58eeb2d7 QY |
11568 | if (env.idle == CPU_NEWLY_IDLE || |
11569 | env.migration_type == migrate_misfit) | |
3f130a37 VS |
11570 | goto out; |
11571 | ||
1e3c88bd | 11572 | /* tune up the balancing interval */ |
47b7aee1 VS |
11573 | if ((env.flags & LBF_ALL_PINNED && |
11574 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
11575 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 11576 | sd->balance_interval *= 2; |
1e3c88bd | 11577 | out: |
1e3c88bd PZ |
11578 | return ld_moved; |
11579 | } | |
11580 | ||
52a08ef1 JL |
11581 | static inline unsigned long |
11582 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
11583 | { | |
11584 | unsigned long interval = sd->balance_interval; | |
11585 | ||
11586 | if (cpu_busy) | |
11587 | interval *= sd->busy_factor; | |
11588 | ||
11589 | /* scale ms to jiffies */ | |
11590 | interval = msecs_to_jiffies(interval); | |
e4d32e4d VG |
11591 | |
11592 | /* | |
11593 | * Reduce likelihood of busy balancing at higher domains racing with | |
11594 | * balancing at lower domains by preventing their balancing periods | |
11595 | * from being multiples of each other. | |
11596 | */ | |
11597 | if (cpu_busy) | |
11598 | interval -= 1; | |
11599 | ||
52a08ef1 JL |
11600 | interval = clamp(interval, 1UL, max_load_balance_interval); |
11601 | ||
11602 | return interval; | |
11603 | } | |
11604 | ||
11605 | static inline void | |
31851a98 | 11606 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
11607 | { |
11608 | unsigned long interval, next; | |
11609 | ||
31851a98 LY |
11610 | /* used by idle balance, so cpu_busy = 0 */ |
11611 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
11612 | next = sd->last_balance + interval; |
11613 | ||
11614 | if (time_after(*next_balance, next)) | |
11615 | *next_balance = next; | |
11616 | } | |
11617 | ||
1e3c88bd | 11618 | /* |
97fb7a0a | 11619 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
11620 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
11621 | * least 1 task to be running on each physical CPU where possible, and | |
11622 | * avoids physical / logical imbalances. | |
1e3c88bd | 11623 | */ |
969c7921 | 11624 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 11625 | { |
969c7921 TH |
11626 | struct rq *busiest_rq = data; |
11627 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 11628 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 11629 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 11630 | struct sched_domain *sd; |
e5673f28 | 11631 | struct task_struct *p = NULL; |
8a8c69c3 | 11632 | struct rq_flags rf; |
969c7921 | 11633 | |
8a8c69c3 | 11634 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
11635 | /* |
11636 | * Between queueing the stop-work and running it is a hole in which | |
11637 | * CPUs can become inactive. We should not move tasks from or to | |
11638 | * inactive CPUs. | |
11639 | */ | |
11640 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
11641 | goto out_unlock; | |
969c7921 | 11642 | |
97fb7a0a | 11643 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
11644 | if (unlikely(busiest_cpu != smp_processor_id() || |
11645 | !busiest_rq->active_balance)) | |
11646 | goto out_unlock; | |
1e3c88bd PZ |
11647 | |
11648 | /* Is there any task to move? */ | |
11649 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 11650 | goto out_unlock; |
1e3c88bd PZ |
11651 | |
11652 | /* | |
11653 | * This condition is "impossible", if it occurs | |
11654 | * we need to fix it. Originally reported by | |
97fb7a0a | 11655 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd | 11656 | */ |
09348d75 | 11657 | WARN_ON_ONCE(busiest_rq == target_rq); |
1e3c88bd | 11658 | |
1e3c88bd | 11659 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 11660 | rcu_read_lock(); |
1e3c88bd | 11661 | for_each_domain(target_cpu, sd) { |
e669ac8a VS |
11662 | if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
11663 | break; | |
1e3c88bd PZ |
11664 | } |
11665 | ||
11666 | if (likely(sd)) { | |
8e45cb54 PZ |
11667 | struct lb_env env = { |
11668 | .sd = sd, | |
ddcdf6e7 PZ |
11669 | .dst_cpu = target_cpu, |
11670 | .dst_rq = target_rq, | |
11671 | .src_cpu = busiest_rq->cpu, | |
11672 | .src_rq = busiest_rq, | |
8e45cb54 | 11673 | .idle = CPU_IDLE, |
23fb06d9 | 11674 | .flags = LBF_ACTIVE_LB, |
8e45cb54 PZ |
11675 | }; |
11676 | ||
ae92882e | 11677 | schedstat_inc(sd->alb_count); |
3bed5e21 | 11678 | update_rq_clock(busiest_rq); |
1e3c88bd | 11679 | |
e5673f28 | 11680 | p = detach_one_task(&env); |
d02c0711 | 11681 | if (p) { |
ae92882e | 11682 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
11683 | /* Active balancing done, reset the failure counter. */ |
11684 | sd->nr_balance_failed = 0; | |
11685 | } else { | |
ae92882e | 11686 | schedstat_inc(sd->alb_failed); |
d02c0711 | 11687 | } |
1e3c88bd | 11688 | } |
dce840a0 | 11689 | rcu_read_unlock(); |
969c7921 TH |
11690 | out_unlock: |
11691 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 11692 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
11693 | |
11694 | if (p) | |
11695 | attach_one_task(target_rq, p); | |
11696 | ||
11697 | local_irq_enable(); | |
11698 | ||
969c7921 | 11699 | return 0; |
1e3c88bd PZ |
11700 | } |
11701 | ||
214c1b7f IM |
11702 | /* |
11703 | * This flag serializes load-balancing passes over large domains | |
11704 | * (above the NODE topology level) - only one load-balancing instance | |
11705 | * may run at a time, to reduce overhead on very large systems with | |
11706 | * lots of CPUs and large NUMA distances. | |
11707 | * | |
11708 | * - Note that load-balancing passes triggered while another one | |
11709 | * is executing are skipped and not re-tried. | |
11710 | * | |
11711 | * - Also note that this does not serialize rebalance_domains() | |
11712 | * execution, as non-SD_SERIALIZE domains will still be | |
11713 | * load-balanced in parallel. | |
11714 | */ | |
11715 | static atomic_t sched_balance_running = ATOMIC_INIT(0); | |
af3fe03c PZ |
11716 | |
11717 | /* | |
4c3e509e | 11718 | * Scale the max sched_balance_rq interval with the number of CPUs in the system. |
af3fe03c PZ |
11719 | * This trades load-balance latency on larger machines for less cross talk. |
11720 | */ | |
11721 | void update_max_interval(void) | |
11722 | { | |
11723 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
11724 | } | |
11725 | ||
e60b56e4 VG |
11726 | static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost) |
11727 | { | |
11728 | if (cost > sd->max_newidle_lb_cost) { | |
11729 | /* | |
11730 | * Track max cost of a domain to make sure to not delay the | |
11731 | * next wakeup on the CPU. | |
11732 | */ | |
11733 | sd->max_newidle_lb_cost = cost; | |
11734 | sd->last_decay_max_lb_cost = jiffies; | |
11735 | } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) { | |
11736 | /* | |
11737 | * Decay the newidle max times by ~1% per second to ensure that | |
11738 | * it is not outdated and the current max cost is actually | |
11739 | * shorter. | |
11740 | */ | |
11741 | sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256; | |
11742 | sd->last_decay_max_lb_cost = jiffies; | |
11743 | ||
11744 | return true; | |
11745 | } | |
11746 | ||
11747 | return false; | |
11748 | } | |
11749 | ||
af3fe03c PZ |
11750 | /* |
11751 | * It checks each scheduling domain to see if it is due to be balanced, | |
11752 | * and initiates a balancing operation if so. | |
11753 | * | |
11754 | * Balancing parameters are set up in init_sched_domains. | |
11755 | */ | |
14ff4dbd | 11756 | static void sched_balance_domains(struct rq *rq, enum cpu_idle_type idle) |
af3fe03c PZ |
11757 | { |
11758 | int continue_balancing = 1; | |
11759 | int cpu = rq->cpu; | |
323af6de | 11760 | int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
11761 | unsigned long interval; |
11762 | struct sched_domain *sd; | |
11763 | /* Earliest time when we have to do rebalance again */ | |
11764 | unsigned long next_balance = jiffies + 60*HZ; | |
11765 | int update_next_balance = 0; | |
11766 | int need_serialize, need_decay = 0; | |
11767 | u64 max_cost = 0; | |
11768 | ||
11769 | rcu_read_lock(); | |
11770 | for_each_domain(cpu, sd) { | |
11771 | /* | |
11772 | * Decay the newidle max times here because this is a regular | |
e60b56e4 | 11773 | * visit to all the domains. |
af3fe03c | 11774 | */ |
e60b56e4 | 11775 | need_decay = update_newidle_cost(sd, 0); |
af3fe03c PZ |
11776 | max_cost += sd->max_newidle_lb_cost; |
11777 | ||
af3fe03c PZ |
11778 | /* |
11779 | * Stop the load balance at this level. There is another | |
11780 | * CPU in our sched group which is doing load balancing more | |
11781 | * actively. | |
11782 | */ | |
11783 | if (!continue_balancing) { | |
11784 | if (need_decay) | |
11785 | continue; | |
11786 | break; | |
11787 | } | |
11788 | ||
323af6de | 11789 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
11790 | |
11791 | need_serialize = sd->flags & SD_SERIALIZE; | |
11792 | if (need_serialize) { | |
214c1b7f | 11793 | if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1)) |
af3fe03c PZ |
11794 | goto out; |
11795 | } | |
11796 | ||
11797 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
4c3e509e | 11798 | if (sched_balance_rq(cpu, rq, sd, idle, &continue_balancing)) { |
af3fe03c PZ |
11799 | /* |
11800 | * The LBF_DST_PINNED logic could have changed | |
11801 | * env->dst_cpu, so we can't know our idle | |
11802 | * state even if we migrated tasks. Update it. | |
11803 | */ | |
38d707c5 IM |
11804 | idle = idle_cpu(cpu); |
11805 | busy = !idle && !sched_idle_cpu(cpu); | |
af3fe03c PZ |
11806 | } |
11807 | sd->last_balance = jiffies; | |
323af6de | 11808 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
11809 | } |
11810 | if (need_serialize) | |
214c1b7f | 11811 | atomic_set_release(&sched_balance_running, 0); |
af3fe03c PZ |
11812 | out: |
11813 | if (time_after(next_balance, sd->last_balance + interval)) { | |
11814 | next_balance = sd->last_balance + interval; | |
11815 | update_next_balance = 1; | |
11816 | } | |
11817 | } | |
11818 | if (need_decay) { | |
11819 | /* | |
11820 | * Ensure the rq-wide value also decays but keep it at a | |
11821 | * reasonable floor to avoid funnies with rq->avg_idle. | |
11822 | */ | |
11823 | rq->max_idle_balance_cost = | |
11824 | max((u64)sysctl_sched_migration_cost, max_cost); | |
11825 | } | |
11826 | rcu_read_unlock(); | |
11827 | ||
11828 | /* | |
11829 | * next_balance will be updated only when there is a need. | |
11830 | * When the cpu is attached to null domain for ex, it will not be | |
11831 | * updated. | |
11832 | */ | |
7a82e5f5 | 11833 | if (likely(update_next_balance)) |
af3fe03c PZ |
11834 | rq->next_balance = next_balance; |
11835 | ||
af3fe03c PZ |
11836 | } |
11837 | ||
d987fc7f MG |
11838 | static inline int on_null_domain(struct rq *rq) |
11839 | { | |
11840 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
11841 | } | |
11842 | ||
3451d024 | 11843 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 | 11844 | /* |
7ef7145a IM |
11845 | * NOHZ idle load balancing (ILB) details: |
11846 | * | |
11847 | * - When one of the busy CPUs notices that there may be an idle rebalancing | |
83cd4fe2 VP |
11848 | * needed, they will kick the idle load balancer, which then does idle |
11849 | * load balancing for all the idle CPUs. | |
7ef7145a IM |
11850 | * |
11851 | * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED is not set | |
9b019acb | 11852 | * anywhere yet. |
83cd4fe2 | 11853 | */ |
3dd0337d | 11854 | static inline int find_new_ilb(void) |
1e3c88bd | 11855 | { |
031e3bd8 | 11856 | const struct cpumask *hk_mask; |
b6dd6984 | 11857 | int ilb_cpu; |
1e3c88bd | 11858 | |
04d4e665 | 11859 | hk_mask = housekeeping_cpumask(HK_TYPE_MISC); |
1e3c88bd | 11860 | |
b6dd6984 | 11861 | for_each_cpu_and(ilb_cpu, nohz.idle_cpus_mask, hk_mask) { |
45da7a2b | 11862 | |
b6dd6984 | 11863 | if (ilb_cpu == smp_processor_id()) |
45da7a2b PZ |
11864 | continue; |
11865 | ||
b6dd6984 IM |
11866 | if (idle_cpu(ilb_cpu)) |
11867 | return ilb_cpu; | |
9b019acb | 11868 | } |
786d6dc7 | 11869 | |
f4bb5705 | 11870 | return -1; |
1e3c88bd | 11871 | } |
1e3c88bd | 11872 | |
83cd4fe2 | 11873 | /* |
7ef7145a IM |
11874 | * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU |
11875 | * SMP function call (IPI). | |
11876 | * | |
11877 | * We pick the first idle CPU in the HK_TYPE_MISC housekeeping set (if there is one). | |
83cd4fe2 | 11878 | */ |
a4064fb6 | 11879 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
11880 | { |
11881 | int ilb_cpu; | |
11882 | ||
3ea2f097 VG |
11883 | /* |
11884 | * Increase nohz.next_balance only when if full ilb is triggered but | |
11885 | * not if we only update stats. | |
11886 | */ | |
11887 | if (flags & NOHZ_BALANCE_KICK) | |
11888 | nohz.next_balance = jiffies+1; | |
83cd4fe2 | 11889 | |
3dd0337d | 11890 | ilb_cpu = find_new_ilb(); |
f4bb5705 | 11891 | if (ilb_cpu < 0) |
0b005cf5 | 11892 | return; |
83cd4fe2 | 11893 | |
19a1f5ec PZ |
11894 | /* |
11895 | * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets | |
11896 | * the first flag owns it; cleared by nohz_csd_func(). | |
11897 | */ | |
a4064fb6 | 11898 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 11899 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 11900 | return; |
4550487a | 11901 | |
1c792db7 | 11902 | /* |
90b5363a | 11903 | * This way we generate an IPI on the target CPU which |
7ef7145a | 11904 | * is idle, and the softirq performing NOHZ idle load balancing |
1c792db7 SS |
11905 | * will be run before returning from the IPI. |
11906 | */ | |
90b5363a | 11907 | smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); |
4550487a PZ |
11908 | } |
11909 | ||
11910 | /* | |
9f132742 VS |
11911 | * Current decision point for kicking the idle load balancer in the presence |
11912 | * of idle CPUs in the system. | |
4550487a PZ |
11913 | */ |
11914 | static void nohz_balancer_kick(struct rq *rq) | |
11915 | { | |
11916 | unsigned long now = jiffies; | |
11917 | struct sched_domain_shared *sds; | |
11918 | struct sched_domain *sd; | |
11919 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 11920 | unsigned int flags = 0; |
4550487a PZ |
11921 | |
11922 | if (unlikely(rq->idle_balance)) | |
11923 | return; | |
11924 | ||
11925 | /* | |
11926 | * We may be recently in ticked or tickless idle mode. At the first | |
11927 | * busy tick after returning from idle, we will update the busy stats. | |
11928 | */ | |
00357f5e | 11929 | nohz_balance_exit_idle(rq); |
4550487a PZ |
11930 | |
11931 | /* | |
11932 | * None are in tickless mode and hence no need for NOHZ idle load | |
7ef7145a | 11933 | * balancing: |
4550487a PZ |
11934 | */ |
11935 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
11936 | return; | |
11937 | ||
f643ea22 VG |
11938 | if (READ_ONCE(nohz.has_blocked) && |
11939 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
11940 | flags = NOHZ_STATS_KICK; |
11941 | ||
4550487a | 11942 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 11943 | goto out; |
4550487a | 11944 | |
a0fe2cf0 | 11945 | if (rq->nr_running >= 2) { |
efd984c4 | 11946 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
11947 | goto out; |
11948 | } | |
11949 | ||
11950 | rcu_read_lock(); | |
4550487a PZ |
11951 | |
11952 | sd = rcu_dereference(rq->sd); | |
11953 | if (sd) { | |
e25a7a94 | 11954 | /* |
7ef7145a IM |
11955 | * If there's a runnable CFS task and the current CPU has reduced |
11956 | * capacity, kick the ILB to see if there's a better CPU to run on: | |
e25a7a94 VS |
11957 | */ |
11958 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
efd984c4 | 11959 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
11960 | goto unlock; |
11961 | } | |
11962 | } | |
11963 | ||
011b27bb | 11964 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 11965 | if (sd) { |
b9a7b883 VS |
11966 | /* |
11967 | * When ASYM_PACKING; see if there's a more preferred CPU | |
11968 | * currently idle; in which case, kick the ILB to move tasks | |
11969 | * around. | |
eefefa71 | 11970 | * |
b9e6e286 | 11971 | * When balancing between cores, all the SMT siblings of the |
eefefa71 | 11972 | * preferred CPU must be idle. |
b9a7b883 | 11973 | */ |
7edab78d | 11974 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
45de2062 | 11975 | if (sched_asym(sd, i, cpu)) { |
efd984c4 | 11976 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
11977 | goto unlock; |
11978 | } | |
11979 | } | |
11980 | } | |
b9a7b883 | 11981 | |
a0fe2cf0 VS |
11982 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
11983 | if (sd) { | |
11984 | /* | |
11985 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
11986 | * to run the misfit task on. | |
11987 | */ | |
22d56074 | 11988 | if (check_misfit_status(rq)) { |
efd984c4 | 11989 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
a0fe2cf0 VS |
11990 | goto unlock; |
11991 | } | |
b9a7b883 VS |
11992 | |
11993 | /* | |
11994 | * For asymmetric systems, we do not want to nicely balance | |
11995 | * cache use, instead we want to embrace asymmetry and only | |
11996 | * ensure tasks have enough CPU capacity. | |
11997 | * | |
11998 | * Skip the LLC logic because it's not relevant in that case. | |
11999 | */ | |
12000 | goto unlock; | |
a0fe2cf0 VS |
12001 | } |
12002 | ||
b9a7b883 VS |
12003 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
12004 | if (sds) { | |
e25a7a94 | 12005 | /* |
b9a7b883 | 12006 | * If there is an imbalance between LLC domains (IOW we could |
7ef7145a IM |
12007 | * increase the overall cache utilization), we need a less-loaded LLC |
12008 | * domain to pull some load from. Likewise, we may need to spread | |
b9a7b883 VS |
12009 | * load within the current LLC domain (e.g. packed SMT cores but |
12010 | * other CPUs are idle). We can't really know from here how busy | |
7ef7145a | 12011 | * the others are - so just get a NOHZ balance going if it looks |
b9a7b883 | 12012 | * like this LLC domain has tasks we could move. |
e25a7a94 | 12013 | */ |
b9a7b883 VS |
12014 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
12015 | if (nr_busy > 1) { | |
efd984c4 | 12016 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
b9a7b883 | 12017 | goto unlock; |
4550487a PZ |
12018 | } |
12019 | } | |
12020 | unlock: | |
12021 | rcu_read_unlock(); | |
12022 | out: | |
7fd7a9e0 VS |
12023 | if (READ_ONCE(nohz.needs_update)) |
12024 | flags |= NOHZ_NEXT_KICK; | |
12025 | ||
a4064fb6 PZ |
12026 | if (flags) |
12027 | kick_ilb(flags); | |
83cd4fe2 VP |
12028 | } |
12029 | ||
00357f5e | 12030 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 12031 | { |
00357f5e | 12032 | struct sched_domain *sd; |
a22e47a4 | 12033 | |
00357f5e PZ |
12034 | rcu_read_lock(); |
12035 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 12036 | |
00357f5e PZ |
12037 | if (!sd || !sd->nohz_idle) |
12038 | goto unlock; | |
12039 | sd->nohz_idle = 0; | |
12040 | ||
12041 | atomic_inc(&sd->shared->nr_busy_cpus); | |
12042 | unlock: | |
12043 | rcu_read_unlock(); | |
71325960 SS |
12044 | } |
12045 | ||
00357f5e PZ |
12046 | void nohz_balance_exit_idle(struct rq *rq) |
12047 | { | |
12048 | SCHED_WARN_ON(rq != this_rq()); | |
12049 | ||
12050 | if (likely(!rq->nohz_tick_stopped)) | |
12051 | return; | |
12052 | ||
12053 | rq->nohz_tick_stopped = 0; | |
12054 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
12055 | atomic_dec(&nohz.nr_cpus); | |
12056 | ||
12057 | set_cpu_sd_state_busy(rq->cpu); | |
12058 | } | |
12059 | ||
12060 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
12061 | { |
12062 | struct sched_domain *sd; | |
69e1e811 | 12063 | |
69e1e811 | 12064 | rcu_read_lock(); |
0e369d75 | 12065 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
12066 | |
12067 | if (!sd || sd->nohz_idle) | |
12068 | goto unlock; | |
12069 | sd->nohz_idle = 1; | |
12070 | ||
0e369d75 | 12071 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 12072 | unlock: |
69e1e811 SS |
12073 | rcu_read_unlock(); |
12074 | } | |
12075 | ||
1e3c88bd | 12076 | /* |
97fb7a0a | 12077 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 12078 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 12079 | */ |
c1cc017c | 12080 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 12081 | { |
00357f5e PZ |
12082 | struct rq *rq = cpu_rq(cpu); |
12083 | ||
12084 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
12085 | ||
97fb7a0a | 12086 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
12087 | if (!cpu_active(cpu)) |
12088 | return; | |
12089 | ||
387bc8b5 | 12090 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
04d4e665 | 12091 | if (!housekeeping_cpu(cpu, HK_TYPE_SCHED)) |
387bc8b5 FW |
12092 | return; |
12093 | ||
f643ea22 VG |
12094 | /* |
12095 | * Can be set safely without rq->lock held | |
12096 | * If a clear happens, it will have evaluated last additions because | |
12097 | * rq->lock is held during the check and the clear | |
12098 | */ | |
12099 | rq->has_blocked_load = 1; | |
12100 | ||
12101 | /* | |
12102 | * The tick is still stopped but load could have been added in the | |
12103 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
12104 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
12105 | * of nohz.has_blocked can only happen after checking the new load | |
12106 | */ | |
00357f5e | 12107 | if (rq->nohz_tick_stopped) |
f643ea22 | 12108 | goto out; |
1e3c88bd | 12109 | |
97fb7a0a | 12110 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 12111 | if (on_null_domain(rq)) |
d987fc7f MG |
12112 | return; |
12113 | ||
00357f5e PZ |
12114 | rq->nohz_tick_stopped = 1; |
12115 | ||
c1cc017c AS |
12116 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
12117 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 12118 | |
f643ea22 VG |
12119 | /* |
12120 | * Ensures that if nohz_idle_balance() fails to observe our | |
12121 | * @idle_cpus_mask store, it must observe the @has_blocked | |
7fd7a9e0 | 12122 | * and @needs_update stores. |
f643ea22 VG |
12123 | */ |
12124 | smp_mb__after_atomic(); | |
12125 | ||
00357f5e | 12126 | set_cpu_sd_state_idle(cpu); |
f643ea22 | 12127 | |
7fd7a9e0 | 12128 | WRITE_ONCE(nohz.needs_update, 1); |
f643ea22 VG |
12129 | out: |
12130 | /* | |
12131 | * Each time a cpu enter idle, we assume that it has blocked load and | |
b9e6e286 | 12132 | * enable the periodic update of the load of idle CPUs |
f643ea22 VG |
12133 | */ |
12134 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 12135 | } |
1e3c88bd | 12136 | |
3f5ad914 Y |
12137 | static bool update_nohz_stats(struct rq *rq) |
12138 | { | |
12139 | unsigned int cpu = rq->cpu; | |
12140 | ||
12141 | if (!rq->has_blocked_load) | |
12142 | return false; | |
12143 | ||
12144 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) | |
12145 | return false; | |
12146 | ||
12147 | if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) | |
12148 | return true; | |
12149 | ||
391b7a53 | 12150 | sched_balance_update_blocked_averages(cpu); |
3f5ad914 Y |
12151 | |
12152 | return rq->has_blocked_load; | |
12153 | } | |
12154 | ||
1e3c88bd | 12155 | /* |
b9e6e286 | 12156 | * Internal function that runs load balance for all idle CPUs. The load balance |
31e77c93 VG |
12157 | * can be a simple update of blocked load or a complete load balance with |
12158 | * tasks movement depending of flags. | |
1e3c88bd | 12159 | */ |
d985ee9f | 12160 | static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags) |
83cd4fe2 | 12161 | { |
c5afb6a8 | 12162 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
12163 | unsigned long now = jiffies; |
12164 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 12165 | bool has_blocked_load = false; |
c5afb6a8 | 12166 | int update_next_balance = 0; |
b7031a02 | 12167 | int this_cpu = this_rq->cpu; |
b7031a02 PZ |
12168 | int balance_cpu; |
12169 | struct rq *rq; | |
83cd4fe2 | 12170 | |
b7031a02 | 12171 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 12172 | |
f643ea22 VG |
12173 | /* |
12174 | * We assume there will be no idle load after this update and clear | |
12175 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
7fd7a9e0 | 12176 | * set the has_blocked flag and trigger another update of idle load. |
f643ea22 VG |
12177 | * Because a cpu that becomes idle, is added to idle_cpus_mask before |
12178 | * setting the flag, we are sure to not clear the state and not | |
12179 | * check the load of an idle cpu. | |
7fd7a9e0 VS |
12180 | * |
12181 | * Same applies to idle_cpus_mask vs needs_update. | |
f643ea22 | 12182 | */ |
efd984c4 VS |
12183 | if (flags & NOHZ_STATS_KICK) |
12184 | WRITE_ONCE(nohz.has_blocked, 0); | |
7fd7a9e0 VS |
12185 | if (flags & NOHZ_NEXT_KICK) |
12186 | WRITE_ONCE(nohz.needs_update, 0); | |
f643ea22 VG |
12187 | |
12188 | /* | |
12189 | * Ensures that if we miss the CPU, we must see the has_blocked | |
12190 | * store from nohz_balance_enter_idle(). | |
12191 | */ | |
12192 | smp_mb(); | |
12193 | ||
7a82e5f5 VG |
12194 | /* |
12195 | * Start with the next CPU after this_cpu so we will end with this_cpu and let a | |
12196 | * chance for other idle cpu to pull load. | |
12197 | */ | |
12198 | for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) { | |
12199 | if (!idle_cpu(balance_cpu)) | |
83cd4fe2 VP |
12200 | continue; |
12201 | ||
12202 | /* | |
97fb7a0a IM |
12203 | * If this CPU gets work to do, stop the load balancing |
12204 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
12205 | * balancing owner will pick it up. |
12206 | */ | |
f643ea22 | 12207 | if (need_resched()) { |
efd984c4 VS |
12208 | if (flags & NOHZ_STATS_KICK) |
12209 | has_blocked_load = true; | |
7fd7a9e0 VS |
12210 | if (flags & NOHZ_NEXT_KICK) |
12211 | WRITE_ONCE(nohz.needs_update, 1); | |
f643ea22 VG |
12212 | goto abort; |
12213 | } | |
83cd4fe2 | 12214 | |
5ed4f1d9 VG |
12215 | rq = cpu_rq(balance_cpu); |
12216 | ||
efd984c4 VS |
12217 | if (flags & NOHZ_STATS_KICK) |
12218 | has_blocked_load |= update_nohz_stats(rq); | |
f643ea22 | 12219 | |
ed61bbc6 TC |
12220 | /* |
12221 | * If time for next balance is due, | |
12222 | * do the balance. | |
12223 | */ | |
12224 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
12225 | struct rq_flags rf; |
12226 | ||
31e77c93 | 12227 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 12228 | update_rq_clock(rq); |
31e77c93 | 12229 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 12230 | |
b7031a02 | 12231 | if (flags & NOHZ_BALANCE_KICK) |
14ff4dbd | 12232 | sched_balance_domains(rq, CPU_IDLE); |
ed61bbc6 | 12233 | } |
83cd4fe2 | 12234 | |
c5afb6a8 VG |
12235 | if (time_after(next_balance, rq->next_balance)) { |
12236 | next_balance = rq->next_balance; | |
12237 | update_next_balance = 1; | |
12238 | } | |
83cd4fe2 | 12239 | } |
c5afb6a8 | 12240 | |
3ea2f097 VG |
12241 | /* |
12242 | * next_balance will be updated only when there is a need. | |
12243 | * When the CPU is attached to null domain for ex, it will not be | |
12244 | * updated. | |
12245 | */ | |
12246 | if (likely(update_next_balance)) | |
12247 | nohz.next_balance = next_balance; | |
12248 | ||
efd984c4 VS |
12249 | if (flags & NOHZ_STATS_KICK) |
12250 | WRITE_ONCE(nohz.next_blocked, | |
12251 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
f643ea22 VG |
12252 | |
12253 | abort: | |
12254 | /* There is still blocked load, enable periodic update */ | |
12255 | if (has_blocked_load) | |
12256 | WRITE_ONCE(nohz.has_blocked, 1); | |
31e77c93 VG |
12257 | } |
12258 | ||
12259 | /* | |
12260 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
b9e6e286 | 12261 | * rebalancing for all the CPUs for whom scheduler ticks are stopped. |
31e77c93 VG |
12262 | */ |
12263 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
12264 | { | |
19a1f5ec | 12265 | unsigned int flags = this_rq->nohz_idle_balance; |
31e77c93 | 12266 | |
19a1f5ec | 12267 | if (!flags) |
31e77c93 VG |
12268 | return false; |
12269 | ||
19a1f5ec | 12270 | this_rq->nohz_idle_balance = 0; |
31e77c93 | 12271 | |
19a1f5ec | 12272 | if (idle != CPU_IDLE) |
31e77c93 VG |
12273 | return false; |
12274 | ||
d985ee9f | 12275 | _nohz_idle_balance(this_rq, flags); |
31e77c93 | 12276 | |
b7031a02 | 12277 | return true; |
83cd4fe2 | 12278 | } |
31e77c93 | 12279 | |
c6f88654 | 12280 | /* |
fb064e5a JFG |
12281 | * Check if we need to directly run the ILB for updating blocked load before |
12282 | * entering idle state. Here we run ILB directly without issuing IPIs. | |
12283 | * | |
12284 | * Note that when this function is called, the tick may not yet be stopped on | |
12285 | * this CPU yet. nohz.idle_cpus_mask is updated only when tick is stopped and | |
12286 | * cleared on the next busy tick. In other words, nohz.idle_cpus_mask updates | |
12287 | * don't align with CPUs enter/exit idle to avoid bottlenecks due to high idle | |
12288 | * entry/exit rate (usec). So it is possible that _nohz_idle_balance() is | |
12289 | * called from this function on (this) CPU that's not yet in the mask. That's | |
12290 | * OK because the goal of nohz_run_idle_balance() is to run ILB only for | |
12291 | * updating the blocked load of already idle CPUs without waking up one of | |
b9e6e286 | 12292 | * those idle CPUs and outside the preempt disable / IRQ off phase of the local |
fb064e5a | 12293 | * cpu about to enter idle, because it can take a long time. |
c6f88654 VG |
12294 | */ |
12295 | void nohz_run_idle_balance(int cpu) | |
12296 | { | |
12297 | unsigned int flags; | |
12298 | ||
12299 | flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu)); | |
12300 | ||
12301 | /* | |
12302 | * Update the blocked load only if no SCHED_SOFTIRQ is about to happen | |
b9e6e286 | 12303 | * (i.e. NOHZ_STATS_KICK set) and will do the same. |
c6f88654 VG |
12304 | */ |
12305 | if ((flags == NOHZ_NEWILB_KICK) && !need_resched()) | |
d985ee9f | 12306 | _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK); |
c6f88654 VG |
12307 | } |
12308 | ||
31e77c93 VG |
12309 | static void nohz_newidle_balance(struct rq *this_rq) |
12310 | { | |
12311 | int this_cpu = this_rq->cpu; | |
12312 | ||
12313 | /* | |
12314 | * This CPU doesn't want to be disturbed by scheduler | |
12315 | * housekeeping | |
12316 | */ | |
04d4e665 | 12317 | if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED)) |
31e77c93 VG |
12318 | return; |
12319 | ||
12320 | /* Will wake up very soon. No time for doing anything else*/ | |
12321 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
12322 | return; | |
12323 | ||
12324 | /* Don't need to update blocked load of idle CPUs*/ | |
12325 | if (!READ_ONCE(nohz.has_blocked) || | |
12326 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
12327 | return; | |
12328 | ||
31e77c93 | 12329 | /* |
c6f88654 VG |
12330 | * Set the need to trigger ILB in order to update blocked load |
12331 | * before entering idle state. | |
31e77c93 | 12332 | */ |
c6f88654 | 12333 | atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu)); |
31e77c93 VG |
12334 | } |
12335 | ||
dd707247 PZ |
12336 | #else /* !CONFIG_NO_HZ_COMMON */ |
12337 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
12338 | ||
31e77c93 | 12339 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
12340 | { |
12341 | return false; | |
12342 | } | |
31e77c93 VG |
12343 | |
12344 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 12345 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 12346 | |
47ea5412 | 12347 | /* |
7d058285 | 12348 | * sched_balance_newidle is called by schedule() if this_cpu is about to become |
47ea5412 | 12349 | * idle. Attempts to pull tasks from other CPUs. |
7277a34c PZ |
12350 | * |
12351 | * Returns: | |
12352 | * < 0 - we released the lock and there are !fair tasks present | |
12353 | * 0 - failed, no new tasks | |
12354 | * > 0 - success, new (fair) tasks present | |
47ea5412 | 12355 | */ |
7d058285 | 12356 | static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf) |
47ea5412 PZ |
12357 | { |
12358 | unsigned long next_balance = jiffies + HZ; | |
12359 | int this_cpu = this_rq->cpu; | |
c829d681 | 12360 | int continue_balancing = 1; |
9e9af819 | 12361 | u64 t0, t1, curr_cost = 0; |
47ea5412 PZ |
12362 | struct sched_domain *sd; |
12363 | int pulled_task = 0; | |
47ea5412 | 12364 | |
5ba553ef | 12365 | update_misfit_status(NULL, this_rq); |
e5e678e4 RR |
12366 | |
12367 | /* | |
12368 | * There is a task waiting to run. No need to search for one. | |
12369 | * Return 0; the task will be enqueued when switching to idle. | |
12370 | */ | |
12371 | if (this_rq->ttwu_pending) | |
12372 | return 0; | |
12373 | ||
47ea5412 | 12374 | /* |
c829d681 SH |
12375 | * We must set idle_stamp _before_ calling sched_balance_rq() |
12376 | * for CPU_NEWLY_IDLE, such that we measure the this duration | |
12377 | * as idle time. | |
47ea5412 PZ |
12378 | */ |
12379 | this_rq->idle_stamp = rq_clock(this_rq); | |
12380 | ||
12381 | /* | |
12382 | * Do not pull tasks towards !active CPUs... | |
12383 | */ | |
12384 | if (!cpu_active(this_cpu)) | |
12385 | return 0; | |
12386 | ||
12387 | /* | |
12388 | * This is OK, because current is on_cpu, which avoids it being picked | |
12389 | * for load-balance and preemption/IRQs are still disabled avoiding | |
12390 | * further scheduler activity on it and we're being very careful to | |
12391 | * re-start the picking loop. | |
12392 | */ | |
12393 | rq_unpin_lock(this_rq, rf); | |
12394 | ||
9d783c8d VG |
12395 | rcu_read_lock(); |
12396 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
12397 | ||
76cc4f91 | 12398 | if (!get_rd_overloaded(this_rq->rd) || |
9d783c8d | 12399 | (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) { |
31e77c93 | 12400 | |
47ea5412 PZ |
12401 | if (sd) |
12402 | update_next_balance(sd, &next_balance); | |
12403 | rcu_read_unlock(); | |
12404 | ||
12405 | goto out; | |
12406 | } | |
9d783c8d | 12407 | rcu_read_unlock(); |
47ea5412 | 12408 | |
5cb9eaa3 | 12409 | raw_spin_rq_unlock(this_rq); |
47ea5412 | 12410 | |
9e9af819 | 12411 | t0 = sched_clock_cpu(this_cpu); |
391b7a53 | 12412 | sched_balance_update_blocked_averages(this_cpu); |
9e9af819 | 12413 | |
47ea5412 PZ |
12414 | rcu_read_lock(); |
12415 | for_each_domain(this_cpu, sd) { | |
9e9af819 | 12416 | u64 domain_cost; |
47ea5412 | 12417 | |
8ea9183d VG |
12418 | update_next_balance(sd, &next_balance); |
12419 | ||
12420 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) | |
47ea5412 | 12421 | break; |
47ea5412 PZ |
12422 | |
12423 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
47ea5412 | 12424 | |
4c3e509e | 12425 | pulled_task = sched_balance_rq(this_cpu, this_rq, |
47ea5412 PZ |
12426 | sd, CPU_NEWLY_IDLE, |
12427 | &continue_balancing); | |
12428 | ||
9e9af819 VG |
12429 | t1 = sched_clock_cpu(this_cpu); |
12430 | domain_cost = t1 - t0; | |
e60b56e4 | 12431 | update_newidle_cost(sd, domain_cost); |
47ea5412 PZ |
12432 | |
12433 | curr_cost += domain_cost; | |
9e9af819 | 12434 | t0 = t1; |
47ea5412 PZ |
12435 | } |
12436 | ||
47ea5412 PZ |
12437 | /* |
12438 | * Stop searching for tasks to pull if there are | |
12439 | * now runnable tasks on this rq. | |
12440 | */ | |
c829d681 | 12441 | if (pulled_task || !continue_balancing) |
47ea5412 PZ |
12442 | break; |
12443 | } | |
12444 | rcu_read_unlock(); | |
12445 | ||
5cb9eaa3 | 12446 | raw_spin_rq_lock(this_rq); |
47ea5412 PZ |
12447 | |
12448 | if (curr_cost > this_rq->max_idle_balance_cost) | |
12449 | this_rq->max_idle_balance_cost = curr_cost; | |
12450 | ||
12451 | /* | |
12452 | * While browsing the domains, we released the rq lock, a task could | |
12453 | * have been enqueued in the meantime. Since we're not going idle, | |
12454 | * pretend we pulled a task. | |
12455 | */ | |
12456 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
12457 | pulled_task = 1; | |
12458 | ||
47ea5412 PZ |
12459 | /* Is there a task of a high priority class? */ |
12460 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
12461 | pulled_task = -1; | |
12462 | ||
6553fc18 VG |
12463 | out: |
12464 | /* Move the next balance forward */ | |
12465 | if (time_after(this_rq->next_balance, next_balance)) | |
12466 | this_rq->next_balance = next_balance; | |
12467 | ||
47ea5412 PZ |
12468 | if (pulled_task) |
12469 | this_rq->idle_stamp = 0; | |
0826530d VG |
12470 | else |
12471 | nohz_newidle_balance(this_rq); | |
47ea5412 PZ |
12472 | |
12473 | rq_repin_lock(this_rq, rf); | |
12474 | ||
12475 | return pulled_task; | |
12476 | } | |
12477 | ||
83cd4fe2 | 12478 | /* |
3dc6f6c8 IM |
12479 | * This softirq handler is triggered via SCHED_SOFTIRQ from two places: |
12480 | * | |
12481 | * - directly from the local scheduler_tick() for periodic load balancing | |
12482 | * | |
12483 | * - indirectly from a remote scheduler_tick() for NOHZ idle balancing | |
12484 | * through the SMP cross-call nohz_csd_func() | |
83cd4fe2 | 12485 | */ |
70a27d6d | 12486 | static __latent_entropy void sched_balance_softirq(struct softirq_action *h) |
1e3c88bd | 12487 | { |
208cb16b | 12488 | struct rq *this_rq = this_rq(); |
38d707c5 | 12489 | enum cpu_idle_type idle = this_rq->idle_balance; |
1e3c88bd | 12490 | /* |
3a5fe930 | 12491 | * If this CPU has a pending NOHZ_BALANCE_KICK, then do the |
97fb7a0a | 12492 | * balancing on behalf of the other idle CPUs whose ticks are |
14ff4dbd | 12493 | * stopped. Do nohz_idle_balance *before* sched_balance_domains to |
97fb7a0a | 12494 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
12495 | * load balance only within the local sched_domain hierarchy |
12496 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 12497 | */ |
b7031a02 PZ |
12498 | if (nohz_idle_balance(this_rq, idle)) |
12499 | return; | |
12500 | ||
12501 | /* normal load balance */ | |
391b7a53 | 12502 | sched_balance_update_blocked_averages(this_rq->cpu); |
14ff4dbd | 12503 | sched_balance_domains(this_rq, idle); |
1e3c88bd PZ |
12504 | } |
12505 | ||
1e3c88bd PZ |
12506 | /* |
12507 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 12508 | */ |
983be062 | 12509 | void sched_balance_trigger(struct rq *rq) |
1e3c88bd | 12510 | { |
e0b257c3 AMB |
12511 | /* |
12512 | * Don't need to rebalance while attached to NULL domain or | |
12513 | * runqueue CPU is not active | |
12514 | */ | |
12515 | if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq)))) | |
c726099e DL |
12516 | return; |
12517 | ||
12518 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 12519 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
12520 | |
12521 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
12522 | } |
12523 | ||
0bcdcf28 CE |
12524 | static void rq_online_fair(struct rq *rq) |
12525 | { | |
12526 | update_sysctl(); | |
0e59bdae KT |
12527 | |
12528 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
12529 | } |
12530 | ||
12531 | static void rq_offline_fair(struct rq *rq) | |
12532 | { | |
12533 | update_sysctl(); | |
a4c96ae3 PB |
12534 | |
12535 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
12536 | unthrottle_offline_cfs_rqs(rq); | |
f60a631a VG |
12537 | |
12538 | /* Ensure that we remove rq contribution to group share: */ | |
12539 | clear_tg_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
12540 | } |
12541 | ||
55e12e5e | 12542 | #endif /* CONFIG_SMP */ |
e1d1484f | 12543 | |
8039e96f VP |
12544 | #ifdef CONFIG_SCHED_CORE |
12545 | static inline bool | |
12546 | __entity_slice_used(struct sched_entity *se, int min_nr_tasks) | |
12547 | { | |
8039e96f | 12548 | u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime; |
147f3efa | 12549 | u64 slice = se->slice; |
8039e96f VP |
12550 | |
12551 | return (rtime * min_nr_tasks > slice); | |
12552 | } | |
12553 | ||
12554 | #define MIN_NR_TASKS_DURING_FORCEIDLE 2 | |
12555 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) | |
12556 | { | |
12557 | if (!sched_core_enabled(rq)) | |
12558 | return; | |
12559 | ||
12560 | /* | |
12561 | * If runqueue has only one task which used up its slice and | |
12562 | * if the sibling is forced idle, then trigger schedule to | |
12563 | * give forced idle task a chance. | |
12564 | * | |
12565 | * sched_slice() considers only this active rq and it gets the | |
12566 | * whole slice. But during force idle, we have siblings acting | |
12567 | * like a single runqueue and hence we need to consider runnable | |
cc00c198 | 12568 | * tasks on this CPU and the forced idle CPU. Ideally, we should |
8039e96f | 12569 | * go through the forced idle rq, but that would be a perf hit. |
cc00c198 | 12570 | * We can assume that the forced idle CPU has at least |
8039e96f | 12571 | * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check |
cc00c198 | 12572 | * if we need to give up the CPU. |
8039e96f | 12573 | */ |
4feee7d1 | 12574 | if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 && |
8039e96f VP |
12575 | __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE)) |
12576 | resched_curr(rq); | |
12577 | } | |
c6047c2e JFG |
12578 | |
12579 | /* | |
12580 | * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed. | |
12581 | */ | |
904cbab7 MWO |
12582 | static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq, |
12583 | bool forceidle) | |
c6047c2e JFG |
12584 | { |
12585 | for_each_sched_entity(se) { | |
12586 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
12587 | ||
12588 | if (forceidle) { | |
12589 | if (cfs_rq->forceidle_seq == fi_seq) | |
12590 | break; | |
12591 | cfs_rq->forceidle_seq = fi_seq; | |
12592 | } | |
12593 | ||
12594 | cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime; | |
12595 | } | |
12596 | } | |
12597 | ||
12598 | void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi) | |
12599 | { | |
12600 | struct sched_entity *se = &p->se; | |
12601 | ||
12602 | if (p->sched_class != &fair_sched_class) | |
12603 | return; | |
12604 | ||
12605 | se_fi_update(se, rq->core->core_forceidle_seq, in_fi); | |
12606 | } | |
12607 | ||
904cbab7 MWO |
12608 | bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b, |
12609 | bool in_fi) | |
c6047c2e JFG |
12610 | { |
12611 | struct rq *rq = task_rq(a); | |
904cbab7 MWO |
12612 | const struct sched_entity *sea = &a->se; |
12613 | const struct sched_entity *seb = &b->se; | |
c6047c2e JFG |
12614 | struct cfs_rq *cfs_rqa; |
12615 | struct cfs_rq *cfs_rqb; | |
12616 | s64 delta; | |
12617 | ||
12618 | SCHED_WARN_ON(task_rq(b)->core != rq->core); | |
12619 | ||
12620 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
12621 | /* | |
12622 | * Find an se in the hierarchy for tasks a and b, such that the se's | |
12623 | * are immediate siblings. | |
12624 | */ | |
12625 | while (sea->cfs_rq->tg != seb->cfs_rq->tg) { | |
12626 | int sea_depth = sea->depth; | |
12627 | int seb_depth = seb->depth; | |
12628 | ||
12629 | if (sea_depth >= seb_depth) | |
12630 | sea = parent_entity(sea); | |
12631 | if (sea_depth <= seb_depth) | |
12632 | seb = parent_entity(seb); | |
12633 | } | |
12634 | ||
12635 | se_fi_update(sea, rq->core->core_forceidle_seq, in_fi); | |
12636 | se_fi_update(seb, rq->core->core_forceidle_seq, in_fi); | |
12637 | ||
12638 | cfs_rqa = sea->cfs_rq; | |
12639 | cfs_rqb = seb->cfs_rq; | |
12640 | #else | |
12641 | cfs_rqa = &task_rq(a)->cfs; | |
12642 | cfs_rqb = &task_rq(b)->cfs; | |
12643 | #endif | |
12644 | ||
12645 | /* | |
12646 | * Find delta after normalizing se's vruntime with its cfs_rq's | |
12647 | * min_vruntime_fi, which would have been updated in prior calls | |
12648 | * to se_fi_update(). | |
12649 | */ | |
12650 | delta = (s64)(sea->vruntime - seb->vruntime) + | |
12651 | (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi); | |
12652 | ||
12653 | return delta > 0; | |
12654 | } | |
530bfad1 HJ |
12655 | |
12656 | static int task_is_throttled_fair(struct task_struct *p, int cpu) | |
12657 | { | |
12658 | struct cfs_rq *cfs_rq; | |
12659 | ||
12660 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
12661 | cfs_rq = task_group(p)->cfs_rq[cpu]; | |
12662 | #else | |
12663 | cfs_rq = &cpu_rq(cpu)->cfs; | |
12664 | #endif | |
12665 | return throttled_hierarchy(cfs_rq); | |
12666 | } | |
8039e96f VP |
12667 | #else |
12668 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {} | |
12669 | #endif | |
12670 | ||
bf0f6f24 | 12671 | /* |
d84b3131 FW |
12672 | * scheduler tick hitting a task of our scheduling class. |
12673 | * | |
12674 | * NOTE: This function can be called remotely by the tick offload that | |
12675 | * goes along full dynticks. Therefore no local assumption can be made | |
12676 | * and everything must be accessed through the @rq and @curr passed in | |
12677 | * parameters. | |
bf0f6f24 | 12678 | */ |
8f4d37ec | 12679 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
12680 | { |
12681 | struct cfs_rq *cfs_rq; | |
12682 | struct sched_entity *se = &curr->se; | |
12683 | ||
12684 | for_each_sched_entity(se) { | |
12685 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 12686 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 12687 | } |
18bf2805 | 12688 | |
b52da86e | 12689 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 12690 | task_tick_numa(rq, curr); |
3b1baa64 MR |
12691 | |
12692 | update_misfit_status(curr, rq); | |
be3a51e6 | 12693 | check_update_overutilized_status(task_rq(curr)); |
8039e96f VP |
12694 | |
12695 | task_tick_core(rq, curr); | |
bf0f6f24 IM |
12696 | } |
12697 | ||
12698 | /* | |
cd29fe6f PZ |
12699 | * called on fork with the child task as argument from the parent's context |
12700 | * - child not yet on the tasklist | |
12701 | * - preemption disabled | |
bf0f6f24 | 12702 | */ |
cd29fe6f | 12703 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 12704 | { |
4fc420c9 | 12705 | struct sched_entity *se = &p->se, *curr; |
e8f331bc | 12706 | struct cfs_rq *cfs_rq; |
cd29fe6f | 12707 | struct rq *rq = this_rq(); |
8a8c69c3 | 12708 | struct rq_flags rf; |
bf0f6f24 | 12709 | |
8a8c69c3 | 12710 | rq_lock(rq, &rf); |
861d034e PZ |
12711 | update_rq_clock(rq); |
12712 | ||
22d56074 QY |
12713 | set_task_max_allowed_capacity(p); |
12714 | ||
4fc420c9 DN |
12715 | cfs_rq = task_cfs_rq(current); |
12716 | curr = cfs_rq->curr; | |
e8f331bc | 12717 | if (curr) |
e210bffd | 12718 | update_curr(cfs_rq); |
d07f09a1 | 12719 | place_entity(cfs_rq, se, ENQUEUE_INITIAL); |
8a8c69c3 | 12720 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
12721 | } |
12722 | ||
cb469845 SR |
12723 | /* |
12724 | * Priority of the task has changed. Check to see if we preempt | |
12725 | * the current task. | |
12726 | */ | |
da7a735e PZ |
12727 | static void |
12728 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 12729 | { |
da0c1e65 | 12730 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
12731 | return; |
12732 | ||
7c2e8bbd FW |
12733 | if (rq->cfs.nr_running == 1) |
12734 | return; | |
12735 | ||
cb469845 SR |
12736 | /* |
12737 | * Reschedule if we are currently running on this runqueue and | |
12738 | * our priority decreased, or if we are not currently running on | |
12739 | * this runqueue and our priority is higher than the current's | |
12740 | */ | |
65bcf072 | 12741 | if (task_current(rq, p)) { |
cb469845 | 12742 | if (p->prio > oldprio) |
8875125e | 12743 | resched_curr(rq); |
cb469845 | 12744 | } else |
e23edc86 | 12745 | wakeup_preempt(rq, p, 0); |
cb469845 SR |
12746 | } |
12747 | ||
09a43ace VG |
12748 | #ifdef CONFIG_FAIR_GROUP_SCHED |
12749 | /* | |
12750 | * Propagate the changes of the sched_entity across the tg tree to make it | |
12751 | * visible to the root | |
12752 | */ | |
12753 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
12754 | { | |
51bf903b CZ |
12755 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
12756 | ||
12757 | if (cfs_rq_throttled(cfs_rq)) | |
12758 | return; | |
09a43ace | 12759 | |
51bf903b CZ |
12760 | if (!throttled_hierarchy(cfs_rq)) |
12761 | list_add_leaf_cfs_rq(cfs_rq); | |
0258bdfa | 12762 | |
09a43ace VG |
12763 | /* Start to propagate at parent */ |
12764 | se = se->parent; | |
12765 | ||
12766 | for_each_sched_entity(se) { | |
12767 | cfs_rq = cfs_rq_of(se); | |
12768 | ||
51bf903b | 12769 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace | 12770 | |
51bf903b | 12771 | if (cfs_rq_throttled(cfs_rq)) |
0258bdfa | 12772 | break; |
51bf903b CZ |
12773 | |
12774 | if (!throttled_hierarchy(cfs_rq)) | |
12775 | list_add_leaf_cfs_rq(cfs_rq); | |
09a43ace VG |
12776 | } |
12777 | } | |
12778 | #else | |
12779 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
12780 | #endif | |
12781 | ||
df217913 | 12782 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 12783 | { |
daa59407 BP |
12784 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
12785 | ||
7e2edaf6 CZ |
12786 | #ifdef CONFIG_SMP |
12787 | /* | |
12788 | * In case the task sched_avg hasn't been attached: | |
12789 | * - A forked task which hasn't been woken up by wake_up_new_task(). | |
12790 | * - A task which has been woken up by try_to_wake_up() but is | |
12791 | * waiting for actually being woken up by sched_ttwu_pending(). | |
12792 | */ | |
12793 | if (!se->avg.last_update_time) | |
12794 | return; | |
12795 | #endif | |
12796 | ||
9d89c257 | 12797 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 12798 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 12799 | detach_entity_load_avg(cfs_rq, se); |
fe749158 | 12800 | update_tg_load_avg(cfs_rq); |
09a43ace | 12801 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
12802 | } |
12803 | ||
df217913 | 12804 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 12805 | { |
daa59407 | 12806 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a | 12807 | |
df217913 | 12808 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 12809 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
a4f9a0e5 | 12810 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 12811 | update_tg_load_avg(cfs_rq); |
09a43ace | 12812 | propagate_entity_cfs_rq(se); |
df217913 VG |
12813 | } |
12814 | ||
12815 | static void detach_task_cfs_rq(struct task_struct *p) | |
12816 | { | |
12817 | struct sched_entity *se = &p->se; | |
df217913 VG |
12818 | |
12819 | detach_entity_cfs_rq(se); | |
12820 | } | |
12821 | ||
12822 | static void attach_task_cfs_rq(struct task_struct *p) | |
12823 | { | |
12824 | struct sched_entity *se = &p->se; | |
df217913 VG |
12825 | |
12826 | attach_entity_cfs_rq(se); | |
daa59407 | 12827 | } |
6efdb105 | 12828 | |
daa59407 BP |
12829 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
12830 | { | |
12831 | detach_task_cfs_rq(p); | |
12832 | } | |
12833 | ||
12834 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
12835 | { | |
12836 | attach_task_cfs_rq(p); | |
7855a35a | 12837 | |
22d56074 QY |
12838 | set_task_max_allowed_capacity(p); |
12839 | ||
daa59407 | 12840 | if (task_on_rq_queued(p)) { |
7855a35a | 12841 | /* |
daa59407 BP |
12842 | * We were most likely switched from sched_rt, so |
12843 | * kick off the schedule if running, otherwise just see | |
12844 | * if we can still preempt the current task. | |
7855a35a | 12845 | */ |
65bcf072 | 12846 | if (task_current(rq, p)) |
daa59407 BP |
12847 | resched_curr(rq); |
12848 | else | |
e23edc86 | 12849 | wakeup_preempt(rq, p, 0); |
7855a35a | 12850 | } |
cb469845 SR |
12851 | } |
12852 | ||
83b699ed SV |
12853 | /* Account for a task changing its policy or group. |
12854 | * | |
12855 | * This routine is mostly called to set cfs_rq->curr field when a task | |
12856 | * migrates between groups/classes. | |
12857 | */ | |
a0e813f2 | 12858 | static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) |
83b699ed | 12859 | { |
03b7fad1 PZ |
12860 | struct sched_entity *se = &p->se; |
12861 | ||
12862 | #ifdef CONFIG_SMP | |
12863 | if (task_on_rq_queued(p)) { | |
12864 | /* | |
12865 | * Move the next running task to the front of the list, so our | |
12866 | * cfs_tasks list becomes MRU one. | |
12867 | */ | |
12868 | list_move(&se->group_node, &rq->cfs_tasks); | |
12869 | } | |
12870 | #endif | |
83b699ed | 12871 | |
ec12cb7f PT |
12872 | for_each_sched_entity(se) { |
12873 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
12874 | ||
12875 | set_next_entity(cfs_rq, se); | |
12876 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
12877 | account_cfs_rq_runtime(cfs_rq, 0); | |
12878 | } | |
83b699ed SV |
12879 | } |
12880 | ||
029632fb PZ |
12881 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
12882 | { | |
bfb06889 | 12883 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
d05b4305 | 12884 | u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20))); |
141965c7 | 12885 | #ifdef CONFIG_SMP |
2a2f5d4e | 12886 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 12887 | #endif |
029632fb PZ |
12888 | } |
12889 | ||
810b3817 | 12890 | #ifdef CONFIG_FAIR_GROUP_SCHED |
39c42611 | 12891 | static void task_change_group_fair(struct task_struct *p) |
810b3817 | 12892 | { |
df16b71c CZ |
12893 | /* |
12894 | * We couldn't detach or attach a forked task which | |
12895 | * hasn't been woken up by wake_up_new_task(). | |
12896 | */ | |
12897 | if (READ_ONCE(p->__state) == TASK_NEW) | |
12898 | return; | |
12899 | ||
daa59407 | 12900 | detach_task_cfs_rq(p); |
6efdb105 BP |
12901 | |
12902 | #ifdef CONFIG_SMP | |
12903 | /* Tell se's cfs_rq has been changed -- migrated */ | |
12904 | p->se.avg.last_update_time = 0; | |
12905 | #endif | |
5d6da83c | 12906 | set_task_rq(p, task_cpu(p)); |
daa59407 | 12907 | attach_task_cfs_rq(p); |
810b3817 | 12908 | } |
029632fb PZ |
12909 | |
12910 | void free_fair_sched_group(struct task_group *tg) | |
12911 | { | |
12912 | int i; | |
12913 | ||
029632fb PZ |
12914 | for_each_possible_cpu(i) { |
12915 | if (tg->cfs_rq) | |
12916 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 12917 | if (tg->se) |
029632fb PZ |
12918 | kfree(tg->se[i]); |
12919 | } | |
12920 | ||
12921 | kfree(tg->cfs_rq); | |
12922 | kfree(tg->se); | |
12923 | } | |
12924 | ||
12925 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
12926 | { | |
029632fb | 12927 | struct sched_entity *se; |
b7fa30c9 | 12928 | struct cfs_rq *cfs_rq; |
029632fb PZ |
12929 | int i; |
12930 | ||
6396bb22 | 12931 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
12932 | if (!tg->cfs_rq) |
12933 | goto err; | |
6396bb22 | 12934 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
12935 | if (!tg->se) |
12936 | goto err; | |
12937 | ||
12938 | tg->shares = NICE_0_LOAD; | |
12939 | ||
c98c1827 | 12940 | init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent)); |
029632fb PZ |
12941 | |
12942 | for_each_possible_cpu(i) { | |
12943 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
12944 | GFP_KERNEL, cpu_to_node(i)); | |
12945 | if (!cfs_rq) | |
12946 | goto err; | |
12947 | ||
ceeadb83 | 12948 | se = kzalloc_node(sizeof(struct sched_entity_stats), |
029632fb PZ |
12949 | GFP_KERNEL, cpu_to_node(i)); |
12950 | if (!se) | |
12951 | goto err_free_rq; | |
12952 | ||
12953 | init_cfs_rq(cfs_rq); | |
12954 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 12955 | init_entity_runnable_average(se); |
029632fb PZ |
12956 | } |
12957 | ||
12958 | return 1; | |
12959 | ||
12960 | err_free_rq: | |
12961 | kfree(cfs_rq); | |
12962 | err: | |
12963 | return 0; | |
12964 | } | |
12965 | ||
8663e24d PZ |
12966 | void online_fair_sched_group(struct task_group *tg) |
12967 | { | |
12968 | struct sched_entity *se; | |
a46d14ec | 12969 | struct rq_flags rf; |
8663e24d PZ |
12970 | struct rq *rq; |
12971 | int i; | |
12972 | ||
12973 | for_each_possible_cpu(i) { | |
12974 | rq = cpu_rq(i); | |
12975 | se = tg->se[i]; | |
a46d14ec | 12976 | rq_lock_irq(rq, &rf); |
4126bad6 | 12977 | update_rq_clock(rq); |
d0326691 | 12978 | attach_entity_cfs_rq(se); |
55e16d30 | 12979 | sync_throttle(tg, i); |
a46d14ec | 12980 | rq_unlock_irq(rq, &rf); |
8663e24d PZ |
12981 | } |
12982 | } | |
12983 | ||
6fe1f348 | 12984 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 12985 | { |
029632fb | 12986 | unsigned long flags; |
6fe1f348 PZ |
12987 | struct rq *rq; |
12988 | int cpu; | |
029632fb | 12989 | |
b027789e MK |
12990 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); |
12991 | ||
6fe1f348 PZ |
12992 | for_each_possible_cpu(cpu) { |
12993 | if (tg->se[cpu]) | |
12994 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 12995 | |
6fe1f348 PZ |
12996 | /* |
12997 | * Only empty task groups can be destroyed; so we can speculatively | |
12998 | * check on_list without danger of it being re-added. | |
12999 | */ | |
13000 | if (!tg->cfs_rq[cpu]->on_list) | |
13001 | continue; | |
13002 | ||
13003 | rq = cpu_rq(cpu); | |
13004 | ||
5cb9eaa3 | 13005 | raw_spin_rq_lock_irqsave(rq, flags); |
6fe1f348 | 13006 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); |
5cb9eaa3 | 13007 | raw_spin_rq_unlock_irqrestore(rq, flags); |
6fe1f348 | 13008 | } |
029632fb PZ |
13009 | } |
13010 | ||
13011 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
13012 | struct sched_entity *se, int cpu, | |
13013 | struct sched_entity *parent) | |
13014 | { | |
13015 | struct rq *rq = cpu_rq(cpu); | |
13016 | ||
13017 | cfs_rq->tg = tg; | |
13018 | cfs_rq->rq = rq; | |
029632fb PZ |
13019 | init_cfs_rq_runtime(cfs_rq); |
13020 | ||
13021 | tg->cfs_rq[cpu] = cfs_rq; | |
13022 | tg->se[cpu] = se; | |
13023 | ||
13024 | /* se could be NULL for root_task_group */ | |
13025 | if (!se) | |
13026 | return; | |
13027 | ||
fed14d45 | 13028 | if (!parent) { |
029632fb | 13029 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
13030 | se->depth = 0; |
13031 | } else { | |
029632fb | 13032 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
13033 | se->depth = parent->depth + 1; |
13034 | } | |
029632fb PZ |
13035 | |
13036 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
13037 | /* guarantee group entities always have weight */ |
13038 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
13039 | se->parent = parent; |
13040 | } | |
13041 | ||
13042 | static DEFINE_MUTEX(shares_mutex); | |
13043 | ||
30400039 | 13044 | static int __sched_group_set_shares(struct task_group *tg, unsigned long shares) |
029632fb PZ |
13045 | { |
13046 | int i; | |
029632fb | 13047 | |
30400039 JD |
13048 | lockdep_assert_held(&shares_mutex); |
13049 | ||
029632fb PZ |
13050 | /* |
13051 | * We can't change the weight of the root cgroup. | |
13052 | */ | |
13053 | if (!tg->se[0]) | |
13054 | return -EINVAL; | |
13055 | ||
13056 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
13057 | ||
029632fb | 13058 | if (tg->shares == shares) |
30400039 | 13059 | return 0; |
029632fb PZ |
13060 | |
13061 | tg->shares = shares; | |
13062 | for_each_possible_cpu(i) { | |
13063 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
13064 | struct sched_entity *se = tg->se[i]; |
13065 | struct rq_flags rf; | |
029632fb | 13066 | |
029632fb | 13067 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 13068 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 13069 | update_rq_clock(rq); |
89ee048f | 13070 | for_each_sched_entity(se) { |
88c0616e | 13071 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 13072 | update_cfs_group(se); |
89ee048f | 13073 | } |
8a8c69c3 | 13074 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
13075 | } |
13076 | ||
30400039 JD |
13077 | return 0; |
13078 | } | |
13079 | ||
13080 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
13081 | { | |
13082 | int ret; | |
13083 | ||
13084 | mutex_lock(&shares_mutex); | |
13085 | if (tg_is_idle(tg)) | |
13086 | ret = -EINVAL; | |
13087 | else | |
13088 | ret = __sched_group_set_shares(tg, shares); | |
13089 | mutex_unlock(&shares_mutex); | |
13090 | ||
13091 | return ret; | |
13092 | } | |
13093 | ||
13094 | int sched_group_set_idle(struct task_group *tg, long idle) | |
13095 | { | |
13096 | int i; | |
13097 | ||
13098 | if (tg == &root_task_group) | |
13099 | return -EINVAL; | |
13100 | ||
13101 | if (idle < 0 || idle > 1) | |
13102 | return -EINVAL; | |
13103 | ||
13104 | mutex_lock(&shares_mutex); | |
13105 | ||
13106 | if (tg->idle == idle) { | |
13107 | mutex_unlock(&shares_mutex); | |
13108 | return 0; | |
13109 | } | |
13110 | ||
13111 | tg->idle = idle; | |
13112 | ||
13113 | for_each_possible_cpu(i) { | |
13114 | struct rq *rq = cpu_rq(i); | |
13115 | struct sched_entity *se = tg->se[i]; | |
a480adde | 13116 | struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i]; |
30400039 JD |
13117 | bool was_idle = cfs_rq_is_idle(grp_cfs_rq); |
13118 | long idle_task_delta; | |
13119 | struct rq_flags rf; | |
13120 | ||
13121 | rq_lock_irqsave(rq, &rf); | |
13122 | ||
13123 | grp_cfs_rq->idle = idle; | |
13124 | if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq))) | |
13125 | goto next_cpu; | |
13126 | ||
a480adde JD |
13127 | if (se->on_rq) { |
13128 | parent_cfs_rq = cfs_rq_of(se); | |
13129 | if (cfs_rq_is_idle(grp_cfs_rq)) | |
13130 | parent_cfs_rq->idle_nr_running++; | |
13131 | else | |
13132 | parent_cfs_rq->idle_nr_running--; | |
13133 | } | |
13134 | ||
30400039 JD |
13135 | idle_task_delta = grp_cfs_rq->h_nr_running - |
13136 | grp_cfs_rq->idle_h_nr_running; | |
13137 | if (!cfs_rq_is_idle(grp_cfs_rq)) | |
13138 | idle_task_delta *= -1; | |
13139 | ||
13140 | for_each_sched_entity(se) { | |
13141 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
13142 | ||
13143 | if (!se->on_rq) | |
13144 | break; | |
13145 | ||
13146 | cfs_rq->idle_h_nr_running += idle_task_delta; | |
13147 | ||
13148 | /* Already accounted at parent level and above. */ | |
13149 | if (cfs_rq_is_idle(cfs_rq)) | |
13150 | break; | |
13151 | } | |
13152 | ||
13153 | next_cpu: | |
13154 | rq_unlock_irqrestore(rq, &rf); | |
13155 | } | |
13156 | ||
13157 | /* Idle groups have minimum weight. */ | |
13158 | if (tg_is_idle(tg)) | |
13159 | __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO)); | |
13160 | else | |
13161 | __sched_group_set_shares(tg, NICE_0_LOAD); | |
13162 | ||
029632fb PZ |
13163 | mutex_unlock(&shares_mutex); |
13164 | return 0; | |
13165 | } | |
30400039 | 13166 | |
029632fb PZ |
13167 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
13168 | ||
810b3817 | 13169 | |
6d686f45 | 13170 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
13171 | { |
13172 | struct sched_entity *se = &task->se; | |
0d721cea PW |
13173 | unsigned int rr_interval = 0; |
13174 | ||
13175 | /* | |
13176 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
13177 | * idle runqueue: | |
13178 | */ | |
0d721cea | 13179 | if (rq->cfs.load.weight) |
147f3efa | 13180 | rr_interval = NS_TO_JIFFIES(se->slice); |
0d721cea PW |
13181 | |
13182 | return rr_interval; | |
13183 | } | |
13184 | ||
bf0f6f24 IM |
13185 | /* |
13186 | * All the scheduling class methods: | |
13187 | */ | |
43c31ac0 PZ |
13188 | DEFINE_SCHED_CLASS(fair) = { |
13189 | ||
bf0f6f24 IM |
13190 | .enqueue_task = enqueue_task_fair, |
13191 | .dequeue_task = dequeue_task_fair, | |
13192 | .yield_task = yield_task_fair, | |
d95f4122 | 13193 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 13194 | |
e23edc86 | 13195 | .wakeup_preempt = check_preempt_wakeup_fair, |
bf0f6f24 | 13196 | |
98c2f700 | 13197 | .pick_next_task = __pick_next_task_fair, |
bf0f6f24 | 13198 | .put_prev_task = put_prev_task_fair, |
03b7fad1 | 13199 | .set_next_task = set_next_task_fair, |
bf0f6f24 | 13200 | |
681f3e68 | 13201 | #ifdef CONFIG_SMP |
6e2df058 | 13202 | .balance = balance_fair, |
21f56ffe | 13203 | .pick_task = pick_task_fair, |
4ce72a2c | 13204 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 13205 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 13206 | |
0bcdcf28 CE |
13207 | .rq_online = rq_online_fair, |
13208 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 13209 | |
12695578 | 13210 | .task_dead = task_dead_fair, |
22d56074 | 13211 | .set_cpus_allowed = set_cpus_allowed_fair, |
681f3e68 | 13212 | #endif |
bf0f6f24 | 13213 | |
bf0f6f24 | 13214 | .task_tick = task_tick_fair, |
cd29fe6f | 13215 | .task_fork = task_fork_fair, |
cb469845 SR |
13216 | |
13217 | .prio_changed = prio_changed_fair, | |
da7a735e | 13218 | .switched_from = switched_from_fair, |
cb469845 | 13219 | .switched_to = switched_to_fair, |
810b3817 | 13220 | |
0d721cea PW |
13221 | .get_rr_interval = get_rr_interval_fair, |
13222 | ||
6e998916 SG |
13223 | .update_curr = update_curr_fair, |
13224 | ||
810b3817 | 13225 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 13226 | .task_change_group = task_change_group_fair, |
810b3817 | 13227 | #endif |
982d9cdc | 13228 | |
530bfad1 HJ |
13229 | #ifdef CONFIG_SCHED_CORE |
13230 | .task_is_throttled = task_is_throttled_fair, | |
13231 | #endif | |
13232 | ||
982d9cdc PB |
13233 | #ifdef CONFIG_UCLAMP_TASK |
13234 | .uclamp_enabled = 1, | |
13235 | #endif | |
bf0f6f24 IM |
13236 | }; |
13237 | ||
13238 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 13239 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 13240 | { |
039ae8bc | 13241 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 13242 | |
5973e5b9 | 13243 | rcu_read_lock(); |
039ae8bc | 13244 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 13245 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 13246 | rcu_read_unlock(); |
bf0f6f24 | 13247 | } |
397f2378 SD |
13248 | |
13249 | #ifdef CONFIG_NUMA_BALANCING | |
13250 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
13251 | { | |
13252 | int node; | |
13253 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 13254 | struct numa_group *ng; |
397f2378 | 13255 | |
cb361d8c JH |
13256 | rcu_read_lock(); |
13257 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
13258 | for_each_online_node(node) { |
13259 | if (p->numa_faults) { | |
13260 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
13261 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
13262 | } | |
cb361d8c JH |
13263 | if (ng) { |
13264 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
13265 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
13266 | } |
13267 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
13268 | } | |
cb361d8c | 13269 | rcu_read_unlock(); |
397f2378 SD |
13270 | } |
13271 | #endif /* CONFIG_NUMA_BALANCING */ | |
13272 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
13273 | |
13274 | __init void init_sched_fair_class(void) | |
13275 | { | |
13276 | #ifdef CONFIG_SMP | |
18c31c97 BH |
13277 | int i; |
13278 | ||
13279 | for_each_possible_cpu(i) { | |
13280 | zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i)); | |
13281 | zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i)); | |
f8858d96 SH |
13282 | zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i), |
13283 | GFP_KERNEL, cpu_to_node(i)); | |
8ad075c2 JD |
13284 | |
13285 | #ifdef CONFIG_CFS_BANDWIDTH | |
13286 | INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i)); | |
13287 | INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list); | |
13288 | #endif | |
18c31c97 BH |
13289 | } |
13290 | ||
70a27d6d | 13291 | open_softirq(SCHED_SOFTIRQ, sched_balance_softirq); |
029632fb | 13292 | |
3451d024 | 13293 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 13294 | nohz.next_balance = jiffies; |
f643ea22 | 13295 | nohz.next_blocked = jiffies; |
029632fb | 13296 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
13297 | #endif |
13298 | #endif /* SMP */ | |
13299 | ||
13300 | } |