]>
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> |
c4ad6fcb IM |
50 | |
51 | #include <asm/switch_to.h> | |
52 | ||
53 | #include <linux/sched/cond_resched.h> | |
54 | ||
325ea10c | 55 | #include "sched.h" |
b9e9c6ca IM |
56 | #include "stats.h" |
57 | #include "autogroup.h" | |
029632fb | 58 | |
bf0f6f24 | 59 | /* |
21805085 | 60 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 61 | * |
21805085 | 62 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
63 | * 'timeslice length' - timeslices in CFS are of variable length |
64 | * and have no persistent notion like in traditional, time-slice | |
65 | * based scheduling concepts. | |
bf0f6f24 | 66 | * |
d274a4ce IM |
67 | * (to see the precise effective timeslice length of your workload, |
68 | * run vmstat and monitor the context-switches (cs) field) | |
2b4d5b25 IM |
69 | * |
70 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 71 | */ |
2b4d5b25 | 72 | unsigned int sysctl_sched_latency = 6000000ULL; |
ed8885a1 | 73 | static unsigned int normalized_sysctl_sched_latency = 6000000ULL; |
2bd8e6d4 | 74 | |
1983a922 CE |
75 | /* |
76 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
77 | * |
78 | * Options are: | |
2b4d5b25 IM |
79 | * |
80 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
81 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
82 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
83 | * | |
84 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 85 | */ |
8a99b683 | 86 | unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 87 | |
2bd8e6d4 | 88 | /* |
b2be5e96 | 89 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 90 | * |
864616ee | 91 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 92 | */ |
ed8885a1 MS |
93 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
94 | static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 | 95 | |
51ce83ed JD |
96 | /* |
97 | * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks. | |
98 | * Applies only when SCHED_IDLE tasks compete with normal tasks. | |
99 | * | |
100 | * (default: 0.75 msec) | |
101 | */ | |
102 | unsigned int sysctl_sched_idle_min_granularity = 750000ULL; | |
103 | ||
21805085 | 104 | /* |
2b4d5b25 | 105 | * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity |
b2be5e96 | 106 | */ |
0bf377bb | 107 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
108 | |
109 | /* | |
2bba22c5 | 110 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 111 | * parent will (try to) run first. |
21805085 | 112 | */ |
2bba22c5 | 113 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 114 | |
bf0f6f24 IM |
115 | /* |
116 | * SCHED_OTHER wake-up granularity. | |
bf0f6f24 IM |
117 | * |
118 | * This option delays the preemption effects of decoupled workloads | |
119 | * and reduces their over-scheduling. Synchronous workloads will still | |
120 | * have immediate wakeup/sleep latencies. | |
2b4d5b25 IM |
121 | * |
122 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 123 | */ |
ed8885a1 MS |
124 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
125 | static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
bf0f6f24 | 126 | |
2b4d5b25 | 127 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 128 | |
05289b90 TG |
129 | int sched_thermal_decay_shift; |
130 | static int __init setup_sched_thermal_decay_shift(char *str) | |
131 | { | |
132 | int _shift = 0; | |
133 | ||
134 | if (kstrtoint(str, 0, &_shift)) | |
135 | pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n"); | |
136 | ||
137 | sched_thermal_decay_shift = clamp(_shift, 0, 10); | |
138 | return 1; | |
139 | } | |
140 | __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift); | |
141 | ||
afe06efd TC |
142 | #ifdef CONFIG_SMP |
143 | /* | |
97fb7a0a | 144 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
145 | */ |
146 | int __weak arch_asym_cpu_priority(int cpu) | |
147 | { | |
148 | return -cpu; | |
149 | } | |
6d101ba6 OJ |
150 | |
151 | /* | |
60e17f5c | 152 | * The margin used when comparing utilization with CPU capacity. |
6d101ba6 OJ |
153 | * |
154 | * (default: ~20%) | |
155 | */ | |
60e17f5c VK |
156 | #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) |
157 | ||
4aed8aa4 VS |
158 | /* |
159 | * The margin used when comparing CPU capacities. | |
160 | * is 'cap1' noticeably greater than 'cap2' | |
161 | * | |
162 | * (default: ~5%) | |
163 | */ | |
164 | #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078) | |
afe06efd TC |
165 | #endif |
166 | ||
ec12cb7f PT |
167 | #ifdef CONFIG_CFS_BANDWIDTH |
168 | /* | |
169 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
170 | * each time a cfs_rq requests quota. | |
171 | * | |
172 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
173 | * to consumption or the quota being specified to be smaller than the slice) | |
174 | * we will always only issue the remaining available time. | |
175 | * | |
2b4d5b25 IM |
176 | * (default: 5 msec, units: microseconds) |
177 | */ | |
d4ae80ff ZN |
178 | static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; |
179 | #endif | |
180 | ||
0dff89c4 KW |
181 | #ifdef CONFIG_NUMA_BALANCING |
182 | /* Restrict the NUMA promotion throughput (MB/s) for each target node. */ | |
183 | static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536; | |
184 | #endif | |
185 | ||
d4ae80ff ZN |
186 | #ifdef CONFIG_SYSCTL |
187 | static struct ctl_table sched_fair_sysctls[] = { | |
188 | { | |
189 | .procname = "sched_child_runs_first", | |
190 | .data = &sysctl_sched_child_runs_first, | |
191 | .maxlen = sizeof(unsigned int), | |
192 | .mode = 0644, | |
193 | .proc_handler = proc_dointvec, | |
194 | }, | |
195 | #ifdef CONFIG_CFS_BANDWIDTH | |
196 | { | |
197 | .procname = "sched_cfs_bandwidth_slice_us", | |
198 | .data = &sysctl_sched_cfs_bandwidth_slice, | |
199 | .maxlen = sizeof(unsigned int), | |
200 | .mode = 0644, | |
201 | .proc_handler = proc_dointvec_minmax, | |
202 | .extra1 = SYSCTL_ONE, | |
203 | }, | |
204 | #endif | |
0dff89c4 KW |
205 | #ifdef CONFIG_NUMA_BALANCING |
206 | { | |
207 | .procname = "numa_balancing_promote_rate_limit_MBps", | |
208 | .data = &sysctl_numa_balancing_promote_rate_limit, | |
209 | .maxlen = sizeof(unsigned int), | |
210 | .mode = 0644, | |
211 | .proc_handler = proc_dointvec_minmax, | |
212 | .extra1 = SYSCTL_ZERO, | |
213 | }, | |
214 | #endif /* CONFIG_NUMA_BALANCING */ | |
d4ae80ff ZN |
215 | {} |
216 | }; | |
217 | ||
218 | static int __init sched_fair_sysctl_init(void) | |
219 | { | |
220 | register_sysctl_init("kernel", sched_fair_sysctls); | |
221 | return 0; | |
222 | } | |
223 | late_initcall(sched_fair_sysctl_init); | |
ec12cb7f PT |
224 | #endif |
225 | ||
8527632d PG |
226 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
227 | { | |
228 | lw->weight += inc; | |
229 | lw->inv_weight = 0; | |
230 | } | |
231 | ||
232 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
233 | { | |
234 | lw->weight -= dec; | |
235 | lw->inv_weight = 0; | |
236 | } | |
237 | ||
238 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
239 | { | |
240 | lw->weight = w; | |
241 | lw->inv_weight = 0; | |
242 | } | |
243 | ||
029632fb PZ |
244 | /* |
245 | * Increase the granularity value when there are more CPUs, | |
246 | * because with more CPUs the 'effective latency' as visible | |
247 | * to users decreases. But the relationship is not linear, | |
248 | * so pick a second-best guess by going with the log2 of the | |
249 | * number of CPUs. | |
250 | * | |
251 | * This idea comes from the SD scheduler of Con Kolivas: | |
252 | */ | |
58ac93e4 | 253 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 254 | { |
58ac93e4 | 255 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
256 | unsigned int factor; |
257 | ||
258 | switch (sysctl_sched_tunable_scaling) { | |
259 | case SCHED_TUNABLESCALING_NONE: | |
260 | factor = 1; | |
261 | break; | |
262 | case SCHED_TUNABLESCALING_LINEAR: | |
263 | factor = cpus; | |
264 | break; | |
265 | case SCHED_TUNABLESCALING_LOG: | |
266 | default: | |
267 | factor = 1 + ilog2(cpus); | |
268 | break; | |
269 | } | |
270 | ||
271 | return factor; | |
272 | } | |
273 | ||
274 | static void update_sysctl(void) | |
275 | { | |
276 | unsigned int factor = get_update_sysctl_factor(); | |
277 | ||
278 | #define SET_SYSCTL(name) \ | |
279 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
280 | SET_SYSCTL(sched_min_granularity); | |
281 | SET_SYSCTL(sched_latency); | |
282 | SET_SYSCTL(sched_wakeup_granularity); | |
283 | #undef SET_SYSCTL | |
284 | } | |
285 | ||
f38f12d1 | 286 | void __init sched_init_granularity(void) |
029632fb PZ |
287 | { |
288 | update_sysctl(); | |
289 | } | |
290 | ||
9dbdb155 | 291 | #define WMULT_CONST (~0U) |
029632fb PZ |
292 | #define WMULT_SHIFT 32 |
293 | ||
9dbdb155 PZ |
294 | static void __update_inv_weight(struct load_weight *lw) |
295 | { | |
296 | unsigned long w; | |
297 | ||
298 | if (likely(lw->inv_weight)) | |
299 | return; | |
300 | ||
301 | w = scale_load_down(lw->weight); | |
302 | ||
303 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
304 | lw->inv_weight = 1; | |
305 | else if (unlikely(!w)) | |
306 | lw->inv_weight = WMULT_CONST; | |
307 | else | |
308 | lw->inv_weight = WMULT_CONST / w; | |
309 | } | |
029632fb PZ |
310 | |
311 | /* | |
9dbdb155 PZ |
312 | * delta_exec * weight / lw.weight |
313 | * OR | |
314 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
315 | * | |
1c3de5e1 | 316 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
317 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
318 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
319 | * | |
320 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
321 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 322 | */ |
9dbdb155 | 323 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 324 | { |
9dbdb155 | 325 | u64 fact = scale_load_down(weight); |
1e17fb8e | 326 | u32 fact_hi = (u32)(fact >> 32); |
9dbdb155 | 327 | int shift = WMULT_SHIFT; |
1e17fb8e | 328 | int fs; |
029632fb | 329 | |
9dbdb155 | 330 | __update_inv_weight(lw); |
029632fb | 331 | |
1e17fb8e CC |
332 | if (unlikely(fact_hi)) { |
333 | fs = fls(fact_hi); | |
334 | shift -= fs; | |
335 | fact >>= fs; | |
029632fb PZ |
336 | } |
337 | ||
2eeb01a2 | 338 | fact = mul_u32_u32(fact, lw->inv_weight); |
029632fb | 339 | |
1e17fb8e CC |
340 | fact_hi = (u32)(fact >> 32); |
341 | if (fact_hi) { | |
342 | fs = fls(fact_hi); | |
343 | shift -= fs; | |
344 | fact >>= fs; | |
9dbdb155 | 345 | } |
029632fb | 346 | |
9dbdb155 | 347 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
348 | } |
349 | ||
350 | ||
351 | const struct sched_class fair_sched_class; | |
a4c2f00f | 352 | |
bf0f6f24 IM |
353 | /************************************************************** |
354 | * CFS operations on generic schedulable entities: | |
355 | */ | |
356 | ||
62160e3f | 357 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8f48894f | 358 | |
b758149c PZ |
359 | /* Walk up scheduling entities hierarchy */ |
360 | #define for_each_sched_entity(se) \ | |
361 | for (; se; se = se->parent) | |
362 | ||
f6783319 | 363 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 364 | { |
5d299eab PZ |
365 | struct rq *rq = rq_of(cfs_rq); |
366 | int cpu = cpu_of(rq); | |
367 | ||
368 | if (cfs_rq->on_list) | |
f6783319 | 369 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
370 | |
371 | cfs_rq->on_list = 1; | |
372 | ||
373 | /* | |
374 | * Ensure we either appear before our parent (if already | |
375 | * enqueued) or force our parent to appear after us when it is | |
376 | * enqueued. The fact that we always enqueue bottom-up | |
377 | * reduces this to two cases and a special case for the root | |
378 | * cfs_rq. Furthermore, it also means that we will always reset | |
379 | * tmp_alone_branch either when the branch is connected | |
380 | * to a tree or when we reach the top of the tree | |
381 | */ | |
382 | if (cfs_rq->tg->parent && | |
383 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 384 | /* |
5d299eab PZ |
385 | * If parent is already on the list, we add the child |
386 | * just before. Thanks to circular linked property of | |
387 | * the list, this means to put the child at the tail | |
388 | * of the list that starts by parent. | |
67e86250 | 389 | */ |
5d299eab PZ |
390 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
391 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
392 | /* | |
393 | * The branch is now connected to its tree so we can | |
394 | * reset tmp_alone_branch to the beginning of the | |
395 | * list. | |
396 | */ | |
397 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 398 | return true; |
5d299eab | 399 | } |
3d4b47b4 | 400 | |
5d299eab PZ |
401 | if (!cfs_rq->tg->parent) { |
402 | /* | |
403 | * cfs rq without parent should be put | |
404 | * at the tail of the list. | |
405 | */ | |
406 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
407 | &rq->leaf_cfs_rq_list); | |
408 | /* | |
409 | * We have reach the top of a tree so we can reset | |
410 | * tmp_alone_branch to the beginning of the list. | |
411 | */ | |
412 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 413 | return true; |
3d4b47b4 | 414 | } |
5d299eab PZ |
415 | |
416 | /* | |
417 | * The parent has not already been added so we want to | |
418 | * make sure that it will be put after us. | |
419 | * tmp_alone_branch points to the begin of the branch | |
420 | * where we will add parent. | |
421 | */ | |
422 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
423 | /* | |
424 | * update tmp_alone_branch to points to the new begin | |
425 | * of the branch | |
426 | */ | |
427 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 428 | return false; |
3d4b47b4 PZ |
429 | } |
430 | ||
431 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
432 | { | |
433 | if (cfs_rq->on_list) { | |
31bc6aea VG |
434 | struct rq *rq = rq_of(cfs_rq); |
435 | ||
436 | /* | |
437 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
438 | * it can happen the tmp_alone_branch points the a leaf that | |
439 | * we finally want to del. In this case, tmp_alone_branch moves | |
440 | * to the prev element but it will point to rq->leaf_cfs_rq_list | |
441 | * at the end of the enqueue. | |
442 | */ | |
443 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
444 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
445 | ||
3d4b47b4 PZ |
446 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
447 | cfs_rq->on_list = 0; | |
448 | } | |
449 | } | |
450 | ||
5d299eab PZ |
451 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
452 | { | |
453 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
454 | } | |
455 | ||
039ae8bc VG |
456 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
457 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ | |
458 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
459 | leaf_cfs_rq_list) | |
b758149c PZ |
460 | |
461 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 462 | static inline struct cfs_rq * |
b758149c PZ |
463 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
464 | { | |
465 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 466 | return se->cfs_rq; |
b758149c | 467 | |
fed14d45 | 468 | return NULL; |
b758149c PZ |
469 | } |
470 | ||
471 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
472 | { | |
473 | return se->parent; | |
474 | } | |
475 | ||
464b7527 PZ |
476 | static void |
477 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
478 | { | |
479 | int se_depth, pse_depth; | |
480 | ||
481 | /* | |
482 | * preemption test can be made between sibling entities who are in the | |
483 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
484 | * both tasks until we find their ancestors who are siblings of common | |
485 | * parent. | |
486 | */ | |
487 | ||
488 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
489 | se_depth = (*se)->depth; |
490 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
491 | |
492 | while (se_depth > pse_depth) { | |
493 | se_depth--; | |
494 | *se = parent_entity(*se); | |
495 | } | |
496 | ||
497 | while (pse_depth > se_depth) { | |
498 | pse_depth--; | |
499 | *pse = parent_entity(*pse); | |
500 | } | |
501 | ||
502 | while (!is_same_group(*se, *pse)) { | |
503 | *se = parent_entity(*se); | |
504 | *pse = parent_entity(*pse); | |
505 | } | |
506 | } | |
507 | ||
30400039 JD |
508 | static int tg_is_idle(struct task_group *tg) |
509 | { | |
510 | return tg->idle > 0; | |
511 | } | |
512 | ||
513 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
514 | { | |
515 | return cfs_rq->idle > 0; | |
516 | } | |
517 | ||
518 | static int se_is_idle(struct sched_entity *se) | |
519 | { | |
520 | if (entity_is_task(se)) | |
521 | return task_has_idle_policy(task_of(se)); | |
522 | return cfs_rq_is_idle(group_cfs_rq(se)); | |
523 | } | |
524 | ||
8f48894f PZ |
525 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
526 | ||
b758149c PZ |
527 | #define for_each_sched_entity(se) \ |
528 | for (; se; se = NULL) | |
bf0f6f24 | 529 | |
f6783319 | 530 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 531 | { |
f6783319 | 532 | return true; |
3d4b47b4 PZ |
533 | } |
534 | ||
535 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
536 | { | |
537 | } | |
538 | ||
5d299eab PZ |
539 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
540 | { | |
541 | } | |
542 | ||
039ae8bc VG |
543 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
544 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 545 | |
b758149c PZ |
546 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
547 | { | |
548 | return NULL; | |
549 | } | |
550 | ||
464b7527 PZ |
551 | static inline void |
552 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
553 | { | |
554 | } | |
555 | ||
366e7ad6 | 556 | static inline int tg_is_idle(struct task_group *tg) |
30400039 JD |
557 | { |
558 | return 0; | |
559 | } | |
560 | ||
561 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
562 | { | |
563 | return 0; | |
564 | } | |
565 | ||
566 | static int se_is_idle(struct sched_entity *se) | |
567 | { | |
568 | return 0; | |
569 | } | |
570 | ||
b758149c PZ |
571 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
572 | ||
6c16a6dc | 573 | static __always_inline |
9dbdb155 | 574 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
575 | |
576 | /************************************************************** | |
577 | * Scheduling class tree data structure manipulation methods: | |
578 | */ | |
579 | ||
1bf08230 | 580 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 581 | { |
1bf08230 | 582 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 583 | if (delta > 0) |
1bf08230 | 584 | max_vruntime = vruntime; |
02e0431a | 585 | |
1bf08230 | 586 | return max_vruntime; |
02e0431a PZ |
587 | } |
588 | ||
0702e3eb | 589 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
590 | { |
591 | s64 delta = (s64)(vruntime - min_vruntime); | |
592 | if (delta < 0) | |
593 | min_vruntime = vruntime; | |
594 | ||
595 | return min_vruntime; | |
596 | } | |
597 | ||
bf9be9a1 | 598 | static inline bool entity_before(struct sched_entity *a, |
54fdc581 FC |
599 | struct sched_entity *b) |
600 | { | |
601 | return (s64)(a->vruntime - b->vruntime) < 0; | |
602 | } | |
603 | ||
bf9be9a1 PZ |
604 | #define __node_2_se(node) \ |
605 | rb_entry((node), struct sched_entity, run_node) | |
606 | ||
1af5f730 PZ |
607 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
608 | { | |
b60205c7 | 609 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 610 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 611 | |
1af5f730 PZ |
612 | u64 vruntime = cfs_rq->min_vruntime; |
613 | ||
b60205c7 PZ |
614 | if (curr) { |
615 | if (curr->on_rq) | |
616 | vruntime = curr->vruntime; | |
617 | else | |
618 | curr = NULL; | |
619 | } | |
1af5f730 | 620 | |
bfb06889 | 621 | if (leftmost) { /* non-empty tree */ |
bf9be9a1 | 622 | struct sched_entity *se = __node_2_se(leftmost); |
1af5f730 | 623 | |
b60205c7 | 624 | if (!curr) |
1af5f730 PZ |
625 | vruntime = se->vruntime; |
626 | else | |
627 | vruntime = min_vruntime(vruntime, se->vruntime); | |
628 | } | |
629 | ||
1bf08230 | 630 | /* ensure we never gain time by being placed backwards. */ |
d05b4305 VD |
631 | u64_u32_store(cfs_rq->min_vruntime, |
632 | max_vruntime(cfs_rq->min_vruntime, vruntime)); | |
1af5f730 PZ |
633 | } |
634 | ||
bf9be9a1 PZ |
635 | static inline bool __entity_less(struct rb_node *a, const struct rb_node *b) |
636 | { | |
637 | return entity_before(__node_2_se(a), __node_2_se(b)); | |
638 | } | |
639 | ||
bf0f6f24 IM |
640 | /* |
641 | * Enqueue an entity into the rb-tree: | |
642 | */ | |
0702e3eb | 643 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 644 | { |
bf9be9a1 | 645 | rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less); |
bf0f6f24 IM |
646 | } |
647 | ||
0702e3eb | 648 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 649 | { |
bfb06889 | 650 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
651 | } |
652 | ||
029632fb | 653 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 654 | { |
bfb06889 | 655 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
656 | |
657 | if (!left) | |
658 | return NULL; | |
659 | ||
bf9be9a1 | 660 | return __node_2_se(left); |
bf0f6f24 IM |
661 | } |
662 | ||
ac53db59 RR |
663 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
664 | { | |
665 | struct rb_node *next = rb_next(&se->run_node); | |
666 | ||
667 | if (!next) | |
668 | return NULL; | |
669 | ||
bf9be9a1 | 670 | return __node_2_se(next); |
ac53db59 RR |
671 | } |
672 | ||
673 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 674 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 675 | { |
bfb06889 | 676 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 677 | |
70eee74b BS |
678 | if (!last) |
679 | return NULL; | |
7eee3e67 | 680 | |
bf9be9a1 | 681 | return __node_2_se(last); |
aeb73b04 PZ |
682 | } |
683 | ||
bf0f6f24 IM |
684 | /************************************************************** |
685 | * Scheduling class statistics methods: | |
686 | */ | |
687 | ||
8a99b683 | 688 | int sched_update_scaling(void) |
b2be5e96 | 689 | { |
58ac93e4 | 690 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 | 691 | |
b2be5e96 PZ |
692 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
693 | sysctl_sched_min_granularity); | |
694 | ||
acb4a848 CE |
695 | #define WRT_SYSCTL(name) \ |
696 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
697 | WRT_SYSCTL(sched_min_granularity); | |
698 | WRT_SYSCTL(sched_latency); | |
699 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
700 | #undef WRT_SYSCTL |
701 | ||
b2be5e96 PZ |
702 | return 0; |
703 | } | |
704 | #endif | |
647e7cac | 705 | |
a7be37ac | 706 | /* |
f9c0b095 | 707 | * delta /= w |
a7be37ac | 708 | */ |
9dbdb155 | 709 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 710 | { |
f9c0b095 | 711 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 712 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
713 | |
714 | return delta; | |
715 | } | |
716 | ||
647e7cac IM |
717 | /* |
718 | * The idea is to set a period in which each task runs once. | |
719 | * | |
532b1858 | 720 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
721 | * this period because otherwise the slices get too small. |
722 | * | |
723 | * p = (nr <= nl) ? l : l*nr/nl | |
724 | */ | |
4d78e7b6 PZ |
725 | static u64 __sched_period(unsigned long nr_running) |
726 | { | |
8e2b0bf3 BF |
727 | if (unlikely(nr_running > sched_nr_latency)) |
728 | return nr_running * sysctl_sched_min_granularity; | |
729 | else | |
730 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
731 | } |
732 | ||
51ce83ed JD |
733 | static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq); |
734 | ||
647e7cac IM |
735 | /* |
736 | * We calculate the wall-time slice from the period by taking a part | |
737 | * proportional to the weight. | |
738 | * | |
f9c0b095 | 739 | * s = p*P[w/rw] |
647e7cac | 740 | */ |
6d0f0ebd | 741 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 742 | { |
0c2de3f0 | 743 | unsigned int nr_running = cfs_rq->nr_running; |
51ce83ed JD |
744 | struct sched_entity *init_se = se; |
745 | unsigned int min_gran; | |
0c2de3f0 PZ |
746 | u64 slice; |
747 | ||
748 | if (sched_feat(ALT_PERIOD)) | |
749 | nr_running = rq_of(cfs_rq)->cfs.h_nr_running; | |
750 | ||
751 | slice = __sched_period(nr_running + !se->on_rq); | |
f9c0b095 | 752 | |
0a582440 | 753 | for_each_sched_entity(se) { |
6272d68c | 754 | struct load_weight *load; |
3104bf03 | 755 | struct load_weight lw; |
51ce83ed | 756 | struct cfs_rq *qcfs_rq; |
6272d68c | 757 | |
51ce83ed JD |
758 | qcfs_rq = cfs_rq_of(se); |
759 | load = &qcfs_rq->load; | |
f9c0b095 | 760 | |
0a582440 | 761 | if (unlikely(!se->on_rq)) { |
51ce83ed | 762 | lw = qcfs_rq->load; |
0a582440 MG |
763 | |
764 | update_load_add(&lw, se->load.weight); | |
765 | load = &lw; | |
766 | } | |
9dbdb155 | 767 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 | 768 | } |
0c2de3f0 | 769 | |
51ce83ed JD |
770 | if (sched_feat(BASE_SLICE)) { |
771 | if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq)) | |
772 | min_gran = sysctl_sched_idle_min_granularity; | |
773 | else | |
774 | min_gran = sysctl_sched_min_granularity; | |
775 | ||
776 | slice = max_t(u64, slice, min_gran); | |
777 | } | |
0c2de3f0 | 778 | |
0a582440 | 779 | return slice; |
bf0f6f24 IM |
780 | } |
781 | ||
647e7cac | 782 | /* |
660cc00f | 783 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 784 | * |
f9c0b095 | 785 | * vs = s/w |
647e7cac | 786 | */ |
f9c0b095 | 787 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 788 | { |
f9c0b095 | 789 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
790 | } |
791 | ||
c0796298 | 792 | #include "pelt.h" |
23127296 | 793 | #ifdef CONFIG_SMP |
283e2ed3 | 794 | |
772bd008 | 795 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 796 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 797 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 798 | |
540247fb YD |
799 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
800 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 801 | { |
540247fb | 802 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 803 | |
f207934f PZ |
804 | memset(sa, 0, sizeof(*sa)); |
805 | ||
b5a9b340 | 806 | /* |
dfcb245e | 807 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 808 | * they get a chance to stabilize to their real load level. |
dfcb245e | 809 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
810 | * nothing has been attached to the task group yet. |
811 | */ | |
812 | if (entity_is_task(se)) | |
0dacee1b | 813 | sa->load_avg = scale_load_down(se->load.weight); |
f207934f | 814 | |
9d89c257 | 815 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 816 | } |
7ea241af | 817 | |
2b8c41da YD |
818 | /* |
819 | * With new tasks being created, their initial util_avgs are extrapolated | |
820 | * based on the cfs_rq's current util_avg: | |
821 | * | |
822 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
823 | * | |
824 | * However, in many cases, the above util_avg does not give a desired | |
825 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
826 | * as when the series is a harmonic series. | |
827 | * | |
828 | * To solve this problem, we also cap the util_avg of successive tasks to | |
829 | * only 1/2 of the left utilization budget: | |
830 | * | |
8fe5c5a9 | 831 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 832 | * |
8fe5c5a9 | 833 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 834 | * |
8fe5c5a9 QP |
835 | * For example, for a CPU with 1024 of capacity, a simplest series from |
836 | * the beginning would be like: | |
2b8c41da YD |
837 | * |
838 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
839 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
840 | * | |
841 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
842 | * if util_avg > util_avg_cap. | |
843 | */ | |
d0fe0b9c | 844 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 845 | { |
d0fe0b9c | 846 | struct sched_entity *se = &p->se; |
2b8c41da YD |
847 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
848 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 849 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 850 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da | 851 | |
d0fe0b9c DE |
852 | if (p->sched_class != &fair_sched_class) { |
853 | /* | |
854 | * For !fair tasks do: | |
855 | * | |
856 | update_cfs_rq_load_avg(now, cfs_rq); | |
a4f9a0e5 | 857 | attach_entity_load_avg(cfs_rq, se); |
d0fe0b9c DE |
858 | switched_from_fair(rq, p); |
859 | * | |
860 | * such that the next switched_to_fair() has the | |
861 | * expected state. | |
862 | */ | |
863 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
864 | return; | |
7dc603c9 | 865 | } |
e4fe074d CZ |
866 | |
867 | if (cap > 0) { | |
868 | if (cfs_rq->avg.util_avg != 0) { | |
869 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
870 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
871 | ||
872 | if (sa->util_avg > cap) | |
873 | sa->util_avg = cap; | |
874 | } else { | |
875 | sa->util_avg = cap; | |
876 | } | |
877 | } | |
878 | ||
879 | sa->runnable_avg = sa->util_avg; | |
2b8c41da YD |
880 | } |
881 | ||
7dc603c9 | 882 | #else /* !CONFIG_SMP */ |
540247fb | 883 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
884 | { |
885 | } | |
d0fe0b9c | 886 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
887 | { |
888 | } | |
fe749158 | 889 | static void update_tg_load_avg(struct cfs_rq *cfs_rq) |
3d30544f PZ |
890 | { |
891 | } | |
7dc603c9 | 892 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 893 | |
bf0f6f24 | 894 | /* |
9dbdb155 | 895 | * Update the current task's runtime statistics. |
bf0f6f24 | 896 | */ |
b7cc0896 | 897 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 898 | { |
429d43bc | 899 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 900 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 901 | u64 delta_exec; |
bf0f6f24 IM |
902 | |
903 | if (unlikely(!curr)) | |
904 | return; | |
905 | ||
9dbdb155 PZ |
906 | delta_exec = now - curr->exec_start; |
907 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 908 | return; |
bf0f6f24 | 909 | |
8ebc91d9 | 910 | curr->exec_start = now; |
d842de87 | 911 | |
ceeadb83 YS |
912 | if (schedstat_enabled()) { |
913 | struct sched_statistics *stats; | |
914 | ||
915 | stats = __schedstats_from_se(curr); | |
916 | __schedstat_set(stats->exec_max, | |
917 | max(delta_exec, stats->exec_max)); | |
918 | } | |
9dbdb155 PZ |
919 | |
920 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 921 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
922 | |
923 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
924 | update_min_vruntime(cfs_rq); | |
925 | ||
d842de87 SV |
926 | if (entity_is_task(curr)) { |
927 | struct task_struct *curtask = task_of(curr); | |
928 | ||
f977bb49 | 929 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 930 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 931 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 932 | } |
ec12cb7f PT |
933 | |
934 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
935 | } |
936 | ||
6e998916 SG |
937 | static void update_curr_fair(struct rq *rq) |
938 | { | |
939 | update_curr(cfs_rq_of(&rq->curr->se)); | |
940 | } | |
941 | ||
bf0f6f24 | 942 | static inline void |
60f2415e | 943 | update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 944 | { |
ceeadb83 | 945 | struct sched_statistics *stats; |
60f2415e | 946 | struct task_struct *p = NULL; |
4fa8d299 JP |
947 | |
948 | if (!schedstat_enabled()) | |
949 | return; | |
950 | ||
ceeadb83 YS |
951 | stats = __schedstats_from_se(se); |
952 | ||
60f2415e YS |
953 | if (entity_is_task(se)) |
954 | p = task_of(se); | |
3ea94de1 | 955 | |
60f2415e | 956 | __update_stats_wait_start(rq_of(cfs_rq), p, stats); |
bf0f6f24 IM |
957 | } |
958 | ||
4fa8d299 | 959 | static inline void |
60f2415e | 960 | update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3ea94de1 | 961 | { |
ceeadb83 YS |
962 | struct sched_statistics *stats; |
963 | struct task_struct *p = NULL; | |
cb251765 | 964 | |
4fa8d299 JP |
965 | if (!schedstat_enabled()) |
966 | return; | |
967 | ||
ceeadb83 YS |
968 | stats = __schedstats_from_se(se); |
969 | ||
b9c88f75 | 970 | /* |
971 | * When the sched_schedstat changes from 0 to 1, some sched se | |
972 | * maybe already in the runqueue, the se->statistics.wait_start | |
973 | * will be 0.So it will let the delta wrong. We need to avoid this | |
974 | * scenario. | |
975 | */ | |
ceeadb83 | 976 | if (unlikely(!schedstat_val(stats->wait_start))) |
b9c88f75 | 977 | return; |
978 | ||
60f2415e | 979 | if (entity_is_task(se)) |
3ea94de1 | 980 | p = task_of(se); |
3ea94de1 | 981 | |
60f2415e | 982 | __update_stats_wait_end(rq_of(cfs_rq), p, stats); |
3ea94de1 | 983 | } |
3ea94de1 | 984 | |
4fa8d299 | 985 | static inline void |
60f2415e | 986 | update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1a3d027c | 987 | { |
ceeadb83 | 988 | struct sched_statistics *stats; |
1a3d027c | 989 | struct task_struct *tsk = NULL; |
4fa8d299 JP |
990 | |
991 | if (!schedstat_enabled()) | |
992 | return; | |
993 | ||
ceeadb83 YS |
994 | stats = __schedstats_from_se(se); |
995 | ||
1a3d027c JP |
996 | if (entity_is_task(se)) |
997 | tsk = task_of(se); | |
998 | ||
60f2415e | 999 | __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats); |
3ea94de1 | 1000 | } |
3ea94de1 | 1001 | |
bf0f6f24 IM |
1002 | /* |
1003 | * Task is being enqueued - update stats: | |
1004 | */ | |
cb251765 | 1005 | static inline void |
60f2415e | 1006 | update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1007 | { |
4fa8d299 JP |
1008 | if (!schedstat_enabled()) |
1009 | return; | |
1010 | ||
bf0f6f24 IM |
1011 | /* |
1012 | * Are we enqueueing a waiting task? (for current tasks | |
1013 | * a dequeue/enqueue event is a NOP) | |
1014 | */ | |
429d43bc | 1015 | if (se != cfs_rq->curr) |
60f2415e | 1016 | update_stats_wait_start_fair(cfs_rq, se); |
1a3d027c JP |
1017 | |
1018 | if (flags & ENQUEUE_WAKEUP) | |
60f2415e | 1019 | update_stats_enqueue_sleeper_fair(cfs_rq, se); |
bf0f6f24 IM |
1020 | } |
1021 | ||
bf0f6f24 | 1022 | static inline void |
60f2415e | 1023 | update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1024 | { |
4fa8d299 JP |
1025 | |
1026 | if (!schedstat_enabled()) | |
1027 | return; | |
1028 | ||
bf0f6f24 IM |
1029 | /* |
1030 | * Mark the end of the wait period if dequeueing a | |
1031 | * waiting task: | |
1032 | */ | |
429d43bc | 1033 | if (se != cfs_rq->curr) |
60f2415e | 1034 | update_stats_wait_end_fair(cfs_rq, se); |
cb251765 | 1035 | |
4fa8d299 JP |
1036 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1037 | struct task_struct *tsk = task_of(se); | |
2f064a59 | 1038 | unsigned int state; |
cb251765 | 1039 | |
2f064a59 PZ |
1040 | /* XXX racy against TTWU */ |
1041 | state = READ_ONCE(tsk->__state); | |
1042 | if (state & TASK_INTERRUPTIBLE) | |
ceeadb83 | 1043 | __schedstat_set(tsk->stats.sleep_start, |
4fa8d299 | 1044 | rq_clock(rq_of(cfs_rq))); |
2f064a59 | 1045 | if (state & TASK_UNINTERRUPTIBLE) |
ceeadb83 | 1046 | __schedstat_set(tsk->stats.block_start, |
4fa8d299 | 1047 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1048 | } |
cb251765 MG |
1049 | } |
1050 | ||
bf0f6f24 IM |
1051 | /* |
1052 | * We are picking a new current task - update its stats: | |
1053 | */ | |
1054 | static inline void | |
79303e9e | 1055 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1056 | { |
1057 | /* | |
1058 | * We are starting a new run period: | |
1059 | */ | |
78becc27 | 1060 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1061 | } |
1062 | ||
bf0f6f24 IM |
1063 | /************************************************** |
1064 | * Scheduling class queueing methods: | |
1065 | */ | |
1066 | ||
cb29a5c1 MG |
1067 | #ifdef CONFIG_NUMA |
1068 | #define NUMA_IMBALANCE_MIN 2 | |
1069 | ||
1070 | static inline long | |
1071 | adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr) | |
1072 | { | |
1073 | /* | |
1074 | * Allow a NUMA imbalance if busy CPUs is less than the maximum | |
1075 | * threshold. Above this threshold, individual tasks may be contending | |
1076 | * for both memory bandwidth and any shared HT resources. This is an | |
1077 | * approximation as the number of running tasks may not be related to | |
1078 | * the number of busy CPUs due to sched_setaffinity. | |
1079 | */ | |
1080 | if (dst_running > imb_numa_nr) | |
1081 | return imbalance; | |
1082 | ||
1083 | /* | |
1084 | * Allow a small imbalance based on a simple pair of communicating | |
1085 | * tasks that remain local when the destination is lightly loaded. | |
1086 | */ | |
1087 | if (imbalance <= NUMA_IMBALANCE_MIN) | |
1088 | return 0; | |
1089 | ||
1090 | return imbalance; | |
1091 | } | |
1092 | #endif /* CONFIG_NUMA */ | |
1093 | ||
cbee9f88 PZ |
1094 | #ifdef CONFIG_NUMA_BALANCING |
1095 | /* | |
598f0ec0 MG |
1096 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1097 | * calculated based on the tasks virtual memory size and | |
1098 | * numa_balancing_scan_size. | |
cbee9f88 | 1099 | */ |
598f0ec0 MG |
1100 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1101 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1102 | |
1103 | /* Portion of address space to scan in MB */ | |
1104 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1105 | |
4b96a29b PZ |
1106 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1107 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1108 | ||
33024536 HY |
1109 | /* The page with hint page fault latency < threshold in ms is considered hot */ |
1110 | unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC; | |
1111 | ||
b5dd77c8 | 1112 | struct numa_group { |
c45a7795 | 1113 | refcount_t refcount; |
b5dd77c8 RR |
1114 | |
1115 | spinlock_t lock; /* nr_tasks, tasks */ | |
1116 | int nr_tasks; | |
1117 | pid_t gid; | |
1118 | int active_nodes; | |
1119 | ||
1120 | struct rcu_head rcu; | |
1121 | unsigned long total_faults; | |
1122 | unsigned long max_faults_cpu; | |
1123 | /* | |
5b763a14 BR |
1124 | * faults[] array is split into two regions: faults_mem and faults_cpu. |
1125 | * | |
b5dd77c8 RR |
1126 | * Faults_cpu is used to decide whether memory should move |
1127 | * towards the CPU. As a consequence, these stats are weighted | |
1128 | * more by CPU use than by memory faults. | |
1129 | */ | |
04f5c362 | 1130 | unsigned long faults[]; |
b5dd77c8 RR |
1131 | }; |
1132 | ||
cb361d8c JH |
1133 | /* |
1134 | * For functions that can be called in multiple contexts that permit reading | |
1135 | * ->numa_group (see struct task_struct for locking rules). | |
1136 | */ | |
1137 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1138 | { | |
1139 | return rcu_dereference_check(p->numa_group, p == current || | |
9ef7e7e3 | 1140 | (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu))); |
cb361d8c JH |
1141 | } |
1142 | ||
1143 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1144 | { | |
1145 | return rcu_dereference_protected(p->numa_group, p == current); | |
1146 | } | |
1147 | ||
b5dd77c8 RR |
1148 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1149 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1150 | ||
598f0ec0 MG |
1151 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1152 | { | |
1153 | unsigned long rss = 0; | |
1154 | unsigned long nr_scan_pages; | |
1155 | ||
1156 | /* | |
1157 | * Calculations based on RSS as non-present and empty pages are skipped | |
1158 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1159 | * on resident pages | |
1160 | */ | |
1161 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1162 | rss = get_mm_rss(p->mm); | |
1163 | if (!rss) | |
1164 | rss = nr_scan_pages; | |
1165 | ||
1166 | rss = round_up(rss, nr_scan_pages); | |
1167 | return rss / nr_scan_pages; | |
1168 | } | |
1169 | ||
3b03706f | 1170 | /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ |
598f0ec0 MG |
1171 | #define MAX_SCAN_WINDOW 2560 |
1172 | ||
1173 | static unsigned int task_scan_min(struct task_struct *p) | |
1174 | { | |
316c1608 | 1175 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1176 | unsigned int scan, floor; |
1177 | unsigned int windows = 1; | |
1178 | ||
64192658 KT |
1179 | if (scan_size < MAX_SCAN_WINDOW) |
1180 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1181 | floor = 1000 / windows; |
1182 | ||
1183 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1184 | return max_t(unsigned int, floor, scan); | |
1185 | } | |
1186 | ||
b5dd77c8 RR |
1187 | static unsigned int task_scan_start(struct task_struct *p) |
1188 | { | |
1189 | unsigned long smin = task_scan_min(p); | |
1190 | unsigned long period = smin; | |
cb361d8c | 1191 | struct numa_group *ng; |
b5dd77c8 RR |
1192 | |
1193 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1194 | rcu_read_lock(); |
1195 | ng = rcu_dereference(p->numa_group); | |
1196 | if (ng) { | |
b5dd77c8 RR |
1197 | unsigned long shared = group_faults_shared(ng); |
1198 | unsigned long private = group_faults_priv(ng); | |
1199 | ||
c45a7795 | 1200 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1201 | period *= shared + 1; |
1202 | period /= private + shared + 1; | |
1203 | } | |
cb361d8c | 1204 | rcu_read_unlock(); |
b5dd77c8 RR |
1205 | |
1206 | return max(smin, period); | |
1207 | } | |
1208 | ||
598f0ec0 MG |
1209 | static unsigned int task_scan_max(struct task_struct *p) |
1210 | { | |
b5dd77c8 RR |
1211 | unsigned long smin = task_scan_min(p); |
1212 | unsigned long smax; | |
cb361d8c | 1213 | struct numa_group *ng; |
598f0ec0 MG |
1214 | |
1215 | /* Watch for min being lower than max due to floor calculations */ | |
1216 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1217 | |
1218 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1219 | ng = deref_curr_numa_group(p); |
1220 | if (ng) { | |
b5dd77c8 RR |
1221 | unsigned long shared = group_faults_shared(ng); |
1222 | unsigned long private = group_faults_priv(ng); | |
1223 | unsigned long period = smax; | |
1224 | ||
c45a7795 | 1225 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1226 | period *= shared + 1; |
1227 | period /= private + shared + 1; | |
1228 | ||
1229 | smax = max(smax, period); | |
1230 | } | |
1231 | ||
598f0ec0 MG |
1232 | return max(smin, smax); |
1233 | } | |
1234 | ||
0ec8aa00 PZ |
1235 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1236 | { | |
98fa15f3 | 1237 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1238 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1239 | } | |
1240 | ||
1241 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1242 | { | |
98fa15f3 | 1243 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1244 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1245 | } | |
1246 | ||
be1e4e76 RR |
1247 | /* Shared or private faults. */ |
1248 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1249 | ||
1250 | /* Memory and CPU locality */ | |
1251 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1252 | ||
1253 | /* Averaged statistics, and temporary buffers. */ | |
1254 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1255 | ||
e29cf08b MG |
1256 | pid_t task_numa_group_id(struct task_struct *p) |
1257 | { | |
cb361d8c JH |
1258 | struct numa_group *ng; |
1259 | pid_t gid = 0; | |
1260 | ||
1261 | rcu_read_lock(); | |
1262 | ng = rcu_dereference(p->numa_group); | |
1263 | if (ng) | |
1264 | gid = ng->gid; | |
1265 | rcu_read_unlock(); | |
1266 | ||
1267 | return gid; | |
e29cf08b MG |
1268 | } |
1269 | ||
44dba3d5 | 1270 | /* |
97fb7a0a | 1271 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1272 | * occupy the first half of the array. The second half of the |
1273 | * array is for current counters, which are averaged into the | |
1274 | * first set by task_numa_placement. | |
1275 | */ | |
1276 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1277 | { |
44dba3d5 | 1278 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1279 | } |
1280 | ||
1281 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1282 | { | |
44dba3d5 | 1283 | if (!p->numa_faults) |
ac8e895b MG |
1284 | return 0; |
1285 | ||
44dba3d5 IM |
1286 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1287 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1288 | } |
1289 | ||
83e1d2cd MG |
1290 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1291 | { | |
cb361d8c JH |
1292 | struct numa_group *ng = deref_task_numa_group(p); |
1293 | ||
1294 | if (!ng) | |
83e1d2cd MG |
1295 | return 0; |
1296 | ||
cb361d8c JH |
1297 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1298 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1299 | } |
1300 | ||
20e07dea RR |
1301 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1302 | { | |
5b763a14 BR |
1303 | return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] + |
1304 | group->faults[task_faults_idx(NUMA_CPU, nid, 1)]; | |
20e07dea RR |
1305 | } |
1306 | ||
b5dd77c8 RR |
1307 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1308 | { | |
1309 | unsigned long faults = 0; | |
1310 | int node; | |
1311 | ||
1312 | for_each_online_node(node) { | |
1313 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1314 | } | |
1315 | ||
1316 | return faults; | |
1317 | } | |
1318 | ||
1319 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1320 | { | |
1321 | unsigned long faults = 0; | |
1322 | int node; | |
1323 | ||
1324 | for_each_online_node(node) { | |
1325 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1326 | } | |
1327 | ||
1328 | return faults; | |
1329 | } | |
1330 | ||
4142c3eb RR |
1331 | /* |
1332 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1333 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1334 | * between these nodes are slowed down, to allow things to settle down. | |
1335 | */ | |
1336 | #define ACTIVE_NODE_FRACTION 3 | |
1337 | ||
1338 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1339 | { | |
1340 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1341 | } | |
1342 | ||
6c6b1193 RR |
1343 | /* Handle placement on systems where not all nodes are directly connected. */ |
1344 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
0fb3978b | 1345 | int lim_dist, bool task) |
6c6b1193 RR |
1346 | { |
1347 | unsigned long score = 0; | |
0fb3978b | 1348 | int node, max_dist; |
6c6b1193 RR |
1349 | |
1350 | /* | |
1351 | * All nodes are directly connected, and the same distance | |
1352 | * from each other. No need for fancy placement algorithms. | |
1353 | */ | |
1354 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1355 | return 0; | |
1356 | ||
0fb3978b HY |
1357 | /* sched_max_numa_distance may be changed in parallel. */ |
1358 | max_dist = READ_ONCE(sched_max_numa_distance); | |
6c6b1193 RR |
1359 | /* |
1360 | * This code is called for each node, introducing N^2 complexity, | |
1361 | * which should be ok given the number of nodes rarely exceeds 8. | |
1362 | */ | |
1363 | for_each_online_node(node) { | |
1364 | unsigned long faults; | |
1365 | int dist = node_distance(nid, node); | |
1366 | ||
1367 | /* | |
1368 | * The furthest away nodes in the system are not interesting | |
1369 | * for placement; nid was already counted. | |
1370 | */ | |
0fb3978b | 1371 | if (dist >= max_dist || node == nid) |
6c6b1193 RR |
1372 | continue; |
1373 | ||
1374 | /* | |
1375 | * On systems with a backplane NUMA topology, compare groups | |
1376 | * of nodes, and move tasks towards the group with the most | |
1377 | * memory accesses. When comparing two nodes at distance | |
1378 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1379 | * of each group. Skip other nodes. | |
1380 | */ | |
0fb3978b | 1381 | if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist) |
6c6b1193 RR |
1382 | continue; |
1383 | ||
1384 | /* Add up the faults from nearby nodes. */ | |
1385 | if (task) | |
1386 | faults = task_faults(p, node); | |
1387 | else | |
1388 | faults = group_faults(p, node); | |
1389 | ||
1390 | /* | |
1391 | * On systems with a glueless mesh NUMA topology, there are | |
1392 | * no fixed "groups of nodes". Instead, nodes that are not | |
1393 | * directly connected bounce traffic through intermediate | |
1394 | * nodes; a numa_group can occupy any set of nodes. | |
1395 | * The further away a node is, the less the faults count. | |
1396 | * This seems to result in good task placement. | |
1397 | */ | |
1398 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
0fb3978b HY |
1399 | faults *= (max_dist - dist); |
1400 | faults /= (max_dist - LOCAL_DISTANCE); | |
6c6b1193 RR |
1401 | } |
1402 | ||
1403 | score += faults; | |
1404 | } | |
1405 | ||
1406 | return score; | |
1407 | } | |
1408 | ||
83e1d2cd MG |
1409 | /* |
1410 | * These return the fraction of accesses done by a particular task, or | |
1411 | * task group, on a particular numa node. The group weight is given a | |
1412 | * larger multiplier, in order to group tasks together that are almost | |
1413 | * evenly spread out between numa nodes. | |
1414 | */ | |
7bd95320 RR |
1415 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1416 | int dist) | |
83e1d2cd | 1417 | { |
7bd95320 | 1418 | unsigned long faults, total_faults; |
83e1d2cd | 1419 | |
44dba3d5 | 1420 | if (!p->numa_faults) |
83e1d2cd MG |
1421 | return 0; |
1422 | ||
1423 | total_faults = p->total_numa_faults; | |
1424 | ||
1425 | if (!total_faults) | |
1426 | return 0; | |
1427 | ||
7bd95320 | 1428 | faults = task_faults(p, nid); |
6c6b1193 RR |
1429 | faults += score_nearby_nodes(p, nid, dist, true); |
1430 | ||
7bd95320 | 1431 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1432 | } |
1433 | ||
7bd95320 RR |
1434 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1435 | int dist) | |
83e1d2cd | 1436 | { |
cb361d8c | 1437 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1438 | unsigned long faults, total_faults; |
1439 | ||
cb361d8c | 1440 | if (!ng) |
7bd95320 RR |
1441 | return 0; |
1442 | ||
cb361d8c | 1443 | total_faults = ng->total_faults; |
7bd95320 RR |
1444 | |
1445 | if (!total_faults) | |
83e1d2cd MG |
1446 | return 0; |
1447 | ||
7bd95320 | 1448 | faults = group_faults(p, nid); |
6c6b1193 RR |
1449 | faults += score_nearby_nodes(p, nid, dist, false); |
1450 | ||
7bd95320 | 1451 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1452 | } |
1453 | ||
33024536 HY |
1454 | /* |
1455 | * If memory tiering mode is enabled, cpupid of slow memory page is | |
1456 | * used to record scan time instead of CPU and PID. When tiering mode | |
1457 | * is disabled at run time, the scan time (in cpupid) will be | |
1458 | * interpreted as CPU and PID. So CPU needs to be checked to avoid to | |
1459 | * access out of array bound. | |
1460 | */ | |
1461 | static inline bool cpupid_valid(int cpupid) | |
1462 | { | |
1463 | return cpupid_to_cpu(cpupid) < nr_cpu_ids; | |
1464 | } | |
1465 | ||
1466 | /* | |
1467 | * For memory tiering mode, if there are enough free pages (more than | |
1468 | * enough watermark defined here) in fast memory node, to take full | |
1469 | * advantage of fast memory capacity, all recently accessed slow | |
1470 | * memory pages will be migrated to fast memory node without | |
1471 | * considering hot threshold. | |
1472 | */ | |
1473 | static bool pgdat_free_space_enough(struct pglist_data *pgdat) | |
1474 | { | |
1475 | int z; | |
1476 | unsigned long enough_wmark; | |
1477 | ||
1478 | enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT, | |
1479 | pgdat->node_present_pages >> 4); | |
1480 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1481 | struct zone *zone = pgdat->node_zones + z; | |
1482 | ||
1483 | if (!populated_zone(zone)) | |
1484 | continue; | |
1485 | ||
1486 | if (zone_watermark_ok(zone, 0, | |
1487 | wmark_pages(zone, WMARK_PROMO) + enough_wmark, | |
1488 | ZONE_MOVABLE, 0)) | |
1489 | return true; | |
1490 | } | |
1491 | return false; | |
1492 | } | |
1493 | ||
1494 | /* | |
1495 | * For memory tiering mode, when page tables are scanned, the scan | |
1496 | * time will be recorded in struct page in addition to make page | |
1497 | * PROT_NONE for slow memory page. So when the page is accessed, in | |
1498 | * hint page fault handler, the hint page fault latency is calculated | |
1499 | * via, | |
1500 | * | |
1501 | * hint page fault latency = hint page fault time - scan time | |
1502 | * | |
1503 | * The smaller the hint page fault latency, the higher the possibility | |
1504 | * for the page to be hot. | |
1505 | */ | |
1506 | static int numa_hint_fault_latency(struct page *page) | |
1507 | { | |
1508 | int last_time, time; | |
1509 | ||
1510 | time = jiffies_to_msecs(jiffies); | |
1511 | last_time = xchg_page_access_time(page, time); | |
1512 | ||
1513 | return (time - last_time) & PAGE_ACCESS_TIME_MASK; | |
1514 | } | |
1515 | ||
c6833e10 HY |
1516 | /* |
1517 | * For memory tiering mode, too high promotion/demotion throughput may | |
1518 | * hurt application latency. So we provide a mechanism to rate limit | |
1519 | * the number of pages that are tried to be promoted. | |
1520 | */ | |
1521 | static bool numa_promotion_rate_limit(struct pglist_data *pgdat, | |
1522 | unsigned long rate_limit, int nr) | |
1523 | { | |
1524 | unsigned long nr_cand; | |
1525 | unsigned int now, start; | |
1526 | ||
1527 | now = jiffies_to_msecs(jiffies); | |
1528 | mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr); | |
1529 | nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE); | |
1530 | start = pgdat->nbp_rl_start; | |
1531 | if (now - start > MSEC_PER_SEC && | |
1532 | cmpxchg(&pgdat->nbp_rl_start, start, now) == start) | |
1533 | pgdat->nbp_rl_nr_cand = nr_cand; | |
1534 | if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit) | |
1535 | return true; | |
1536 | return false; | |
1537 | } | |
1538 | ||
c959924b HY |
1539 | #define NUMA_MIGRATION_ADJUST_STEPS 16 |
1540 | ||
1541 | static void numa_promotion_adjust_threshold(struct pglist_data *pgdat, | |
1542 | unsigned long rate_limit, | |
1543 | unsigned int ref_th) | |
1544 | { | |
1545 | unsigned int now, start, th_period, unit_th, th; | |
1546 | unsigned long nr_cand, ref_cand, diff_cand; | |
1547 | ||
1548 | now = jiffies_to_msecs(jiffies); | |
1549 | th_period = sysctl_numa_balancing_scan_period_max; | |
1550 | start = pgdat->nbp_th_start; | |
1551 | if (now - start > th_period && | |
1552 | cmpxchg(&pgdat->nbp_th_start, start, now) == start) { | |
1553 | ref_cand = rate_limit * | |
1554 | sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC; | |
1555 | nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE); | |
1556 | diff_cand = nr_cand - pgdat->nbp_th_nr_cand; | |
1557 | unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS; | |
1558 | th = pgdat->nbp_threshold ? : ref_th; | |
1559 | if (diff_cand > ref_cand * 11 / 10) | |
1560 | th = max(th - unit_th, unit_th); | |
1561 | else if (diff_cand < ref_cand * 9 / 10) | |
1562 | th = min(th + unit_th, ref_th * 2); | |
1563 | pgdat->nbp_th_nr_cand = nr_cand; | |
1564 | pgdat->nbp_threshold = th; | |
1565 | } | |
1566 | } | |
1567 | ||
10f39042 RR |
1568 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1569 | int src_nid, int dst_cpu) | |
1570 | { | |
cb361d8c | 1571 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1572 | int dst_nid = cpu_to_node(dst_cpu); |
1573 | int last_cpupid, this_cpupid; | |
1574 | ||
33024536 HY |
1575 | /* |
1576 | * The pages in slow memory node should be migrated according | |
1577 | * to hot/cold instead of private/shared. | |
1578 | */ | |
1579 | if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING && | |
1580 | !node_is_toptier(src_nid)) { | |
1581 | struct pglist_data *pgdat; | |
c959924b HY |
1582 | unsigned long rate_limit; |
1583 | unsigned int latency, th, def_th; | |
33024536 HY |
1584 | |
1585 | pgdat = NODE_DATA(dst_nid); | |
c959924b HY |
1586 | if (pgdat_free_space_enough(pgdat)) { |
1587 | /* workload changed, reset hot threshold */ | |
1588 | pgdat->nbp_threshold = 0; | |
33024536 | 1589 | return true; |
c959924b HY |
1590 | } |
1591 | ||
1592 | def_th = sysctl_numa_balancing_hot_threshold; | |
1593 | rate_limit = sysctl_numa_balancing_promote_rate_limit << \ | |
1594 | (20 - PAGE_SHIFT); | |
1595 | numa_promotion_adjust_threshold(pgdat, rate_limit, def_th); | |
33024536 | 1596 | |
c959924b | 1597 | th = pgdat->nbp_threshold ? : def_th; |
33024536 HY |
1598 | latency = numa_hint_fault_latency(page); |
1599 | if (latency >= th) | |
1600 | return false; | |
1601 | ||
c6833e10 HY |
1602 | return !numa_promotion_rate_limit(pgdat, rate_limit, |
1603 | thp_nr_pages(page)); | |
33024536 HY |
1604 | } |
1605 | ||
10f39042 | 1606 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); |
37355bdc MG |
1607 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1608 | ||
33024536 HY |
1609 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) && |
1610 | !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid)) | |
1611 | return false; | |
1612 | ||
37355bdc MG |
1613 | /* |
1614 | * Allow first faults or private faults to migrate immediately early in | |
1615 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1616 | * two full passes of the "multi-stage node selection" test that is | |
1617 | * executed below. | |
1618 | */ | |
98fa15f3 | 1619 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1620 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1621 | return true; | |
10f39042 RR |
1622 | |
1623 | /* | |
1624 | * Multi-stage node selection is used in conjunction with a periodic | |
1625 | * migration fault to build a temporal task<->page relation. By using | |
1626 | * a two-stage filter we remove short/unlikely relations. | |
1627 | * | |
1628 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1629 | * a task's usage of a particular page (n_p) per total usage of this | |
1630 | * page (n_t) (in a given time-span) to a probability. | |
1631 | * | |
1632 | * Our periodic faults will sample this probability and getting the | |
1633 | * same result twice in a row, given these samples are fully | |
1634 | * independent, is then given by P(n)^2, provided our sample period | |
1635 | * is sufficiently short compared to the usage pattern. | |
1636 | * | |
1637 | * This quadric squishes small probabilities, making it less likely we | |
1638 | * act on an unlikely task<->page relation. | |
1639 | */ | |
10f39042 RR |
1640 | if (!cpupid_pid_unset(last_cpupid) && |
1641 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1642 | return false; | |
1643 | ||
1644 | /* Always allow migrate on private faults */ | |
1645 | if (cpupid_match_pid(p, last_cpupid)) | |
1646 | return true; | |
1647 | ||
1648 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1649 | if (!ng) | |
1650 | return true; | |
1651 | ||
1652 | /* | |
4142c3eb RR |
1653 | * Destination node is much more heavily used than the source |
1654 | * node? Allow migration. | |
10f39042 | 1655 | */ |
4142c3eb RR |
1656 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1657 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1658 | return true; |
1659 | ||
1660 | /* | |
4142c3eb RR |
1661 | * Distribute memory according to CPU & memory use on each node, |
1662 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1663 | * | |
1664 | * faults_cpu(dst) 3 faults_cpu(src) | |
1665 | * --------------- * - > --------------- | |
1666 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1667 | */ |
4142c3eb RR |
1668 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1669 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1670 | } |
1671 | ||
6499b1b2 VG |
1672 | /* |
1673 | * 'numa_type' describes the node at the moment of load balancing. | |
1674 | */ | |
1675 | enum numa_type { | |
1676 | /* The node has spare capacity that can be used to run more tasks. */ | |
1677 | node_has_spare = 0, | |
1678 | /* | |
1679 | * The node is fully used and the tasks don't compete for more CPU | |
1680 | * cycles. Nevertheless, some tasks might wait before running. | |
1681 | */ | |
1682 | node_fully_busy, | |
1683 | /* | |
1684 | * The node is overloaded and can't provide expected CPU cycles to all | |
1685 | * tasks. | |
1686 | */ | |
1687 | node_overloaded | |
1688 | }; | |
58d081b5 | 1689 | |
fb13c7ee | 1690 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1691 | struct numa_stats { |
1692 | unsigned long load; | |
8e0e0eda | 1693 | unsigned long runnable; |
6499b1b2 | 1694 | unsigned long util; |
fb13c7ee | 1695 | /* Total compute capacity of CPUs on a node */ |
5ef20ca1 | 1696 | unsigned long compute_capacity; |
6499b1b2 VG |
1697 | unsigned int nr_running; |
1698 | unsigned int weight; | |
1699 | enum numa_type node_type; | |
ff7db0bf | 1700 | int idle_cpu; |
58d081b5 | 1701 | }; |
e6628d5b | 1702 | |
ff7db0bf MG |
1703 | static inline bool is_core_idle(int cpu) |
1704 | { | |
1705 | #ifdef CONFIG_SCHED_SMT | |
1706 | int sibling; | |
1707 | ||
1708 | for_each_cpu(sibling, cpu_smt_mask(cpu)) { | |
1709 | if (cpu == sibling) | |
1710 | continue; | |
1711 | ||
1c6829cf | 1712 | if (!idle_cpu(sibling)) |
ff7db0bf MG |
1713 | return false; |
1714 | } | |
1715 | #endif | |
1716 | ||
1717 | return true; | |
1718 | } | |
1719 | ||
58d081b5 MG |
1720 | struct task_numa_env { |
1721 | struct task_struct *p; | |
e6628d5b | 1722 | |
58d081b5 MG |
1723 | int src_cpu, src_nid; |
1724 | int dst_cpu, dst_nid; | |
e496132e | 1725 | int imb_numa_nr; |
e6628d5b | 1726 | |
58d081b5 | 1727 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1728 | |
40ea2b42 | 1729 | int imbalance_pct; |
7bd95320 | 1730 | int dist; |
fb13c7ee MG |
1731 | |
1732 | struct task_struct *best_task; | |
1733 | long best_imp; | |
58d081b5 MG |
1734 | int best_cpu; |
1735 | }; | |
1736 | ||
6499b1b2 | 1737 | static unsigned long cpu_load(struct rq *rq); |
8e0e0eda | 1738 | static unsigned long cpu_runnable(struct rq *rq); |
6499b1b2 VG |
1739 | |
1740 | static inline enum | |
1741 | numa_type numa_classify(unsigned int imbalance_pct, | |
1742 | struct numa_stats *ns) | |
1743 | { | |
1744 | if ((ns->nr_running > ns->weight) && | |
8e0e0eda VG |
1745 | (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || |
1746 | ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) | |
6499b1b2 VG |
1747 | return node_overloaded; |
1748 | ||
1749 | if ((ns->nr_running < ns->weight) || | |
8e0e0eda VG |
1750 | (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && |
1751 | ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) | |
6499b1b2 VG |
1752 | return node_has_spare; |
1753 | ||
1754 | return node_fully_busy; | |
1755 | } | |
1756 | ||
76c389ab VS |
1757 | #ifdef CONFIG_SCHED_SMT |
1758 | /* Forward declarations of select_idle_sibling helpers */ | |
398ba2b0 | 1759 | static inline bool test_idle_cores(int cpu); |
ff7db0bf MG |
1760 | static inline int numa_idle_core(int idle_core, int cpu) |
1761 | { | |
ff7db0bf | 1762 | if (!static_branch_likely(&sched_smt_present) || |
398ba2b0 | 1763 | idle_core >= 0 || !test_idle_cores(cpu)) |
ff7db0bf MG |
1764 | return idle_core; |
1765 | ||
1766 | /* | |
1767 | * Prefer cores instead of packing HT siblings | |
1768 | * and triggering future load balancing. | |
1769 | */ | |
1770 | if (is_core_idle(cpu)) | |
1771 | idle_core = cpu; | |
ff7db0bf MG |
1772 | |
1773 | return idle_core; | |
1774 | } | |
76c389ab VS |
1775 | #else |
1776 | static inline int numa_idle_core(int idle_core, int cpu) | |
1777 | { | |
1778 | return idle_core; | |
1779 | } | |
1780 | #endif | |
ff7db0bf | 1781 | |
6499b1b2 | 1782 | /* |
ff7db0bf MG |
1783 | * Gather all necessary information to make NUMA balancing placement |
1784 | * decisions that are compatible with standard load balancer. This | |
1785 | * borrows code and logic from update_sg_lb_stats but sharing a | |
1786 | * common implementation is impractical. | |
6499b1b2 VG |
1787 | */ |
1788 | static void update_numa_stats(struct task_numa_env *env, | |
ff7db0bf MG |
1789 | struct numa_stats *ns, int nid, |
1790 | bool find_idle) | |
6499b1b2 | 1791 | { |
ff7db0bf | 1792 | int cpu, idle_core = -1; |
6499b1b2 VG |
1793 | |
1794 | memset(ns, 0, sizeof(*ns)); | |
ff7db0bf MG |
1795 | ns->idle_cpu = -1; |
1796 | ||
0621df31 | 1797 | rcu_read_lock(); |
6499b1b2 VG |
1798 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
1799 | struct rq *rq = cpu_rq(cpu); | |
1800 | ||
1801 | ns->load += cpu_load(rq); | |
8e0e0eda | 1802 | ns->runnable += cpu_runnable(rq); |
82762d2a | 1803 | ns->util += cpu_util_cfs(cpu); |
6499b1b2 VG |
1804 | ns->nr_running += rq->cfs.h_nr_running; |
1805 | ns->compute_capacity += capacity_of(cpu); | |
ff7db0bf MG |
1806 | |
1807 | if (find_idle && !rq->nr_running && idle_cpu(cpu)) { | |
1808 | if (READ_ONCE(rq->numa_migrate_on) || | |
1809 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) | |
1810 | continue; | |
1811 | ||
1812 | if (ns->idle_cpu == -1) | |
1813 | ns->idle_cpu = cpu; | |
1814 | ||
1815 | idle_core = numa_idle_core(idle_core, cpu); | |
1816 | } | |
6499b1b2 | 1817 | } |
0621df31 | 1818 | rcu_read_unlock(); |
6499b1b2 VG |
1819 | |
1820 | ns->weight = cpumask_weight(cpumask_of_node(nid)); | |
1821 | ||
1822 | ns->node_type = numa_classify(env->imbalance_pct, ns); | |
ff7db0bf MG |
1823 | |
1824 | if (idle_core >= 0) | |
1825 | ns->idle_cpu = idle_core; | |
6499b1b2 VG |
1826 | } |
1827 | ||
fb13c7ee MG |
1828 | static void task_numa_assign(struct task_numa_env *env, |
1829 | struct task_struct *p, long imp) | |
1830 | { | |
a4739eca SD |
1831 | struct rq *rq = cpu_rq(env->dst_cpu); |
1832 | ||
5fb52dd9 MG |
1833 | /* Check if run-queue part of active NUMA balance. */ |
1834 | if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { | |
1835 | int cpu; | |
1836 | int start = env->dst_cpu; | |
1837 | ||
1838 | /* Find alternative idle CPU. */ | |
1839 | for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { | |
1840 | if (cpu == env->best_cpu || !idle_cpu(cpu) || | |
1841 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { | |
1842 | continue; | |
1843 | } | |
1844 | ||
1845 | env->dst_cpu = cpu; | |
1846 | rq = cpu_rq(env->dst_cpu); | |
1847 | if (!xchg(&rq->numa_migrate_on, 1)) | |
1848 | goto assign; | |
1849 | } | |
1850 | ||
1851 | /* Failed to find an alternative idle CPU */ | |
a4739eca | 1852 | return; |
5fb52dd9 | 1853 | } |
a4739eca | 1854 | |
5fb52dd9 | 1855 | assign: |
a4739eca SD |
1856 | /* |
1857 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1858 | * found a better CPU to move/swap. | |
1859 | */ | |
5fb52dd9 | 1860 | if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { |
a4739eca SD |
1861 | rq = cpu_rq(env->best_cpu); |
1862 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1863 | } | |
1864 | ||
fb13c7ee MG |
1865 | if (env->best_task) |
1866 | put_task_struct(env->best_task); | |
bac78573 ON |
1867 | if (p) |
1868 | get_task_struct(p); | |
fb13c7ee MG |
1869 | |
1870 | env->best_task = p; | |
1871 | env->best_imp = imp; | |
1872 | env->best_cpu = env->dst_cpu; | |
1873 | } | |
1874 | ||
28a21745 | 1875 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1876 | struct task_numa_env *env) |
1877 | { | |
e4991b24 RR |
1878 | long imb, old_imb; |
1879 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1880 | long src_capacity, dst_capacity; |
1881 | ||
1882 | /* | |
1883 | * The load is corrected for the CPU capacity available on each node. | |
1884 | * | |
1885 | * src_load dst_load | |
1886 | * ------------ vs --------- | |
1887 | * src_capacity dst_capacity | |
1888 | */ | |
1889 | src_capacity = env->src_stats.compute_capacity; | |
1890 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1891 | |
5f95ba7a | 1892 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1893 | |
28a21745 | 1894 | orig_src_load = env->src_stats.load; |
e4991b24 | 1895 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1896 | |
5f95ba7a | 1897 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1898 | |
1899 | /* Would this change make things worse? */ | |
1900 | return (imb > old_imb); | |
e63da036 RR |
1901 | } |
1902 | ||
6fd98e77 SD |
1903 | /* |
1904 | * Maximum NUMA importance can be 1998 (2*999); | |
1905 | * SMALLIMP @ 30 would be close to 1998/64. | |
1906 | * Used to deter task migration. | |
1907 | */ | |
1908 | #define SMALLIMP 30 | |
1909 | ||
fb13c7ee MG |
1910 | /* |
1911 | * This checks if the overall compute and NUMA accesses of the system would | |
1912 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1913 | * into account that it might be best if task running on the dst_cpu should | |
1914 | * be exchanged with the source task | |
1915 | */ | |
a0f03b61 | 1916 | static bool task_numa_compare(struct task_numa_env *env, |
305c1fac | 1917 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1918 | { |
cb361d8c | 1919 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 1920 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 1921 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 1922 | struct task_struct *cur; |
28a21745 | 1923 | long src_load, dst_load; |
7bd95320 | 1924 | int dist = env->dist; |
cb361d8c JH |
1925 | long moveimp = imp; |
1926 | long load; | |
a0f03b61 | 1927 | bool stopsearch = false; |
fb13c7ee | 1928 | |
a4739eca | 1929 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
a0f03b61 | 1930 | return false; |
a4739eca | 1931 | |
fb13c7ee | 1932 | rcu_read_lock(); |
154abafc | 1933 | cur = rcu_dereference(dst_rq->curr); |
bac78573 | 1934 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) |
fb13c7ee MG |
1935 | cur = NULL; |
1936 | ||
7af68335 PZ |
1937 | /* |
1938 | * Because we have preemption enabled we can get migrated around and | |
1939 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1940 | */ | |
a0f03b61 MG |
1941 | if (cur == env->p) { |
1942 | stopsearch = true; | |
7af68335 | 1943 | goto unlock; |
a0f03b61 | 1944 | } |
7af68335 | 1945 | |
305c1fac | 1946 | if (!cur) { |
6fd98e77 | 1947 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1948 | goto assign; |
1949 | else | |
1950 | goto unlock; | |
1951 | } | |
1952 | ||
88cca72c MG |
1953 | /* Skip this swap candidate if cannot move to the source cpu. */ |
1954 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) | |
1955 | goto unlock; | |
1956 | ||
1957 | /* | |
1958 | * Skip this swap candidate if it is not moving to its preferred | |
1959 | * node and the best task is. | |
1960 | */ | |
1961 | if (env->best_task && | |
1962 | env->best_task->numa_preferred_nid == env->src_nid && | |
1963 | cur->numa_preferred_nid != env->src_nid) { | |
1964 | goto unlock; | |
1965 | } | |
1966 | ||
fb13c7ee MG |
1967 | /* |
1968 | * "imp" is the fault differential for the source task between the | |
1969 | * source and destination node. Calculate the total differential for | |
1970 | * the source task and potential destination task. The more negative | |
305c1fac | 1971 | * the value is, the more remote accesses that would be expected to |
fb13c7ee | 1972 | * be incurred if the tasks were swapped. |
88cca72c | 1973 | * |
305c1fac SD |
1974 | * If dst and source tasks are in the same NUMA group, or not |
1975 | * in any group then look only at task weights. | |
1976 | */ | |
cb361d8c JH |
1977 | cur_ng = rcu_dereference(cur->numa_group); |
1978 | if (cur_ng == p_ng) { | |
13ede331 MG |
1979 | /* |
1980 | * Do not swap within a group or between tasks that have | |
1981 | * no group if there is spare capacity. Swapping does | |
1982 | * not address the load imbalance and helps one task at | |
1983 | * the cost of punishing another. | |
1984 | */ | |
1985 | if (env->dst_stats.node_type == node_has_spare) | |
1986 | goto unlock; | |
1987 | ||
305c1fac SD |
1988 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1989 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1990 | /* |
305c1fac SD |
1991 | * Add some hysteresis to prevent swapping the |
1992 | * tasks within a group over tiny differences. | |
887c290e | 1993 | */ |
cb361d8c | 1994 | if (cur_ng) |
305c1fac SD |
1995 | imp -= imp / 16; |
1996 | } else { | |
1997 | /* | |
1998 | * Compare the group weights. If a task is all by itself | |
1999 | * (not part of a group), use the task weight instead. | |
2000 | */ | |
cb361d8c | 2001 | if (cur_ng && p_ng) |
305c1fac SD |
2002 | imp += group_weight(cur, env->src_nid, dist) - |
2003 | group_weight(cur, env->dst_nid, dist); | |
2004 | else | |
2005 | imp += task_weight(cur, env->src_nid, dist) - | |
2006 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
2007 | } |
2008 | ||
88cca72c MG |
2009 | /* Discourage picking a task already on its preferred node */ |
2010 | if (cur->numa_preferred_nid == env->dst_nid) | |
2011 | imp -= imp / 16; | |
2012 | ||
2013 | /* | |
2014 | * Encourage picking a task that moves to its preferred node. | |
2015 | * This potentially makes imp larger than it's maximum of | |
2016 | * 1998 (see SMALLIMP and task_weight for why) but in this | |
2017 | * case, it does not matter. | |
2018 | */ | |
2019 | if (cur->numa_preferred_nid == env->src_nid) | |
2020 | imp += imp / 8; | |
2021 | ||
305c1fac | 2022 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 2023 | imp = moveimp; |
305c1fac | 2024 | cur = NULL; |
fb13c7ee | 2025 | goto assign; |
305c1fac | 2026 | } |
fb13c7ee | 2027 | |
88cca72c MG |
2028 | /* |
2029 | * Prefer swapping with a task moving to its preferred node over a | |
2030 | * task that is not. | |
2031 | */ | |
2032 | if (env->best_task && cur->numa_preferred_nid == env->src_nid && | |
2033 | env->best_task->numa_preferred_nid != env->src_nid) { | |
2034 | goto assign; | |
2035 | } | |
2036 | ||
6fd98e77 SD |
2037 | /* |
2038 | * If the NUMA importance is less than SMALLIMP, | |
2039 | * task migration might only result in ping pong | |
2040 | * of tasks and also hurt performance due to cache | |
2041 | * misses. | |
2042 | */ | |
2043 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
2044 | goto unlock; | |
2045 | ||
fb13c7ee MG |
2046 | /* |
2047 | * In the overloaded case, try and keep the load balanced. | |
2048 | */ | |
305c1fac SD |
2049 | load = task_h_load(env->p) - task_h_load(cur); |
2050 | if (!load) | |
2051 | goto assign; | |
2052 | ||
e720fff6 PZ |
2053 | dst_load = env->dst_stats.load + load; |
2054 | src_load = env->src_stats.load - load; | |
fb13c7ee | 2055 | |
28a21745 | 2056 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
2057 | goto unlock; |
2058 | ||
305c1fac | 2059 | assign: |
ff7db0bf | 2060 | /* Evaluate an idle CPU for a task numa move. */ |
10e2f1ac | 2061 | if (!cur) { |
ff7db0bf MG |
2062 | int cpu = env->dst_stats.idle_cpu; |
2063 | ||
2064 | /* Nothing cached so current CPU went idle since the search. */ | |
2065 | if (cpu < 0) | |
2066 | cpu = env->dst_cpu; | |
2067 | ||
10e2f1ac | 2068 | /* |
ff7db0bf MG |
2069 | * If the CPU is no longer truly idle and the previous best CPU |
2070 | * is, keep using it. | |
10e2f1ac | 2071 | */ |
ff7db0bf MG |
2072 | if (!idle_cpu(cpu) && env->best_cpu >= 0 && |
2073 | idle_cpu(env->best_cpu)) { | |
2074 | cpu = env->best_cpu; | |
2075 | } | |
2076 | ||
ff7db0bf | 2077 | env->dst_cpu = cpu; |
10e2f1ac | 2078 | } |
ba7e5a27 | 2079 | |
fb13c7ee | 2080 | task_numa_assign(env, cur, imp); |
a0f03b61 MG |
2081 | |
2082 | /* | |
2083 | * If a move to idle is allowed because there is capacity or load | |
2084 | * balance improves then stop the search. While a better swap | |
2085 | * candidate may exist, a search is not free. | |
2086 | */ | |
2087 | if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) | |
2088 | stopsearch = true; | |
2089 | ||
2090 | /* | |
2091 | * If a swap candidate must be identified and the current best task | |
2092 | * moves its preferred node then stop the search. | |
2093 | */ | |
2094 | if (!maymove && env->best_task && | |
2095 | env->best_task->numa_preferred_nid == env->src_nid) { | |
2096 | stopsearch = true; | |
2097 | } | |
fb13c7ee MG |
2098 | unlock: |
2099 | rcu_read_unlock(); | |
a0f03b61 MG |
2100 | |
2101 | return stopsearch; | |
fb13c7ee MG |
2102 | } |
2103 | ||
887c290e RR |
2104 | static void task_numa_find_cpu(struct task_numa_env *env, |
2105 | long taskimp, long groupimp) | |
2c8a50aa | 2106 | { |
305c1fac | 2107 | bool maymove = false; |
2c8a50aa MG |
2108 | int cpu; |
2109 | ||
305c1fac | 2110 | /* |
fb86f5b2 MG |
2111 | * If dst node has spare capacity, then check if there is an |
2112 | * imbalance that would be overruled by the load balancer. | |
305c1fac | 2113 | */ |
fb86f5b2 MG |
2114 | if (env->dst_stats.node_type == node_has_spare) { |
2115 | unsigned int imbalance; | |
2116 | int src_running, dst_running; | |
2117 | ||
2118 | /* | |
2119 | * Would movement cause an imbalance? Note that if src has | |
2120 | * more running tasks that the imbalance is ignored as the | |
2121 | * move improves the imbalance from the perspective of the | |
2122 | * CPU load balancer. | |
2123 | * */ | |
2124 | src_running = env->src_stats.nr_running - 1; | |
2125 | dst_running = env->dst_stats.nr_running + 1; | |
2126 | imbalance = max(0, dst_running - src_running); | |
7d2b5dd0 | 2127 | imbalance = adjust_numa_imbalance(imbalance, dst_running, |
e496132e | 2128 | env->imb_numa_nr); |
fb86f5b2 MG |
2129 | |
2130 | /* Use idle CPU if there is no imbalance */ | |
ff7db0bf | 2131 | if (!imbalance) { |
fb86f5b2 | 2132 | maymove = true; |
ff7db0bf MG |
2133 | if (env->dst_stats.idle_cpu >= 0) { |
2134 | env->dst_cpu = env->dst_stats.idle_cpu; | |
2135 | task_numa_assign(env, NULL, 0); | |
2136 | return; | |
2137 | } | |
2138 | } | |
fb86f5b2 MG |
2139 | } else { |
2140 | long src_load, dst_load, load; | |
2141 | /* | |
2142 | * If the improvement from just moving env->p direction is better | |
2143 | * than swapping tasks around, check if a move is possible. | |
2144 | */ | |
2145 | load = task_h_load(env->p); | |
2146 | dst_load = env->dst_stats.load + load; | |
2147 | src_load = env->src_stats.load - load; | |
2148 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
2149 | } | |
305c1fac | 2150 | |
2c8a50aa MG |
2151 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
2152 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 2153 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
2154 | continue; |
2155 | ||
2156 | env->dst_cpu = cpu; | |
a0f03b61 MG |
2157 | if (task_numa_compare(env, taskimp, groupimp, maymove)) |
2158 | break; | |
2c8a50aa MG |
2159 | } |
2160 | } | |
2161 | ||
58d081b5 MG |
2162 | static int task_numa_migrate(struct task_struct *p) |
2163 | { | |
58d081b5 MG |
2164 | struct task_numa_env env = { |
2165 | .p = p, | |
fb13c7ee | 2166 | |
58d081b5 | 2167 | .src_cpu = task_cpu(p), |
b32e86b4 | 2168 | .src_nid = task_node(p), |
fb13c7ee MG |
2169 | |
2170 | .imbalance_pct = 112, | |
2171 | ||
2172 | .best_task = NULL, | |
2173 | .best_imp = 0, | |
4142c3eb | 2174 | .best_cpu = -1, |
58d081b5 | 2175 | }; |
cb361d8c | 2176 | unsigned long taskweight, groupweight; |
58d081b5 | 2177 | struct sched_domain *sd; |
cb361d8c JH |
2178 | long taskimp, groupimp; |
2179 | struct numa_group *ng; | |
a4739eca | 2180 | struct rq *best_rq; |
7bd95320 | 2181 | int nid, ret, dist; |
e6628d5b | 2182 | |
58d081b5 | 2183 | /* |
fb13c7ee MG |
2184 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
2185 | * imbalance and would be the first to start moving tasks about. | |
2186 | * | |
2187 | * And we want to avoid any moving of tasks about, as that would create | |
2188 | * random movement of tasks -- counter the numa conditions we're trying | |
2189 | * to satisfy here. | |
58d081b5 MG |
2190 | */ |
2191 | rcu_read_lock(); | |
fb13c7ee | 2192 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
e496132e | 2193 | if (sd) { |
46a73e8a | 2194 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; |
e496132e MG |
2195 | env.imb_numa_nr = sd->imb_numa_nr; |
2196 | } | |
e6628d5b MG |
2197 | rcu_read_unlock(); |
2198 | ||
46a73e8a RR |
2199 | /* |
2200 | * Cpusets can break the scheduler domain tree into smaller | |
2201 | * balance domains, some of which do not cross NUMA boundaries. | |
2202 | * Tasks that are "trapped" in such domains cannot be migrated | |
2203 | * elsewhere, so there is no point in (re)trying. | |
2204 | */ | |
2205 | if (unlikely(!sd)) { | |
8cd45eee | 2206 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
2207 | return -EINVAL; |
2208 | } | |
2209 | ||
2c8a50aa | 2210 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
2211 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
2212 | taskweight = task_weight(p, env.src_nid, dist); | |
2213 | groupweight = group_weight(p, env.src_nid, dist); | |
ff7db0bf | 2214 | update_numa_stats(&env, &env.src_stats, env.src_nid, false); |
7bd95320 RR |
2215 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
2216 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
ff7db0bf | 2217 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
58d081b5 | 2218 | |
a43455a1 | 2219 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 2220 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 2221 | |
9de05d48 RR |
2222 | /* |
2223 | * Look at other nodes in these cases: | |
2224 | * - there is no space available on the preferred_nid | |
2225 | * - the task is part of a numa_group that is interleaved across | |
2226 | * multiple NUMA nodes; in order to better consolidate the group, | |
2227 | * we need to check other locations. | |
2228 | */ | |
cb361d8c JH |
2229 | ng = deref_curr_numa_group(p); |
2230 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
5c7b1aaf | 2231 | for_each_node_state(nid, N_CPU) { |
2c8a50aa MG |
2232 | if (nid == env.src_nid || nid == p->numa_preferred_nid) |
2233 | continue; | |
58d081b5 | 2234 | |
7bd95320 | 2235 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
2236 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
2237 | dist != env.dist) { | |
2238 | taskweight = task_weight(p, env.src_nid, dist); | |
2239 | groupweight = group_weight(p, env.src_nid, dist); | |
2240 | } | |
7bd95320 | 2241 | |
83e1d2cd | 2242 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
2243 | taskimp = task_weight(p, nid, dist) - taskweight; |
2244 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 2245 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
2246 | continue; |
2247 | ||
7bd95320 | 2248 | env.dist = dist; |
2c8a50aa | 2249 | env.dst_nid = nid; |
ff7db0bf | 2250 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
2d4056fa | 2251 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
2252 | } |
2253 | } | |
2254 | ||
68d1b02a RR |
2255 | /* |
2256 | * If the task is part of a workload that spans multiple NUMA nodes, | |
2257 | * and is migrating into one of the workload's active nodes, remember | |
2258 | * this node as the task's preferred numa node, so the workload can | |
2259 | * settle down. | |
2260 | * A task that migrated to a second choice node will be better off | |
2261 | * trying for a better one later. Do not set the preferred node here. | |
2262 | */ | |
cb361d8c | 2263 | if (ng) { |
db015dae RR |
2264 | if (env.best_cpu == -1) |
2265 | nid = env.src_nid; | |
2266 | else | |
8cd45eee | 2267 | nid = cpu_to_node(env.best_cpu); |
db015dae | 2268 | |
8cd45eee SD |
2269 | if (nid != p->numa_preferred_nid) |
2270 | sched_setnuma(p, nid); | |
db015dae RR |
2271 | } |
2272 | ||
2273 | /* No better CPU than the current one was found. */ | |
f22aef4a | 2274 | if (env.best_cpu == -1) { |
b2b2042b | 2275 | trace_sched_stick_numa(p, env.src_cpu, NULL, -1); |
db015dae | 2276 | return -EAGAIN; |
f22aef4a | 2277 | } |
0ec8aa00 | 2278 | |
a4739eca | 2279 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 2280 | if (env.best_task == NULL) { |
286549dc | 2281 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 2282 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc | 2283 | if (ret != 0) |
b2b2042b | 2284 | trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); |
fb13c7ee MG |
2285 | return ret; |
2286 | } | |
2287 | ||
0ad4e3df | 2288 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 2289 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 2290 | |
286549dc | 2291 | if (ret != 0) |
b2b2042b | 2292 | trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); |
fb13c7ee MG |
2293 | put_task_struct(env.best_task); |
2294 | return ret; | |
e6628d5b MG |
2295 | } |
2296 | ||
6b9a7460 MG |
2297 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
2298 | static void numa_migrate_preferred(struct task_struct *p) | |
2299 | { | |
5085e2a3 RR |
2300 | unsigned long interval = HZ; |
2301 | ||
2739d3ee | 2302 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 2303 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
2304 | return; |
2305 | ||
2739d3ee | 2306 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 2307 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 2308 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
2309 | |
2310 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 2311 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
2312 | return; |
2313 | ||
2314 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 2315 | task_numa_migrate(p); |
6b9a7460 MG |
2316 | } |
2317 | ||
20e07dea | 2318 | /* |
7d380f24 | 2319 | * Find out how many nodes the workload is actively running on. Do this by |
20e07dea RR |
2320 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
2321 | * be different from the set of nodes where the workload's memory is currently | |
2322 | * located. | |
20e07dea | 2323 | */ |
4142c3eb | 2324 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
2325 | { |
2326 | unsigned long faults, max_faults = 0; | |
4142c3eb | 2327 | int nid, active_nodes = 0; |
20e07dea | 2328 | |
5c7b1aaf | 2329 | for_each_node_state(nid, N_CPU) { |
20e07dea RR |
2330 | faults = group_faults_cpu(numa_group, nid); |
2331 | if (faults > max_faults) | |
2332 | max_faults = faults; | |
2333 | } | |
2334 | ||
5c7b1aaf | 2335 | for_each_node_state(nid, N_CPU) { |
20e07dea | 2336 | faults = group_faults_cpu(numa_group, nid); |
4142c3eb RR |
2337 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
2338 | active_nodes++; | |
20e07dea | 2339 | } |
4142c3eb RR |
2340 | |
2341 | numa_group->max_faults_cpu = max_faults; | |
2342 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
2343 | } |
2344 | ||
04bb2f94 RR |
2345 | /* |
2346 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
2347 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
2348 | * period will be for the next scan window. If local/(local+remote) ratio is |
2349 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
2350 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
2351 | */ |
2352 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 2353 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
2354 | |
2355 | /* | |
2356 | * Increase the scan period (slow down scanning) if the majority of | |
2357 | * our memory is already on our local node, or if the majority of | |
2358 | * the page accesses are shared with other processes. | |
2359 | * Otherwise, decrease the scan period. | |
2360 | */ | |
2361 | static void update_task_scan_period(struct task_struct *p, | |
2362 | unsigned long shared, unsigned long private) | |
2363 | { | |
2364 | unsigned int period_slot; | |
37ec97de | 2365 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
2366 | int diff; |
2367 | ||
2368 | unsigned long remote = p->numa_faults_locality[0]; | |
2369 | unsigned long local = p->numa_faults_locality[1]; | |
2370 | ||
2371 | /* | |
2372 | * If there were no record hinting faults then either the task is | |
7d380f24 | 2373 | * completely idle or all activity is in areas that are not of interest |
074c2381 MG |
2374 | * to automatic numa balancing. Related to that, if there were failed |
2375 | * migration then it implies we are migrating too quickly or the local | |
2376 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 2377 | */ |
074c2381 | 2378 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
2379 | p->numa_scan_period = min(p->numa_scan_period_max, |
2380 | p->numa_scan_period << 1); | |
2381 | ||
2382 | p->mm->numa_next_scan = jiffies + | |
2383 | msecs_to_jiffies(p->numa_scan_period); | |
2384 | ||
2385 | return; | |
2386 | } | |
2387 | ||
2388 | /* | |
2389 | * Prepare to scale scan period relative to the current period. | |
2390 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
2391 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
2392 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
2393 | */ | |
2394 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
2395 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
2396 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
2397 | ||
2398 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2399 | /* | |
2400 | * Most memory accesses are local. There is no need to | |
2401 | * do fast NUMA scanning, since memory is already local. | |
2402 | */ | |
2403 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
2404 | if (!slot) | |
2405 | slot = 1; | |
2406 | diff = slot * period_slot; | |
2407 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2408 | /* | |
2409 | * Most memory accesses are shared with other tasks. | |
2410 | * There is no point in continuing fast NUMA scanning, | |
2411 | * since other tasks may just move the memory elsewhere. | |
2412 | */ | |
2413 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
2414 | if (!slot) |
2415 | slot = 1; | |
2416 | diff = slot * period_slot; | |
2417 | } else { | |
04bb2f94 | 2418 | /* |
37ec97de RR |
2419 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
2420 | * yet they are not on the local NUMA node. Speed up | |
2421 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 2422 | */ |
37ec97de RR |
2423 | int ratio = max(lr_ratio, ps_ratio); |
2424 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
2425 | } |
2426 | ||
2427 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2428 | task_scan_min(p), task_scan_max(p)); | |
2429 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2430 | } | |
2431 | ||
7e2703e6 RR |
2432 | /* |
2433 | * Get the fraction of time the task has been running since the last | |
2434 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2435 | * decays those on a 32ms period, which is orders of magnitude off | |
2436 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2437 | * stats only if the task is so new there are no NUMA statistics yet. | |
2438 | */ | |
2439 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2440 | { | |
2441 | u64 runtime, delta, now; | |
2442 | /* Use the start of this time slice to avoid calculations. */ | |
2443 | now = p->se.exec_start; | |
2444 | runtime = p->se.sum_exec_runtime; | |
2445 | ||
2446 | if (p->last_task_numa_placement) { | |
2447 | delta = runtime - p->last_sum_exec_runtime; | |
2448 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2449 | |
2450 | /* Avoid time going backwards, prevent potential divide error: */ | |
2451 | if (unlikely((s64)*period < 0)) | |
2452 | *period = 0; | |
7e2703e6 | 2453 | } else { |
c7b50216 | 2454 | delta = p->se.avg.load_sum; |
9d89c257 | 2455 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2456 | } |
2457 | ||
2458 | p->last_sum_exec_runtime = runtime; | |
2459 | p->last_task_numa_placement = now; | |
2460 | ||
2461 | return delta; | |
2462 | } | |
2463 | ||
54009416 RR |
2464 | /* |
2465 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2466 | * be done in a way that produces consistent results with group_weight, | |
2467 | * otherwise workloads might not converge. | |
2468 | */ | |
2469 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2470 | { | |
2471 | nodemask_t nodes; | |
2472 | int dist; | |
2473 | ||
2474 | /* Direct connections between all NUMA nodes. */ | |
2475 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2476 | return nid; | |
2477 | ||
2478 | /* | |
2479 | * On a system with glueless mesh NUMA topology, group_weight | |
2480 | * scores nodes according to the number of NUMA hinting faults on | |
2481 | * both the node itself, and on nearby nodes. | |
2482 | */ | |
2483 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2484 | unsigned long score, max_score = 0; | |
2485 | int node, max_node = nid; | |
2486 | ||
2487 | dist = sched_max_numa_distance; | |
2488 | ||
5c7b1aaf | 2489 | for_each_node_state(node, N_CPU) { |
54009416 RR |
2490 | score = group_weight(p, node, dist); |
2491 | if (score > max_score) { | |
2492 | max_score = score; | |
2493 | max_node = node; | |
2494 | } | |
2495 | } | |
2496 | return max_node; | |
2497 | } | |
2498 | ||
2499 | /* | |
2500 | * Finding the preferred nid in a system with NUMA backplane | |
2501 | * interconnect topology is more involved. The goal is to locate | |
2502 | * tasks from numa_groups near each other in the system, and | |
2503 | * untangle workloads from different sides of the system. This requires | |
2504 | * searching down the hierarchy of node groups, recursively searching | |
2505 | * inside the highest scoring group of nodes. The nodemask tricks | |
2506 | * keep the complexity of the search down. | |
2507 | */ | |
5c7b1aaf | 2508 | nodes = node_states[N_CPU]; |
54009416 RR |
2509 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { |
2510 | unsigned long max_faults = 0; | |
81907478 | 2511 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2512 | int a, b; |
2513 | ||
2514 | /* Are there nodes at this distance from each other? */ | |
2515 | if (!find_numa_distance(dist)) | |
2516 | continue; | |
2517 | ||
2518 | for_each_node_mask(a, nodes) { | |
2519 | unsigned long faults = 0; | |
2520 | nodemask_t this_group; | |
2521 | nodes_clear(this_group); | |
2522 | ||
2523 | /* Sum group's NUMA faults; includes a==b case. */ | |
2524 | for_each_node_mask(b, nodes) { | |
2525 | if (node_distance(a, b) < dist) { | |
2526 | faults += group_faults(p, b); | |
2527 | node_set(b, this_group); | |
2528 | node_clear(b, nodes); | |
2529 | } | |
2530 | } | |
2531 | ||
2532 | /* Remember the top group. */ | |
2533 | if (faults > max_faults) { | |
2534 | max_faults = faults; | |
2535 | max_group = this_group; | |
2536 | /* | |
2537 | * subtle: at the smallest distance there is | |
2538 | * just one node left in each "group", the | |
2539 | * winner is the preferred nid. | |
2540 | */ | |
2541 | nid = a; | |
2542 | } | |
2543 | } | |
2544 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2545 | if (!max_faults) |
2546 | break; | |
54009416 RR |
2547 | nodes = max_group; |
2548 | } | |
2549 | return nid; | |
2550 | } | |
2551 | ||
cbee9f88 PZ |
2552 | static void task_numa_placement(struct task_struct *p) |
2553 | { | |
98fa15f3 | 2554 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2555 | unsigned long max_faults = 0; |
04bb2f94 | 2556 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2557 | unsigned long total_faults; |
2558 | u64 runtime, period; | |
7dbd13ed | 2559 | spinlock_t *group_lock = NULL; |
cb361d8c | 2560 | struct numa_group *ng; |
cbee9f88 | 2561 | |
7e5a2c17 JL |
2562 | /* |
2563 | * The p->mm->numa_scan_seq field gets updated without | |
2564 | * exclusive access. Use READ_ONCE() here to ensure | |
2565 | * that the field is read in a single access: | |
2566 | */ | |
316c1608 | 2567 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2568 | if (p->numa_scan_seq == seq) |
2569 | return; | |
2570 | p->numa_scan_seq = seq; | |
598f0ec0 | 2571 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2572 | |
7e2703e6 RR |
2573 | total_faults = p->numa_faults_locality[0] + |
2574 | p->numa_faults_locality[1]; | |
2575 | runtime = numa_get_avg_runtime(p, &period); | |
2576 | ||
7dbd13ed | 2577 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2578 | ng = deref_curr_numa_group(p); |
2579 | if (ng) { | |
2580 | group_lock = &ng->lock; | |
60e69eed | 2581 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2582 | } |
2583 | ||
688b7585 MG |
2584 | /* Find the node with the highest number of faults */ |
2585 | for_each_online_node(nid) { | |
44dba3d5 IM |
2586 | /* Keep track of the offsets in numa_faults array */ |
2587 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2588 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2589 | int priv; |
745d6147 | 2590 | |
be1e4e76 | 2591 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2592 | long diff, f_diff, f_weight; |
8c8a743c | 2593 | |
44dba3d5 IM |
2594 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2595 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2596 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2597 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2598 | |
ac8e895b | 2599 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2600 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2601 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2602 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2603 | |
7e2703e6 RR |
2604 | /* |
2605 | * Normalize the faults_from, so all tasks in a group | |
2606 | * count according to CPU use, instead of by the raw | |
2607 | * number of faults. Tasks with little runtime have | |
2608 | * little over-all impact on throughput, and thus their | |
2609 | * faults are less important. | |
2610 | */ | |
2611 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2612 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2613 | (total_faults + 1); |
44dba3d5 IM |
2614 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2615 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2616 | |
44dba3d5 IM |
2617 | p->numa_faults[mem_idx] += diff; |
2618 | p->numa_faults[cpu_idx] += f_diff; | |
2619 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2620 | p->total_numa_faults += diff; |
cb361d8c | 2621 | if (ng) { |
44dba3d5 IM |
2622 | /* |
2623 | * safe because we can only change our own group | |
2624 | * | |
2625 | * mem_idx represents the offset for a given | |
2626 | * nid and priv in a specific region because it | |
2627 | * is at the beginning of the numa_faults array. | |
2628 | */ | |
cb361d8c | 2629 | ng->faults[mem_idx] += diff; |
5b763a14 | 2630 | ng->faults[cpu_idx] += f_diff; |
cb361d8c JH |
2631 | ng->total_faults += diff; |
2632 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2633 | } |
ac8e895b MG |
2634 | } |
2635 | ||
cb361d8c | 2636 | if (!ng) { |
f03bb676 SD |
2637 | if (faults > max_faults) { |
2638 | max_faults = faults; | |
2639 | max_nid = nid; | |
2640 | } | |
2641 | } else if (group_faults > max_faults) { | |
2642 | max_faults = group_faults; | |
688b7585 MG |
2643 | max_nid = nid; |
2644 | } | |
83e1d2cd MG |
2645 | } |
2646 | ||
5c7b1aaf | 2647 | /* Cannot migrate task to CPU-less node */ |
ab31c7fd | 2648 | if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) { |
5c7b1aaf HY |
2649 | int near_nid = max_nid; |
2650 | int distance, near_distance = INT_MAX; | |
2651 | ||
2652 | for_each_node_state(nid, N_CPU) { | |
2653 | distance = node_distance(max_nid, nid); | |
2654 | if (distance < near_distance) { | |
2655 | near_nid = nid; | |
2656 | near_distance = distance; | |
2657 | } | |
2658 | } | |
2659 | max_nid = near_nid; | |
2660 | } | |
2661 | ||
cb361d8c JH |
2662 | if (ng) { |
2663 | numa_group_count_active_nodes(ng); | |
60e69eed | 2664 | spin_unlock_irq(group_lock); |
f03bb676 | 2665 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2666 | } |
2667 | ||
bb97fc31 RR |
2668 | if (max_faults) { |
2669 | /* Set the new preferred node */ | |
2670 | if (max_nid != p->numa_preferred_nid) | |
2671 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2672 | } |
30619c89 SD |
2673 | |
2674 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2675 | } |
2676 | ||
8c8a743c PZ |
2677 | static inline int get_numa_group(struct numa_group *grp) |
2678 | { | |
c45a7795 | 2679 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2680 | } |
2681 | ||
2682 | static inline void put_numa_group(struct numa_group *grp) | |
2683 | { | |
c45a7795 | 2684 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2685 | kfree_rcu(grp, rcu); |
2686 | } | |
2687 | ||
3e6a9418 MG |
2688 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2689 | int *priv) | |
8c8a743c PZ |
2690 | { |
2691 | struct numa_group *grp, *my_grp; | |
2692 | struct task_struct *tsk; | |
2693 | bool join = false; | |
2694 | int cpu = cpupid_to_cpu(cpupid); | |
2695 | int i; | |
2696 | ||
cb361d8c | 2697 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2698 | unsigned int size = sizeof(struct numa_group) + |
7a2341fc BR |
2699 | NR_NUMA_HINT_FAULT_STATS * |
2700 | nr_node_ids * sizeof(unsigned long); | |
8c8a743c PZ |
2701 | |
2702 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2703 | if (!grp) | |
2704 | return; | |
2705 | ||
c45a7795 | 2706 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2707 | grp->active_nodes = 1; |
2708 | grp->max_faults_cpu = 0; | |
8c8a743c | 2709 | spin_lock_init(&grp->lock); |
e29cf08b | 2710 | grp->gid = p->pid; |
8c8a743c | 2711 | |
be1e4e76 | 2712 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2713 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2714 | |
989348b5 | 2715 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2716 | |
8c8a743c PZ |
2717 | grp->nr_tasks++; |
2718 | rcu_assign_pointer(p->numa_group, grp); | |
2719 | } | |
2720 | ||
2721 | rcu_read_lock(); | |
316c1608 | 2722 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2723 | |
2724 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2725 | goto no_join; |
8c8a743c PZ |
2726 | |
2727 | grp = rcu_dereference(tsk->numa_group); | |
2728 | if (!grp) | |
3354781a | 2729 | goto no_join; |
8c8a743c | 2730 | |
cb361d8c | 2731 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2732 | if (grp == my_grp) |
3354781a | 2733 | goto no_join; |
8c8a743c PZ |
2734 | |
2735 | /* | |
2736 | * Only join the other group if its bigger; if we're the bigger group, | |
2737 | * the other task will join us. | |
2738 | */ | |
2739 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2740 | goto no_join; |
8c8a743c PZ |
2741 | |
2742 | /* | |
2743 | * Tie-break on the grp address. | |
2744 | */ | |
2745 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2746 | goto no_join; |
8c8a743c | 2747 | |
dabe1d99 RR |
2748 | /* Always join threads in the same process. */ |
2749 | if (tsk->mm == current->mm) | |
2750 | join = true; | |
2751 | ||
2752 | /* Simple filter to avoid false positives due to PID collisions */ | |
2753 | if (flags & TNF_SHARED) | |
2754 | join = true; | |
8c8a743c | 2755 | |
3e6a9418 MG |
2756 | /* Update priv based on whether false sharing was detected */ |
2757 | *priv = !join; | |
2758 | ||
dabe1d99 | 2759 | if (join && !get_numa_group(grp)) |
3354781a | 2760 | goto no_join; |
8c8a743c | 2761 | |
8c8a743c PZ |
2762 | rcu_read_unlock(); |
2763 | ||
2764 | if (!join) | |
2765 | return; | |
2766 | ||
09348d75 | 2767 | WARN_ON_ONCE(irqs_disabled()); |
60e69eed | 2768 | double_lock_irq(&my_grp->lock, &grp->lock); |
989348b5 | 2769 | |
be1e4e76 | 2770 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2771 | my_grp->faults[i] -= p->numa_faults[i]; |
2772 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2773 | } |
989348b5 MG |
2774 | my_grp->total_faults -= p->total_numa_faults; |
2775 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2776 | |
8c8a743c PZ |
2777 | my_grp->nr_tasks--; |
2778 | grp->nr_tasks++; | |
2779 | ||
2780 | spin_unlock(&my_grp->lock); | |
60e69eed | 2781 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2782 | |
2783 | rcu_assign_pointer(p->numa_group, grp); | |
2784 | ||
2785 | put_numa_group(my_grp); | |
3354781a PZ |
2786 | return; |
2787 | ||
2788 | no_join: | |
2789 | rcu_read_unlock(); | |
2790 | return; | |
8c8a743c PZ |
2791 | } |
2792 | ||
16d51a59 | 2793 | /* |
3b03706f | 2794 | * Get rid of NUMA statistics associated with a task (either current or dead). |
16d51a59 JH |
2795 | * If @final is set, the task is dead and has reached refcount zero, so we can |
2796 | * safely free all relevant data structures. Otherwise, there might be | |
2797 | * concurrent reads from places like load balancing and procfs, and we should | |
2798 | * reset the data back to default state without freeing ->numa_faults. | |
2799 | */ | |
2800 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 2801 | { |
cb361d8c JH |
2802 | /* safe: p either is current or is being freed by current */ |
2803 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 2804 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
2805 | unsigned long flags; |
2806 | int i; | |
8c8a743c | 2807 | |
16d51a59 JH |
2808 | if (!numa_faults) |
2809 | return; | |
2810 | ||
8c8a743c | 2811 | if (grp) { |
e9dd685c | 2812 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2813 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2814 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2815 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2816 | |
8c8a743c | 2817 | grp->nr_tasks--; |
e9dd685c | 2818 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2819 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2820 | put_numa_group(grp); |
2821 | } | |
2822 | ||
16d51a59 JH |
2823 | if (final) { |
2824 | p->numa_faults = NULL; | |
2825 | kfree(numa_faults); | |
2826 | } else { | |
2827 | p->total_numa_faults = 0; | |
2828 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
2829 | numa_faults[i] = 0; | |
2830 | } | |
8c8a743c PZ |
2831 | } |
2832 | ||
cbee9f88 PZ |
2833 | /* |
2834 | * Got a PROT_NONE fault for a page on @node. | |
2835 | */ | |
58b46da3 | 2836 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2837 | { |
2838 | struct task_struct *p = current; | |
6688cc05 | 2839 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2840 | int cpu_node = task_node(current); |
792568ec | 2841 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2842 | struct numa_group *ng; |
ac8e895b | 2843 | int priv; |
cbee9f88 | 2844 | |
2a595721 | 2845 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2846 | return; |
2847 | ||
9ff1d9ff MG |
2848 | /* for example, ksmd faulting in a user's mm */ |
2849 | if (!p->mm) | |
2850 | return; | |
2851 | ||
33024536 HY |
2852 | /* |
2853 | * NUMA faults statistics are unnecessary for the slow memory | |
2854 | * node for memory tiering mode. | |
2855 | */ | |
2856 | if (!node_is_toptier(mem_node) && | |
2857 | (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING || | |
2858 | !cpupid_valid(last_cpupid))) | |
2859 | return; | |
2860 | ||
f809ca9a | 2861 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2862 | if (unlikely(!p->numa_faults)) { |
2863 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2864 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2865 | |
44dba3d5 IM |
2866 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2867 | if (!p->numa_faults) | |
f809ca9a | 2868 | return; |
745d6147 | 2869 | |
83e1d2cd | 2870 | p->total_numa_faults = 0; |
04bb2f94 | 2871 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2872 | } |
cbee9f88 | 2873 | |
8c8a743c PZ |
2874 | /* |
2875 | * First accesses are treated as private, otherwise consider accesses | |
2876 | * to be private if the accessing pid has not changed | |
2877 | */ | |
2878 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2879 | priv = 1; | |
2880 | } else { | |
2881 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2882 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2883 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2884 | } |
2885 | ||
792568ec RR |
2886 | /* |
2887 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2888 | * occurs wholly within the set of nodes that the workload is | |
2889 | * actively using should be counted as local. This allows the | |
2890 | * scan rate to slow down when a workload has settled down. | |
2891 | */ | |
cb361d8c | 2892 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
2893 | if (!priv && !local && ng && ng->active_nodes > 1 && |
2894 | numa_is_active_node(cpu_node, ng) && | |
2895 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2896 | local = 1; |
2897 | ||
2739d3ee | 2898 | /* |
e1ff516a YW |
2899 | * Retry to migrate task to preferred node periodically, in case it |
2900 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2901 | */ |
b6a60cf3 SD |
2902 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2903 | task_numa_placement(p); | |
6b9a7460 | 2904 | numa_migrate_preferred(p); |
b6a60cf3 | 2905 | } |
6b9a7460 | 2906 | |
b32e86b4 IM |
2907 | if (migrated) |
2908 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2909 | if (flags & TNF_MIGRATE_FAIL) |
2910 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2911 | |
44dba3d5 IM |
2912 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2913 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2914 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2915 | } |
2916 | ||
6e5fb223 PZ |
2917 | static void reset_ptenuma_scan(struct task_struct *p) |
2918 | { | |
7e5a2c17 JL |
2919 | /* |
2920 | * We only did a read acquisition of the mmap sem, so | |
2921 | * p->mm->numa_scan_seq is written to without exclusive access | |
2922 | * and the update is not guaranteed to be atomic. That's not | |
2923 | * much of an issue though, since this is just used for | |
2924 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2925 | * expensive, to avoid any form of compiler optimizations: | |
2926 | */ | |
316c1608 | 2927 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2928 | p->mm->numa_scan_offset = 0; |
2929 | } | |
2930 | ||
cbee9f88 PZ |
2931 | /* |
2932 | * The expensive part of numa migration is done from task_work context. | |
2933 | * Triggered from task_tick_numa(). | |
2934 | */ | |
9434f9f5 | 2935 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
2936 | { |
2937 | unsigned long migrate, next_scan, now = jiffies; | |
2938 | struct task_struct *p = current; | |
2939 | struct mm_struct *mm = p->mm; | |
51170840 | 2940 | u64 runtime = p->se.sum_exec_runtime; |
0cd4d02c | 2941 | MA_STATE(mas, &mm->mm_mt, 0, 0); |
6e5fb223 | 2942 | struct vm_area_struct *vma; |
9f40604c | 2943 | unsigned long start, end; |
598f0ec0 | 2944 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2945 | long pages, virtpages; |
cbee9f88 | 2946 | |
9148a3a1 | 2947 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 2948 | |
b34920d4 | 2949 | work->next = work; |
cbee9f88 PZ |
2950 | /* |
2951 | * Who cares about NUMA placement when they're dying. | |
2952 | * | |
2953 | * NOTE: make sure not to dereference p->mm before this check, | |
2954 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2955 | * without p->mm even though we still had it when we enqueued this | |
2956 | * work. | |
2957 | */ | |
2958 | if (p->flags & PF_EXITING) | |
2959 | return; | |
2960 | ||
930aa174 | 2961 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2962 | mm->numa_next_scan = now + |
2963 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2964 | } |
2965 | ||
cbee9f88 PZ |
2966 | /* |
2967 | * Enforce maximal scan/migration frequency.. | |
2968 | */ | |
2969 | migrate = mm->numa_next_scan; | |
2970 | if (time_before(now, migrate)) | |
2971 | return; | |
2972 | ||
598f0ec0 MG |
2973 | if (p->numa_scan_period == 0) { |
2974 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2975 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2976 | } |
cbee9f88 | 2977 | |
fb003b80 | 2978 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
8baceabc | 2979 | if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan)) |
cbee9f88 PZ |
2980 | return; |
2981 | ||
19a78d11 PZ |
2982 | /* |
2983 | * Delay this task enough that another task of this mm will likely win | |
2984 | * the next time around. | |
2985 | */ | |
2986 | p->node_stamp += 2 * TICK_NSEC; | |
2987 | ||
9f40604c MG |
2988 | start = mm->numa_scan_offset; |
2989 | pages = sysctl_numa_balancing_scan_size; | |
2990 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2991 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2992 | if (!pages) |
2993 | return; | |
cbee9f88 | 2994 | |
4620f8c1 | 2995 | |
d8ed45c5 | 2996 | if (!mmap_read_trylock(mm)) |
8655d549 | 2997 | return; |
0cd4d02c MWO |
2998 | mas_set(&mas, start); |
2999 | vma = mas_find(&mas, ULONG_MAX); | |
6e5fb223 PZ |
3000 | if (!vma) { |
3001 | reset_ptenuma_scan(p); | |
9f40604c | 3002 | start = 0; |
0cd4d02c MWO |
3003 | mas_set(&mas, start); |
3004 | vma = mas_find(&mas, ULONG_MAX); | |
6e5fb223 | 3005 | } |
0cd4d02c MWO |
3006 | |
3007 | for (; vma; vma = mas_find(&mas, ULONG_MAX)) { | |
6b79c57b | 3008 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 3009 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 3010 | continue; |
6b79c57b | 3011 | } |
6e5fb223 | 3012 | |
4591ce4f MG |
3013 | /* |
3014 | * Shared library pages mapped by multiple processes are not | |
3015 | * migrated as it is expected they are cache replicated. Avoid | |
3016 | * hinting faults in read-only file-backed mappings or the vdso | |
3017 | * as migrating the pages will be of marginal benefit. | |
3018 | */ | |
3019 | if (!vma->vm_mm || | |
3020 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
3021 | continue; | |
3022 | ||
3c67f474 MG |
3023 | /* |
3024 | * Skip inaccessible VMAs to avoid any confusion between | |
3025 | * PROT_NONE and NUMA hinting ptes | |
3026 | */ | |
3122e80e | 3027 | if (!vma_is_accessible(vma)) |
3c67f474 | 3028 | continue; |
4591ce4f | 3029 | |
9f40604c MG |
3030 | do { |
3031 | start = max(start, vma->vm_start); | |
3032 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
3033 | end = min(end, vma->vm_end); | |
4620f8c1 | 3034 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
3035 | |
3036 | /* | |
4620f8c1 RR |
3037 | * Try to scan sysctl_numa_balancing_size worth of |
3038 | * hpages that have at least one present PTE that | |
3039 | * is not already pte-numa. If the VMA contains | |
3040 | * areas that are unused or already full of prot_numa | |
3041 | * PTEs, scan up to virtpages, to skip through those | |
3042 | * areas faster. | |
598f0ec0 MG |
3043 | */ |
3044 | if (nr_pte_updates) | |
3045 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 3046 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 3047 | |
9f40604c | 3048 | start = end; |
4620f8c1 | 3049 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 3050 | goto out; |
3cf1962c RR |
3051 | |
3052 | cond_resched(); | |
9f40604c | 3053 | } while (end != vma->vm_end); |
cbee9f88 | 3054 | } |
6e5fb223 | 3055 | |
9f40604c | 3056 | out: |
6e5fb223 | 3057 | /* |
c69307d5 PZ |
3058 | * It is possible to reach the end of the VMA list but the last few |
3059 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
3060 | * would find the !migratable VMA on the next scan but not reset the | |
3061 | * scanner to the start so check it now. | |
6e5fb223 PZ |
3062 | */ |
3063 | if (vma) | |
9f40604c | 3064 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
3065 | else |
3066 | reset_ptenuma_scan(p); | |
d8ed45c5 | 3067 | mmap_read_unlock(mm); |
51170840 RR |
3068 | |
3069 | /* | |
3070 | * Make sure tasks use at least 32x as much time to run other code | |
3071 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
3072 | * Usually update_task_scan_period slows down scanning enough; on an | |
3073 | * overloaded system we need to limit overhead on a per task basis. | |
3074 | */ | |
3075 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
3076 | u64 diff = p->se.sum_exec_runtime - runtime; | |
3077 | p->node_stamp += 32 * diff; | |
3078 | } | |
cbee9f88 PZ |
3079 | } |
3080 | ||
d35927a1 VS |
3081 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
3082 | { | |
3083 | int mm_users = 0; | |
3084 | struct mm_struct *mm = p->mm; | |
3085 | ||
3086 | if (mm) { | |
3087 | mm_users = atomic_read(&mm->mm_users); | |
3088 | if (mm_users == 1) { | |
3089 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
3090 | mm->numa_scan_seq = 0; | |
3091 | } | |
3092 | } | |
3093 | p->node_stamp = 0; | |
3094 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
3095 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
70ce3ea9 | 3096 | p->numa_migrate_retry = 0; |
b34920d4 | 3097 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
3098 | p->numa_work.next = &p->numa_work; |
3099 | p->numa_faults = NULL; | |
12bf8a7e HW |
3100 | p->numa_pages_migrated = 0; |
3101 | p->total_numa_faults = 0; | |
d35927a1 VS |
3102 | RCU_INIT_POINTER(p->numa_group, NULL); |
3103 | p->last_task_numa_placement = 0; | |
3104 | p->last_sum_exec_runtime = 0; | |
3105 | ||
b34920d4 VS |
3106 | init_task_work(&p->numa_work, task_numa_work); |
3107 | ||
d35927a1 VS |
3108 | /* New address space, reset the preferred nid */ |
3109 | if (!(clone_flags & CLONE_VM)) { | |
3110 | p->numa_preferred_nid = NUMA_NO_NODE; | |
3111 | return; | |
3112 | } | |
3113 | ||
3114 | /* | |
3115 | * New thread, keep existing numa_preferred_nid which should be copied | |
3116 | * already by arch_dup_task_struct but stagger when scans start. | |
3117 | */ | |
3118 | if (mm) { | |
3119 | unsigned int delay; | |
3120 | ||
3121 | delay = min_t(unsigned int, task_scan_max(current), | |
3122 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
3123 | delay += 2 * TICK_NSEC; | |
3124 | p->node_stamp = delay; | |
3125 | } | |
3126 | } | |
3127 | ||
cbee9f88 PZ |
3128 | /* |
3129 | * Drive the periodic memory faults.. | |
3130 | */ | |
b1546edc | 3131 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
3132 | { |
3133 | struct callback_head *work = &curr->numa_work; | |
3134 | u64 period, now; | |
3135 | ||
3136 | /* | |
3137 | * We don't care about NUMA placement if we don't have memory. | |
3138 | */ | |
b3f9916d | 3139 | if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) |
cbee9f88 PZ |
3140 | return; |
3141 | ||
3142 | /* | |
3143 | * Using runtime rather than walltime has the dual advantage that | |
3144 | * we (mostly) drive the selection from busy threads and that the | |
3145 | * task needs to have done some actual work before we bother with | |
3146 | * NUMA placement. | |
3147 | */ | |
3148 | now = curr->se.sum_exec_runtime; | |
3149 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
3150 | ||
25b3e5a3 | 3151 | if (now > curr->node_stamp + period) { |
4b96a29b | 3152 | if (!curr->node_stamp) |
b5dd77c8 | 3153 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 3154 | curr->node_stamp += period; |
cbee9f88 | 3155 | |
b34920d4 | 3156 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
91989c70 | 3157 | task_work_add(curr, work, TWA_RESUME); |
cbee9f88 PZ |
3158 | } |
3159 | } | |
3fed382b | 3160 | |
3f9672ba SD |
3161 | static void update_scan_period(struct task_struct *p, int new_cpu) |
3162 | { | |
3163 | int src_nid = cpu_to_node(task_cpu(p)); | |
3164 | int dst_nid = cpu_to_node(new_cpu); | |
3165 | ||
05cbdf4f MG |
3166 | if (!static_branch_likely(&sched_numa_balancing)) |
3167 | return; | |
3168 | ||
3f9672ba SD |
3169 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
3170 | return; | |
3171 | ||
05cbdf4f MG |
3172 | if (src_nid == dst_nid) |
3173 | return; | |
3174 | ||
3175 | /* | |
3176 | * Allow resets if faults have been trapped before one scan | |
3177 | * has completed. This is most likely due to a new task that | |
3178 | * is pulled cross-node due to wakeups or load balancing. | |
3179 | */ | |
3180 | if (p->numa_scan_seq) { | |
3181 | /* | |
3182 | * Avoid scan adjustments if moving to the preferred | |
3183 | * node or if the task was not previously running on | |
3184 | * the preferred node. | |
3185 | */ | |
3186 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
3187 | (p->numa_preferred_nid != NUMA_NO_NODE && |
3188 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
3189 | return; |
3190 | } | |
3191 | ||
3192 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
3193 | } |
3194 | ||
cbee9f88 PZ |
3195 | #else |
3196 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
3197 | { | |
3198 | } | |
0ec8aa00 PZ |
3199 | |
3200 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
3201 | { | |
3202 | } | |
3203 | ||
3204 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
3205 | { | |
3206 | } | |
3fed382b | 3207 | |
3f9672ba SD |
3208 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
3209 | { | |
3210 | } | |
3211 | ||
cbee9f88 PZ |
3212 | #endif /* CONFIG_NUMA_BALANCING */ |
3213 | ||
30cfdcfc DA |
3214 | static void |
3215 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3216 | { | |
3217 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 3218 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3219 | if (entity_is_task(se)) { |
3220 | struct rq *rq = rq_of(cfs_rq); | |
3221 | ||
3222 | account_numa_enqueue(rq, task_of(se)); | |
3223 | list_add(&se->group_node, &rq->cfs_tasks); | |
3224 | } | |
367456c7 | 3225 | #endif |
30cfdcfc | 3226 | cfs_rq->nr_running++; |
a480adde JD |
3227 | if (se_is_idle(se)) |
3228 | cfs_rq->idle_nr_running++; | |
30cfdcfc DA |
3229 | } |
3230 | ||
3231 | static void | |
3232 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3233 | { | |
3234 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 3235 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3236 | if (entity_is_task(se)) { |
3237 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 3238 | list_del_init(&se->group_node); |
0ec8aa00 | 3239 | } |
bfdb198c | 3240 | #endif |
30cfdcfc | 3241 | cfs_rq->nr_running--; |
a480adde JD |
3242 | if (se_is_idle(se)) |
3243 | cfs_rq->idle_nr_running--; | |
30cfdcfc DA |
3244 | } |
3245 | ||
8d5b9025 PZ |
3246 | /* |
3247 | * Signed add and clamp on underflow. | |
3248 | * | |
3249 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3250 | * memory. This allows lockless observations without ever seeing the negative | |
3251 | * values. | |
3252 | */ | |
3253 | #define add_positive(_ptr, _val) do { \ | |
3254 | typeof(_ptr) ptr = (_ptr); \ | |
3255 | typeof(_val) val = (_val); \ | |
3256 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3257 | \ | |
3258 | res = var + val; \ | |
3259 | \ | |
3260 | if (val < 0 && res > var) \ | |
3261 | res = 0; \ | |
3262 | \ | |
3263 | WRITE_ONCE(*ptr, res); \ | |
3264 | } while (0) | |
3265 | ||
3266 | /* | |
3267 | * Unsigned subtract and clamp on underflow. | |
3268 | * | |
3269 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3270 | * memory. This allows lockless observations without ever seeing the negative | |
3271 | * values. | |
3272 | */ | |
3273 | #define sub_positive(_ptr, _val) do { \ | |
3274 | typeof(_ptr) ptr = (_ptr); \ | |
3275 | typeof(*ptr) val = (_val); \ | |
3276 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3277 | res = var - val; \ | |
3278 | if (res > var) \ | |
3279 | res = 0; \ | |
3280 | WRITE_ONCE(*ptr, res); \ | |
3281 | } while (0) | |
3282 | ||
b5c0ce7b PB |
3283 | /* |
3284 | * Remove and clamp on negative, from a local variable. | |
3285 | * | |
3286 | * A variant of sub_positive(), which does not use explicit load-store | |
3287 | * and is thus optimized for local variable updates. | |
3288 | */ | |
3289 | #define lsub_positive(_ptr, _val) do { \ | |
3290 | typeof(_ptr) ptr = (_ptr); \ | |
3291 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
3292 | } while (0) | |
3293 | ||
8d5b9025 | 3294 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
3295 | static inline void |
3296 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3297 | { | |
3298 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3299 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
3300 | } | |
3301 | ||
3302 | static inline void | |
3303 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3304 | { | |
3305 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
2d02fa8c VG |
3306 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); |
3307 | /* See update_cfs_rq_load_avg() */ | |
3308 | cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, | |
3309 | cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); | |
8d5b9025 PZ |
3310 | } |
3311 | #else | |
3312 | static inline void | |
8d5b9025 PZ |
3313 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } |
3314 | static inline void | |
3315 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
3316 | #endif | |
3317 | ||
9059393e | 3318 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
0dacee1b | 3319 | unsigned long weight) |
9059393e VG |
3320 | { |
3321 | if (se->on_rq) { | |
3322 | /* commit outstanding execution time */ | |
3323 | if (cfs_rq->curr == se) | |
3324 | update_curr(cfs_rq); | |
1724b95b | 3325 | update_load_sub(&cfs_rq->load, se->load.weight); |
9059393e VG |
3326 | } |
3327 | dequeue_load_avg(cfs_rq, se); | |
3328 | ||
3329 | update_load_set(&se->load, weight); | |
3330 | ||
3331 | #ifdef CONFIG_SMP | |
1ea6c46a | 3332 | do { |
87e867b4 | 3333 | u32 divider = get_pelt_divider(&se->avg); |
1ea6c46a PZ |
3334 | |
3335 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
1ea6c46a | 3336 | } while (0); |
9059393e VG |
3337 | #endif |
3338 | ||
3339 | enqueue_load_avg(cfs_rq, se); | |
0dacee1b | 3340 | if (se->on_rq) |
1724b95b | 3341 | update_load_add(&cfs_rq->load, se->load.weight); |
0dacee1b | 3342 | |
9059393e VG |
3343 | } |
3344 | ||
3345 | void reweight_task(struct task_struct *p, int prio) | |
3346 | { | |
3347 | struct sched_entity *se = &p->se; | |
3348 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3349 | struct load_weight *load = &se->load; | |
3350 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
3351 | ||
0dacee1b | 3352 | reweight_entity(cfs_rq, se, weight); |
9059393e VG |
3353 | load->inv_weight = sched_prio_to_wmult[prio]; |
3354 | } | |
3355 | ||
51bf903b CZ |
3356 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
3357 | ||
3ff6dcac | 3358 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 3359 | #ifdef CONFIG_SMP |
cef27403 PZ |
3360 | /* |
3361 | * All this does is approximate the hierarchical proportion which includes that | |
3362 | * global sum we all love to hate. | |
3363 | * | |
3364 | * That is, the weight of a group entity, is the proportional share of the | |
3365 | * group weight based on the group runqueue weights. That is: | |
3366 | * | |
3367 | * tg->weight * grq->load.weight | |
3368 | * ge->load.weight = ----------------------------- (1) | |
08f7c2f4 | 3369 | * \Sum grq->load.weight |
cef27403 PZ |
3370 | * |
3371 | * Now, because computing that sum is prohibitively expensive to compute (been | |
3372 | * there, done that) we approximate it with this average stuff. The average | |
3373 | * moves slower and therefore the approximation is cheaper and more stable. | |
3374 | * | |
3375 | * So instead of the above, we substitute: | |
3376 | * | |
3377 | * grq->load.weight -> grq->avg.load_avg (2) | |
3378 | * | |
3379 | * which yields the following: | |
3380 | * | |
3381 | * tg->weight * grq->avg.load_avg | |
3382 | * ge->load.weight = ------------------------------ (3) | |
08f7c2f4 | 3383 | * tg->load_avg |
cef27403 PZ |
3384 | * |
3385 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
3386 | * | |
3387 | * That is shares_avg, and it is right (given the approximation (2)). | |
3388 | * | |
3389 | * The problem with it is that because the average is slow -- it was designed | |
3390 | * to be exactly that of course -- this leads to transients in boundary | |
3391 | * conditions. In specific, the case where the group was idle and we start the | |
3392 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
3393 | * yielding bad latency etc.. | |
3394 | * | |
3395 | * Now, in that special case (1) reduces to: | |
3396 | * | |
3397 | * tg->weight * grq->load.weight | |
17de4ee0 | 3398 | * ge->load.weight = ----------------------------- = tg->weight (4) |
08f7c2f4 | 3399 | * grp->load.weight |
cef27403 PZ |
3400 | * |
3401 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
3402 | * | |
3403 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
3404 | * UP case, like: | |
3405 | * | |
3406 | * ge->load.weight = | |
3407 | * | |
3408 | * tg->weight * grq->load.weight | |
3409 | * --------------------------------------------------- (5) | |
3410 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
3411 | * | |
17de4ee0 PZ |
3412 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
3413 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
3414 | * | |
3415 | * | |
3416 | * tg->weight * grq->load.weight | |
3417 | * ge->load.weight = ----------------------------- (6) | |
08f7c2f4 | 3418 | * tg_load_avg' |
17de4ee0 PZ |
3419 | * |
3420 | * Where: | |
3421 | * | |
3422 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
3423 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
3424 | * |
3425 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
3426 | * (4) while in the normal case it approaches (3). It consistently | |
3427 | * overestimates the ge->load.weight and therefore: | |
3428 | * | |
3429 | * \Sum ge->load.weight >= tg->weight | |
3430 | * | |
3431 | * hence icky! | |
3432 | */ | |
2c8e4dce | 3433 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 3434 | { |
7c80cfc9 PZ |
3435 | long tg_weight, tg_shares, load, shares; |
3436 | struct task_group *tg = cfs_rq->tg; | |
3437 | ||
3438 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 3439 | |
3d4b60d3 | 3440 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 3441 | |
ea1dc6fc | 3442 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 3443 | |
ea1dc6fc PZ |
3444 | /* Ensure tg_weight >= load */ |
3445 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
3446 | tg_weight += load; | |
3ff6dcac | 3447 | |
7c80cfc9 | 3448 | shares = (tg_shares * load); |
cf5f0acf PZ |
3449 | if (tg_weight) |
3450 | shares /= tg_weight; | |
3ff6dcac | 3451 | |
b8fd8423 DE |
3452 | /* |
3453 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
3454 | * of a group with small tg->shares value. It is a floor value which is | |
3455 | * assigned as a minimum load.weight to the sched_entity representing | |
3456 | * the group on a CPU. | |
3457 | * | |
3458 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
3459 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
3460 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
3461 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
3462 | * instead of 0. | |
3463 | */ | |
7c80cfc9 | 3464 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 3465 | } |
387f77cc | 3466 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 3467 | |
1ea6c46a PZ |
3468 | /* |
3469 | * Recomputes the group entity based on the current state of its group | |
3470 | * runqueue. | |
3471 | */ | |
3472 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 3473 | { |
1ea6c46a | 3474 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
0dacee1b | 3475 | long shares; |
2069dd75 | 3476 | |
1ea6c46a | 3477 | if (!gcfs_rq) |
89ee048f VG |
3478 | return; |
3479 | ||
1ea6c46a | 3480 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3481 | return; |
89ee048f | 3482 | |
3ff6dcac | 3483 | #ifndef CONFIG_SMP |
0dacee1b | 3484 | shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
3485 | |
3486 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 3487 | return; |
7c80cfc9 | 3488 | #else |
2c8e4dce | 3489 | shares = calc_group_shares(gcfs_rq); |
3ff6dcac | 3490 | #endif |
2069dd75 | 3491 | |
0dacee1b | 3492 | reweight_entity(cfs_rq_of(se), se, shares); |
2069dd75 | 3493 | } |
89ee048f | 3494 | |
2069dd75 | 3495 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3496 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3497 | { |
3498 | } | |
3499 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3500 | ||
ea14b57e | 3501 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3502 | { |
43964409 LT |
3503 | struct rq *rq = rq_of(cfs_rq); |
3504 | ||
a4f9a0e5 | 3505 | if (&rq->cfs == cfs_rq) { |
a030d738 VK |
3506 | /* |
3507 | * There are a few boundary cases this might miss but it should | |
3508 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3509 | * a real problem. |
a030d738 VK |
3510 | * |
3511 | * It will not get called when we go idle, because the idle | |
3512 | * thread is a different class (!fair), nor will the utilization | |
3513 | * number include things like RT tasks. | |
3514 | * | |
3515 | * As is, the util number is not freq-invariant (we'd have to | |
3516 | * implement arch_scale_freq_capacity() for that). | |
3517 | * | |
82762d2a | 3518 | * See cpu_util_cfs(). |
a030d738 | 3519 | */ |
ea14b57e | 3520 | cpufreq_update_util(rq, flags); |
a030d738 VK |
3521 | } |
3522 | } | |
3523 | ||
141965c7 | 3524 | #ifdef CONFIG_SMP |
e2f3e35f VD |
3525 | static inline bool load_avg_is_decayed(struct sched_avg *sa) |
3526 | { | |
3527 | if (sa->load_sum) | |
3528 | return false; | |
3529 | ||
3530 | if (sa->util_sum) | |
3531 | return false; | |
3532 | ||
3533 | if (sa->runnable_sum) | |
3534 | return false; | |
3535 | ||
3536 | /* | |
3537 | * _avg must be null when _sum are null because _avg = _sum / divider | |
3538 | * Make sure that rounding and/or propagation of PELT values never | |
3539 | * break this. | |
3540 | */ | |
3541 | SCHED_WARN_ON(sa->load_avg || | |
3542 | sa->util_avg || | |
3543 | sa->runnable_avg); | |
3544 | ||
3545 | return true; | |
3546 | } | |
3547 | ||
d05b4305 VD |
3548 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3549 | { | |
3550 | return u64_u32_load_copy(cfs_rq->avg.last_update_time, | |
3551 | cfs_rq->last_update_time_copy); | |
3552 | } | |
c566e8e9 | 3553 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fdaba61e RR |
3554 | /* |
3555 | * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list | |
3556 | * immediately before a parent cfs_rq, and cfs_rqs are removed from the list | |
3557 | * bottom-up, we only have to test whether the cfs_rq before us on the list | |
3558 | * is our child. | |
3559 | * If cfs_rq is not on the list, test whether a child needs its to be added to | |
3560 | * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details). | |
3561 | */ | |
3562 | static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq) | |
3563 | { | |
3564 | struct cfs_rq *prev_cfs_rq; | |
3565 | struct list_head *prev; | |
3566 | ||
3567 | if (cfs_rq->on_list) { | |
3568 | prev = cfs_rq->leaf_cfs_rq_list.prev; | |
3569 | } else { | |
3570 | struct rq *rq = rq_of(cfs_rq); | |
3571 | ||
3572 | prev = rq->tmp_alone_branch; | |
3573 | } | |
3574 | ||
3575 | prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list); | |
3576 | ||
3577 | return (prev_cfs_rq->tg->parent == cfs_rq->tg); | |
3578 | } | |
a7b359fc OU |
3579 | |
3580 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) | |
3581 | { | |
3582 | if (cfs_rq->load.weight) | |
3583 | return false; | |
3584 | ||
e2f3e35f | 3585 | if (!load_avg_is_decayed(&cfs_rq->avg)) |
a7b359fc OU |
3586 | return false; |
3587 | ||
fdaba61e RR |
3588 | if (child_cfs_rq_on_list(cfs_rq)) |
3589 | return false; | |
3590 | ||
a7b359fc OU |
3591 | return true; |
3592 | } | |
3593 | ||
7c3edd2c PZ |
3594 | /** |
3595 | * update_tg_load_avg - update the tg's load avg | |
3596 | * @cfs_rq: the cfs_rq whose avg changed | |
7c3edd2c PZ |
3597 | * |
3598 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3599 | * However, because tg->load_avg is a global value there are performance | |
3600 | * considerations. | |
3601 | * | |
3602 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3603 | * differential update where we store the last value we propagated. This in | |
3604 | * turn allows skipping updates if the differential is 'small'. | |
3605 | * | |
815abf5a | 3606 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3607 | */ |
fe749158 | 3608 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) |
bb17f655 | 3609 | { |
9d89c257 | 3610 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3611 | |
aa0b7ae0 WL |
3612 | /* |
3613 | * No need to update load_avg for root_task_group as it is not used. | |
3614 | */ | |
3615 | if (cfs_rq->tg == &root_task_group) | |
3616 | return; | |
3617 | ||
fe749158 | 3618 | if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
9d89c257 YD |
3619 | atomic_long_add(delta, &cfs_rq->tg->load_avg); |
3620 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3621 | } |
8165e145 | 3622 | } |
f5f9739d | 3623 | |
ad936d86 | 3624 | /* |
97fb7a0a | 3625 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3626 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3627 | * including the state of rq->lock, should be made. | |
3628 | */ | |
3629 | void set_task_rq_fair(struct sched_entity *se, | |
3630 | struct cfs_rq *prev, struct cfs_rq *next) | |
3631 | { | |
0ccb977f PZ |
3632 | u64 p_last_update_time; |
3633 | u64 n_last_update_time; | |
3634 | ||
ad936d86 BP |
3635 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3636 | return; | |
3637 | ||
3638 | /* | |
3639 | * We are supposed to update the task to "current" time, then its up to | |
3640 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3641 | * getting what current time is, so simply throw away the out-of-date | |
3642 | * time. This will result in the wakee task is less decayed, but giving | |
3643 | * the wakee more load sounds not bad. | |
3644 | */ | |
0ccb977f PZ |
3645 | if (!(se->avg.last_update_time && prev)) |
3646 | return; | |
ad936d86 | 3647 | |
d05b4305 VD |
3648 | p_last_update_time = cfs_rq_last_update_time(prev); |
3649 | n_last_update_time = cfs_rq_last_update_time(next); | |
ad936d86 | 3650 | |
23127296 | 3651 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 3652 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 3653 | } |
09a43ace | 3654 | |
0e2d2aaa PZ |
3655 | /* |
3656 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3657 | * propagate its contribution. The key to this propagation is the invariant | |
3658 | * that for each group: | |
3659 | * | |
3660 | * ge->avg == grq->avg (1) | |
3661 | * | |
3662 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3663 | * represent the very same entity, just at different points in the hierarchy. | |
3664 | * | |
9f683953 VG |
3665 | * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial |
3666 | * and simply copies the running/runnable sum over (but still wrong, because | |
3667 | * the group entity and group rq do not have their PELT windows aligned). | |
0e2d2aaa | 3668 | * |
0dacee1b | 3669 | * However, update_tg_cfs_load() is more complex. So we have: |
0e2d2aaa PZ |
3670 | * |
3671 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3672 | * | |
3673 | * And since, like util, the runnable part should be directly transferable, | |
3674 | * the following would _appear_ to be the straight forward approach: | |
3675 | * | |
a4c3c049 | 3676 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3677 | * |
3678 | * And per (1) we have: | |
3679 | * | |
a4c3c049 | 3680 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3681 | * |
3682 | * Which gives: | |
3683 | * | |
3684 | * ge->load.weight * grq->avg.load_avg | |
3685 | * ge->avg.load_avg = ----------------------------------- (4) | |
3686 | * grq->load.weight | |
3687 | * | |
3688 | * Except that is wrong! | |
3689 | * | |
3690 | * Because while for entities historical weight is not important and we | |
3691 | * really only care about our future and therefore can consider a pure | |
3692 | * runnable sum, runqueues can NOT do this. | |
3693 | * | |
3694 | * We specifically want runqueues to have a load_avg that includes | |
3695 | * historical weights. Those represent the blocked load, the load we expect | |
3696 | * to (shortly) return to us. This only works by keeping the weights as | |
3697 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3698 | * | |
a4c3c049 VG |
3699 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3700 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3701 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3702 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3703 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3704 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3705 | * |
a4c3c049 | 3706 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3707 | * |
a4c3c049 | 3708 | * Given the constraint: |
0e2d2aaa | 3709 | * |
a4c3c049 | 3710 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3711 | * |
a4c3c049 VG |
3712 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3713 | * overlap. | |
0e2d2aaa | 3714 | * |
a4c3c049 | 3715 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3716 | * |
a4c3c049 | 3717 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3718 | * |
a4c3c049 | 3719 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3720 | * |
0e2d2aaa | 3721 | */ |
09a43ace | 3722 | static inline void |
0e2d2aaa | 3723 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3724 | { |
7ceb7710 VG |
3725 | long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg; |
3726 | u32 new_sum, divider; | |
09a43ace VG |
3727 | |
3728 | /* Nothing to update */ | |
7ceb7710 | 3729 | if (!delta_avg) |
09a43ace VG |
3730 | return; |
3731 | ||
87e867b4 VG |
3732 | /* |
3733 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3734 | * See ___update_load_avg() for details. | |
3735 | */ | |
3736 | divider = get_pelt_divider(&cfs_rq->avg); | |
3737 | ||
7ceb7710 | 3738 | |
09a43ace VG |
3739 | /* Set new sched_entity's utilization */ |
3740 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
7ceb7710 VG |
3741 | new_sum = se->avg.util_avg * divider; |
3742 | delta_sum = (long)new_sum - (long)se->avg.util_sum; | |
3743 | se->avg.util_sum = new_sum; | |
09a43ace VG |
3744 | |
3745 | /* Update parent cfs_rq utilization */ | |
7ceb7710 VG |
3746 | add_positive(&cfs_rq->avg.util_avg, delta_avg); |
3747 | add_positive(&cfs_rq->avg.util_sum, delta_sum); | |
3748 | ||
3749 | /* See update_cfs_rq_load_avg() */ | |
3750 | cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, | |
3751 | cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); | |
09a43ace VG |
3752 | } |
3753 | ||
9f683953 VG |
3754 | static inline void |
3755 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) | |
3756 | { | |
95246d1e VG |
3757 | long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; |
3758 | u32 new_sum, divider; | |
9f683953 VG |
3759 | |
3760 | /* Nothing to update */ | |
95246d1e | 3761 | if (!delta_avg) |
9f683953 VG |
3762 | return; |
3763 | ||
87e867b4 VG |
3764 | /* |
3765 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3766 | * See ___update_load_avg() for details. | |
3767 | */ | |
3768 | divider = get_pelt_divider(&cfs_rq->avg); | |
3769 | ||
9f683953 VG |
3770 | /* Set new sched_entity's runnable */ |
3771 | se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; | |
95246d1e VG |
3772 | new_sum = se->avg.runnable_avg * divider; |
3773 | delta_sum = (long)new_sum - (long)se->avg.runnable_sum; | |
3774 | se->avg.runnable_sum = new_sum; | |
9f683953 VG |
3775 | |
3776 | /* Update parent cfs_rq runnable */ | |
95246d1e VG |
3777 | add_positive(&cfs_rq->avg.runnable_avg, delta_avg); |
3778 | add_positive(&cfs_rq->avg.runnable_sum, delta_sum); | |
3779 | /* See update_cfs_rq_load_avg() */ | |
3780 | cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, | |
3781 | cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); | |
9f683953 VG |
3782 | } |
3783 | ||
09a43ace | 3784 | static inline void |
0dacee1b | 3785 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3786 | { |
2d02fa8c | 3787 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
0dacee1b VG |
3788 | unsigned long load_avg; |
3789 | u64 load_sum = 0; | |
2d02fa8c | 3790 | s64 delta_sum; |
95d68593 | 3791 | u32 divider; |
09a43ace | 3792 | |
0e2d2aaa PZ |
3793 | if (!runnable_sum) |
3794 | return; | |
09a43ace | 3795 | |
0e2d2aaa | 3796 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3797 | |
95d68593 VG |
3798 | /* |
3799 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3800 | * See ___update_load_avg() for details. | |
3801 | */ | |
87e867b4 | 3802 | divider = get_pelt_divider(&cfs_rq->avg); |
95d68593 | 3803 | |
a4c3c049 VG |
3804 | if (runnable_sum >= 0) { |
3805 | /* | |
3806 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3807 | * the CPU is saturated running == runnable. | |
3808 | */ | |
3809 | runnable_sum += se->avg.load_sum; | |
95d68593 | 3810 | runnable_sum = min_t(long, runnable_sum, divider); |
a4c3c049 VG |
3811 | } else { |
3812 | /* | |
3813 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3814 | * assuming all tasks are equally runnable. | |
3815 | */ | |
3816 | if (scale_load_down(gcfs_rq->load.weight)) { | |
2d02fa8c | 3817 | load_sum = div_u64(gcfs_rq->avg.load_sum, |
a4c3c049 VG |
3818 | scale_load_down(gcfs_rq->load.weight)); |
3819 | } | |
3820 | ||
3821 | /* But make sure to not inflate se's runnable */ | |
3822 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3823 | } | |
3824 | ||
3825 | /* | |
3826 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
3827 | * Rescale running sum to be in the same range as runnable sum |
3828 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
3829 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 3830 | */ |
23127296 | 3831 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
3832 | runnable_sum = max(runnable_sum, running_sum); |
3833 | ||
2d02fa8c VG |
3834 | load_sum = se_weight(se) * runnable_sum; |
3835 | load_avg = div_u64(load_sum, divider); | |
83c5e9d5 | 3836 | |
2d02fa8c VG |
3837 | delta_avg = load_avg - se->avg.load_avg; |
3838 | if (!delta_avg) | |
83c5e9d5 | 3839 | return; |
09a43ace | 3840 | |
2d02fa8c | 3841 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
7c7ad626 | 3842 | |
2d02fa8c VG |
3843 | se->avg.load_sum = runnable_sum; |
3844 | se->avg.load_avg = load_avg; | |
3845 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
3846 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
3847 | /* See update_cfs_rq_load_avg() */ | |
3848 | cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, | |
3849 | cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); | |
09a43ace VG |
3850 | } |
3851 | ||
0e2d2aaa | 3852 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3853 | { |
0e2d2aaa PZ |
3854 | cfs_rq->propagate = 1; |
3855 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3856 | } |
3857 | ||
3858 | /* Update task and its cfs_rq load average */ | |
3859 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3860 | { | |
0e2d2aaa | 3861 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3862 | |
3863 | if (entity_is_task(se)) | |
3864 | return 0; | |
3865 | ||
0e2d2aaa PZ |
3866 | gcfs_rq = group_cfs_rq(se); |
3867 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3868 | return 0; |
3869 | ||
0e2d2aaa PZ |
3870 | gcfs_rq->propagate = 0; |
3871 | ||
09a43ace VG |
3872 | cfs_rq = cfs_rq_of(se); |
3873 | ||
0e2d2aaa | 3874 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3875 | |
0e2d2aaa | 3876 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
9f683953 | 3877 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); |
0dacee1b | 3878 | update_tg_cfs_load(cfs_rq, se, gcfs_rq); |
09a43ace | 3879 | |
ba19f51f | 3880 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 3881 | trace_pelt_se_tp(se); |
ba19f51f | 3882 | |
09a43ace VG |
3883 | return 1; |
3884 | } | |
3885 | ||
bc427898 VG |
3886 | /* |
3887 | * Check if we need to update the load and the utilization of a blocked | |
3888 | * group_entity: | |
3889 | */ | |
3890 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3891 | { | |
3892 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3893 | ||
3894 | /* | |
3895 | * If sched_entity still have not zero load or utilization, we have to | |
3896 | * decay it: | |
3897 | */ | |
3898 | if (se->avg.load_avg || se->avg.util_avg) | |
3899 | return false; | |
3900 | ||
3901 | /* | |
3902 | * If there is a pending propagation, we have to update the load and | |
3903 | * the utilization of the sched_entity: | |
3904 | */ | |
0e2d2aaa | 3905 | if (gcfs_rq->propagate) |
bc427898 VG |
3906 | return false; |
3907 | ||
3908 | /* | |
3909 | * Otherwise, the load and the utilization of the sched_entity is | |
3910 | * already zero and there is no pending propagation, so it will be a | |
3911 | * waste of time to try to decay it: | |
3912 | */ | |
3913 | return true; | |
3914 | } | |
3915 | ||
6e83125c | 3916 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3917 | |
fe749158 | 3918 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} |
09a43ace VG |
3919 | |
3920 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3921 | { | |
3922 | return 0; | |
3923 | } | |
3924 | ||
0e2d2aaa | 3925 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3926 | |
6e83125c | 3927 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3928 | |
e2f3e35f VD |
3929 | #ifdef CONFIG_NO_HZ_COMMON |
3930 | static inline void migrate_se_pelt_lag(struct sched_entity *se) | |
3931 | { | |
3932 | u64 throttled = 0, now, lut; | |
3933 | struct cfs_rq *cfs_rq; | |
3934 | struct rq *rq; | |
3935 | bool is_idle; | |
3936 | ||
3937 | if (load_avg_is_decayed(&se->avg)) | |
3938 | return; | |
3939 | ||
3940 | cfs_rq = cfs_rq_of(se); | |
3941 | rq = rq_of(cfs_rq); | |
3942 | ||
3943 | rcu_read_lock(); | |
3944 | is_idle = is_idle_task(rcu_dereference(rq->curr)); | |
3945 | rcu_read_unlock(); | |
3946 | ||
3947 | /* | |
3948 | * The lag estimation comes with a cost we don't want to pay all the | |
3949 | * time. Hence, limiting to the case where the source CPU is idle and | |
3950 | * we know we are at the greatest risk to have an outdated clock. | |
3951 | */ | |
3952 | if (!is_idle) | |
3953 | return; | |
3954 | ||
3955 | /* | |
3956 | * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where: | |
3957 | * | |
3958 | * last_update_time (the cfs_rq's last_update_time) | |
3959 | * = cfs_rq_clock_pelt()@cfs_rq_idle | |
3960 | * = rq_clock_pelt()@cfs_rq_idle | |
3961 | * - cfs->throttled_clock_pelt_time@cfs_rq_idle | |
3962 | * | |
3963 | * cfs_idle_lag (delta between rq's update and cfs_rq's update) | |
3964 | * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle | |
3965 | * | |
3966 | * rq_idle_lag (delta between now and rq's update) | |
3967 | * = sched_clock_cpu() - rq_clock()@rq_idle | |
3968 | * | |
3969 | * We can then write: | |
3970 | * | |
3971 | * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time + | |
3972 | * sched_clock_cpu() - rq_clock()@rq_idle | |
3973 | * Where: | |
3974 | * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle | |
3975 | * rq_clock()@rq_idle is rq->clock_idle | |
3976 | * cfs->throttled_clock_pelt_time@cfs_rq_idle | |
3977 | * is cfs_rq->throttled_pelt_idle | |
3978 | */ | |
3979 | ||
3980 | #ifdef CONFIG_CFS_BANDWIDTH | |
3981 | throttled = u64_u32_load(cfs_rq->throttled_pelt_idle); | |
3982 | /* The clock has been stopped for throttling */ | |
3983 | if (throttled == U64_MAX) | |
3984 | return; | |
3985 | #endif | |
3986 | now = u64_u32_load(rq->clock_pelt_idle); | |
3987 | /* | |
3988 | * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case | |
3989 | * is observed the old clock_pelt_idle value and the new clock_idle, | |
3990 | * which lead to an underestimation. The opposite would lead to an | |
3991 | * overestimation. | |
3992 | */ | |
3993 | smp_rmb(); | |
3994 | lut = cfs_rq_last_update_time(cfs_rq); | |
3995 | ||
3996 | now -= throttled; | |
3997 | if (now < lut) | |
3998 | /* | |
3999 | * cfs_rq->avg.last_update_time is more recent than our | |
4000 | * estimation, let's use it. | |
4001 | */ | |
4002 | now = lut; | |
4003 | else | |
4004 | now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle); | |
4005 | ||
4006 | __update_load_avg_blocked_se(now, se); | |
4007 | } | |
4008 | #else | |
4009 | static void migrate_se_pelt_lag(struct sched_entity *se) {} | |
4010 | #endif | |
4011 | ||
3d30544f PZ |
4012 | /** |
4013 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 4014 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 4015 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
4016 | * |
4017 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
d6531ab6 | 4018 | * avg. The immediate corollary is that all (fair) tasks must be attached. |
3d30544f PZ |
4019 | * |
4020 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
4021 | * | |
a315da5e | 4022 | * Return: true if the load decayed or we removed load. |
7c3edd2c PZ |
4023 | * |
4024 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
4025 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 4026 | */ |
a2c6c91f | 4027 | static inline int |
3a123bbb | 4028 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 4029 | { |
9f683953 | 4030 | unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; |
9d89c257 | 4031 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 4032 | int decayed = 0; |
2dac754e | 4033 | |
2a2f5d4e PZ |
4034 | if (cfs_rq->removed.nr) { |
4035 | unsigned long r; | |
87e867b4 | 4036 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
2a2f5d4e PZ |
4037 | |
4038 | raw_spin_lock(&cfs_rq->removed.lock); | |
4039 | swap(cfs_rq->removed.util_avg, removed_util); | |
4040 | swap(cfs_rq->removed.load_avg, removed_load); | |
9f683953 | 4041 | swap(cfs_rq->removed.runnable_avg, removed_runnable); |
2a2f5d4e PZ |
4042 | cfs_rq->removed.nr = 0; |
4043 | raw_spin_unlock(&cfs_rq->removed.lock); | |
4044 | ||
2a2f5d4e | 4045 | r = removed_load; |
89741892 | 4046 | sub_positive(&sa->load_avg, r); |
2d02fa8c VG |
4047 | sub_positive(&sa->load_sum, r * divider); |
4048 | /* See sa->util_sum below */ | |
4049 | sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER); | |
2dac754e | 4050 | |
2a2f5d4e | 4051 | r = removed_util; |
89741892 | 4052 | sub_positive(&sa->util_avg, r); |
98b0d890 VG |
4053 | sub_positive(&sa->util_sum, r * divider); |
4054 | /* | |
4055 | * Because of rounding, se->util_sum might ends up being +1 more than | |
4056 | * cfs->util_sum. Although this is not a problem by itself, detaching | |
4057 | * a lot of tasks with the rounding problem between 2 updates of | |
4058 | * util_avg (~1ms) can make cfs->util_sum becoming null whereas | |
4059 | * cfs_util_avg is not. | |
4060 | * Check that util_sum is still above its lower bound for the new | |
4061 | * util_avg. Given that period_contrib might have moved since the last | |
4062 | * sync, we are only sure that util_sum must be above or equal to | |
4063 | * util_avg * minimum possible divider | |
4064 | */ | |
4065 | sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER); | |
2a2f5d4e | 4066 | |
9f683953 VG |
4067 | r = removed_runnable; |
4068 | sub_positive(&sa->runnable_avg, r); | |
95246d1e VG |
4069 | sub_positive(&sa->runnable_sum, r * divider); |
4070 | /* See sa->util_sum above */ | |
4071 | sa->runnable_sum = max_t(u32, sa->runnable_sum, | |
4072 | sa->runnable_avg * PELT_MIN_DIVIDER); | |
9f683953 VG |
4073 | |
4074 | /* | |
4075 | * removed_runnable is the unweighted version of removed_load so we | |
4076 | * can use it to estimate removed_load_sum. | |
4077 | */ | |
4078 | add_tg_cfs_propagate(cfs_rq, | |
4079 | -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); | |
2a2f5d4e PZ |
4080 | |
4081 | decayed = 1; | |
9d89c257 | 4082 | } |
36ee28e4 | 4083 | |
23127296 | 4084 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
d05b4305 VD |
4085 | u64_u32_store_copy(sa->last_update_time, |
4086 | cfs_rq->last_update_time_copy, | |
4087 | sa->last_update_time); | |
2a2f5d4e | 4088 | return decayed; |
21e96f88 SM |
4089 | } |
4090 | ||
3d30544f PZ |
4091 | /** |
4092 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
4093 | * @cfs_rq: cfs_rq to attach to | |
4094 | * @se: sched_entity to attach | |
4095 | * | |
4096 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
4097 | * cfs_rq->avg.last_update_time being current. | |
4098 | */ | |
a4f9a0e5 | 4099 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
a05e8c51 | 4100 | { |
95d68593 VG |
4101 | /* |
4102 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
4103 | * See ___update_load_avg() for details. | |
4104 | */ | |
87e867b4 | 4105 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
f207934f PZ |
4106 | |
4107 | /* | |
4108 | * When we attach the @se to the @cfs_rq, we must align the decay | |
4109 | * window because without that, really weird and wonderful things can | |
4110 | * happen. | |
4111 | * | |
4112 | * XXX illustrate | |
4113 | */ | |
a05e8c51 | 4114 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
4115 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
4116 | ||
4117 | /* | |
4118 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
4119 | * period_contrib. This isn't strictly correct, but since we're | |
4120 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
4121 | * _sum a little. | |
4122 | */ | |
4123 | se->avg.util_sum = se->avg.util_avg * divider; | |
4124 | ||
9f683953 VG |
4125 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
4126 | ||
40f5aa4c | 4127 | se->avg.load_sum = se->avg.load_avg * divider; |
4128 | if (se_weight(se) < se->avg.load_sum) | |
4129 | se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se)); | |
4130 | else | |
4131 | se->avg.load_sum = 1; | |
f207934f | 4132 | |
8d5b9025 | 4133 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
4134 | cfs_rq->avg.util_avg += se->avg.util_avg; |
4135 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
9f683953 VG |
4136 | cfs_rq->avg.runnable_avg += se->avg.runnable_avg; |
4137 | cfs_rq->avg.runnable_sum += se->avg.runnable_sum; | |
0e2d2aaa PZ |
4138 | |
4139 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 4140 | |
a4f9a0e5 | 4141 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
4142 | |
4143 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
4144 | } |
4145 | ||
3d30544f PZ |
4146 | /** |
4147 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
4148 | * @cfs_rq: cfs_rq to detach from | |
4149 | * @se: sched_entity to detach | |
4150 | * | |
4151 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
4152 | * cfs_rq->avg.last_update_time being current. | |
4153 | */ | |
a05e8c51 BP |
4154 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4155 | { | |
8d5b9025 | 4156 | dequeue_load_avg(cfs_rq, se); |
89741892 | 4157 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
7ceb7710 VG |
4158 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); |
4159 | /* See update_cfs_rq_load_avg() */ | |
4160 | cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, | |
4161 | cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); | |
4162 | ||
9f683953 | 4163 | sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); |
95246d1e VG |
4164 | sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); |
4165 | /* See update_cfs_rq_load_avg() */ | |
4166 | cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, | |
4167 | cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); | |
0e2d2aaa PZ |
4168 | |
4169 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 4170 | |
ea14b57e | 4171 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
4172 | |
4173 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
4174 | } |
4175 | ||
b382a531 PZ |
4176 | /* |
4177 | * Optional action to be done while updating the load average | |
4178 | */ | |
4179 | #define UPDATE_TG 0x1 | |
4180 | #define SKIP_AGE_LOAD 0x2 | |
4181 | #define DO_ATTACH 0x4 | |
e1f078f5 | 4182 | #define DO_DETACH 0x8 |
b382a531 PZ |
4183 | |
4184 | /* Update task and its cfs_rq load average */ | |
4185 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
4186 | { | |
23127296 | 4187 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
4188 | int decayed; |
4189 | ||
4190 | /* | |
4191 | * Track task load average for carrying it to new CPU after migrated, and | |
4192 | * track group sched_entity load average for task_h_load calc in migration | |
4193 | */ | |
4194 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 4195 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
4196 | |
4197 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
4198 | decayed |= propagate_entity_load_avg(se); | |
4199 | ||
4200 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
4201 | ||
ea14b57e PZ |
4202 | /* |
4203 | * DO_ATTACH means we're here from enqueue_entity(). | |
4204 | * !last_update_time means we've passed through | |
4205 | * migrate_task_rq_fair() indicating we migrated. | |
4206 | * | |
4207 | * IOW we're enqueueing a task on a new CPU. | |
4208 | */ | |
a4f9a0e5 | 4209 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 4210 | update_tg_load_avg(cfs_rq); |
b382a531 | 4211 | |
e1f078f5 CZ |
4212 | } else if (flags & DO_DETACH) { |
4213 | /* | |
4214 | * DO_DETACH means we're here from dequeue_entity() | |
4215 | * and we are migrating task out of the CPU. | |
4216 | */ | |
4217 | detach_entity_load_avg(cfs_rq, se); | |
4218 | update_tg_load_avg(cfs_rq); | |
bef69dd8 VG |
4219 | } else if (decayed) { |
4220 | cfs_rq_util_change(cfs_rq, 0); | |
4221 | ||
4222 | if (flags & UPDATE_TG) | |
fe749158 | 4223 | update_tg_load_avg(cfs_rq); |
bef69dd8 | 4224 | } |
b382a531 PZ |
4225 | } |
4226 | ||
104cb16d MR |
4227 | /* |
4228 | * Synchronize entity load avg of dequeued entity without locking | |
4229 | * the previous rq. | |
4230 | */ | |
71b47eaf | 4231 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
4232 | { |
4233 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4234 | u64 last_update_time; | |
4235 | ||
4236 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 4237 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
4238 | } |
4239 | ||
0905f04e YD |
4240 | /* |
4241 | * Task first catches up with cfs_rq, and then subtract | |
4242 | * itself from the cfs_rq (task must be off the queue now). | |
4243 | */ | |
71b47eaf | 4244 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
4245 | { |
4246 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 4247 | unsigned long flags; |
0905f04e YD |
4248 | |
4249 | /* | |
7dc603c9 | 4250 | * tasks cannot exit without having gone through wake_up_new_task() -> |
d6531ab6 CZ |
4251 | * enqueue_task_fair() which will have added things to the cfs_rq, |
4252 | * so we can remove unconditionally. | |
0905f04e | 4253 | */ |
0905f04e | 4254 | |
104cb16d | 4255 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
4256 | |
4257 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
4258 | ++cfs_rq->removed.nr; | |
4259 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
4260 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
9f683953 | 4261 | cfs_rq->removed.runnable_avg += se->avg.runnable_avg; |
2a2f5d4e | 4262 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 4263 | } |
642dbc39 | 4264 | |
9f683953 VG |
4265 | static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) |
4266 | { | |
4267 | return cfs_rq->avg.runnable_avg; | |
4268 | } | |
4269 | ||
7ea241af YD |
4270 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) |
4271 | { | |
4272 | return cfs_rq->avg.load_avg; | |
4273 | } | |
4274 | ||
d91cecc1 CY |
4275 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); |
4276 | ||
7f65ea42 PB |
4277 | static inline unsigned long task_util(struct task_struct *p) |
4278 | { | |
4279 | return READ_ONCE(p->se.avg.util_avg); | |
4280 | } | |
4281 | ||
4282 | static inline unsigned long _task_util_est(struct task_struct *p) | |
4283 | { | |
4284 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
4285 | ||
68d7a190 | 4286 | return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED)); |
7f65ea42 PB |
4287 | } |
4288 | ||
4289 | static inline unsigned long task_util_est(struct task_struct *p) | |
4290 | { | |
4291 | return max(task_util(p), _task_util_est(p)); | |
4292 | } | |
4293 | ||
a7008c07 | 4294 | #ifdef CONFIG_UCLAMP_TASK |
d81304bc QY |
4295 | static inline unsigned long uclamp_task_util(struct task_struct *p, |
4296 | unsigned long uclamp_min, | |
4297 | unsigned long uclamp_max) | |
a7008c07 | 4298 | { |
d81304bc | 4299 | return clamp(task_util_est(p), uclamp_min, uclamp_max); |
a7008c07 VS |
4300 | } |
4301 | #else | |
d81304bc QY |
4302 | static inline unsigned long uclamp_task_util(struct task_struct *p, |
4303 | unsigned long uclamp_min, | |
4304 | unsigned long uclamp_max) | |
a7008c07 VS |
4305 | { |
4306 | return task_util_est(p); | |
4307 | } | |
4308 | #endif | |
4309 | ||
7f65ea42 PB |
4310 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, |
4311 | struct task_struct *p) | |
4312 | { | |
4313 | unsigned int enqueued; | |
4314 | ||
4315 | if (!sched_feat(UTIL_EST)) | |
4316 | return; | |
4317 | ||
4318 | /* Update root cfs_rq's estimated utilization */ | |
4319 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 4320 | enqueued += _task_util_est(p); |
7f65ea42 | 4321 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
4581bea8 VD |
4322 | |
4323 | trace_sched_util_est_cfs_tp(cfs_rq); | |
7f65ea42 PB |
4324 | } |
4325 | ||
8c1f560c XY |
4326 | static inline void util_est_dequeue(struct cfs_rq *cfs_rq, |
4327 | struct task_struct *p) | |
4328 | { | |
4329 | unsigned int enqueued; | |
4330 | ||
4331 | if (!sched_feat(UTIL_EST)) | |
4332 | return; | |
4333 | ||
4334 | /* Update root cfs_rq's estimated utilization */ | |
4335 | enqueued = cfs_rq->avg.util_est.enqueued; | |
4336 | enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); | |
4337 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); | |
4338 | ||
4339 | trace_sched_util_est_cfs_tp(cfs_rq); | |
4340 | } | |
4341 | ||
b89997aa VD |
4342 | #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100) |
4343 | ||
7f65ea42 PB |
4344 | /* |
4345 | * Check if a (signed) value is within a specified (unsigned) margin, | |
4346 | * based on the observation that: | |
4347 | * | |
4348 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
4349 | * | |
3b03706f | 4350 | * NOTE: this only works when value + margin < INT_MAX. |
7f65ea42 PB |
4351 | */ |
4352 | static inline bool within_margin(int value, int margin) | |
4353 | { | |
4354 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
4355 | } | |
4356 | ||
8c1f560c XY |
4357 | static inline void util_est_update(struct cfs_rq *cfs_rq, |
4358 | struct task_struct *p, | |
4359 | bool task_sleep) | |
7f65ea42 | 4360 | { |
b89997aa | 4361 | long last_ewma_diff, last_enqueued_diff; |
7f65ea42 PB |
4362 | struct util_est ue; |
4363 | ||
4364 | if (!sched_feat(UTIL_EST)) | |
4365 | return; | |
4366 | ||
7f65ea42 PB |
4367 | /* |
4368 | * Skip update of task's estimated utilization when the task has not | |
4369 | * yet completed an activation, e.g. being migrated. | |
4370 | */ | |
4371 | if (!task_sleep) | |
4372 | return; | |
4373 | ||
d519329f PB |
4374 | /* |
4375 | * If the PELT values haven't changed since enqueue time, | |
4376 | * skip the util_est update. | |
4377 | */ | |
4378 | ue = p->se.avg.util_est; | |
4379 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
4380 | return; | |
4381 | ||
b89997aa VD |
4382 | last_enqueued_diff = ue.enqueued; |
4383 | ||
b8c96361 PB |
4384 | /* |
4385 | * Reset EWMA on utilization increases, the moving average is used only | |
4386 | * to smooth utilization decreases. | |
4387 | */ | |
68d7a190 | 4388 | ue.enqueued = task_util(p); |
b8c96361 PB |
4389 | if (sched_feat(UTIL_EST_FASTUP)) { |
4390 | if (ue.ewma < ue.enqueued) { | |
4391 | ue.ewma = ue.enqueued; | |
4392 | goto done; | |
4393 | } | |
4394 | } | |
4395 | ||
7f65ea42 | 4396 | /* |
b89997aa | 4397 | * Skip update of task's estimated utilization when its members are |
7f65ea42 PB |
4398 | * already ~1% close to its last activation value. |
4399 | */ | |
7f65ea42 | 4400 | last_ewma_diff = ue.enqueued - ue.ewma; |
b89997aa VD |
4401 | last_enqueued_diff -= ue.enqueued; |
4402 | if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) { | |
4403 | if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN)) | |
4404 | goto done; | |
4405 | ||
7f65ea42 | 4406 | return; |
b89997aa | 4407 | } |
7f65ea42 | 4408 | |
10a35e68 VG |
4409 | /* |
4410 | * To avoid overestimation of actual task utilization, skip updates if | |
4411 | * we cannot grant there is idle time in this CPU. | |
4412 | */ | |
8c1f560c | 4413 | if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq)))) |
10a35e68 VG |
4414 | return; |
4415 | ||
7f65ea42 PB |
4416 | /* |
4417 | * Update Task's estimated utilization | |
4418 | * | |
4419 | * When *p completes an activation we can consolidate another sample | |
4420 | * of the task size. This is done by storing the current PELT value | |
4421 | * as ue.enqueued and by using this value to update the Exponential | |
4422 | * Weighted Moving Average (EWMA): | |
4423 | * | |
4424 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
4425 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
4426 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
4427 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
4428 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
4429 | * | |
4430 | * Where 'w' is the weight of new samples, which is configured to be | |
4431 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
4432 | */ | |
4433 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
4434 | ue.ewma += last_ewma_diff; | |
4435 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
b8c96361 | 4436 | done: |
68d7a190 | 4437 | ue.enqueued |= UTIL_AVG_UNCHANGED; |
7f65ea42 | 4438 | WRITE_ONCE(p->se.avg.util_est, ue); |
4581bea8 VD |
4439 | |
4440 | trace_sched_util_est_se_tp(&p->se); | |
7f65ea42 PB |
4441 | } |
4442 | ||
48d5e9da QY |
4443 | static inline int util_fits_cpu(unsigned long util, |
4444 | unsigned long uclamp_min, | |
4445 | unsigned long uclamp_max, | |
4446 | int cpu) | |
4447 | { | |
4448 | unsigned long capacity_orig, capacity_orig_thermal; | |
4449 | unsigned long capacity = capacity_of(cpu); | |
4450 | bool fits, uclamp_max_fits; | |
4451 | ||
4452 | /* | |
4453 | * Check if the real util fits without any uclamp boost/cap applied. | |
4454 | */ | |
4455 | fits = fits_capacity(util, capacity); | |
4456 | ||
4457 | if (!uclamp_is_used()) | |
4458 | return fits; | |
4459 | ||
4460 | /* | |
4461 | * We must use capacity_orig_of() for comparing against uclamp_min and | |
4462 | * uclamp_max. We only care about capacity pressure (by using | |
4463 | * capacity_of()) for comparing against the real util. | |
4464 | * | |
4465 | * If a task is boosted to 1024 for example, we don't want a tiny | |
4466 | * pressure to skew the check whether it fits a CPU or not. | |
4467 | * | |
4468 | * Similarly if a task is capped to capacity_orig_of(little_cpu), it | |
4469 | * should fit a little cpu even if there's some pressure. | |
4470 | * | |
4471 | * Only exception is for thermal pressure since it has a direct impact | |
4472 | * on available OPP of the system. | |
4473 | * | |
4474 | * We honour it for uclamp_min only as a drop in performance level | |
4475 | * could result in not getting the requested minimum performance level. | |
4476 | * | |
4477 | * For uclamp_max, we can tolerate a drop in performance level as the | |
4478 | * goal is to cap the task. So it's okay if it's getting less. | |
48d5e9da | 4479 | */ |
b3740796 VG |
4480 | capacity_orig = capacity_orig_of(cpu); |
4481 | capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu); | |
48d5e9da QY |
4482 | |
4483 | /* | |
4484 | * We want to force a task to fit a cpu as implied by uclamp_max. | |
4485 | * But we do have some corner cases to cater for.. | |
4486 | * | |
4487 | * | |
4488 | * C=z | |
4489 | * | ___ | |
4490 | * | C=y | | | |
4491 | * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max | |
4492 | * | C=x | | | | | |
4493 | * | ___ | | | | | |
4494 | * | | | | | | | (util somewhere in this region) | |
4495 | * | | | | | | | | |
4496 | * | | | | | | | | |
4497 | * +---------------------------------------- | |
4498 | * cpu0 cpu1 cpu2 | |
4499 | * | |
4500 | * In the above example if a task is capped to a specific performance | |
4501 | * point, y, then when: | |
4502 | * | |
4503 | * * util = 80% of x then it does not fit on cpu0 and should migrate | |
4504 | * to cpu1 | |
4505 | * * util = 80% of y then it is forced to fit on cpu1 to honour | |
4506 | * uclamp_max request. | |
4507 | * | |
4508 | * which is what we're enforcing here. A task always fits if | |
4509 | * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig, | |
4510 | * the normal upmigration rules should withhold still. | |
4511 | * | |
4512 | * Only exception is when we are on max capacity, then we need to be | |
4513 | * careful not to block overutilized state. This is so because: | |
4514 | * | |
4515 | * 1. There's no concept of capping at max_capacity! We can't go | |
4516 | * beyond this performance level anyway. | |
4517 | * 2. The system is being saturated when we're operating near | |
4518 | * max capacity, it doesn't make sense to block overutilized. | |
4519 | */ | |
4520 | uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE); | |
4521 | uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig); | |
4522 | fits = fits || uclamp_max_fits; | |
4523 | ||
4524 | /* | |
4525 | * | |
4526 | * C=z | |
4527 | * | ___ (region a, capped, util >= uclamp_max) | |
4528 | * | C=y | | | |
4529 | * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max | |
4530 | * | C=x | | | | | |
4531 | * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max) | |
4532 | * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min | |
4533 | * | | | | | | | | |
4534 | * | | | | | | | (region c, boosted, util < uclamp_min) | |
4535 | * +---------------------------------------- | |
4536 | * cpu0 cpu1 cpu2 | |
4537 | * | |
4538 | * a) If util > uclamp_max, then we're capped, we don't care about | |
4539 | * actual fitness value here. We only care if uclamp_max fits | |
4540 | * capacity without taking margin/pressure into account. | |
4541 | * See comment above. | |
4542 | * | |
4543 | * b) If uclamp_min <= util <= uclamp_max, then the normal | |
4544 | * fits_capacity() rules apply. Except we need to ensure that we | |
4545 | * enforce we remain within uclamp_max, see comment above. | |
4546 | * | |
4547 | * c) If util < uclamp_min, then we are boosted. Same as (b) but we | |
4548 | * need to take into account the boosted value fits the CPU without | |
4549 | * taking margin/pressure into account. | |
4550 | * | |
4551 | * Cases (a) and (b) are handled in the 'fits' variable already. We | |
4552 | * just need to consider an extra check for case (c) after ensuring we | |
4553 | * handle the case uclamp_min > uclamp_max. | |
4554 | */ | |
4555 | uclamp_min = min(uclamp_min, uclamp_max); | |
b40e128f VG |
4556 | if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal)) |
4557 | return -1; | |
48d5e9da QY |
4558 | |
4559 | return fits; | |
4560 | } | |
4561 | ||
b48e16a6 | 4562 | static inline int task_fits_cpu(struct task_struct *p, int cpu) |
3b1baa64 | 4563 | { |
b48e16a6 QY |
4564 | unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN); |
4565 | unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX); | |
4566 | unsigned long util = task_util_est(p); | |
b40e128f VG |
4567 | /* |
4568 | * Return true only if the cpu fully fits the task requirements, which | |
4569 | * include the utilization but also the performance hints. | |
4570 | */ | |
4571 | return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0); | |
3b1baa64 MR |
4572 | } |
4573 | ||
4574 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
4575 | { | |
740cf8a7 | 4576 | if (!sched_asym_cpucap_active()) |
3b1baa64 MR |
4577 | return; |
4578 | ||
0ae78eec | 4579 | if (!p || p->nr_cpus_allowed == 1) { |
3b1baa64 MR |
4580 | rq->misfit_task_load = 0; |
4581 | return; | |
4582 | } | |
4583 | ||
b48e16a6 | 4584 | if (task_fits_cpu(p, cpu_of(rq))) { |
3b1baa64 MR |
4585 | rq->misfit_task_load = 0; |
4586 | return; | |
4587 | } | |
4588 | ||
01cfcde9 VG |
4589 | /* |
4590 | * Make sure that misfit_task_load will not be null even if | |
4591 | * task_h_load() returns 0. | |
4592 | */ | |
4593 | rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); | |
3b1baa64 MR |
4594 | } |
4595 | ||
38033c37 PZ |
4596 | #else /* CONFIG_SMP */ |
4597 | ||
a7b359fc OU |
4598 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
4599 | { | |
4600 | return true; | |
4601 | } | |
4602 | ||
d31b1a66 VG |
4603 | #define UPDATE_TG 0x0 |
4604 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 4605 | #define DO_ATTACH 0x0 |
e1f078f5 | 4606 | #define DO_DETACH 0x0 |
d31b1a66 | 4607 | |
88c0616e | 4608 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 4609 | { |
ea14b57e | 4610 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
4611 | } |
4612 | ||
9d89c257 | 4613 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 4614 | |
a05e8c51 | 4615 | static inline void |
a4f9a0e5 | 4616 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} |
a05e8c51 BP |
4617 | static inline void |
4618 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
4619 | ||
d91cecc1 | 4620 | static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
4621 | { |
4622 | return 0; | |
4623 | } | |
4624 | ||
7f65ea42 PB |
4625 | static inline void |
4626 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
4627 | ||
4628 | static inline void | |
8c1f560c XY |
4629 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {} |
4630 | ||
4631 | static inline void | |
4632 | util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p, | |
4633 | bool task_sleep) {} | |
3b1baa64 | 4634 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 4635 | |
38033c37 | 4636 | #endif /* CONFIG_SMP */ |
9d85f21c | 4637 | |
ddc97297 PZ |
4638 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4639 | { | |
4640 | #ifdef CONFIG_SCHED_DEBUG | |
4641 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
4642 | ||
4643 | if (d < 0) | |
4644 | d = -d; | |
4645 | ||
4646 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 4647 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
4648 | #endif |
4649 | } | |
4650 | ||
e50c4a23 VG |
4651 | static inline bool entity_is_long_sleeper(struct sched_entity *se) |
4652 | { | |
4653 | struct cfs_rq *cfs_rq; | |
4654 | u64 sleep_time; | |
4655 | ||
4656 | if (se->exec_start == 0) | |
4657 | return false; | |
4658 | ||
4659 | cfs_rq = cfs_rq_of(se); | |
4660 | ||
4661 | sleep_time = rq_clock_task(rq_of(cfs_rq)); | |
4662 | ||
4663 | /* Happen while migrating because of clock task divergence */ | |
4664 | if (sleep_time <= se->exec_start) | |
4665 | return false; | |
4666 | ||
4667 | sleep_time -= se->exec_start; | |
4668 | if (sleep_time > ((1ULL << 63) / scale_load_down(NICE_0_LOAD))) | |
4669 | return true; | |
4670 | ||
4671 | return false; | |
4672 | } | |
4673 | ||
aeb73b04 PZ |
4674 | static void |
4675 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
4676 | { | |
1af5f730 | 4677 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 4678 | |
2cb8600e PZ |
4679 | /* |
4680 | * The 'current' period is already promised to the current tasks, | |
4681 | * however the extra weight of the new task will slow them down a | |
4682 | * little, place the new task so that it fits in the slot that | |
4683 | * stays open at the end. | |
4684 | */ | |
94dfb5e7 | 4685 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 4686 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 4687 | |
a2e7a7eb | 4688 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 4689 | if (!initial) { |
2cae3948 JD |
4690 | unsigned long thresh; |
4691 | ||
4692 | if (se_is_idle(se)) | |
4693 | thresh = sysctl_sched_min_granularity; | |
4694 | else | |
4695 | thresh = sysctl_sched_latency; | |
a7be37ac | 4696 | |
a2e7a7eb MG |
4697 | /* |
4698 | * Halve their sleep time's effect, to allow | |
4699 | * for a gentler effect of sleepers: | |
4700 | */ | |
4701 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
4702 | thresh >>= 1; | |
51e0304c | 4703 | |
a2e7a7eb | 4704 | vruntime -= thresh; |
aeb73b04 PZ |
4705 | } |
4706 | ||
361ec2a1 ZQ |
4707 | /* |
4708 | * Pull vruntime of the entity being placed to the base level of | |
e50c4a23 VG |
4709 | * cfs_rq, to prevent boosting it if placed backwards. |
4710 | * However, min_vruntime can advance much faster than real time, with | |
4711 | * the extreme being when an entity with the minimal weight always runs | |
4712 | * on the cfs_rq. If the waking entity slept for a long time, its | |
4713 | * vruntime difference from min_vruntime may overflow s64 and their | |
4714 | * comparison may get inversed, so ignore the entity's original | |
4715 | * vruntime in that case. | |
4716 | * The maximal vruntime speedup is given by the ratio of normal to | |
4717 | * minimal weight: scale_load_down(NICE_0_LOAD) / MIN_SHARES. | |
4718 | * When placing a migrated waking entity, its exec_start has been set | |
4719 | * from a different rq. In order to take into account a possible | |
4720 | * divergence between new and prev rq's clocks task because of irq and | |
4721 | * stolen time, we take an additional margin. | |
4722 | * So, cutting off on the sleep time of | |
4723 | * 2^63 / scale_load_down(NICE_0_LOAD) ~ 104 days | |
4724 | * should be safe. | |
4725 | */ | |
4726 | if (entity_is_long_sleeper(se)) | |
361ec2a1 ZQ |
4727 | se->vruntime = vruntime; |
4728 | else | |
4729 | se->vruntime = max_vruntime(se->vruntime, vruntime); | |
aeb73b04 PZ |
4730 | } |
4731 | ||
d3d9dc33 PT |
4732 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
4733 | ||
fe61468b | 4734 | static inline bool cfs_bandwidth_used(void); |
b5179ac7 PZ |
4735 | |
4736 | /* | |
4737 | * MIGRATION | |
4738 | * | |
4739 | * dequeue | |
4740 | * update_curr() | |
4741 | * update_min_vruntime() | |
4742 | * vruntime -= min_vruntime | |
4743 | * | |
4744 | * enqueue | |
4745 | * update_curr() | |
4746 | * update_min_vruntime() | |
4747 | * vruntime += min_vruntime | |
4748 | * | |
4749 | * this way the vruntime transition between RQs is done when both | |
4750 | * min_vruntime are up-to-date. | |
4751 | * | |
4752 | * WAKEUP (remote) | |
4753 | * | |
59efa0ba | 4754 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
4755 | * vruntime -= min_vruntime |
4756 | * | |
4757 | * enqueue | |
4758 | * update_curr() | |
4759 | * update_min_vruntime() | |
4760 | * vruntime += min_vruntime | |
4761 | * | |
4762 | * this way we don't have the most up-to-date min_vruntime on the originating | |
4763 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
4764 | */ | |
4765 | ||
bf0f6f24 | 4766 | static void |
88ec22d3 | 4767 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4768 | { |
2f950354 PZ |
4769 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
4770 | bool curr = cfs_rq->curr == se; | |
4771 | ||
88ec22d3 | 4772 | /* |
2f950354 PZ |
4773 | * If we're the current task, we must renormalise before calling |
4774 | * update_curr(). | |
88ec22d3 | 4775 | */ |
2f950354 | 4776 | if (renorm && curr) |
88ec22d3 PZ |
4777 | se->vruntime += cfs_rq->min_vruntime; |
4778 | ||
2f950354 PZ |
4779 | update_curr(cfs_rq); |
4780 | ||
bf0f6f24 | 4781 | /* |
2f950354 PZ |
4782 | * Otherwise, renormalise after, such that we're placed at the current |
4783 | * moment in time, instead of some random moment in the past. Being | |
4784 | * placed in the past could significantly boost this task to the | |
4785 | * fairness detriment of existing tasks. | |
bf0f6f24 | 4786 | */ |
2f950354 PZ |
4787 | if (renorm && !curr) |
4788 | se->vruntime += cfs_rq->min_vruntime; | |
4789 | ||
89ee048f VG |
4790 | /* |
4791 | * When enqueuing a sched_entity, we must: | |
4792 | * - Update loads to have both entity and cfs_rq synced with now. | |
859f2062 CZ |
4793 | * - For group_entity, update its runnable_weight to reflect the new |
4794 | * h_nr_running of its group cfs_rq. | |
89ee048f VG |
4795 | * - For group_entity, update its weight to reflect the new share of |
4796 | * its group cfs_rq | |
4797 | * - Add its new weight to cfs_rq->load.weight | |
4798 | */ | |
b382a531 | 4799 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
9f683953 | 4800 | se_update_runnable(se); |
1ea6c46a | 4801 | update_cfs_group(se); |
17bc14b7 | 4802 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 4803 | |
1a3d027c | 4804 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 4805 | place_entity(cfs_rq, se, 0); |
e50c4a23 VG |
4806 | /* Entity has migrated, no longer consider this task hot */ |
4807 | if (flags & ENQUEUE_MIGRATED) | |
4808 | se->exec_start = 0; | |
bf0f6f24 | 4809 | |
cb251765 | 4810 | check_schedstat_required(); |
60f2415e | 4811 | update_stats_enqueue_fair(cfs_rq, se, flags); |
4fa8d299 | 4812 | check_spread(cfs_rq, se); |
2f950354 | 4813 | if (!curr) |
83b699ed | 4814 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4815 | se->on_rq = 1; |
3d4b47b4 | 4816 | |
51bf903b | 4817 | if (cfs_rq->nr_running == 1) { |
d3d9dc33 | 4818 | check_enqueue_throttle(cfs_rq); |
51bf903b CZ |
4819 | if (!throttled_hierarchy(cfs_rq)) |
4820 | list_add_leaf_cfs_rq(cfs_rq); | |
4821 | } | |
bf0f6f24 IM |
4822 | } |
4823 | ||
2c13c919 | 4824 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4825 | { |
2c13c919 RR |
4826 | for_each_sched_entity(se) { |
4827 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4828 | if (cfs_rq->last != se) |
2c13c919 | 4829 | break; |
f1044799 PZ |
4830 | |
4831 | cfs_rq->last = NULL; | |
2c13c919 RR |
4832 | } |
4833 | } | |
2002c695 | 4834 | |
2c13c919 RR |
4835 | static void __clear_buddies_next(struct sched_entity *se) |
4836 | { | |
4837 | for_each_sched_entity(se) { | |
4838 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4839 | if (cfs_rq->next != se) |
2c13c919 | 4840 | break; |
f1044799 PZ |
4841 | |
4842 | cfs_rq->next = NULL; | |
2c13c919 | 4843 | } |
2002c695 PZ |
4844 | } |
4845 | ||
ac53db59 RR |
4846 | static void __clear_buddies_skip(struct sched_entity *se) |
4847 | { | |
4848 | for_each_sched_entity(se) { | |
4849 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4850 | if (cfs_rq->skip != se) |
ac53db59 | 4851 | break; |
f1044799 PZ |
4852 | |
4853 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4854 | } |
4855 | } | |
4856 | ||
a571bbea PZ |
4857 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4858 | { | |
2c13c919 RR |
4859 | if (cfs_rq->last == se) |
4860 | __clear_buddies_last(se); | |
4861 | ||
4862 | if (cfs_rq->next == se) | |
4863 | __clear_buddies_next(se); | |
ac53db59 RR |
4864 | |
4865 | if (cfs_rq->skip == se) | |
4866 | __clear_buddies_skip(se); | |
a571bbea PZ |
4867 | } |
4868 | ||
6c16a6dc | 4869 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4870 | |
bf0f6f24 | 4871 | static void |
371fd7e7 | 4872 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4873 | { |
e1f078f5 CZ |
4874 | int action = UPDATE_TG; |
4875 | ||
4876 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se))) | |
4877 | action |= DO_DETACH; | |
4878 | ||
a2a2d680 DA |
4879 | /* |
4880 | * Update run-time statistics of the 'current'. | |
4881 | */ | |
4882 | update_curr(cfs_rq); | |
89ee048f VG |
4883 | |
4884 | /* | |
4885 | * When dequeuing a sched_entity, we must: | |
4886 | * - Update loads to have both entity and cfs_rq synced with now. | |
859f2062 CZ |
4887 | * - For group_entity, update its runnable_weight to reflect the new |
4888 | * h_nr_running of its group cfs_rq. | |
dfcb245e | 4889 | * - Subtract its previous weight from cfs_rq->load.weight. |
89ee048f VG |
4890 | * - For group entity, update its weight to reflect the new share |
4891 | * of its group cfs_rq. | |
4892 | */ | |
e1f078f5 | 4893 | update_load_avg(cfs_rq, se, action); |
9f683953 | 4894 | se_update_runnable(se); |
a2a2d680 | 4895 | |
60f2415e | 4896 | update_stats_dequeue_fair(cfs_rq, se, flags); |
67e9fb2a | 4897 | |
2002c695 | 4898 | clear_buddies(cfs_rq, se); |
4793241b | 4899 | |
83b699ed | 4900 | if (se != cfs_rq->curr) |
30cfdcfc | 4901 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4902 | se->on_rq = 0; |
30cfdcfc | 4903 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4904 | |
4905 | /* | |
b60205c7 PZ |
4906 | * Normalize after update_curr(); which will also have moved |
4907 | * min_vruntime if @se is the one holding it back. But before doing | |
4908 | * update_min_vruntime() again, which will discount @se's position and | |
4909 | * can move min_vruntime forward still more. | |
88ec22d3 | 4910 | */ |
371fd7e7 | 4911 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4912 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4913 | |
d8b4986d PT |
4914 | /* return excess runtime on last dequeue */ |
4915 | return_cfs_rq_runtime(cfs_rq); | |
4916 | ||
1ea6c46a | 4917 | update_cfs_group(se); |
b60205c7 PZ |
4918 | |
4919 | /* | |
4920 | * Now advance min_vruntime if @se was the entity holding it back, | |
4921 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4922 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4923 | * further than we started -- ie. we'll be penalized. | |
4924 | */ | |
9845c49c | 4925 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4926 | update_min_vruntime(cfs_rq); |
e2f3e35f VD |
4927 | |
4928 | if (cfs_rq->nr_running == 0) | |
4929 | update_idle_cfs_rq_clock_pelt(cfs_rq); | |
bf0f6f24 IM |
4930 | } |
4931 | ||
4932 | /* | |
4933 | * Preempt the current task with a newly woken task if needed: | |
4934 | */ | |
7c92e54f | 4935 | static void |
2e09bf55 | 4936 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4937 | { |
11697830 | 4938 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4939 | struct sched_entity *se; |
4940 | s64 delta; | |
11697830 | 4941 | |
6d0f0ebd | 4942 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4943 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4944 | if (delta_exec > ideal_runtime) { |
8875125e | 4945 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4946 | /* |
4947 | * The current task ran long enough, ensure it doesn't get | |
4948 | * re-elected due to buddy favours. | |
4949 | */ | |
4950 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4951 | return; |
4952 | } | |
4953 | ||
4954 | /* | |
4955 | * Ensure that a task that missed wakeup preemption by a | |
4956 | * narrow margin doesn't have to wait for a full slice. | |
4957 | * This also mitigates buddy induced latencies under load. | |
4958 | */ | |
f685ceac MG |
4959 | if (delta_exec < sysctl_sched_min_granularity) |
4960 | return; | |
4961 | ||
f4cfb33e WX |
4962 | se = __pick_first_entity(cfs_rq); |
4963 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4964 | |
f4cfb33e WX |
4965 | if (delta < 0) |
4966 | return; | |
d7d82944 | 4967 | |
f4cfb33e | 4968 | if (delta > ideal_runtime) |
8875125e | 4969 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4970 | } |
4971 | ||
83b699ed | 4972 | static void |
8494f412 | 4973 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4974 | { |
21f56ffe PZ |
4975 | clear_buddies(cfs_rq, se); |
4976 | ||
83b699ed SV |
4977 | /* 'current' is not kept within the tree. */ |
4978 | if (se->on_rq) { | |
4979 | /* | |
4980 | * Any task has to be enqueued before it get to execute on | |
4981 | * a CPU. So account for the time it spent waiting on the | |
4982 | * runqueue. | |
4983 | */ | |
60f2415e | 4984 | update_stats_wait_end_fair(cfs_rq, se); |
83b699ed | 4985 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4986 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4987 | } |
4988 | ||
79303e9e | 4989 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4990 | cfs_rq->curr = se; |
4fa8d299 | 4991 | |
eba1ed4b IM |
4992 | /* |
4993 | * Track our maximum slice length, if the CPU's load is at | |
4994 | * least twice that of our own weight (i.e. dont track it | |
4995 | * when there are only lesser-weight tasks around): | |
4996 | */ | |
f2bedc47 DE |
4997 | if (schedstat_enabled() && |
4998 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
ceeadb83 YS |
4999 | struct sched_statistics *stats; |
5000 | ||
5001 | stats = __schedstats_from_se(se); | |
5002 | __schedstat_set(stats->slice_max, | |
5003 | max((u64)stats->slice_max, | |
a2dcb276 | 5004 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); |
eba1ed4b | 5005 | } |
4fa8d299 | 5006 | |
4a55b450 | 5007 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
5008 | } |
5009 | ||
3f3a4904 PZ |
5010 | static int |
5011 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
5012 | ||
ac53db59 RR |
5013 | /* |
5014 | * Pick the next process, keeping these things in mind, in this order: | |
5015 | * 1) keep things fair between processes/task groups | |
5016 | * 2) pick the "next" process, since someone really wants that to run | |
5017 | * 3) pick the "last" process, for cache locality | |
5018 | * 4) do not run the "skip" process, if something else is available | |
5019 | */ | |
678d5718 PZ |
5020 | static struct sched_entity * |
5021 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 5022 | { |
678d5718 PZ |
5023 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
5024 | struct sched_entity *se; | |
5025 | ||
5026 | /* | |
5027 | * If curr is set we have to see if its left of the leftmost entity | |
5028 | * still in the tree, provided there was anything in the tree at all. | |
5029 | */ | |
5030 | if (!left || (curr && entity_before(curr, left))) | |
5031 | left = curr; | |
5032 | ||
5033 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 5034 | |
ac53db59 RR |
5035 | /* |
5036 | * Avoid running the skip buddy, if running something else can | |
5037 | * be done without getting too unfair. | |
5038 | */ | |
21f56ffe | 5039 | if (cfs_rq->skip && cfs_rq->skip == se) { |
678d5718 PZ |
5040 | struct sched_entity *second; |
5041 | ||
5042 | if (se == curr) { | |
5043 | second = __pick_first_entity(cfs_rq); | |
5044 | } else { | |
5045 | second = __pick_next_entity(se); | |
5046 | if (!second || (curr && entity_before(curr, second))) | |
5047 | second = curr; | |
5048 | } | |
5049 | ||
ac53db59 RR |
5050 | if (second && wakeup_preempt_entity(second, left) < 1) |
5051 | se = second; | |
5052 | } | |
aa2ac252 | 5053 | |
9abb8973 PO |
5054 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { |
5055 | /* | |
5056 | * Someone really wants this to run. If it's not unfair, run it. | |
5057 | */ | |
ac53db59 | 5058 | se = cfs_rq->next; |
9abb8973 PO |
5059 | } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { |
5060 | /* | |
5061 | * Prefer last buddy, try to return the CPU to a preempted task. | |
5062 | */ | |
5063 | se = cfs_rq->last; | |
5064 | } | |
ac53db59 | 5065 | |
4793241b | 5066 | return se; |
aa2ac252 PZ |
5067 | } |
5068 | ||
678d5718 | 5069 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 5070 | |
ab6cde26 | 5071 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
5072 | { |
5073 | /* | |
5074 | * If still on the runqueue then deactivate_task() | |
5075 | * was not called and update_curr() has to be done: | |
5076 | */ | |
5077 | if (prev->on_rq) | |
b7cc0896 | 5078 | update_curr(cfs_rq); |
bf0f6f24 | 5079 | |
d3d9dc33 PT |
5080 | /* throttle cfs_rqs exceeding runtime */ |
5081 | check_cfs_rq_runtime(cfs_rq); | |
5082 | ||
4fa8d299 | 5083 | check_spread(cfs_rq, prev); |
cb251765 | 5084 | |
30cfdcfc | 5085 | if (prev->on_rq) { |
60f2415e | 5086 | update_stats_wait_start_fair(cfs_rq, prev); |
30cfdcfc DA |
5087 | /* Put 'current' back into the tree. */ |
5088 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 5089 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 5090 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 5091 | } |
429d43bc | 5092 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
5093 | } |
5094 | ||
8f4d37ec PZ |
5095 | static void |
5096 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 5097 | { |
bf0f6f24 | 5098 | /* |
30cfdcfc | 5099 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 5100 | */ |
30cfdcfc | 5101 | update_curr(cfs_rq); |
bf0f6f24 | 5102 | |
9d85f21c PT |
5103 | /* |
5104 | * Ensure that runnable average is periodically updated. | |
5105 | */ | |
88c0616e | 5106 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 5107 | update_cfs_group(curr); |
9d85f21c | 5108 | |
8f4d37ec PZ |
5109 | #ifdef CONFIG_SCHED_HRTICK |
5110 | /* | |
5111 | * queued ticks are scheduled to match the slice, so don't bother | |
5112 | * validating it and just reschedule. | |
5113 | */ | |
983ed7a6 | 5114 | if (queued) { |
8875125e | 5115 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
5116 | return; |
5117 | } | |
8f4d37ec PZ |
5118 | /* |
5119 | * don't let the period tick interfere with the hrtick preemption | |
5120 | */ | |
5121 | if (!sched_feat(DOUBLE_TICK) && | |
5122 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
5123 | return; | |
5124 | #endif | |
5125 | ||
2c2efaed | 5126 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 5127 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
5128 | } |
5129 | ||
ab84d31e PT |
5130 | |
5131 | /************************************************** | |
5132 | * CFS bandwidth control machinery | |
5133 | */ | |
5134 | ||
5135 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 5136 | |
e9666d10 | 5137 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 5138 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
5139 | |
5140 | static inline bool cfs_bandwidth_used(void) | |
5141 | { | |
c5905afb | 5142 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
5143 | } |
5144 | ||
1ee14e6c | 5145 | void cfs_bandwidth_usage_inc(void) |
029632fb | 5146 | { |
ce48c146 | 5147 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
5148 | } |
5149 | ||
5150 | void cfs_bandwidth_usage_dec(void) | |
5151 | { | |
ce48c146 | 5152 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 5153 | } |
e9666d10 | 5154 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
5155 | static bool cfs_bandwidth_used(void) |
5156 | { | |
5157 | return true; | |
5158 | } | |
5159 | ||
1ee14e6c BS |
5160 | void cfs_bandwidth_usage_inc(void) {} |
5161 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 5162 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 5163 | |
ab84d31e PT |
5164 | /* |
5165 | * default period for cfs group bandwidth. | |
5166 | * default: 0.1s, units: nanoseconds | |
5167 | */ | |
5168 | static inline u64 default_cfs_period(void) | |
5169 | { | |
5170 | return 100000000ULL; | |
5171 | } | |
ec12cb7f PT |
5172 | |
5173 | static inline u64 sched_cfs_bandwidth_slice(void) | |
5174 | { | |
5175 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
5176 | } | |
5177 | ||
a9cf55b2 | 5178 | /* |
763a9ec0 QC |
5179 | * Replenish runtime according to assigned quota. We use sched_clock_cpu |
5180 | * directly instead of rq->clock to avoid adding additional synchronization | |
5181 | * around rq->lock. | |
a9cf55b2 PT |
5182 | * |
5183 | * requires cfs_b->lock | |
5184 | */ | |
029632fb | 5185 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 | 5186 | { |
bcb1704a HC |
5187 | s64 runtime; |
5188 | ||
f4183717 HC |
5189 | if (unlikely(cfs_b->quota == RUNTIME_INF)) |
5190 | return; | |
5191 | ||
5192 | cfs_b->runtime += cfs_b->quota; | |
bcb1704a HC |
5193 | runtime = cfs_b->runtime_snap - cfs_b->runtime; |
5194 | if (runtime > 0) { | |
5195 | cfs_b->burst_time += runtime; | |
5196 | cfs_b->nr_burst++; | |
5197 | } | |
5198 | ||
f4183717 | 5199 | cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst); |
bcb1704a | 5200 | cfs_b->runtime_snap = cfs_b->runtime; |
a9cf55b2 PT |
5201 | } |
5202 | ||
029632fb PZ |
5203 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5204 | { | |
5205 | return &tg->cfs_bandwidth; | |
5206 | } | |
5207 | ||
85dac906 | 5208 | /* returns 0 on failure to allocate runtime */ |
e98fa02c PT |
5209 | static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, |
5210 | struct cfs_rq *cfs_rq, u64 target_runtime) | |
ec12cb7f | 5211 | { |
e98fa02c PT |
5212 | u64 min_amount, amount = 0; |
5213 | ||
5214 | lockdep_assert_held(&cfs_b->lock); | |
ec12cb7f PT |
5215 | |
5216 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
e98fa02c | 5217 | min_amount = target_runtime - cfs_rq->runtime_remaining; |
ec12cb7f | 5218 | |
ec12cb7f PT |
5219 | if (cfs_b->quota == RUNTIME_INF) |
5220 | amount = min_amount; | |
58088ad0 | 5221 | else { |
77a4d1a1 | 5222 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
5223 | |
5224 | if (cfs_b->runtime > 0) { | |
5225 | amount = min(cfs_b->runtime, min_amount); | |
5226 | cfs_b->runtime -= amount; | |
5227 | cfs_b->idle = 0; | |
5228 | } | |
ec12cb7f | 5229 | } |
ec12cb7f PT |
5230 | |
5231 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
5232 | |
5233 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
5234 | } |
5235 | ||
e98fa02c PT |
5236 | /* returns 0 on failure to allocate runtime */ |
5237 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5238 | { | |
5239 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5240 | int ret; | |
5241 | ||
5242 | raw_spin_lock(&cfs_b->lock); | |
5243 | ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); | |
5244 | raw_spin_unlock(&cfs_b->lock); | |
5245 | ||
5246 | return ret; | |
5247 | } | |
5248 | ||
9dbdb155 | 5249 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
5250 | { |
5251 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 5252 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
5253 | |
5254 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
5255 | return; |
5256 | ||
5e2d2cc2 L |
5257 | if (cfs_rq->throttled) |
5258 | return; | |
85dac906 PT |
5259 | /* |
5260 | * if we're unable to extend our runtime we resched so that the active | |
5261 | * hierarchy can be throttled | |
5262 | */ | |
5263 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 5264 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
5265 | } |
5266 | ||
6c16a6dc | 5267 | static __always_inline |
9dbdb155 | 5268 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 5269 | { |
56f570e5 | 5270 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
5271 | return; |
5272 | ||
5273 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
5274 | } | |
5275 | ||
85dac906 PT |
5276 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
5277 | { | |
56f570e5 | 5278 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
5279 | } |
5280 | ||
64660c86 PT |
5281 | /* check whether cfs_rq, or any parent, is throttled */ |
5282 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5283 | { | |
56f570e5 | 5284 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
5285 | } |
5286 | ||
5287 | /* | |
5288 | * Ensure that neither of the group entities corresponding to src_cpu or | |
5289 | * dest_cpu are members of a throttled hierarchy when performing group | |
5290 | * load-balance operations. | |
5291 | */ | |
5292 | static inline int throttled_lb_pair(struct task_group *tg, | |
5293 | int src_cpu, int dest_cpu) | |
5294 | { | |
5295 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
5296 | ||
5297 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
5298 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
5299 | ||
5300 | return throttled_hierarchy(src_cfs_rq) || | |
5301 | throttled_hierarchy(dest_cfs_rq); | |
5302 | } | |
5303 | ||
64660c86 PT |
5304 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
5305 | { | |
5306 | struct rq *rq = data; | |
5307 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
5308 | ||
5309 | cfs_rq->throttle_count--; | |
64660c86 | 5310 | if (!cfs_rq->throttle_count) { |
64eaf507 CZ |
5311 | cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) - |
5312 | cfs_rq->throttled_clock_pelt; | |
31bc6aea | 5313 | |
a7b359fc | 5314 | /* Add cfs_rq with load or one or more already running entities to the list */ |
0a00a354 | 5315 | if (!cfs_rq_is_decayed(cfs_rq)) |
31bc6aea | 5316 | list_add_leaf_cfs_rq(cfs_rq); |
64660c86 | 5317 | } |
64660c86 PT |
5318 | |
5319 | return 0; | |
5320 | } | |
5321 | ||
5322 | static int tg_throttle_down(struct task_group *tg, void *data) | |
5323 | { | |
5324 | struct rq *rq = data; | |
5325 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
5326 | ||
82958366 | 5327 | /* group is entering throttled state, stop time */ |
31bc6aea | 5328 | if (!cfs_rq->throttle_count) { |
64eaf507 | 5329 | cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq); |
31bc6aea VG |
5330 | list_del_leaf_cfs_rq(cfs_rq); |
5331 | } | |
64660c86 PT |
5332 | cfs_rq->throttle_count++; |
5333 | ||
5334 | return 0; | |
5335 | } | |
5336 | ||
e98fa02c | 5337 | static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
5338 | { |
5339 | struct rq *rq = rq_of(cfs_rq); | |
5340 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5341 | struct sched_entity *se; | |
43e9f7f2 | 5342 | long task_delta, idle_task_delta, dequeue = 1; |
e98fa02c PT |
5343 | |
5344 | raw_spin_lock(&cfs_b->lock); | |
5345 | /* This will start the period timer if necessary */ | |
5346 | if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { | |
5347 | /* | |
5348 | * We have raced with bandwidth becoming available, and if we | |
5349 | * actually throttled the timer might not unthrottle us for an | |
5350 | * entire period. We additionally needed to make sure that any | |
5351 | * subsequent check_cfs_rq_runtime calls agree not to throttle | |
5352 | * us, as we may commit to do cfs put_prev+pick_next, so we ask | |
5353 | * for 1ns of runtime rather than just check cfs_b. | |
5354 | */ | |
5355 | dequeue = 0; | |
5356 | } else { | |
5357 | list_add_tail_rcu(&cfs_rq->throttled_list, | |
5358 | &cfs_b->throttled_cfs_rq); | |
5359 | } | |
5360 | raw_spin_unlock(&cfs_b->lock); | |
5361 | ||
5362 | if (!dequeue) | |
5363 | return false; /* Throttle no longer required. */ | |
85dac906 PT |
5364 | |
5365 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
5366 | ||
f1b17280 | 5367 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
5368 | rcu_read_lock(); |
5369 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
5370 | rcu_read_unlock(); | |
85dac906 PT |
5371 | |
5372 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 5373 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
5374 | for_each_sched_entity(se) { |
5375 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
5376 | /* throttled entity or throttle-on-deactivate */ | |
5377 | if (!se->on_rq) | |
b6d37a76 | 5378 | goto done; |
85dac906 | 5379 | |
b6d37a76 | 5380 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
6212437f | 5381 | |
30400039 JD |
5382 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5383 | idle_task_delta = cfs_rq->h_nr_running; | |
5384 | ||
85dac906 | 5385 | qcfs_rq->h_nr_running -= task_delta; |
43e9f7f2 | 5386 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 | 5387 | |
b6d37a76 PW |
5388 | if (qcfs_rq->load.weight) { |
5389 | /* Avoid re-evaluating load for this entity: */ | |
5390 | se = parent_entity(se); | |
5391 | break; | |
5392 | } | |
5393 | } | |
5394 | ||
5395 | for_each_sched_entity(se) { | |
5396 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
5397 | /* throttled entity or throttle-on-deactivate */ | |
5398 | if (!se->on_rq) | |
5399 | goto done; | |
5400 | ||
5401 | update_load_avg(qcfs_rq, se, 0); | |
5402 | se_update_runnable(se); | |
5403 | ||
30400039 JD |
5404 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5405 | idle_task_delta = cfs_rq->h_nr_running; | |
5406 | ||
b6d37a76 PW |
5407 | qcfs_rq->h_nr_running -= task_delta; |
5408 | qcfs_rq->idle_h_nr_running -= idle_task_delta; | |
85dac906 PT |
5409 | } |
5410 | ||
b6d37a76 PW |
5411 | /* At this point se is NULL and we are at root level*/ |
5412 | sub_nr_running(rq, task_delta); | |
85dac906 | 5413 | |
b6d37a76 | 5414 | done: |
c06f04c7 | 5415 | /* |
e98fa02c PT |
5416 | * Note: distribution will already see us throttled via the |
5417 | * throttled-list. rq->lock protects completion. | |
c06f04c7 | 5418 | */ |
e98fa02c PT |
5419 | cfs_rq->throttled = 1; |
5420 | cfs_rq->throttled_clock = rq_clock(rq); | |
5421 | return true; | |
85dac906 PT |
5422 | } |
5423 | ||
029632fb | 5424 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
5425 | { |
5426 | struct rq *rq = rq_of(cfs_rq); | |
5427 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5428 | struct sched_entity *se; | |
43e9f7f2 | 5429 | long task_delta, idle_task_delta; |
671fd9da | 5430 | |
22b958d8 | 5431 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
5432 | |
5433 | cfs_rq->throttled = 0; | |
1a55af2e FW |
5434 | |
5435 | update_rq_clock(rq); | |
5436 | ||
671fd9da | 5437 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 5438 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
5439 | list_del_rcu(&cfs_rq->throttled_list); |
5440 | raw_spin_unlock(&cfs_b->lock); | |
5441 | ||
64660c86 PT |
5442 | /* update hierarchical throttle state */ |
5443 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
5444 | ||
2630cde2 | 5445 | if (!cfs_rq->load.weight) { |
51bf903b CZ |
5446 | if (!cfs_rq->on_list) |
5447 | return; | |
5448 | /* | |
5449 | * Nothing to run but something to decay (on_list)? | |
5450 | * Complete the branch. | |
5451 | */ | |
5452 | for_each_sched_entity(se) { | |
5453 | if (list_add_leaf_cfs_rq(cfs_rq_of(se))) | |
5454 | break; | |
5455 | } | |
5456 | goto unthrottle_throttle; | |
2630cde2 | 5457 | } |
671fd9da PT |
5458 | |
5459 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 5460 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da | 5461 | for_each_sched_entity(se) { |
30400039 JD |
5462 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
5463 | ||
671fd9da | 5464 | if (se->on_rq) |
39f23ce0 | 5465 | break; |
30400039 JD |
5466 | enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP); |
5467 | ||
5468 | if (cfs_rq_is_idle(group_cfs_rq(se))) | |
5469 | idle_task_delta = cfs_rq->h_nr_running; | |
39f23ce0 | 5470 | |
30400039 JD |
5471 | qcfs_rq->h_nr_running += task_delta; |
5472 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
5473 | |
5474 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 5475 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 VG |
5476 | goto unthrottle_throttle; |
5477 | } | |
671fd9da | 5478 | |
39f23ce0 | 5479 | for_each_sched_entity(se) { |
30400039 | 5480 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
39f23ce0 | 5481 | |
30400039 | 5482 | update_load_avg(qcfs_rq, se, UPDATE_TG); |
39f23ce0 | 5483 | se_update_runnable(se); |
6212437f | 5484 | |
30400039 JD |
5485 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5486 | idle_task_delta = cfs_rq->h_nr_running; | |
671fd9da | 5487 | |
30400039 JD |
5488 | qcfs_rq->h_nr_running += task_delta; |
5489 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
5490 | |
5491 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 5492 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 | 5493 | goto unthrottle_throttle; |
671fd9da PT |
5494 | } |
5495 | ||
39f23ce0 VG |
5496 | /* At this point se is NULL and we are at root level*/ |
5497 | add_nr_running(rq, task_delta); | |
671fd9da | 5498 | |
39f23ce0 | 5499 | unthrottle_throttle: |
fe61468b VG |
5500 | assert_list_leaf_cfs_rq(rq); |
5501 | ||
97fb7a0a | 5502 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 5503 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 5504 | resched_curr(rq); |
671fd9da PT |
5505 | } |
5506 | ||
26a8b127 | 5507 | static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) |
671fd9da PT |
5508 | { |
5509 | struct cfs_rq *cfs_rq; | |
26a8b127 | 5510 | u64 runtime, remaining = 1; |
671fd9da PT |
5511 | |
5512 | rcu_read_lock(); | |
5513 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
5514 | throttled_list) { | |
5515 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 5516 | struct rq_flags rf; |
671fd9da | 5517 | |
c0ad4aa4 | 5518 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
5519 | if (!cfs_rq_throttled(cfs_rq)) |
5520 | goto next; | |
5521 | ||
5e2d2cc2 L |
5522 | /* By the above check, this should never be true */ |
5523 | SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); | |
5524 | ||
26a8b127 | 5525 | raw_spin_lock(&cfs_b->lock); |
671fd9da | 5526 | runtime = -cfs_rq->runtime_remaining + 1; |
26a8b127 HC |
5527 | if (runtime > cfs_b->runtime) |
5528 | runtime = cfs_b->runtime; | |
5529 | cfs_b->runtime -= runtime; | |
5530 | remaining = cfs_b->runtime; | |
5531 | raw_spin_unlock(&cfs_b->lock); | |
671fd9da PT |
5532 | |
5533 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
5534 | |
5535 | /* we check whether we're throttled above */ | |
5536 | if (cfs_rq->runtime_remaining > 0) | |
5537 | unthrottle_cfs_rq(cfs_rq); | |
5538 | ||
5539 | next: | |
c0ad4aa4 | 5540 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
5541 | |
5542 | if (!remaining) | |
5543 | break; | |
5544 | } | |
5545 | rcu_read_unlock(); | |
671fd9da PT |
5546 | } |
5547 | ||
58088ad0 PT |
5548 | /* |
5549 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
5550 | * cfs_rqs as appropriate. If there has been no activity within the last | |
5551 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
5552 | * used to track this state. | |
5553 | */ | |
c0ad4aa4 | 5554 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 5555 | { |
51f2176d | 5556 | int throttled; |
58088ad0 | 5557 | |
58088ad0 PT |
5558 | /* no need to continue the timer with no bandwidth constraint */ |
5559 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 5560 | goto out_deactivate; |
58088ad0 | 5561 | |
671fd9da | 5562 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 5563 | cfs_b->nr_periods += overrun; |
671fd9da | 5564 | |
f4183717 HC |
5565 | /* Refill extra burst quota even if cfs_b->idle */ |
5566 | __refill_cfs_bandwidth_runtime(cfs_b); | |
5567 | ||
51f2176d BS |
5568 | /* |
5569 | * idle depends on !throttled (for the case of a large deficit), and if | |
5570 | * we're going inactive then everything else can be deferred | |
5571 | */ | |
5572 | if (cfs_b->idle && !throttled) | |
5573 | goto out_deactivate; | |
a9cf55b2 | 5574 | |
671fd9da PT |
5575 | if (!throttled) { |
5576 | /* mark as potentially idle for the upcoming period */ | |
5577 | cfs_b->idle = 1; | |
51f2176d | 5578 | return 0; |
671fd9da PT |
5579 | } |
5580 | ||
e8da1b18 NR |
5581 | /* account preceding periods in which throttling occurred */ |
5582 | cfs_b->nr_throttled += overrun; | |
5583 | ||
671fd9da | 5584 | /* |
26a8b127 | 5585 | * This check is repeated as we release cfs_b->lock while we unthrottle. |
671fd9da | 5586 | */ |
ab93a4bc | 5587 | while (throttled && cfs_b->runtime > 0) { |
c0ad4aa4 | 5588 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 5589 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
26a8b127 | 5590 | distribute_cfs_runtime(cfs_b); |
c0ad4aa4 | 5591 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da PT |
5592 | |
5593 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
5594 | } | |
58088ad0 | 5595 | |
671fd9da PT |
5596 | /* |
5597 | * While we are ensured activity in the period following an | |
5598 | * unthrottle, this also covers the case in which the new bandwidth is | |
5599 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
5600 | * timer to remain active while there are any throttled entities.) | |
5601 | */ | |
5602 | cfs_b->idle = 0; | |
58088ad0 | 5603 | |
51f2176d BS |
5604 | return 0; |
5605 | ||
5606 | out_deactivate: | |
51f2176d | 5607 | return 1; |
58088ad0 | 5608 | } |
d3d9dc33 | 5609 | |
d8b4986d PT |
5610 | /* a cfs_rq won't donate quota below this amount */ |
5611 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
5612 | /* minimum remaining period time to redistribute slack quota */ | |
5613 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
5614 | /* how long we wait to gather additional slack before distributing */ | |
5615 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
5616 | ||
db06e78c BS |
5617 | /* |
5618 | * Are we near the end of the current quota period? | |
5619 | * | |
5620 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 5621 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
5622 | * migrate_hrtimers, base is never cleared, so we are fine. |
5623 | */ | |
d8b4986d PT |
5624 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
5625 | { | |
5626 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
72d0ad7c | 5627 | s64 remaining; |
d8b4986d PT |
5628 | |
5629 | /* if the call-back is running a quota refresh is already occurring */ | |
5630 | if (hrtimer_callback_running(refresh_timer)) | |
5631 | return 1; | |
5632 | ||
5633 | /* is a quota refresh about to occur? */ | |
5634 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
72d0ad7c | 5635 | if (remaining < (s64)min_expire) |
d8b4986d PT |
5636 | return 1; |
5637 | ||
5638 | return 0; | |
5639 | } | |
5640 | ||
5641 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
5642 | { | |
5643 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
5644 | ||
5645 | /* if there's a quota refresh soon don't bother with slack */ | |
5646 | if (runtime_refresh_within(cfs_b, min_left)) | |
5647 | return; | |
5648 | ||
66567fcb | 5649 | /* don't push forwards an existing deferred unthrottle */ |
5650 | if (cfs_b->slack_started) | |
5651 | return; | |
5652 | cfs_b->slack_started = true; | |
5653 | ||
4cfafd30 PZ |
5654 | hrtimer_start(&cfs_b->slack_timer, |
5655 | ns_to_ktime(cfs_bandwidth_slack_period), | |
5656 | HRTIMER_MODE_REL); | |
d8b4986d PT |
5657 | } |
5658 | ||
5659 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
5660 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5661 | { | |
5662 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5663 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
5664 | ||
5665 | if (slack_runtime <= 0) | |
5666 | return; | |
5667 | ||
5668 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 5669 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
5670 | cfs_b->runtime += slack_runtime; |
5671 | ||
5672 | /* we are under rq->lock, defer unthrottling using a timer */ | |
5673 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
5674 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
5675 | start_cfs_slack_bandwidth(cfs_b); | |
5676 | } | |
5677 | raw_spin_unlock(&cfs_b->lock); | |
5678 | ||
5679 | /* even if it's not valid for return we don't want to try again */ | |
5680 | cfs_rq->runtime_remaining -= slack_runtime; | |
5681 | } | |
5682 | ||
5683 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5684 | { | |
56f570e5 PT |
5685 | if (!cfs_bandwidth_used()) |
5686 | return; | |
5687 | ||
fccfdc6f | 5688 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
5689 | return; |
5690 | ||
5691 | __return_cfs_rq_runtime(cfs_rq); | |
5692 | } | |
5693 | ||
5694 | /* | |
5695 | * This is done with a timer (instead of inline with bandwidth return) since | |
5696 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
5697 | */ | |
5698 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
5699 | { | |
5700 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 5701 | unsigned long flags; |
d8b4986d PT |
5702 | |
5703 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 5704 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 5705 | cfs_b->slack_started = false; |
baa9be4f | 5706 | |
db06e78c | 5707 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 5708 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 5709 | return; |
db06e78c | 5710 | } |
d8b4986d | 5711 | |
c06f04c7 | 5712 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 5713 | runtime = cfs_b->runtime; |
c06f04c7 | 5714 | |
c0ad4aa4 | 5715 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
5716 | |
5717 | if (!runtime) | |
5718 | return; | |
5719 | ||
26a8b127 | 5720 | distribute_cfs_runtime(cfs_b); |
d8b4986d PT |
5721 | } |
5722 | ||
d3d9dc33 PT |
5723 | /* |
5724 | * When a group wakes up we want to make sure that its quota is not already | |
5725 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
c034f48e | 5726 | * runtime as update_curr() throttling can not trigger until it's on-rq. |
d3d9dc33 PT |
5727 | */ |
5728 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
5729 | { | |
56f570e5 PT |
5730 | if (!cfs_bandwidth_used()) |
5731 | return; | |
5732 | ||
d3d9dc33 PT |
5733 | /* an active group must be handled by the update_curr()->put() path */ |
5734 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
5735 | return; | |
5736 | ||
5737 | /* ensure the group is not already throttled */ | |
5738 | if (cfs_rq_throttled(cfs_rq)) | |
5739 | return; | |
5740 | ||
5741 | /* update runtime allocation */ | |
5742 | account_cfs_rq_runtime(cfs_rq, 0); | |
5743 | if (cfs_rq->runtime_remaining <= 0) | |
5744 | throttle_cfs_rq(cfs_rq); | |
5745 | } | |
5746 | ||
55e16d30 PZ |
5747 | static void sync_throttle(struct task_group *tg, int cpu) |
5748 | { | |
5749 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
5750 | ||
5751 | if (!cfs_bandwidth_used()) | |
5752 | return; | |
5753 | ||
5754 | if (!tg->parent) | |
5755 | return; | |
5756 | ||
5757 | cfs_rq = tg->cfs_rq[cpu]; | |
5758 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
5759 | ||
5760 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
64eaf507 | 5761 | cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu)); |
55e16d30 PZ |
5762 | } |
5763 | ||
d3d9dc33 | 5764 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 5765 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 5766 | { |
56f570e5 | 5767 | if (!cfs_bandwidth_used()) |
678d5718 | 5768 | return false; |
56f570e5 | 5769 | |
d3d9dc33 | 5770 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 5771 | return false; |
d3d9dc33 PT |
5772 | |
5773 | /* | |
5774 | * it's possible for a throttled entity to be forced into a running | |
5775 | * state (e.g. set_curr_task), in this case we're finished. | |
5776 | */ | |
5777 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 5778 | return true; |
d3d9dc33 | 5779 | |
e98fa02c | 5780 | return throttle_cfs_rq(cfs_rq); |
d3d9dc33 | 5781 | } |
029632fb | 5782 | |
029632fb PZ |
5783 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
5784 | { | |
5785 | struct cfs_bandwidth *cfs_b = | |
5786 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 5787 | |
029632fb PZ |
5788 | do_sched_cfs_slack_timer(cfs_b); |
5789 | ||
5790 | return HRTIMER_NORESTART; | |
5791 | } | |
5792 | ||
2e8e1922 PA |
5793 | extern const u64 max_cfs_quota_period; |
5794 | ||
029632fb PZ |
5795 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
5796 | { | |
5797 | struct cfs_bandwidth *cfs_b = | |
5798 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 5799 | unsigned long flags; |
029632fb PZ |
5800 | int overrun; |
5801 | int idle = 0; | |
2e8e1922 | 5802 | int count = 0; |
029632fb | 5803 | |
c0ad4aa4 | 5804 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 5805 | for (;;) { |
77a4d1a1 | 5806 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
5807 | if (!overrun) |
5808 | break; | |
5809 | ||
5a6d6a6c HC |
5810 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
5811 | ||
2e8e1922 PA |
5812 | if (++count > 3) { |
5813 | u64 new, old = ktime_to_ns(cfs_b->period); | |
5814 | ||
4929a4e6 XZ |
5815 | /* |
5816 | * Grow period by a factor of 2 to avoid losing precision. | |
5817 | * Precision loss in the quota/period ratio can cause __cfs_schedulable | |
5818 | * to fail. | |
5819 | */ | |
5820 | new = old * 2; | |
5821 | if (new < max_cfs_quota_period) { | |
5822 | cfs_b->period = ns_to_ktime(new); | |
5823 | cfs_b->quota *= 2; | |
f4183717 | 5824 | cfs_b->burst *= 2; |
4929a4e6 XZ |
5825 | |
5826 | pr_warn_ratelimited( | |
5827 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5828 | smp_processor_id(), | |
5829 | div_u64(new, NSEC_PER_USEC), | |
5830 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5831 | } else { | |
5832 | pr_warn_ratelimited( | |
5833 | "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5834 | smp_processor_id(), | |
5835 | div_u64(old, NSEC_PER_USEC), | |
5836 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5837 | } | |
2e8e1922 PA |
5838 | |
5839 | /* reset count so we don't come right back in here */ | |
5840 | count = 0; | |
5841 | } | |
029632fb | 5842 | } |
4cfafd30 PZ |
5843 | if (idle) |
5844 | cfs_b->period_active = 0; | |
c0ad4aa4 | 5845 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
5846 | |
5847 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
5848 | } | |
5849 | ||
5850 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5851 | { | |
5852 | raw_spin_lock_init(&cfs_b->lock); | |
5853 | cfs_b->runtime = 0; | |
5854 | cfs_b->quota = RUNTIME_INF; | |
5855 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
f4183717 | 5856 | cfs_b->burst = 0; |
029632fb PZ |
5857 | |
5858 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 5859 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5860 | cfs_b->period_timer.function = sched_cfs_period_timer; |
5861 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5862 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
66567fcb | 5863 | cfs_b->slack_started = false; |
029632fb PZ |
5864 | } |
5865 | ||
5866 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5867 | { | |
5868 | cfs_rq->runtime_enabled = 0; | |
5869 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
5870 | } | |
5871 | ||
77a4d1a1 | 5872 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 5873 | { |
4cfafd30 | 5874 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 5875 | |
f1d1be8a XP |
5876 | if (cfs_b->period_active) |
5877 | return; | |
5878 | ||
5879 | cfs_b->period_active = 1; | |
763a9ec0 | 5880 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); |
f1d1be8a | 5881 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5882 | } |
5883 | ||
5884 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5885 | { | |
7f1a169b TH |
5886 | /* init_cfs_bandwidth() was not called */ |
5887 | if (!cfs_b->throttled_cfs_rq.next) | |
5888 | return; | |
5889 | ||
029632fb PZ |
5890 | hrtimer_cancel(&cfs_b->period_timer); |
5891 | hrtimer_cancel(&cfs_b->slack_timer); | |
5892 | } | |
5893 | ||
502ce005 | 5894 | /* |
97fb7a0a | 5895 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
5896 | * |
5897 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5898 | * bits doesn't do much. | |
5899 | */ | |
5900 | ||
3b03706f | 5901 | /* cpu online callback */ |
0e59bdae KT |
5902 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5903 | { | |
502ce005 | 5904 | struct task_group *tg; |
0e59bdae | 5905 | |
5cb9eaa3 | 5906 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
5907 | |
5908 | rcu_read_lock(); | |
5909 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5910 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5911 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5912 | |
5913 | raw_spin_lock(&cfs_b->lock); | |
5914 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5915 | raw_spin_unlock(&cfs_b->lock); | |
5916 | } | |
502ce005 | 5917 | rcu_read_unlock(); |
0e59bdae KT |
5918 | } |
5919 | ||
502ce005 | 5920 | /* cpu offline callback */ |
38dc3348 | 5921 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5922 | { |
502ce005 PZ |
5923 | struct task_group *tg; |
5924 | ||
5cb9eaa3 | 5925 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
5926 | |
5927 | rcu_read_lock(); | |
5928 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5929 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5930 | |
029632fb PZ |
5931 | if (!cfs_rq->runtime_enabled) |
5932 | continue; | |
5933 | ||
5934 | /* | |
5935 | * clock_task is not advancing so we just need to make sure | |
5936 | * there's some valid quota amount | |
5937 | */ | |
51f2176d | 5938 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 5939 | /* |
97fb7a0a | 5940 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
5941 | * in take_cpu_down(), so we prevent new cfs throttling here. |
5942 | */ | |
5943 | cfs_rq->runtime_enabled = 0; | |
5944 | ||
029632fb PZ |
5945 | if (cfs_rq_throttled(cfs_rq)) |
5946 | unthrottle_cfs_rq(cfs_rq); | |
5947 | } | |
502ce005 | 5948 | rcu_read_unlock(); |
029632fb PZ |
5949 | } |
5950 | ||
5951 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f6783319 VG |
5952 | |
5953 | static inline bool cfs_bandwidth_used(void) | |
5954 | { | |
5955 | return false; | |
5956 | } | |
5957 | ||
9dbdb155 | 5958 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5959 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5960 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5961 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5962 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5963 | |
5964 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5965 | { | |
5966 | return 0; | |
5967 | } | |
64660c86 PT |
5968 | |
5969 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5970 | { | |
5971 | return 0; | |
5972 | } | |
5973 | ||
5974 | static inline int throttled_lb_pair(struct task_group *tg, | |
5975 | int src_cpu, int dest_cpu) | |
5976 | { | |
5977 | return 0; | |
5978 | } | |
029632fb PZ |
5979 | |
5980 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5981 | ||
5982 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5983 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5984 | #endif |
5985 | ||
029632fb PZ |
5986 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5987 | { | |
5988 | return NULL; | |
5989 | } | |
5990 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5991 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5992 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5993 | |
5994 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5995 | ||
bf0f6f24 IM |
5996 | /************************************************** |
5997 | * CFS operations on tasks: | |
5998 | */ | |
5999 | ||
8f4d37ec PZ |
6000 | #ifdef CONFIG_SCHED_HRTICK |
6001 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
6002 | { | |
8f4d37ec PZ |
6003 | struct sched_entity *se = &p->se; |
6004 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6005 | ||
9148a3a1 | 6006 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 6007 | |
8bf46a39 | 6008 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
6009 | u64 slice = sched_slice(cfs_rq, se); |
6010 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
6011 | s64 delta = slice - ran; | |
6012 | ||
6013 | if (delta < 0) { | |
65bcf072 | 6014 | if (task_current(rq, p)) |
8875125e | 6015 | resched_curr(rq); |
8f4d37ec PZ |
6016 | return; |
6017 | } | |
31656519 | 6018 | hrtick_start(rq, delta); |
8f4d37ec PZ |
6019 | } |
6020 | } | |
a4c2f00f PZ |
6021 | |
6022 | /* | |
6023 | * called from enqueue/dequeue and updates the hrtick when the | |
6024 | * current task is from our class and nr_running is low enough | |
6025 | * to matter. | |
6026 | */ | |
6027 | static void hrtick_update(struct rq *rq) | |
6028 | { | |
6029 | struct task_struct *curr = rq->curr; | |
6030 | ||
e0ee463c | 6031 | if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
6032 | return; |
6033 | ||
6034 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
6035 | hrtick_start_fair(rq, curr); | |
6036 | } | |
55e12e5e | 6037 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
6038 | static inline void |
6039 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
6040 | { | |
6041 | } | |
a4c2f00f PZ |
6042 | |
6043 | static inline void hrtick_update(struct rq *rq) | |
6044 | { | |
6045 | } | |
8f4d37ec PZ |
6046 | #endif |
6047 | ||
2802bf3c | 6048 | #ifdef CONFIG_SMP |
2802bf3c MR |
6049 | static inline bool cpu_overutilized(int cpu) |
6050 | { | |
c56ab1b3 QY |
6051 | unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN); |
6052 | unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX); | |
6053 | ||
b40e128f | 6054 | /* Return true only if the utilization doesn't fit CPU's capacity */ |
c56ab1b3 | 6055 | return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu); |
2802bf3c MR |
6056 | } |
6057 | ||
6058 | static inline void update_overutilized_status(struct rq *rq) | |
6059 | { | |
f9f240f9 | 6060 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { |
2802bf3c | 6061 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); |
f9f240f9 QY |
6062 | trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); |
6063 | } | |
2802bf3c MR |
6064 | } |
6065 | #else | |
6066 | static inline void update_overutilized_status(struct rq *rq) { } | |
6067 | #endif | |
6068 | ||
323af6de VK |
6069 | /* Runqueue only has SCHED_IDLE tasks enqueued */ |
6070 | static int sched_idle_rq(struct rq *rq) | |
6071 | { | |
6072 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
6073 | rq->nr_running); | |
6074 | } | |
6075 | ||
a480adde JD |
6076 | /* |
6077 | * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use | |
6078 | * of idle_nr_running, which does not consider idle descendants of normal | |
6079 | * entities. | |
6080 | */ | |
6081 | static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq) | |
6082 | { | |
6083 | return cfs_rq->nr_running && | |
6084 | cfs_rq->nr_running == cfs_rq->idle_nr_running; | |
6085 | } | |
6086 | ||
afa70d94 | 6087 | #ifdef CONFIG_SMP |
323af6de VK |
6088 | static int sched_idle_cpu(int cpu) |
6089 | { | |
6090 | return sched_idle_rq(cpu_rq(cpu)); | |
6091 | } | |
afa70d94 | 6092 | #endif |
323af6de | 6093 | |
bf0f6f24 IM |
6094 | /* |
6095 | * The enqueue_task method is called before nr_running is | |
6096 | * increased. Here we update the fair scheduling stats and | |
6097 | * then put the task into the rbtree: | |
6098 | */ | |
ea87bb78 | 6099 | static void |
371fd7e7 | 6100 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
6101 | { |
6102 | struct cfs_rq *cfs_rq; | |
62fb1851 | 6103 | struct sched_entity *se = &p->se; |
43e9f7f2 | 6104 | int idle_h_nr_running = task_has_idle_policy(p); |
8e1ac429 | 6105 | int task_new = !(flags & ENQUEUE_WAKEUP); |
bf0f6f24 | 6106 | |
2539fc82 PB |
6107 | /* |
6108 | * The code below (indirectly) updates schedutil which looks at | |
6109 | * the cfs_rq utilization to select a frequency. | |
6110 | * Let's add the task's estimated utilization to the cfs_rq's | |
6111 | * estimated utilization, before we update schedutil. | |
6112 | */ | |
6113 | util_est_enqueue(&rq->cfs, p); | |
6114 | ||
8c34ab19 RW |
6115 | /* |
6116 | * If in_iowait is set, the code below may not trigger any cpufreq | |
6117 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
6118 | * passed. | |
6119 | */ | |
6120 | if (p->in_iowait) | |
674e7541 | 6121 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 6122 | |
bf0f6f24 | 6123 | for_each_sched_entity(se) { |
62fb1851 | 6124 | if (se->on_rq) |
bf0f6f24 IM |
6125 | break; |
6126 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 6127 | enqueue_entity(cfs_rq, se, flags); |
85dac906 | 6128 | |
953bfcd1 | 6129 | cfs_rq->h_nr_running++; |
43e9f7f2 | 6130 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 6131 | |
30400039 JD |
6132 | if (cfs_rq_is_idle(cfs_rq)) |
6133 | idle_h_nr_running = 1; | |
6134 | ||
6d4d2246 VG |
6135 | /* end evaluation on encountering a throttled cfs_rq */ |
6136 | if (cfs_rq_throttled(cfs_rq)) | |
6137 | goto enqueue_throttle; | |
6138 | ||
88ec22d3 | 6139 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 6140 | } |
8f4d37ec | 6141 | |
2069dd75 | 6142 | for_each_sched_entity(se) { |
0f317143 | 6143 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 6144 | |
88c0616e | 6145 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 6146 | se_update_runnable(se); |
1ea6c46a | 6147 | update_cfs_group(se); |
6d4d2246 VG |
6148 | |
6149 | cfs_rq->h_nr_running++; | |
6150 | cfs_rq->idle_h_nr_running += idle_h_nr_running; | |
5ab297ba | 6151 | |
30400039 JD |
6152 | if (cfs_rq_is_idle(cfs_rq)) |
6153 | idle_h_nr_running = 1; | |
6154 | ||
5ab297ba VG |
6155 | /* end evaluation on encountering a throttled cfs_rq */ |
6156 | if (cfs_rq_throttled(cfs_rq)) | |
6157 | goto enqueue_throttle; | |
2069dd75 PZ |
6158 | } |
6159 | ||
7d148be6 VG |
6160 | /* At this point se is NULL and we are at root level*/ |
6161 | add_nr_running(rq, 1); | |
2802bf3c | 6162 | |
7d148be6 VG |
6163 | /* |
6164 | * Since new tasks are assigned an initial util_avg equal to | |
6165 | * half of the spare capacity of their CPU, tiny tasks have the | |
6166 | * ability to cross the overutilized threshold, which will | |
6167 | * result in the load balancer ruining all the task placement | |
6168 | * done by EAS. As a way to mitigate that effect, do not account | |
6169 | * for the first enqueue operation of new tasks during the | |
6170 | * overutilized flag detection. | |
6171 | * | |
6172 | * A better way of solving this problem would be to wait for | |
6173 | * the PELT signals of tasks to converge before taking them | |
6174 | * into account, but that is not straightforward to implement, | |
6175 | * and the following generally works well enough in practice. | |
6176 | */ | |
8e1ac429 | 6177 | if (!task_new) |
7d148be6 | 6178 | update_overutilized_status(rq); |
cd126afe | 6179 | |
7d148be6 | 6180 | enqueue_throttle: |
5d299eab PZ |
6181 | assert_list_leaf_cfs_rq(rq); |
6182 | ||
a4c2f00f | 6183 | hrtick_update(rq); |
bf0f6f24 IM |
6184 | } |
6185 | ||
2f36825b VP |
6186 | static void set_next_buddy(struct sched_entity *se); |
6187 | ||
bf0f6f24 IM |
6188 | /* |
6189 | * The dequeue_task method is called before nr_running is | |
6190 | * decreased. We remove the task from the rbtree and | |
6191 | * update the fair scheduling stats: | |
6192 | */ | |
371fd7e7 | 6193 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
6194 | { |
6195 | struct cfs_rq *cfs_rq; | |
62fb1851 | 6196 | struct sched_entity *se = &p->se; |
2f36825b | 6197 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 6198 | int idle_h_nr_running = task_has_idle_policy(p); |
323af6de | 6199 | bool was_sched_idle = sched_idle_rq(rq); |
bf0f6f24 | 6200 | |
8c1f560c XY |
6201 | util_est_dequeue(&rq->cfs, p); |
6202 | ||
bf0f6f24 IM |
6203 | for_each_sched_entity(se) { |
6204 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 6205 | dequeue_entity(cfs_rq, se, flags); |
85dac906 | 6206 | |
953bfcd1 | 6207 | cfs_rq->h_nr_running--; |
43e9f7f2 | 6208 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 6209 | |
30400039 JD |
6210 | if (cfs_rq_is_idle(cfs_rq)) |
6211 | idle_h_nr_running = 1; | |
6212 | ||
6d4d2246 VG |
6213 | /* end evaluation on encountering a throttled cfs_rq */ |
6214 | if (cfs_rq_throttled(cfs_rq)) | |
6215 | goto dequeue_throttle; | |
6216 | ||
bf0f6f24 | 6217 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 6218 | if (cfs_rq->load.weight) { |
754bd598 KK |
6219 | /* Avoid re-evaluating load for this entity: */ |
6220 | se = parent_entity(se); | |
2f36825b VP |
6221 | /* |
6222 | * Bias pick_next to pick a task from this cfs_rq, as | |
6223 | * p is sleeping when it is within its sched_slice. | |
6224 | */ | |
754bd598 KK |
6225 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
6226 | set_next_buddy(se); | |
bf0f6f24 | 6227 | break; |
2f36825b | 6228 | } |
371fd7e7 | 6229 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 6230 | } |
8f4d37ec | 6231 | |
2069dd75 | 6232 | for_each_sched_entity(se) { |
0f317143 | 6233 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 6234 | |
88c0616e | 6235 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 6236 | se_update_runnable(se); |
1ea6c46a | 6237 | update_cfs_group(se); |
6d4d2246 VG |
6238 | |
6239 | cfs_rq->h_nr_running--; | |
6240 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; | |
5ab297ba | 6241 | |
30400039 JD |
6242 | if (cfs_rq_is_idle(cfs_rq)) |
6243 | idle_h_nr_running = 1; | |
6244 | ||
5ab297ba VG |
6245 | /* end evaluation on encountering a throttled cfs_rq */ |
6246 | if (cfs_rq_throttled(cfs_rq)) | |
6247 | goto dequeue_throttle; | |
6248 | ||
2069dd75 PZ |
6249 | } |
6250 | ||
423d02e1 PW |
6251 | /* At this point se is NULL and we are at root level*/ |
6252 | sub_nr_running(rq, 1); | |
cd126afe | 6253 | |
323af6de VK |
6254 | /* balance early to pull high priority tasks */ |
6255 | if (unlikely(!was_sched_idle && sched_idle_rq(rq))) | |
6256 | rq->next_balance = jiffies; | |
6257 | ||
423d02e1 | 6258 | dequeue_throttle: |
8c1f560c | 6259 | util_est_update(&rq->cfs, p, task_sleep); |
a4c2f00f | 6260 | hrtick_update(rq); |
bf0f6f24 IM |
6261 | } |
6262 | ||
e7693a36 | 6263 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
6264 | |
6265 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
18c31c97 BH |
6266 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
6267 | static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask); | |
10e2f1ac | 6268 | |
9fd81dd5 | 6269 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
6270 | |
6271 | static struct { | |
6272 | cpumask_var_t idle_cpus_mask; | |
6273 | atomic_t nr_cpus; | |
f643ea22 | 6274 | int has_blocked; /* Idle CPUS has blocked load */ |
7fd7a9e0 | 6275 | int needs_update; /* Newly idle CPUs need their next_balance collated */ |
e022e0d3 | 6276 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 6277 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
6278 | } nohz ____cacheline_aligned; |
6279 | ||
9fd81dd5 | 6280 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 6281 | |
b0fb1eb4 VG |
6282 | static unsigned long cpu_load(struct rq *rq) |
6283 | { | |
6284 | return cfs_rq_load_avg(&rq->cfs); | |
6285 | } | |
6286 | ||
3318544b VG |
6287 | /* |
6288 | * cpu_load_without - compute CPU load without any contributions from *p | |
6289 | * @cpu: the CPU which load is requested | |
6290 | * @p: the task which load should be discounted | |
6291 | * | |
6292 | * The load of a CPU is defined by the load of tasks currently enqueued on that | |
6293 | * CPU as well as tasks which are currently sleeping after an execution on that | |
6294 | * CPU. | |
6295 | * | |
6296 | * This method returns the load of the specified CPU by discounting the load of | |
6297 | * the specified task, whenever the task is currently contributing to the CPU | |
6298 | * load. | |
6299 | */ | |
6300 | static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) | |
6301 | { | |
6302 | struct cfs_rq *cfs_rq; | |
6303 | unsigned int load; | |
6304 | ||
6305 | /* Task has no contribution or is new */ | |
6306 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
6307 | return cpu_load(rq); | |
6308 | ||
6309 | cfs_rq = &rq->cfs; | |
6310 | load = READ_ONCE(cfs_rq->avg.load_avg); | |
6311 | ||
6312 | /* Discount task's util from CPU's util */ | |
6313 | lsub_positive(&load, task_h_load(p)); | |
6314 | ||
6315 | return load; | |
6316 | } | |
6317 | ||
9f683953 VG |
6318 | static unsigned long cpu_runnable(struct rq *rq) |
6319 | { | |
6320 | return cfs_rq_runnable_avg(&rq->cfs); | |
6321 | } | |
6322 | ||
070f5e86 VG |
6323 | static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) |
6324 | { | |
6325 | struct cfs_rq *cfs_rq; | |
6326 | unsigned int runnable; | |
6327 | ||
6328 | /* Task has no contribution or is new */ | |
6329 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
6330 | return cpu_runnable(rq); | |
6331 | ||
6332 | cfs_rq = &rq->cfs; | |
6333 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
6334 | ||
6335 | /* Discount task's runnable from CPU's runnable */ | |
6336 | lsub_positive(&runnable, p->se.avg.runnable_avg); | |
6337 | ||
6338 | return runnable; | |
6339 | } | |
6340 | ||
ced549fa | 6341 | static unsigned long capacity_of(int cpu) |
029632fb | 6342 | { |
ced549fa | 6343 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
6344 | } |
6345 | ||
c58d25f3 PZ |
6346 | static void record_wakee(struct task_struct *p) |
6347 | { | |
6348 | /* | |
6349 | * Only decay a single time; tasks that have less then 1 wakeup per | |
6350 | * jiffy will not have built up many flips. | |
6351 | */ | |
6352 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
6353 | current->wakee_flips >>= 1; | |
6354 | current->wakee_flip_decay_ts = jiffies; | |
6355 | } | |
6356 | ||
6357 | if (current->last_wakee != p) { | |
6358 | current->last_wakee = p; | |
6359 | current->wakee_flips++; | |
6360 | } | |
6361 | } | |
6362 | ||
63b0e9ed MG |
6363 | /* |
6364 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 6365 | * |
63b0e9ed | 6366 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
6367 | * at a frequency roughly N times higher than one of its wakees. |
6368 | * | |
6369 | * In order to determine whether we should let the load spread vs consolidating | |
6370 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
6371 | * partner, and a factor of lls_size higher frequency in the other. | |
6372 | * | |
6373 | * With both conditions met, we can be relatively sure that the relationship is | |
6374 | * non-monogamous, with partner count exceeding socket size. | |
6375 | * | |
6376 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
6377 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
6378 | * socket size. | |
63b0e9ed | 6379 | */ |
62470419 MW |
6380 | static int wake_wide(struct task_struct *p) |
6381 | { | |
63b0e9ed MG |
6382 | unsigned int master = current->wakee_flips; |
6383 | unsigned int slave = p->wakee_flips; | |
17c891ab | 6384 | int factor = __this_cpu_read(sd_llc_size); |
62470419 | 6385 | |
63b0e9ed MG |
6386 | if (master < slave) |
6387 | swap(master, slave); | |
6388 | if (slave < factor || master < slave * factor) | |
6389 | return 0; | |
6390 | return 1; | |
62470419 MW |
6391 | } |
6392 | ||
90001d67 | 6393 | /* |
d153b153 PZ |
6394 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
6395 | * soonest. For the purpose of speed we only consider the waking and previous | |
6396 | * CPU. | |
90001d67 | 6397 | * |
7332dec0 MG |
6398 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
6399 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
6400 | * |
6401 | * wake_affine_weight() - considers the weight to reflect the average | |
6402 | * scheduling latency of the CPUs. This seems to work | |
6403 | * for the overloaded case. | |
90001d67 | 6404 | */ |
3b76c4a3 | 6405 | static int |
89a55f56 | 6406 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 6407 | { |
7332dec0 MG |
6408 | /* |
6409 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
6410 | * context. Only allow the move if cache is shared. Otherwise an | |
6411 | * interrupt intensive workload could force all tasks onto one | |
6412 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
6413 | * |
6414 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
6415 | * There is no guarantee that the cache hot data from an interrupt | |
6416 | * is more important than cache hot data on the prev_cpu and from | |
6417 | * a cpufreq perspective, it's better to have higher utilisation | |
6418 | * on one CPU. | |
7332dec0 | 6419 | */ |
943d355d RJ |
6420 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
6421 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 6422 | |
d153b153 | 6423 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 6424 | return this_cpu; |
90001d67 | 6425 | |
d8fcb81f JL |
6426 | if (available_idle_cpu(prev_cpu)) |
6427 | return prev_cpu; | |
6428 | ||
3b76c4a3 | 6429 | return nr_cpumask_bits; |
90001d67 PZ |
6430 | } |
6431 | ||
3b76c4a3 | 6432 | static int |
f2cdd9cc PZ |
6433 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
6434 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 6435 | { |
90001d67 PZ |
6436 | s64 this_eff_load, prev_eff_load; |
6437 | unsigned long task_load; | |
6438 | ||
11f10e54 | 6439 | this_eff_load = cpu_load(cpu_rq(this_cpu)); |
90001d67 | 6440 | |
90001d67 PZ |
6441 | if (sync) { |
6442 | unsigned long current_load = task_h_load(current); | |
6443 | ||
f2cdd9cc | 6444 | if (current_load > this_eff_load) |
3b76c4a3 | 6445 | return this_cpu; |
90001d67 | 6446 | |
f2cdd9cc | 6447 | this_eff_load -= current_load; |
90001d67 PZ |
6448 | } |
6449 | ||
90001d67 PZ |
6450 | task_load = task_h_load(p); |
6451 | ||
f2cdd9cc PZ |
6452 | this_eff_load += task_load; |
6453 | if (sched_feat(WA_BIAS)) | |
6454 | this_eff_load *= 100; | |
6455 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 6456 | |
11f10e54 | 6457 | prev_eff_load = cpu_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
6458 | prev_eff_load -= task_load; |
6459 | if (sched_feat(WA_BIAS)) | |
6460 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
6461 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 6462 | |
082f764a MG |
6463 | /* |
6464 | * If sync, adjust the weight of prev_eff_load such that if | |
6465 | * prev_eff == this_eff that select_idle_sibling() will consider | |
6466 | * stacking the wakee on top of the waker if no other CPU is | |
6467 | * idle. | |
6468 | */ | |
6469 | if (sync) | |
6470 | prev_eff_load += 1; | |
6471 | ||
6472 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
6473 | } |
6474 | ||
772bd008 | 6475 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 6476 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 6477 | { |
3b76c4a3 | 6478 | int target = nr_cpumask_bits; |
098fb9db | 6479 | |
89a55f56 | 6480 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 6481 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 6482 | |
3b76c4a3 MG |
6483 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
6484 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 6485 | |
ceeadb83 | 6486 | schedstat_inc(p->stats.nr_wakeups_affine_attempts); |
aa535ba3 | 6487 | if (target != this_cpu) |
3b76c4a3 | 6488 | return prev_cpu; |
098fb9db | 6489 | |
3b76c4a3 | 6490 | schedstat_inc(sd->ttwu_move_affine); |
ceeadb83 | 6491 | schedstat_inc(p->stats.nr_wakeups_affine); |
3b76c4a3 | 6492 | return target; |
098fb9db IM |
6493 | } |
6494 | ||
aaee1203 | 6495 | static struct sched_group * |
45da2773 | 6496 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); |
aaee1203 PZ |
6497 | |
6498 | /* | |
97fb7a0a | 6499 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
6500 | */ |
6501 | static int | |
18bd1b4b | 6502 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
6503 | { |
6504 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
6505 | unsigned int min_exit_latency = UINT_MAX; |
6506 | u64 latest_idle_timestamp = 0; | |
6507 | int least_loaded_cpu = this_cpu; | |
17346452 | 6508 | int shallowest_idle_cpu = -1; |
aaee1203 PZ |
6509 | int i; |
6510 | ||
eaecf41f MR |
6511 | /* Check if we have any choice: */ |
6512 | if (group->group_weight == 1) | |
ae4df9d6 | 6513 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 6514 | |
aaee1203 | 6515 | /* Traverse only the allowed CPUs */ |
3bd37062 | 6516 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
97886d9d AL |
6517 | struct rq *rq = cpu_rq(i); |
6518 | ||
6519 | if (!sched_core_cookie_match(rq, p)) | |
6520 | continue; | |
6521 | ||
17346452 VK |
6522 | if (sched_idle_cpu(i)) |
6523 | return i; | |
6524 | ||
943d355d | 6525 | if (available_idle_cpu(i)) { |
83a0a96a NP |
6526 | struct cpuidle_state *idle = idle_get_state(rq); |
6527 | if (idle && idle->exit_latency < min_exit_latency) { | |
6528 | /* | |
6529 | * We give priority to a CPU whose idle state | |
6530 | * has the smallest exit latency irrespective | |
6531 | * of any idle timestamp. | |
6532 | */ | |
6533 | min_exit_latency = idle->exit_latency; | |
6534 | latest_idle_timestamp = rq->idle_stamp; | |
6535 | shallowest_idle_cpu = i; | |
6536 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
6537 | rq->idle_stamp > latest_idle_timestamp) { | |
6538 | /* | |
6539 | * If equal or no active idle state, then | |
6540 | * the most recently idled CPU might have | |
6541 | * a warmer cache. | |
6542 | */ | |
6543 | latest_idle_timestamp = rq->idle_stamp; | |
6544 | shallowest_idle_cpu = i; | |
6545 | } | |
17346452 | 6546 | } else if (shallowest_idle_cpu == -1) { |
11f10e54 | 6547 | load = cpu_load(cpu_rq(i)); |
18cec7e0 | 6548 | if (load < min_load) { |
83a0a96a NP |
6549 | min_load = load; |
6550 | least_loaded_cpu = i; | |
6551 | } | |
e7693a36 GH |
6552 | } |
6553 | } | |
6554 | ||
17346452 | 6555 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 6556 | } |
e7693a36 | 6557 | |
18bd1b4b BJ |
6558 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
6559 | int cpu, int prev_cpu, int sd_flag) | |
6560 | { | |
93f50f90 | 6561 | int new_cpu = cpu; |
18bd1b4b | 6562 | |
3bd37062 | 6563 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
6564 | return prev_cpu; |
6565 | ||
c976a862 | 6566 | /* |
57abff06 | 6567 | * We need task's util for cpu_util_without, sync it up to |
c469933e | 6568 | * prev_cpu's last_update_time. |
c976a862 VK |
6569 | */ |
6570 | if (!(sd_flag & SD_BALANCE_FORK)) | |
6571 | sync_entity_load_avg(&p->se); | |
6572 | ||
18bd1b4b BJ |
6573 | while (sd) { |
6574 | struct sched_group *group; | |
6575 | struct sched_domain *tmp; | |
6576 | int weight; | |
6577 | ||
6578 | if (!(sd->flags & sd_flag)) { | |
6579 | sd = sd->child; | |
6580 | continue; | |
6581 | } | |
6582 | ||
45da2773 | 6583 | group = find_idlest_group(sd, p, cpu); |
18bd1b4b BJ |
6584 | if (!group) { |
6585 | sd = sd->child; | |
6586 | continue; | |
6587 | } | |
6588 | ||
6589 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 6590 | if (new_cpu == cpu) { |
97fb7a0a | 6591 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
6592 | sd = sd->child; |
6593 | continue; | |
6594 | } | |
6595 | ||
97fb7a0a | 6596 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
6597 | cpu = new_cpu; |
6598 | weight = sd->span_weight; | |
6599 | sd = NULL; | |
6600 | for_each_domain(cpu, tmp) { | |
6601 | if (weight <= tmp->span_weight) | |
6602 | break; | |
6603 | if (tmp->flags & sd_flag) | |
6604 | sd = tmp; | |
6605 | } | |
18bd1b4b BJ |
6606 | } |
6607 | ||
6608 | return new_cpu; | |
6609 | } | |
6610 | ||
97886d9d | 6611 | static inline int __select_idle_cpu(int cpu, struct task_struct *p) |
9fe1f127 | 6612 | { |
97886d9d AL |
6613 | if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) && |
6614 | sched_cpu_cookie_match(cpu_rq(cpu), p)) | |
9fe1f127 MG |
6615 | return cpu; |
6616 | ||
6617 | return -1; | |
6618 | } | |
6619 | ||
10e2f1ac | 6620 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 6621 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 6622 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
6623 | |
6624 | static inline void set_idle_cores(int cpu, int val) | |
6625 | { | |
6626 | struct sched_domain_shared *sds; | |
6627 | ||
6628 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
6629 | if (sds) | |
6630 | WRITE_ONCE(sds->has_idle_cores, val); | |
6631 | } | |
6632 | ||
398ba2b0 | 6633 | static inline bool test_idle_cores(int cpu) |
10e2f1ac PZ |
6634 | { |
6635 | struct sched_domain_shared *sds; | |
6636 | ||
c722f35b RR |
6637 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
6638 | if (sds) | |
6639 | return READ_ONCE(sds->has_idle_cores); | |
10e2f1ac | 6640 | |
398ba2b0 | 6641 | return false; |
10e2f1ac PZ |
6642 | } |
6643 | ||
6644 | /* | |
6645 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
6646 | * information in sd_llc_shared->has_idle_cores. | |
6647 | * | |
6648 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
6649 | * state should be fairly cheap. | |
6650 | */ | |
1b568f0a | 6651 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
6652 | { |
6653 | int core = cpu_of(rq); | |
6654 | int cpu; | |
6655 | ||
6656 | rcu_read_lock(); | |
398ba2b0 | 6657 | if (test_idle_cores(core)) |
10e2f1ac PZ |
6658 | goto unlock; |
6659 | ||
6660 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6661 | if (cpu == core) | |
6662 | continue; | |
6663 | ||
943d355d | 6664 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6665 | goto unlock; |
6666 | } | |
6667 | ||
6668 | set_idle_cores(core, 1); | |
6669 | unlock: | |
6670 | rcu_read_unlock(); | |
6671 | } | |
6672 | ||
6673 | /* | |
6674 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
6675 | * there are no idle cores left in the system; tracked through | |
6676 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
6677 | */ | |
9fe1f127 | 6678 | static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) |
10e2f1ac | 6679 | { |
9fe1f127 MG |
6680 | bool idle = true; |
6681 | int cpu; | |
10e2f1ac | 6682 | |
9fe1f127 MG |
6683 | for_each_cpu(cpu, cpu_smt_mask(core)) { |
6684 | if (!available_idle_cpu(cpu)) { | |
6685 | idle = false; | |
6686 | if (*idle_cpu == -1) { | |
6687 | if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) { | |
6688 | *idle_cpu = cpu; | |
6689 | break; | |
6690 | } | |
6691 | continue; | |
bec2860a | 6692 | } |
9fe1f127 | 6693 | break; |
10e2f1ac | 6694 | } |
9fe1f127 MG |
6695 | if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr)) |
6696 | *idle_cpu = cpu; | |
10e2f1ac PZ |
6697 | } |
6698 | ||
9fe1f127 MG |
6699 | if (idle) |
6700 | return core; | |
10e2f1ac | 6701 | |
9fe1f127 | 6702 | cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); |
10e2f1ac PZ |
6703 | return -1; |
6704 | } | |
6705 | ||
c722f35b RR |
6706 | /* |
6707 | * Scan the local SMT mask for idle CPUs. | |
6708 | */ | |
3e6efe87 | 6709 | static int select_idle_smt(struct task_struct *p, int target) |
c722f35b RR |
6710 | { |
6711 | int cpu; | |
6712 | ||
3e6efe87 | 6713 | for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) { |
b9bae704 AW |
6714 | if (cpu == target) |
6715 | continue; | |
c722f35b RR |
6716 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) |
6717 | return cpu; | |
6718 | } | |
6719 | ||
6720 | return -1; | |
6721 | } | |
6722 | ||
10e2f1ac PZ |
6723 | #else /* CONFIG_SCHED_SMT */ |
6724 | ||
9fe1f127 | 6725 | static inline void set_idle_cores(int cpu, int val) |
10e2f1ac | 6726 | { |
9fe1f127 MG |
6727 | } |
6728 | ||
398ba2b0 | 6729 | static inline bool test_idle_cores(int cpu) |
9fe1f127 | 6730 | { |
398ba2b0 | 6731 | return false; |
9fe1f127 MG |
6732 | } |
6733 | ||
6734 | static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) | |
6735 | { | |
97886d9d | 6736 | return __select_idle_cpu(core, p); |
10e2f1ac PZ |
6737 | } |
6738 | ||
3e6efe87 | 6739 | static inline int select_idle_smt(struct task_struct *p, int target) |
c722f35b RR |
6740 | { |
6741 | return -1; | |
6742 | } | |
6743 | ||
10e2f1ac PZ |
6744 | #endif /* CONFIG_SCHED_SMT */ |
6745 | ||
6746 | /* | |
6747 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6748 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6749 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6750 | */ |
c722f35b | 6751 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target) |
10e2f1ac | 6752 | { |
ec4fc801 | 6753 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
9fe1f127 | 6754 | int i, cpu, idle_cpu = -1, nr = INT_MAX; |
70fb5ccf | 6755 | struct sched_domain_shared *sd_share; |
94aafc3e | 6756 | struct rq *this_rq = this_rq(); |
9fe1f127 | 6757 | int this = smp_processor_id(); |
96c1c0cf | 6758 | struct sched_domain *this_sd = NULL; |
94aafc3e | 6759 | u64 time = 0; |
10e2f1ac | 6760 | |
bae4ec13 MG |
6761 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6762 | ||
c722f35b | 6763 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
e6e0dc2d | 6764 | u64 avg_cost, avg_idle, span_avg; |
94aafc3e | 6765 | unsigned long now = jiffies; |
1ad3aaf3 | 6766 | |
96c1c0cf AW |
6767 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6768 | if (!this_sd) | |
6769 | return -1; | |
6770 | ||
e6e0dc2d | 6771 | /* |
94aafc3e PZ |
6772 | * If we're busy, the assumption that the last idle period |
6773 | * predicts the future is flawed; age away the remaining | |
6774 | * predicted idle time. | |
e6e0dc2d | 6775 | */ |
94aafc3e PZ |
6776 | if (unlikely(this_rq->wake_stamp < now)) { |
6777 | while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) { | |
6778 | this_rq->wake_stamp++; | |
6779 | this_rq->wake_avg_idle >>= 1; | |
6780 | } | |
6781 | } | |
6782 | ||
6783 | avg_idle = this_rq->wake_avg_idle; | |
e6e0dc2d | 6784 | avg_cost = this_sd->avg_scan_cost + 1; |
10e2f1ac | 6785 | |
e6e0dc2d | 6786 | span_avg = sd->span_weight * avg_idle; |
1ad3aaf3 PZ |
6787 | if (span_avg > 4*avg_cost) |
6788 | nr = div_u64(span_avg, avg_cost); | |
6789 | else | |
6790 | nr = 4; | |
10e2f1ac | 6791 | |
bae4ec13 MG |
6792 | time = cpu_clock(this); |
6793 | } | |
60588bfa | 6794 | |
70fb5ccf CY |
6795 | if (sched_feat(SIS_UTIL)) { |
6796 | sd_share = rcu_dereference(per_cpu(sd_llc_shared, target)); | |
6797 | if (sd_share) { | |
6798 | /* because !--nr is the condition to stop scan */ | |
6799 | nr = READ_ONCE(sd_share->nr_idle_scan) + 1; | |
6800 | /* overloaded LLC is unlikely to have idle cpu/core */ | |
6801 | if (nr == 1) | |
6802 | return -1; | |
6803 | } | |
6804 | } | |
6805 | ||
56498cfb | 6806 | for_each_cpu_wrap(cpu, cpus, target + 1) { |
c722f35b | 6807 | if (has_idle_core) { |
9fe1f127 MG |
6808 | i = select_idle_core(p, cpu, cpus, &idle_cpu); |
6809 | if ((unsigned int)i < nr_cpumask_bits) | |
6810 | return i; | |
6811 | ||
6812 | } else { | |
6813 | if (!--nr) | |
6814 | return -1; | |
97886d9d | 6815 | idle_cpu = __select_idle_cpu(cpu, p); |
9fe1f127 MG |
6816 | if ((unsigned int)idle_cpu < nr_cpumask_bits) |
6817 | break; | |
6818 | } | |
10e2f1ac PZ |
6819 | } |
6820 | ||
c722f35b | 6821 | if (has_idle_core) |
02dbb724 | 6822 | set_idle_cores(target, false); |
9fe1f127 | 6823 | |
96c1c0cf | 6824 | if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) { |
bae4ec13 | 6825 | time = cpu_clock(this) - time; |
94aafc3e PZ |
6826 | |
6827 | /* | |
6828 | * Account for the scan cost of wakeups against the average | |
6829 | * idle time. | |
6830 | */ | |
6831 | this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time); | |
6832 | ||
bae4ec13 MG |
6833 | update_avg(&this_sd->avg_scan_cost, time); |
6834 | } | |
10e2f1ac | 6835 | |
9fe1f127 | 6836 | return idle_cpu; |
10e2f1ac PZ |
6837 | } |
6838 | ||
b7a33161 MR |
6839 | /* |
6840 | * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which | |
6841 | * the task fits. If no CPU is big enough, but there are idle ones, try to | |
6842 | * maximize capacity. | |
6843 | */ | |
6844 | static int | |
6845 | select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) | |
6846 | { | |
b759caa1 | 6847 | unsigned long task_util, util_min, util_max, best_cap = 0; |
b40e128f | 6848 | int fits, best_fits = 0; |
b7a33161 MR |
6849 | int cpu, best_cpu = -1; |
6850 | struct cpumask *cpus; | |
6851 | ||
ec4fc801 | 6852 | cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
b7a33161 MR |
6853 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6854 | ||
b759caa1 QY |
6855 | task_util = task_util_est(p); |
6856 | util_min = uclamp_eff_value(p, UCLAMP_MIN); | |
6857 | util_max = uclamp_eff_value(p, UCLAMP_MAX); | |
b4c9c9f1 | 6858 | |
b7a33161 MR |
6859 | for_each_cpu_wrap(cpu, cpus, target) { |
6860 | unsigned long cpu_cap = capacity_of(cpu); | |
6861 | ||
6862 | if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) | |
6863 | continue; | |
b40e128f VG |
6864 | |
6865 | fits = util_fits_cpu(task_util, util_min, util_max, cpu); | |
6866 | ||
6867 | /* This CPU fits with all requirements */ | |
6868 | if (fits > 0) | |
b7a33161 | 6869 | return cpu; |
b40e128f VG |
6870 | /* |
6871 | * Only the min performance hint (i.e. uclamp_min) doesn't fit. | |
6872 | * Look for the CPU with best capacity. | |
6873 | */ | |
6874 | else if (fits < 0) | |
6875 | cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu)); | |
b7a33161 | 6876 | |
b40e128f VG |
6877 | /* |
6878 | * First, select CPU which fits better (-1 being better than 0). | |
6879 | * Then, select the one with best capacity at same level. | |
6880 | */ | |
6881 | if ((fits < best_fits) || | |
6882 | ((fits == best_fits) && (cpu_cap > best_cap))) { | |
b7a33161 MR |
6883 | best_cap = cpu_cap; |
6884 | best_cpu = cpu; | |
b40e128f | 6885 | best_fits = fits; |
b7a33161 MR |
6886 | } |
6887 | } | |
6888 | ||
6889 | return best_cpu; | |
6890 | } | |
6891 | ||
a2e7f03e QY |
6892 | static inline bool asym_fits_cpu(unsigned long util, |
6893 | unsigned long util_min, | |
6894 | unsigned long util_max, | |
6895 | int cpu) | |
b4c9c9f1 | 6896 | { |
740cf8a7 | 6897 | if (sched_asym_cpucap_active()) |
b40e128f VG |
6898 | /* |
6899 | * Return true only if the cpu fully fits the task requirements | |
6900 | * which include the utilization and the performance hints. | |
6901 | */ | |
6902 | return (util_fits_cpu(util, util_min, util_max, cpu) > 0); | |
b4c9c9f1 VG |
6903 | |
6904 | return true; | |
6905 | } | |
6906 | ||
10e2f1ac PZ |
6907 | /* |
6908 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6909 | */ |
772bd008 | 6910 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6911 | { |
c722f35b | 6912 | bool has_idle_core = false; |
99bd5e2f | 6913 | struct sched_domain *sd; |
a2e7f03e | 6914 | unsigned long task_util, util_min, util_max; |
32e839dd | 6915 | int i, recent_used_cpu; |
a50bde51 | 6916 | |
b7a33161 | 6917 | /* |
b4c9c9f1 VG |
6918 | * On asymmetric system, update task utilization because we will check |
6919 | * that the task fits with cpu's capacity. | |
b7a33161 | 6920 | */ |
740cf8a7 | 6921 | if (sched_asym_cpucap_active()) { |
b4c9c9f1 | 6922 | sync_entity_load_avg(&p->se); |
a2e7f03e QY |
6923 | task_util = task_util_est(p); |
6924 | util_min = uclamp_eff_value(p, UCLAMP_MIN); | |
6925 | util_max = uclamp_eff_value(p, UCLAMP_MAX); | |
b7a33161 MR |
6926 | } |
6927 | ||
9099a147 | 6928 | /* |
ec4fc801 | 6929 | * per-cpu select_rq_mask usage |
9099a147 PZ |
6930 | */ |
6931 | lockdep_assert_irqs_disabled(); | |
6932 | ||
b4c9c9f1 | 6933 | if ((available_idle_cpu(target) || sched_idle_cpu(target)) && |
a2e7f03e | 6934 | asym_fits_cpu(task_util, util_min, util_max, target)) |
e0a79f52 | 6935 | return target; |
99bd5e2f SS |
6936 | |
6937 | /* | |
97fb7a0a | 6938 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6939 | */ |
3c29e651 | 6940 | if (prev != target && cpus_share_cache(prev, target) && |
b4c9c9f1 | 6941 | (available_idle_cpu(prev) || sched_idle_cpu(prev)) && |
a2e7f03e | 6942 | asym_fits_cpu(task_util, util_min, util_max, prev)) |
772bd008 | 6943 | return prev; |
a50bde51 | 6944 | |
52262ee5 MG |
6945 | /* |
6946 | * Allow a per-cpu kthread to stack with the wakee if the | |
6947 | * kworker thread and the tasks previous CPUs are the same. | |
6948 | * The assumption is that the wakee queued work for the | |
6949 | * per-cpu kthread that is now complete and the wakeup is | |
6950 | * essentially a sync wakeup. An obvious example of this | |
6951 | * pattern is IO completions. | |
6952 | */ | |
6953 | if (is_per_cpu_kthread(current) && | |
8b4e74cc | 6954 | in_task() && |
52262ee5 | 6955 | prev == smp_processor_id() && |
014ba44e | 6956 | this_rq()->nr_running <= 1 && |
a2e7f03e | 6957 | asym_fits_cpu(task_util, util_min, util_max, prev)) { |
52262ee5 MG |
6958 | return prev; |
6959 | } | |
6960 | ||
97fb7a0a | 6961 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd | 6962 | recent_used_cpu = p->recent_used_cpu; |
89aafd67 | 6963 | p->recent_used_cpu = prev; |
32e839dd MG |
6964 | if (recent_used_cpu != prev && |
6965 | recent_used_cpu != target && | |
6966 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 6967 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
f0d1ae05 | 6968 | cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) && |
a2e7f03e | 6969 | asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) { |
32e839dd MG |
6970 | return recent_used_cpu; |
6971 | } | |
6972 | ||
b4c9c9f1 VG |
6973 | /* |
6974 | * For asymmetric CPU capacity systems, our domain of interest is | |
6975 | * sd_asym_cpucapacity rather than sd_llc. | |
6976 | */ | |
740cf8a7 | 6977 | if (sched_asym_cpucap_active()) { |
b4c9c9f1 VG |
6978 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); |
6979 | /* | |
6980 | * On an asymmetric CPU capacity system where an exclusive | |
6981 | * cpuset defines a symmetric island (i.e. one unique | |
6982 | * capacity_orig value through the cpuset), the key will be set | |
6983 | * but the CPUs within that cpuset will not have a domain with | |
6984 | * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric | |
6985 | * capacity path. | |
6986 | */ | |
6987 | if (sd) { | |
6988 | i = select_idle_capacity(p, sd, target); | |
6989 | return ((unsigned)i < nr_cpumask_bits) ? i : target; | |
6990 | } | |
6991 | } | |
6992 | ||
518cd623 | 6993 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6994 | if (!sd) |
6995 | return target; | |
772bd008 | 6996 | |
c722f35b | 6997 | if (sched_smt_active()) { |
398ba2b0 | 6998 | has_idle_core = test_idle_cores(target); |
c722f35b RR |
6999 | |
7000 | if (!has_idle_core && cpus_share_cache(prev, target)) { | |
3e6efe87 | 7001 | i = select_idle_smt(p, prev); |
c722f35b RR |
7002 | if ((unsigned int)i < nr_cpumask_bits) |
7003 | return i; | |
7004 | } | |
7005 | } | |
7006 | ||
7007 | i = select_idle_cpu(p, sd, has_idle_core, target); | |
10e2f1ac PZ |
7008 | if ((unsigned)i < nr_cpumask_bits) |
7009 | return i; | |
7010 | ||
a50bde51 PZ |
7011 | return target; |
7012 | } | |
231678b7 | 7013 | |
104cb16d | 7014 | /* |
4e3c7d33 DE |
7015 | * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu |
7016 | * (@dst_cpu = -1) or migrated to @dst_cpu. | |
390031e4 QP |
7017 | */ |
7018 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
7019 | { | |
7020 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
4e3c7d33 | 7021 | unsigned long util = READ_ONCE(cfs_rq->avg.util_avg); |
390031e4 QP |
7022 | |
7023 | /* | |
4e3c7d33 DE |
7024 | * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its |
7025 | * contribution. If @p migrates from another CPU to @cpu add its | |
7026 | * contribution. In all the other cases @cpu is not impacted by the | |
7027 | * migration so its util_avg is already correct. | |
390031e4 QP |
7028 | */ |
7029 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
736cc6b3 | 7030 | lsub_positive(&util, task_util(p)); |
390031e4 QP |
7031 | else if (task_cpu(p) != cpu && dst_cpu == cpu) |
7032 | util += task_util(p); | |
7033 | ||
7034 | if (sched_feat(UTIL_EST)) { | |
4e3c7d33 DE |
7035 | unsigned long util_est; |
7036 | ||
390031e4 QP |
7037 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); |
7038 | ||
7039 | /* | |
4e3c7d33 DE |
7040 | * During wake-up @p isn't enqueued yet and doesn't contribute |
7041 | * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued. | |
7042 | * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p | |
7043 | * has been enqueued. | |
7044 | * | |
7045 | * During exec (@dst_cpu = -1) @p is enqueued and does | |
7046 | * contribute to cpu_rq(cpu)->cfs.util_est.enqueued. | |
7047 | * Remove it to "simulate" cpu_util without @p's contribution. | |
7048 | * | |
7049 | * Despite the task_on_rq_queued(@p) check there is still a | |
7050 | * small window for a possible race when an exec | |
7051 | * select_task_rq_fair() races with LB's detach_task(). | |
7052 | * | |
7053 | * detach_task() | |
7054 | * deactivate_task() | |
7055 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
7056 | * -------------------------------- A | |
7057 | * dequeue_task() \ | |
7058 | * dequeue_task_fair() + Race Time | |
7059 | * util_est_dequeue() / | |
7060 | * -------------------------------- B | |
7061 | * | |
7062 | * The additional check "current == p" is required to further | |
7063 | * reduce the race window. | |
390031e4 QP |
7064 | */ |
7065 | if (dst_cpu == cpu) | |
7066 | util_est += _task_util_est(p); | |
4e3c7d33 DE |
7067 | else if (unlikely(task_on_rq_queued(p) || current == p)) |
7068 | lsub_positive(&util_est, _task_util_est(p)); | |
390031e4 QP |
7069 | |
7070 | util = max(util, util_est); | |
7071 | } | |
7072 | ||
7073 | return min(util, capacity_orig_of(cpu)); | |
7074 | } | |
7075 | ||
4e3c7d33 DE |
7076 | /* |
7077 | * cpu_util_without: compute cpu utilization without any contributions from *p | |
7078 | * @cpu: the CPU which utilization is requested | |
7079 | * @p: the task which utilization should be discounted | |
7080 | * | |
7081 | * The utilization of a CPU is defined by the utilization of tasks currently | |
7082 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
7083 | * execution on that CPU. | |
7084 | * | |
7085 | * This method returns the utilization of the specified CPU by discounting the | |
7086 | * utilization of the specified task, whenever the task is currently | |
7087 | * contributing to the CPU utilization. | |
7088 | */ | |
7089 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) | |
7090 | { | |
7091 | /* Task has no contribution or is new */ | |
7092 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
7093 | return cpu_util_cfs(cpu); | |
7094 | ||
7095 | return cpu_util_next(cpu, p, -1); | |
7096 | } | |
7097 | ||
390031e4 | 7098 | /* |
3e8c6c9a VD |
7099 | * energy_env - Utilization landscape for energy estimation. |
7100 | * @task_busy_time: Utilization contribution by the task for which we test the | |
7101 | * placement. Given by eenv_task_busy_time(). | |
7102 | * @pd_busy_time: Utilization of the whole perf domain without the task | |
7103 | * contribution. Given by eenv_pd_busy_time(). | |
7104 | * @cpu_cap: Maximum CPU capacity for the perf domain. | |
7105 | * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap). | |
390031e4 | 7106 | */ |
3e8c6c9a VD |
7107 | struct energy_env { |
7108 | unsigned long task_busy_time; | |
7109 | unsigned long pd_busy_time; | |
7110 | unsigned long cpu_cap; | |
7111 | unsigned long pd_cap; | |
7112 | }; | |
7113 | ||
7114 | /* | |
7115 | * Compute the task busy time for compute_energy(). This time cannot be | |
7116 | * injected directly into effective_cpu_util() because of the IRQ scaling. | |
7117 | * The latter only makes sense with the most recent CPUs where the task has | |
7118 | * run. | |
7119 | */ | |
7120 | static inline void eenv_task_busy_time(struct energy_env *eenv, | |
7121 | struct task_struct *p, int prev_cpu) | |
390031e4 | 7122 | { |
3e8c6c9a VD |
7123 | unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu); |
7124 | unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu)); | |
7125 | ||
7126 | if (unlikely(irq >= max_cap)) | |
7127 | busy_time = max_cap; | |
7128 | else | |
7129 | busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap); | |
7130 | ||
7131 | eenv->task_busy_time = busy_time; | |
7132 | } | |
7133 | ||
7134 | /* | |
7135 | * Compute the perf_domain (PD) busy time for compute_energy(). Based on the | |
7136 | * utilization for each @pd_cpus, it however doesn't take into account | |
7137 | * clamping since the ratio (utilization / cpu_capacity) is already enough to | |
7138 | * scale the EM reported power consumption at the (eventually clamped) | |
7139 | * cpu_capacity. | |
7140 | * | |
7141 | * The contribution of the task @p for which we want to estimate the | |
7142 | * energy cost is removed (by cpu_util_next()) and must be calculated | |
7143 | * separately (see eenv_task_busy_time). This ensures: | |
7144 | * | |
7145 | * - A stable PD utilization, no matter which CPU of that PD we want to place | |
7146 | * the task on. | |
7147 | * | |
7148 | * - A fair comparison between CPUs as the task contribution (task_util()) | |
7149 | * will always be the same no matter which CPU utilization we rely on | |
7150 | * (util_avg or util_est). | |
7151 | * | |
7152 | * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't | |
7153 | * exceed @eenv->pd_cap. | |
7154 | */ | |
7155 | static inline void eenv_pd_busy_time(struct energy_env *eenv, | |
7156 | struct cpumask *pd_cpus, | |
7157 | struct task_struct *p) | |
7158 | { | |
7159 | unsigned long busy_time = 0; | |
390031e4 QP |
7160 | int cpu; |
7161 | ||
3e8c6c9a VD |
7162 | for_each_cpu(cpu, pd_cpus) { |
7163 | unsigned long util = cpu_util_next(cpu, p, -1); | |
489f1645 | 7164 | |
3e8c6c9a VD |
7165 | busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL); |
7166 | } | |
0372e1cf | 7167 | |
3e8c6c9a VD |
7168 | eenv->pd_busy_time = min(eenv->pd_cap, busy_time); |
7169 | } | |
af24bde8 | 7170 | |
3e8c6c9a VD |
7171 | /* |
7172 | * Compute the maximum utilization for compute_energy() when the task @p | |
7173 | * is placed on the cpu @dst_cpu. | |
7174 | * | |
7175 | * Returns the maximum utilization among @eenv->cpus. This utilization can't | |
7176 | * exceed @eenv->cpu_cap. | |
7177 | */ | |
7178 | static inline unsigned long | |
7179 | eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus, | |
7180 | struct task_struct *p, int dst_cpu) | |
7181 | { | |
7182 | unsigned long max_util = 0; | |
7183 | int cpu; | |
489f1645 | 7184 | |
3e8c6c9a VD |
7185 | for_each_cpu(cpu, pd_cpus) { |
7186 | struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL; | |
7187 | unsigned long util = cpu_util_next(cpu, p, dst_cpu); | |
7188 | unsigned long cpu_util; | |
af24bde8 | 7189 | |
390031e4 | 7190 | /* |
eb92692b QP |
7191 | * Performance domain frequency: utilization clamping |
7192 | * must be considered since it affects the selection | |
7193 | * of the performance domain frequency. | |
7194 | * NOTE: in case RT tasks are running, by default the | |
7195 | * FREQUENCY_UTIL's utilization can be max OPP. | |
390031e4 | 7196 | */ |
3e8c6c9a VD |
7197 | cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk); |
7198 | max_util = max(max_util, cpu_util); | |
390031e4 QP |
7199 | } |
7200 | ||
3e8c6c9a VD |
7201 | return min(max_util, eenv->cpu_cap); |
7202 | } | |
7203 | ||
7204 | /* | |
7205 | * compute_energy(): Use the Energy Model to estimate the energy that @pd would | |
7206 | * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task | |
7207 | * contribution is ignored. | |
7208 | */ | |
7209 | static inline unsigned long | |
7210 | compute_energy(struct energy_env *eenv, struct perf_domain *pd, | |
7211 | struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu) | |
7212 | { | |
7213 | unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu); | |
7214 | unsigned long busy_time = eenv->pd_busy_time; | |
7215 | ||
7216 | if (dst_cpu >= 0) | |
7217 | busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time); | |
7218 | ||
7219 | return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap); | |
390031e4 QP |
7220 | } |
7221 | ||
732cd75b QP |
7222 | /* |
7223 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
7224 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
7225 | * spare capacity in each performance domain and uses it as a potential | |
7226 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
7227 | * out which of the CPU candidates is the most energy-efficient. | |
7228 | * | |
7229 | * The rationale for this heuristic is as follows. In a performance domain, | |
7230 | * all the most energy efficient CPU candidates (according to the Energy | |
7231 | * Model) are those for which we'll request a low frequency. When there are | |
7232 | * several CPUs for which the frequency request will be the same, we don't | |
7233 | * have enough data to break the tie between them, because the Energy Model | |
7234 | * only includes active power costs. With this model, if we assume that | |
7235 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
7236 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
7237 | * the best candidates of the performance domain. | |
7238 | * | |
7239 | * In practice, it could be preferable from an energy standpoint to pack | |
7240 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
7241 | * but that could also hurt our chances to go cluster idle, and we have no | |
7242 | * ways to tell with the current Energy Model if this is actually a good | |
7243 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
7244 | * cluster-packing, and spreading inside a cluster. That should at least be | |
7245 | * a good thing for latency, and this is consistent with the idea that most | |
7246 | * of the energy savings of EAS come from the asymmetry of the system, and | |
7247 | * not so much from breaking the tie between identical CPUs. That's also the | |
7248 | * reason why EAS is enabled in the topology code only for systems where | |
7249 | * SD_ASYM_CPUCAPACITY is set. | |
7250 | * | |
7251 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
7252 | * they don't have any useful utilization data yet and it's not possible to | |
7253 | * forecast their impact on energy consumption. Consequently, they will be | |
7254 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
7255 | * to be energy-inefficient in some use-cases. The alternative would be to | |
7256 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
7257 | * their util_avg from the parent task, but those heuristics could hurt | |
7258 | * other use-cases too. So, until someone finds a better way to solve this, | |
7259 | * let's keep things simple by re-using the existing slow path. | |
7260 | */ | |
732cd75b QP |
7261 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) |
7262 | { | |
9b340131 | 7263 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
eb92692b | 7264 | unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; |
24422603 QY |
7265 | unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0; |
7266 | unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024; | |
3e8c6c9a | 7267 | struct root_domain *rd = this_rq()->rd; |
b812fc97 | 7268 | int cpu, best_energy_cpu, target = -1; |
b40e128f VG |
7269 | int prev_fits = -1, best_fits = -1; |
7270 | unsigned long best_thermal_cap = 0; | |
7271 | unsigned long prev_thermal_cap = 0; | |
732cd75b | 7272 | struct sched_domain *sd; |
eb92692b | 7273 | struct perf_domain *pd; |
3e8c6c9a | 7274 | struct energy_env eenv; |
732cd75b QP |
7275 | |
7276 | rcu_read_lock(); | |
7277 | pd = rcu_dereference(rd->pd); | |
7278 | if (!pd || READ_ONCE(rd->overutilized)) | |
619e090c | 7279 | goto unlock; |
732cd75b QP |
7280 | |
7281 | /* | |
7282 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
7283 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
7284 | */ | |
7285 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
7286 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
7287 | sd = sd->parent; | |
7288 | if (!sd) | |
619e090c PG |
7289 | goto unlock; |
7290 | ||
7291 | target = prev_cpu; | |
732cd75b QP |
7292 | |
7293 | sync_entity_load_avg(&p->se); | |
d81304bc | 7294 | if (!uclamp_task_util(p, p_util_min, p_util_max)) |
732cd75b QP |
7295 | goto unlock; |
7296 | ||
3e8c6c9a VD |
7297 | eenv_task_busy_time(&eenv, p, prev_cpu); |
7298 | ||
732cd75b | 7299 | for (; pd; pd = pd->next) { |
e26fd28d | 7300 | unsigned long util_min = p_util_min, util_max = p_util_max; |
3e8c6c9a VD |
7301 | unsigned long cpu_cap, cpu_thermal_cap, util; |
7302 | unsigned long cur_delta, max_spare_cap = 0; | |
24422603 | 7303 | unsigned long rq_util_min, rq_util_max; |
ad841e56 | 7304 | unsigned long prev_spare_cap = 0; |
732cd75b | 7305 | int max_spare_cap_cpu = -1; |
b812fc97 | 7306 | unsigned long base_energy; |
b40e128f | 7307 | int fits, max_fits = -1; |
732cd75b | 7308 | |
9b340131 DE |
7309 | cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask); |
7310 | ||
3e8c6c9a VD |
7311 | if (cpumask_empty(cpus)) |
7312 | continue; | |
7313 | ||
7314 | /* Account thermal pressure for the energy estimation */ | |
7315 | cpu = cpumask_first(cpus); | |
7316 | cpu_thermal_cap = arch_scale_cpu_capacity(cpu); | |
7317 | cpu_thermal_cap -= arch_scale_thermal_pressure(cpu); | |
7318 | ||
7319 | eenv.cpu_cap = cpu_thermal_cap; | |
7320 | eenv.pd_cap = 0; | |
7321 | ||
7322 | for_each_cpu(cpu, cpus) { | |
e26fd28d QY |
7323 | struct rq *rq = cpu_rq(cpu); |
7324 | ||
3e8c6c9a VD |
7325 | eenv.pd_cap += cpu_thermal_cap; |
7326 | ||
7327 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
7328 | continue; | |
7329 | ||
3bd37062 | 7330 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
7331 | continue; |
7332 | ||
732cd75b QP |
7333 | util = cpu_util_next(cpu, p, cpu); |
7334 | cpu_cap = capacity_of(cpu); | |
1d42509e VS |
7335 | |
7336 | /* | |
7337 | * Skip CPUs that cannot satisfy the capacity request. | |
7338 | * IOW, placing the task there would make the CPU | |
7339 | * overutilized. Take uclamp into account to see how | |
7340 | * much capacity we can get out of the CPU; this is | |
a5418be9 | 7341 | * aligned with sched_cpu_util(). |
1d42509e | 7342 | */ |
e26fd28d QY |
7343 | if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) { |
7344 | /* | |
7345 | * Open code uclamp_rq_util_with() except for | |
7346 | * the clamp() part. Ie: apply max aggregation | |
7347 | * only. util_fits_cpu() logic requires to | |
7348 | * operate on non clamped util but must use the | |
7349 | * max-aggregated uclamp_{min, max}. | |
7350 | */ | |
7351 | rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN); | |
7352 | rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX); | |
7353 | ||
7354 | util_min = max(rq_util_min, p_util_min); | |
7355 | util_max = max(rq_util_max, p_util_max); | |
24422603 | 7356 | } |
b40e128f VG |
7357 | |
7358 | fits = util_fits_cpu(util, util_min, util_max, cpu); | |
7359 | if (!fits) | |
732cd75b QP |
7360 | continue; |
7361 | ||
3e8c6c9a VD |
7362 | lsub_positive(&cpu_cap, util); |
7363 | ||
732cd75b | 7364 | if (cpu == prev_cpu) { |
8d4c97c1 | 7365 | /* Always use prev_cpu as a candidate. */ |
ad841e56 | 7366 | prev_spare_cap = cpu_cap; |
b40e128f VG |
7367 | prev_fits = fits; |
7368 | } else if ((fits > max_fits) || | |
7369 | ((fits == max_fits) && (cpu_cap > max_spare_cap))) { | |
8d4c97c1 PG |
7370 | /* |
7371 | * Find the CPU with the maximum spare capacity | |
ad841e56 PG |
7372 | * among the remaining CPUs in the performance |
7373 | * domain. | |
8d4c97c1 | 7374 | */ |
3e8c6c9a | 7375 | max_spare_cap = cpu_cap; |
732cd75b | 7376 | max_spare_cap_cpu = cpu; |
b40e128f | 7377 | max_fits = fits; |
732cd75b QP |
7378 | } |
7379 | } | |
7380 | ||
ad841e56 | 7381 | if (max_spare_cap_cpu < 0 && prev_spare_cap == 0) |
8d4c97c1 PG |
7382 | continue; |
7383 | ||
3e8c6c9a | 7384 | eenv_pd_busy_time(&eenv, cpus, p); |
8d4c97c1 | 7385 | /* Compute the 'base' energy of the pd, without @p */ |
b812fc97 | 7386 | base_energy = compute_energy(&eenv, pd, cpus, p, -1); |
8d4c97c1 PG |
7387 | |
7388 | /* Evaluate the energy impact of using prev_cpu. */ | |
ad841e56 | 7389 | if (prev_spare_cap > 0) { |
3e8c6c9a VD |
7390 | prev_delta = compute_energy(&eenv, pd, cpus, p, |
7391 | prev_cpu); | |
7392 | /* CPU utilization has changed */ | |
b812fc97 | 7393 | if (prev_delta < base_energy) |
619e090c | 7394 | goto unlock; |
b812fc97 | 7395 | prev_delta -= base_energy; |
b40e128f | 7396 | prev_thermal_cap = cpu_thermal_cap; |
8d4c97c1 PG |
7397 | best_delta = min(best_delta, prev_delta); |
7398 | } | |
7399 | ||
7400 | /* Evaluate the energy impact of using max_spare_cap_cpu. */ | |
ad841e56 | 7401 | if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) { |
b40e128f VG |
7402 | /* Current best energy cpu fits better */ |
7403 | if (max_fits < best_fits) | |
7404 | continue; | |
7405 | ||
7406 | /* | |
7407 | * Both don't fit performance hint (i.e. uclamp_min) | |
7408 | * but best energy cpu has better capacity. | |
7409 | */ | |
7410 | if ((max_fits < 0) && | |
7411 | (cpu_thermal_cap <= best_thermal_cap)) | |
7412 | continue; | |
7413 | ||
3e8c6c9a VD |
7414 | cur_delta = compute_energy(&eenv, pd, cpus, p, |
7415 | max_spare_cap_cpu); | |
7416 | /* CPU utilization has changed */ | |
b812fc97 | 7417 | if (cur_delta < base_energy) |
619e090c | 7418 | goto unlock; |
b812fc97 | 7419 | cur_delta -= base_energy; |
b40e128f VG |
7420 | |
7421 | /* | |
7422 | * Both fit for the task but best energy cpu has lower | |
7423 | * energy impact. | |
7424 | */ | |
7425 | if ((max_fits > 0) && (best_fits > 0) && | |
7426 | (cur_delta >= best_delta)) | |
7427 | continue; | |
7428 | ||
7429 | best_delta = cur_delta; | |
7430 | best_energy_cpu = max_spare_cap_cpu; | |
7431 | best_fits = max_fits; | |
7432 | best_thermal_cap = cpu_thermal_cap; | |
732cd75b QP |
7433 | } |
7434 | } | |
732cd75b QP |
7435 | rcu_read_unlock(); |
7436 | ||
b40e128f VG |
7437 | if ((best_fits > prev_fits) || |
7438 | ((best_fits > 0) && (best_delta < prev_delta)) || | |
7439 | ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap))) | |
619e090c | 7440 | target = best_energy_cpu; |
732cd75b | 7441 | |
619e090c | 7442 | return target; |
732cd75b | 7443 | |
619e090c | 7444 | unlock: |
732cd75b QP |
7445 | rcu_read_unlock(); |
7446 | ||
619e090c | 7447 | return target; |
732cd75b QP |
7448 | } |
7449 | ||
aaee1203 | 7450 | /* |
de91b9cb | 7451 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
3aef1551 | 7452 | * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE, |
de91b9cb | 7453 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. |
aaee1203 | 7454 | * |
97fb7a0a IM |
7455 | * Balances load by selecting the idlest CPU in the idlest group, or under |
7456 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 7457 | * |
97fb7a0a | 7458 | * Returns the target CPU number. |
aaee1203 | 7459 | */ |
0017d735 | 7460 | static int |
3aef1551 | 7461 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags) |
aaee1203 | 7462 | { |
3aef1551 | 7463 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
f1d88b44 | 7464 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 7465 | int cpu = smp_processor_id(); |
63b0e9ed | 7466 | int new_cpu = prev_cpu; |
99bd5e2f | 7467 | int want_affine = 0; |
3aef1551 VS |
7468 | /* SD_flags and WF_flags share the first nibble */ |
7469 | int sd_flag = wake_flags & 0xF; | |
c88d5910 | 7470 | |
9099a147 PZ |
7471 | /* |
7472 | * required for stable ->cpus_allowed | |
7473 | */ | |
7474 | lockdep_assert_held(&p->pi_lock); | |
dc824eb8 | 7475 | if (wake_flags & WF_TTWU) { |
c58d25f3 | 7476 | record_wakee(p); |
732cd75b | 7477 | |
f8a696f2 | 7478 | if (sched_energy_enabled()) { |
732cd75b QP |
7479 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
7480 | if (new_cpu >= 0) | |
7481 | return new_cpu; | |
7482 | new_cpu = prev_cpu; | |
7483 | } | |
7484 | ||
00061968 | 7485 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 7486 | } |
aaee1203 | 7487 | |
dce840a0 | 7488 | rcu_read_lock(); |
aaee1203 | 7489 | for_each_domain(cpu, tmp) { |
fe3bcfe1 | 7490 | /* |
97fb7a0a | 7491 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 7492 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 7493 | */ |
99bd5e2f SS |
7494 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
7495 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
7496 | if (cpu != prev_cpu) |
7497 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
7498 | ||
7499 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 7500 | break; |
f03542a7 | 7501 | } |
29cd8bae | 7502 | |
2917406c BS |
7503 | /* |
7504 | * Usually only true for WF_EXEC and WF_FORK, as sched_domains | |
7505 | * usually do not have SD_BALANCE_WAKE set. That means wakeup | |
7506 | * will usually go to the fast path. | |
7507 | */ | |
f03542a7 | 7508 | if (tmp->flags & sd_flag) |
29cd8bae | 7509 | sd = tmp; |
63b0e9ed MG |
7510 | else if (!want_affine) |
7511 | break; | |
29cd8bae PZ |
7512 | } |
7513 | ||
f1d88b44 VK |
7514 | if (unlikely(sd)) { |
7515 | /* Slow path */ | |
18bd1b4b | 7516 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
dc824eb8 | 7517 | } else if (wake_flags & WF_TTWU) { /* XXX always ? */ |
f1d88b44 | 7518 | /* Fast path */ |
f1d88b44 | 7519 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
e7693a36 | 7520 | } |
dce840a0 | 7521 | rcu_read_unlock(); |
e7693a36 | 7522 | |
c88d5910 | 7523 | return new_cpu; |
e7693a36 | 7524 | } |
0a74bef8 PT |
7525 | |
7526 | /* | |
97fb7a0a | 7527 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 7528 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 7529 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 7530 | */ |
3f9672ba | 7531 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 7532 | { |
e2f3e35f VD |
7533 | struct sched_entity *se = &p->se; |
7534 | ||
59efa0ba PZ |
7535 | /* |
7536 | * As blocked tasks retain absolute vruntime the migration needs to | |
7537 | * deal with this by subtracting the old and adding the new | |
7538 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
7539 | * the task on the new runqueue. | |
7540 | */ | |
2f064a59 | 7541 | if (READ_ONCE(p->__state) == TASK_WAKING) { |
59efa0ba | 7542 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
59efa0ba | 7543 | |
d05b4305 | 7544 | se->vruntime -= u64_u32_load(cfs_rq->min_vruntime); |
59efa0ba PZ |
7545 | } |
7546 | ||
e1f078f5 | 7547 | if (!task_on_rq_migrating(p)) { |
e2f3e35f VD |
7548 | remove_entity_load_avg(se); |
7549 | ||
144d8487 | 7550 | /* |
e2f3e35f VD |
7551 | * Here, the task's PELT values have been updated according to |
7552 | * the current rq's clock. But if that clock hasn't been | |
7553 | * updated in a while, a substantial idle time will be missed, | |
7554 | * leading to an inflation after wake-up on the new rq. | |
7555 | * | |
7556 | * Estimate the missing time from the cfs_rq last_update_time | |
7557 | * and update sched_avg to improve the PELT continuity after | |
7558 | * migration. | |
144d8487 | 7559 | */ |
e2f3e35f | 7560 | migrate_se_pelt_lag(se); |
144d8487 | 7561 | } |
9d89c257 YD |
7562 | |
7563 | /* Tell new CPU we are migrated */ | |
e2f3e35f | 7564 | se->avg.last_update_time = 0; |
3944a927 | 7565 | |
3f9672ba | 7566 | update_scan_period(p, new_cpu); |
0a74bef8 | 7567 | } |
12695578 YD |
7568 | |
7569 | static void task_dead_fair(struct task_struct *p) | |
7570 | { | |
7571 | remove_entity_load_avg(&p->se); | |
7572 | } | |
6e2df058 PZ |
7573 | |
7574 | static int | |
7575 | balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
7576 | { | |
7577 | if (rq->nr_running) | |
7578 | return 1; | |
7579 | ||
7580 | return newidle_balance(rq, rf) != 0; | |
7581 | } | |
e7693a36 GH |
7582 | #endif /* CONFIG_SMP */ |
7583 | ||
a555e9d8 | 7584 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
7585 | { |
7586 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
7587 | ||
7588 | /* | |
e52fb7c0 PZ |
7589 | * Since its curr running now, convert the gran from real-time |
7590 | * to virtual-time in his units. | |
13814d42 MG |
7591 | * |
7592 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
7593 | * they get preempted easier. That is, if 'se' < 'curr' then | |
7594 | * the resulting gran will be larger, therefore penalizing the | |
7595 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
7596 | * be smaller, again penalizing the lighter task. | |
7597 | * | |
7598 | * This is especially important for buddies when the leftmost | |
7599 | * task is higher priority than the buddy. | |
0bbd3336 | 7600 | */ |
f4ad9bd2 | 7601 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
7602 | } |
7603 | ||
464b7527 PZ |
7604 | /* |
7605 | * Should 'se' preempt 'curr'. | |
7606 | * | |
7607 | * |s1 | |
7608 | * |s2 | |
7609 | * |s3 | |
7610 | * g | |
7611 | * |<--->|c | |
7612 | * | |
7613 | * w(c, s1) = -1 | |
7614 | * w(c, s2) = 0 | |
7615 | * w(c, s3) = 1 | |
7616 | * | |
7617 | */ | |
7618 | static int | |
7619 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
7620 | { | |
7621 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
7622 | ||
7623 | if (vdiff <= 0) | |
7624 | return -1; | |
7625 | ||
a555e9d8 | 7626 | gran = wakeup_gran(se); |
464b7527 PZ |
7627 | if (vdiff > gran) |
7628 | return 1; | |
7629 | ||
7630 | return 0; | |
7631 | } | |
7632 | ||
02479099 PZ |
7633 | static void set_last_buddy(struct sched_entity *se) |
7634 | { | |
c5ae366e DA |
7635 | for_each_sched_entity(se) { |
7636 | if (SCHED_WARN_ON(!se->on_rq)) | |
7637 | return; | |
30400039 JD |
7638 | if (se_is_idle(se)) |
7639 | return; | |
69c80f3e | 7640 | cfs_rq_of(se)->last = se; |
c5ae366e | 7641 | } |
02479099 PZ |
7642 | } |
7643 | ||
7644 | static void set_next_buddy(struct sched_entity *se) | |
7645 | { | |
c5ae366e DA |
7646 | for_each_sched_entity(se) { |
7647 | if (SCHED_WARN_ON(!se->on_rq)) | |
7648 | return; | |
30400039 JD |
7649 | if (se_is_idle(se)) |
7650 | return; | |
69c80f3e | 7651 | cfs_rq_of(se)->next = se; |
c5ae366e | 7652 | } |
02479099 PZ |
7653 | } |
7654 | ||
ac53db59 RR |
7655 | static void set_skip_buddy(struct sched_entity *se) |
7656 | { | |
69c80f3e VP |
7657 | for_each_sched_entity(se) |
7658 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
7659 | } |
7660 | ||
bf0f6f24 IM |
7661 | /* |
7662 | * Preempt the current task with a newly woken task if needed: | |
7663 | */ | |
5a9b86f6 | 7664 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
7665 | { |
7666 | struct task_struct *curr = rq->curr; | |
8651a86c | 7667 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 7668 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 7669 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 7670 | int next_buddy_marked = 0; |
30400039 | 7671 | int cse_is_idle, pse_is_idle; |
bf0f6f24 | 7672 | |
4ae7d5ce IM |
7673 | if (unlikely(se == pse)) |
7674 | return; | |
7675 | ||
5238cdd3 | 7676 | /* |
163122b7 | 7677 | * This is possible from callers such as attach_tasks(), in which we |
3b03706f | 7678 | * unconditionally check_preempt_curr() after an enqueue (which may have |
5238cdd3 PT |
7679 | * lead to a throttle). This both saves work and prevents false |
7680 | * next-buddy nomination below. | |
7681 | */ | |
7682 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
7683 | return; | |
7684 | ||
2f36825b | 7685 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 7686 | set_next_buddy(pse); |
2f36825b VP |
7687 | next_buddy_marked = 1; |
7688 | } | |
57fdc26d | 7689 | |
aec0a514 BR |
7690 | /* |
7691 | * We can come here with TIF_NEED_RESCHED already set from new task | |
7692 | * wake up path. | |
5238cdd3 PT |
7693 | * |
7694 | * Note: this also catches the edge-case of curr being in a throttled | |
7695 | * group (e.g. via set_curr_task), since update_curr() (in the | |
7696 | * enqueue of curr) will have resulted in resched being set. This | |
7697 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
7698 | * below. | |
aec0a514 BR |
7699 | */ |
7700 | if (test_tsk_need_resched(curr)) | |
7701 | return; | |
7702 | ||
a2f5c9ab | 7703 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
7704 | if (unlikely(task_has_idle_policy(curr)) && |
7705 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
7706 | goto preempt; |
7707 | ||
91c234b4 | 7708 | /* |
a2f5c9ab DH |
7709 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
7710 | * is driven by the tick): | |
91c234b4 | 7711 | */ |
8ed92e51 | 7712 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 7713 | return; |
bf0f6f24 | 7714 | |
464b7527 | 7715 | find_matching_se(&se, &pse); |
09348d75 | 7716 | WARN_ON_ONCE(!pse); |
30400039 JD |
7717 | |
7718 | cse_is_idle = se_is_idle(se); | |
7719 | pse_is_idle = se_is_idle(pse); | |
7720 | ||
7721 | /* | |
7722 | * Preempt an idle group in favor of a non-idle group (and don't preempt | |
7723 | * in the inverse case). | |
7724 | */ | |
7725 | if (cse_is_idle && !pse_is_idle) | |
7726 | goto preempt; | |
7727 | if (cse_is_idle != pse_is_idle) | |
7728 | return; | |
7729 | ||
7730 | update_curr(cfs_rq_of(se)); | |
2f36825b VP |
7731 | if (wakeup_preempt_entity(se, pse) == 1) { |
7732 | /* | |
7733 | * Bias pick_next to pick the sched entity that is | |
7734 | * triggering this preemption. | |
7735 | */ | |
7736 | if (!next_buddy_marked) | |
7737 | set_next_buddy(pse); | |
3a7e73a2 | 7738 | goto preempt; |
2f36825b | 7739 | } |
464b7527 | 7740 | |
3a7e73a2 | 7741 | return; |
a65ac745 | 7742 | |
3a7e73a2 | 7743 | preempt: |
8875125e | 7744 | resched_curr(rq); |
3a7e73a2 PZ |
7745 | /* |
7746 | * Only set the backward buddy when the current task is still | |
7747 | * on the rq. This can happen when a wakeup gets interleaved | |
7748 | * with schedule on the ->pre_schedule() or idle_balance() | |
7749 | * point, either of which can * drop the rq lock. | |
7750 | * | |
7751 | * Also, during early boot the idle thread is in the fair class, | |
7752 | * for obvious reasons its a bad idea to schedule back to it. | |
7753 | */ | |
7754 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
7755 | return; | |
7756 | ||
7757 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
7758 | set_last_buddy(se); | |
bf0f6f24 IM |
7759 | } |
7760 | ||
21f56ffe PZ |
7761 | #ifdef CONFIG_SMP |
7762 | static struct task_struct *pick_task_fair(struct rq *rq) | |
7763 | { | |
7764 | struct sched_entity *se; | |
7765 | struct cfs_rq *cfs_rq; | |
7766 | ||
7767 | again: | |
7768 | cfs_rq = &rq->cfs; | |
7769 | if (!cfs_rq->nr_running) | |
7770 | return NULL; | |
7771 | ||
7772 | do { | |
7773 | struct sched_entity *curr = cfs_rq->curr; | |
7774 | ||
7775 | /* When we pick for a remote RQ, we'll not have done put_prev_entity() */ | |
7776 | if (curr) { | |
7777 | if (curr->on_rq) | |
7778 | update_curr(cfs_rq); | |
7779 | else | |
7780 | curr = NULL; | |
7781 | ||
7782 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
7783 | goto again; | |
7784 | } | |
7785 | ||
7786 | se = pick_next_entity(cfs_rq, curr); | |
7787 | cfs_rq = group_cfs_rq(se); | |
7788 | } while (cfs_rq); | |
7789 | ||
7790 | return task_of(se); | |
7791 | } | |
7792 | #endif | |
7793 | ||
5d7d6056 | 7794 | struct task_struct * |
d8ac8971 | 7795 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
7796 | { |
7797 | struct cfs_rq *cfs_rq = &rq->cfs; | |
7798 | struct sched_entity *se; | |
678d5718 | 7799 | struct task_struct *p; |
37e117c0 | 7800 | int new_tasks; |
678d5718 | 7801 | |
6e83125c | 7802 | again: |
6e2df058 | 7803 | if (!sched_fair_runnable(rq)) |
38033c37 | 7804 | goto idle; |
678d5718 | 7805 | |
9674f5ca | 7806 | #ifdef CONFIG_FAIR_GROUP_SCHED |
67692435 | 7807 | if (!prev || prev->sched_class != &fair_sched_class) |
678d5718 PZ |
7808 | goto simple; |
7809 | ||
7810 | /* | |
7811 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
7812 | * likely that a next task is from the same cgroup as the current. | |
7813 | * | |
7814 | * Therefore attempt to avoid putting and setting the entire cgroup | |
7815 | * hierarchy, only change the part that actually changes. | |
7816 | */ | |
7817 | ||
7818 | do { | |
7819 | struct sched_entity *curr = cfs_rq->curr; | |
7820 | ||
7821 | /* | |
7822 | * Since we got here without doing put_prev_entity() we also | |
7823 | * have to consider cfs_rq->curr. If it is still a runnable | |
7824 | * entity, update_curr() will update its vruntime, otherwise | |
7825 | * forget we've ever seen it. | |
7826 | */ | |
54d27365 BS |
7827 | if (curr) { |
7828 | if (curr->on_rq) | |
7829 | update_curr(cfs_rq); | |
7830 | else | |
7831 | curr = NULL; | |
678d5718 | 7832 | |
54d27365 BS |
7833 | /* |
7834 | * This call to check_cfs_rq_runtime() will do the | |
7835 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 7836 | * Therefore the nr_running test will indeed |
54d27365 BS |
7837 | * be correct. |
7838 | */ | |
9674f5ca VK |
7839 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
7840 | cfs_rq = &rq->cfs; | |
7841 | ||
7842 | if (!cfs_rq->nr_running) | |
7843 | goto idle; | |
7844 | ||
54d27365 | 7845 | goto simple; |
9674f5ca | 7846 | } |
54d27365 | 7847 | } |
678d5718 PZ |
7848 | |
7849 | se = pick_next_entity(cfs_rq, curr); | |
7850 | cfs_rq = group_cfs_rq(se); | |
7851 | } while (cfs_rq); | |
7852 | ||
7853 | p = task_of(se); | |
7854 | ||
7855 | /* | |
7856 | * Since we haven't yet done put_prev_entity and if the selected task | |
7857 | * is a different task than we started out with, try and touch the | |
7858 | * least amount of cfs_rqs. | |
7859 | */ | |
7860 | if (prev != p) { | |
7861 | struct sched_entity *pse = &prev->se; | |
7862 | ||
7863 | while (!(cfs_rq = is_same_group(se, pse))) { | |
7864 | int se_depth = se->depth; | |
7865 | int pse_depth = pse->depth; | |
7866 | ||
7867 | if (se_depth <= pse_depth) { | |
7868 | put_prev_entity(cfs_rq_of(pse), pse); | |
7869 | pse = parent_entity(pse); | |
7870 | } | |
7871 | if (se_depth >= pse_depth) { | |
7872 | set_next_entity(cfs_rq_of(se), se); | |
7873 | se = parent_entity(se); | |
7874 | } | |
7875 | } | |
7876 | ||
7877 | put_prev_entity(cfs_rq, pse); | |
7878 | set_next_entity(cfs_rq, se); | |
7879 | } | |
7880 | ||
93824900 | 7881 | goto done; |
678d5718 | 7882 | simple: |
678d5718 | 7883 | #endif |
67692435 PZ |
7884 | if (prev) |
7885 | put_prev_task(rq, prev); | |
606dba2e | 7886 | |
bf0f6f24 | 7887 | do { |
678d5718 | 7888 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 7889 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
7890 | cfs_rq = group_cfs_rq(se); |
7891 | } while (cfs_rq); | |
7892 | ||
8f4d37ec | 7893 | p = task_of(se); |
678d5718 | 7894 | |
13a453c2 | 7895 | done: __maybe_unused; |
93824900 UR |
7896 | #ifdef CONFIG_SMP |
7897 | /* | |
7898 | * Move the next running task to the front of | |
7899 | * the list, so our cfs_tasks list becomes MRU | |
7900 | * one. | |
7901 | */ | |
7902 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
7903 | #endif | |
7904 | ||
e0ee463c | 7905 | if (hrtick_enabled_fair(rq)) |
b39e66ea | 7906 | hrtick_start_fair(rq, p); |
8f4d37ec | 7907 | |
3b1baa64 MR |
7908 | update_misfit_status(p, rq); |
7909 | ||
8f4d37ec | 7910 | return p; |
38033c37 PZ |
7911 | |
7912 | idle: | |
67692435 PZ |
7913 | if (!rf) |
7914 | return NULL; | |
7915 | ||
5ba553ef | 7916 | new_tasks = newidle_balance(rq, rf); |
46f69fa3 | 7917 | |
37e117c0 | 7918 | /* |
5ba553ef | 7919 | * Because newidle_balance() releases (and re-acquires) rq->lock, it is |
37e117c0 PZ |
7920 | * possible for any higher priority task to appear. In that case we |
7921 | * must re-start the pick_next_entity() loop. | |
7922 | */ | |
e4aa358b | 7923 | if (new_tasks < 0) |
37e117c0 PZ |
7924 | return RETRY_TASK; |
7925 | ||
e4aa358b | 7926 | if (new_tasks > 0) |
38033c37 | 7927 | goto again; |
38033c37 | 7928 | |
23127296 VG |
7929 | /* |
7930 | * rq is about to be idle, check if we need to update the | |
7931 | * lost_idle_time of clock_pelt | |
7932 | */ | |
7933 | update_idle_rq_clock_pelt(rq); | |
7934 | ||
38033c37 | 7935 | return NULL; |
bf0f6f24 IM |
7936 | } |
7937 | ||
98c2f700 PZ |
7938 | static struct task_struct *__pick_next_task_fair(struct rq *rq) |
7939 | { | |
7940 | return pick_next_task_fair(rq, NULL, NULL); | |
7941 | } | |
7942 | ||
bf0f6f24 IM |
7943 | /* |
7944 | * Account for a descheduled task: | |
7945 | */ | |
6e2df058 | 7946 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
7947 | { |
7948 | struct sched_entity *se = &prev->se; | |
7949 | struct cfs_rq *cfs_rq; | |
7950 | ||
7951 | for_each_sched_entity(se) { | |
7952 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 7953 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
7954 | } |
7955 | } | |
7956 | ||
ac53db59 RR |
7957 | /* |
7958 | * sched_yield() is very simple | |
7959 | * | |
7960 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
7961 | */ | |
7962 | static void yield_task_fair(struct rq *rq) | |
7963 | { | |
7964 | struct task_struct *curr = rq->curr; | |
7965 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
7966 | struct sched_entity *se = &curr->se; | |
7967 | ||
7968 | /* | |
7969 | * Are we the only task in the tree? | |
7970 | */ | |
7971 | if (unlikely(rq->nr_running == 1)) | |
7972 | return; | |
7973 | ||
7974 | clear_buddies(cfs_rq, se); | |
7975 | ||
7976 | if (curr->policy != SCHED_BATCH) { | |
7977 | update_rq_clock(rq); | |
7978 | /* | |
7979 | * Update run-time statistics of the 'current'. | |
7980 | */ | |
7981 | update_curr(cfs_rq); | |
916671c0 MG |
7982 | /* |
7983 | * Tell update_rq_clock() that we've just updated, | |
7984 | * so we don't do microscopic update in schedule() | |
7985 | * and double the fastpath cost. | |
7986 | */ | |
adcc8da8 | 7987 | rq_clock_skip_update(rq); |
ac53db59 RR |
7988 | } |
7989 | ||
7990 | set_skip_buddy(se); | |
7991 | } | |
7992 | ||
0900acf2 | 7993 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) |
d95f4122 MG |
7994 | { |
7995 | struct sched_entity *se = &p->se; | |
7996 | ||
5238cdd3 PT |
7997 | /* throttled hierarchies are not runnable */ |
7998 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
7999 | return false; |
8000 | ||
8001 | /* Tell the scheduler that we'd really like pse to run next. */ | |
8002 | set_next_buddy(se); | |
8003 | ||
d95f4122 MG |
8004 | yield_task_fair(rq); |
8005 | ||
8006 | return true; | |
8007 | } | |
8008 | ||
681f3e68 | 8009 | #ifdef CONFIG_SMP |
bf0f6f24 | 8010 | /************************************************** |
e9c84cb8 PZ |
8011 | * Fair scheduling class load-balancing methods. |
8012 | * | |
8013 | * BASICS | |
8014 | * | |
8015 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 8016 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
8017 | * time to each task. This is expressed in the following equation: |
8018 | * | |
8019 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
8020 | * | |
97fb7a0a | 8021 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
8022 | * W_i,0 is defined as: |
8023 | * | |
8024 | * W_i,0 = \Sum_j w_i,j (2) | |
8025 | * | |
97fb7a0a | 8026 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 8027 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
8028 | * |
8029 | * The weight average is an exponential decay average of the instantaneous | |
8030 | * weight: | |
8031 | * | |
8032 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
8033 | * | |
97fb7a0a | 8034 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
8035 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
8036 | * can also include other factors [XXX]. | |
8037 | * | |
8038 | * To achieve this balance we define a measure of imbalance which follows | |
8039 | * directly from (1): | |
8040 | * | |
ced549fa | 8041 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
8042 | * |
8043 | * We them move tasks around to minimize the imbalance. In the continuous | |
8044 | * function space it is obvious this converges, in the discrete case we get | |
8045 | * a few fun cases generally called infeasible weight scenarios. | |
8046 | * | |
8047 | * [XXX expand on: | |
8048 | * - infeasible weights; | |
8049 | * - local vs global optima in the discrete case. ] | |
8050 | * | |
8051 | * | |
8052 | * SCHED DOMAINS | |
8053 | * | |
8054 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 8055 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 8056 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 8057 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 8058 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 8059 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
8060 | * the groups. |
8061 | * | |
8062 | * This yields: | |
8063 | * | |
8064 | * log_2 n 1 n | |
8065 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
8066 | * i = 0 2^i 2^i | |
8067 | * `- size of each group | |
97fb7a0a | 8068 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
8069 | * | `- freq |
8070 | * `- sum over all levels | |
8071 | * | |
8072 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
8073 | * this makes (5) the runtime complexity of the balancer. | |
8074 | * | |
8075 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 8076 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
8077 | * |
8078 | * The adjacency matrix of the resulting graph is given by: | |
8079 | * | |
97a7142f | 8080 | * log_2 n |
e9c84cb8 PZ |
8081 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
8082 | * k = 0 | |
8083 | * | |
8084 | * And you'll find that: | |
8085 | * | |
8086 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
8087 | * | |
97fb7a0a | 8088 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
8089 | * The task movement gives a factor of O(m), giving a convergence complexity |
8090 | * of: | |
8091 | * | |
8092 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
8093 | * | |
8094 | * | |
8095 | * WORK CONSERVING | |
8096 | * | |
8097 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 8098 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
8099 | * tree itself instead of relying on other CPUs to bring it work. |
8100 | * | |
8101 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
8102 | * time. | |
8103 | * | |
8104 | * [XXX more?] | |
8105 | * | |
8106 | * | |
8107 | * CGROUPS | |
8108 | * | |
8109 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
8110 | * | |
8111 | * s_k,i | |
8112 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
8113 | * S_k | |
8114 | * | |
8115 | * Where | |
8116 | * | |
8117 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
8118 | * | |
97fb7a0a | 8119 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
8120 | * |
8121 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
8122 | * property. | |
8123 | * | |
8124 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
8125 | * rewrite all of this once again.] | |
97a7142f | 8126 | */ |
bf0f6f24 | 8127 | |
ed387b78 HS |
8128 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
8129 | ||
0ec8aa00 PZ |
8130 | enum fbq_type { regular, remote, all }; |
8131 | ||
0b0695f2 | 8132 | /* |
a9723389 VG |
8133 | * 'group_type' describes the group of CPUs at the moment of load balancing. |
8134 | * | |
0b0695f2 | 8135 | * The enum is ordered by pulling priority, with the group with lowest priority |
a9723389 VG |
8136 | * first so the group_type can simply be compared when selecting the busiest |
8137 | * group. See update_sd_pick_busiest(). | |
0b0695f2 | 8138 | */ |
3b1baa64 | 8139 | enum group_type { |
a9723389 | 8140 | /* The group has spare capacity that can be used to run more tasks. */ |
0b0695f2 | 8141 | group_has_spare = 0, |
a9723389 VG |
8142 | /* |
8143 | * The group is fully used and the tasks don't compete for more CPU | |
8144 | * cycles. Nevertheless, some tasks might wait before running. | |
8145 | */ | |
0b0695f2 | 8146 | group_fully_busy, |
a9723389 | 8147 | /* |
c82a6962 VG |
8148 | * One task doesn't fit with CPU's capacity and must be migrated to a |
8149 | * more powerful CPU. | |
a9723389 | 8150 | */ |
3b1baa64 | 8151 | group_misfit_task, |
a9723389 VG |
8152 | /* |
8153 | * SD_ASYM_PACKING only: One local CPU with higher capacity is available, | |
8154 | * and the task should be migrated to it instead of running on the | |
8155 | * current CPU. | |
8156 | */ | |
0b0695f2 | 8157 | group_asym_packing, |
a9723389 VG |
8158 | /* |
8159 | * The tasks' affinity constraints previously prevented the scheduler | |
8160 | * from balancing the load across the system. | |
8161 | */ | |
3b1baa64 | 8162 | group_imbalanced, |
a9723389 VG |
8163 | /* |
8164 | * The CPU is overloaded and can't provide expected CPU cycles to all | |
8165 | * tasks. | |
8166 | */ | |
0b0695f2 VG |
8167 | group_overloaded |
8168 | }; | |
8169 | ||
8170 | enum migration_type { | |
8171 | migrate_load = 0, | |
8172 | migrate_util, | |
8173 | migrate_task, | |
8174 | migrate_misfit | |
3b1baa64 MR |
8175 | }; |
8176 | ||
ddcdf6e7 | 8177 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 8178 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
8179 | #define LBF_DST_PINNED 0x04 |
8180 | #define LBF_SOME_PINNED 0x08 | |
23fb06d9 | 8181 | #define LBF_ACTIVE_LB 0x10 |
ddcdf6e7 PZ |
8182 | |
8183 | struct lb_env { | |
8184 | struct sched_domain *sd; | |
8185 | ||
ddcdf6e7 | 8186 | struct rq *src_rq; |
85c1e7da | 8187 | int src_cpu; |
ddcdf6e7 PZ |
8188 | |
8189 | int dst_cpu; | |
8190 | struct rq *dst_rq; | |
8191 | ||
88b8dac0 SV |
8192 | struct cpumask *dst_grpmask; |
8193 | int new_dst_cpu; | |
ddcdf6e7 | 8194 | enum cpu_idle_type idle; |
bd939f45 | 8195 | long imbalance; |
b9403130 MW |
8196 | /* The set of CPUs under consideration for load-balancing */ |
8197 | struct cpumask *cpus; | |
8198 | ||
ddcdf6e7 | 8199 | unsigned int flags; |
367456c7 PZ |
8200 | |
8201 | unsigned int loop; | |
8202 | unsigned int loop_break; | |
8203 | unsigned int loop_max; | |
0ec8aa00 PZ |
8204 | |
8205 | enum fbq_type fbq_type; | |
0b0695f2 | 8206 | enum migration_type migration_type; |
163122b7 | 8207 | struct list_head tasks; |
ddcdf6e7 PZ |
8208 | }; |
8209 | ||
029632fb PZ |
8210 | /* |
8211 | * Is this task likely cache-hot: | |
8212 | */ | |
5d5e2b1b | 8213 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
8214 | { |
8215 | s64 delta; | |
8216 | ||
5cb9eaa3 | 8217 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 8218 | |
029632fb PZ |
8219 | if (p->sched_class != &fair_sched_class) |
8220 | return 0; | |
8221 | ||
1da1843f | 8222 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
8223 | return 0; |
8224 | ||
ec73240b JD |
8225 | /* SMT siblings share cache */ |
8226 | if (env->sd->flags & SD_SHARE_CPUCAPACITY) | |
8227 | return 0; | |
8228 | ||
029632fb PZ |
8229 | /* |
8230 | * Buddy candidates are cache hot: | |
8231 | */ | |
5d5e2b1b | 8232 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
8233 | (&p->se == cfs_rq_of(&p->se)->next || |
8234 | &p->se == cfs_rq_of(&p->se)->last)) | |
8235 | return 1; | |
8236 | ||
8237 | if (sysctl_sched_migration_cost == -1) | |
8238 | return 1; | |
97886d9d AL |
8239 | |
8240 | /* | |
8241 | * Don't migrate task if the task's cookie does not match | |
8242 | * with the destination CPU's core cookie. | |
8243 | */ | |
8244 | if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p)) | |
8245 | return 1; | |
8246 | ||
029632fb PZ |
8247 | if (sysctl_sched_migration_cost == 0) |
8248 | return 0; | |
8249 | ||
5d5e2b1b | 8250 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
8251 | |
8252 | return delta < (s64)sysctl_sched_migration_cost; | |
8253 | } | |
8254 | ||
3a7053b3 | 8255 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 8256 | /* |
2a1ed24c SD |
8257 | * Returns 1, if task migration degrades locality |
8258 | * Returns 0, if task migration improves locality i.e migration preferred. | |
8259 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 8260 | */ |
2a1ed24c | 8261 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 8262 | { |
b1ad065e | 8263 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
8264 | unsigned long src_weight, dst_weight; |
8265 | int src_nid, dst_nid, dist; | |
3a7053b3 | 8266 | |
2a595721 | 8267 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
8268 | return -1; |
8269 | ||
c3b9bc5b | 8270 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 8271 | return -1; |
7a0f3083 MG |
8272 | |
8273 | src_nid = cpu_to_node(env->src_cpu); | |
8274 | dst_nid = cpu_to_node(env->dst_cpu); | |
8275 | ||
83e1d2cd | 8276 | if (src_nid == dst_nid) |
2a1ed24c | 8277 | return -1; |
7a0f3083 | 8278 | |
2a1ed24c SD |
8279 | /* Migrating away from the preferred node is always bad. */ |
8280 | if (src_nid == p->numa_preferred_nid) { | |
8281 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
8282 | return 1; | |
8283 | else | |
8284 | return -1; | |
8285 | } | |
b1ad065e | 8286 | |
c1ceac62 RR |
8287 | /* Encourage migration to the preferred node. */ |
8288 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 8289 | return 0; |
b1ad065e | 8290 | |
739294fb | 8291 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 8292 | if (env->idle == CPU_IDLE) |
739294fb RR |
8293 | return -1; |
8294 | ||
f35678b6 | 8295 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 8296 | if (numa_group) { |
f35678b6 SD |
8297 | src_weight = group_weight(p, src_nid, dist); |
8298 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 8299 | } else { |
f35678b6 SD |
8300 | src_weight = task_weight(p, src_nid, dist); |
8301 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
8302 | } |
8303 | ||
f35678b6 | 8304 | return dst_weight < src_weight; |
7a0f3083 MG |
8305 | } |
8306 | ||
3a7053b3 | 8307 | #else |
2a1ed24c | 8308 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
8309 | struct lb_env *env) |
8310 | { | |
2a1ed24c | 8311 | return -1; |
7a0f3083 | 8312 | } |
3a7053b3 MG |
8313 | #endif |
8314 | ||
1e3c88bd PZ |
8315 | /* |
8316 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
8317 | */ | |
8318 | static | |
8e45cb54 | 8319 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 8320 | { |
2a1ed24c | 8321 | int tsk_cache_hot; |
e5673f28 | 8322 | |
5cb9eaa3 | 8323 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 8324 | |
1e3c88bd PZ |
8325 | /* |
8326 | * We do not migrate tasks that are: | |
d3198084 | 8327 | * 1) throttled_lb_pair, or |
3bd37062 | 8328 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
8329 | * 3) running (obviously), or |
8330 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 8331 | */ |
d3198084 JK |
8332 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
8333 | return 0; | |
8334 | ||
9bcb959d | 8335 | /* Disregard pcpu kthreads; they are where they need to be. */ |
3a7956e2 | 8336 | if (kthread_is_per_cpu(p)) |
9bcb959d LC |
8337 | return 0; |
8338 | ||
3bd37062 | 8339 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 8340 | int cpu; |
88b8dac0 | 8341 | |
ceeadb83 | 8342 | schedstat_inc(p->stats.nr_failed_migrations_affine); |
88b8dac0 | 8343 | |
6263322c PZ |
8344 | env->flags |= LBF_SOME_PINNED; |
8345 | ||
88b8dac0 | 8346 | /* |
97fb7a0a | 8347 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
8348 | * our sched_group. We may want to revisit it if we couldn't |
8349 | * meet load balance goals by pulling other tasks on src_cpu. | |
8350 | * | |
23fb06d9 VS |
8351 | * Avoid computing new_dst_cpu |
8352 | * - for NEWLY_IDLE | |
8353 | * - if we have already computed one in current iteration | |
8354 | * - if it's an active balance | |
88b8dac0 | 8355 | */ |
23fb06d9 VS |
8356 | if (env->idle == CPU_NEWLY_IDLE || |
8357 | env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) | |
88b8dac0 SV |
8358 | return 0; |
8359 | ||
97fb7a0a | 8360 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 8361 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 8362 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 8363 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
8364 | env->new_dst_cpu = cpu; |
8365 | break; | |
8366 | } | |
88b8dac0 | 8367 | } |
e02e60c1 | 8368 | |
1e3c88bd PZ |
8369 | return 0; |
8370 | } | |
88b8dac0 | 8371 | |
3b03706f | 8372 | /* Record that we found at least one task that could run on dst_cpu */ |
8e45cb54 | 8373 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 8374 | |
0b9d46fc | 8375 | if (task_on_cpu(env->src_rq, p)) { |
ceeadb83 | 8376 | schedstat_inc(p->stats.nr_failed_migrations_running); |
1e3c88bd PZ |
8377 | return 0; |
8378 | } | |
8379 | ||
8380 | /* | |
8381 | * Aggressive migration if: | |
23fb06d9 VS |
8382 | * 1) active balance |
8383 | * 2) destination numa is preferred | |
8384 | * 3) task is cache cold, or | |
8385 | * 4) too many balance attempts have failed. | |
1e3c88bd | 8386 | */ |
23fb06d9 VS |
8387 | if (env->flags & LBF_ACTIVE_LB) |
8388 | return 1; | |
8389 | ||
2a1ed24c SD |
8390 | tsk_cache_hot = migrate_degrades_locality(p, env); |
8391 | if (tsk_cache_hot == -1) | |
8392 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 8393 | |
2a1ed24c | 8394 | if (tsk_cache_hot <= 0 || |
7a96c231 | 8395 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 8396 | if (tsk_cache_hot == 1) { |
ae92882e | 8397 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
ceeadb83 | 8398 | schedstat_inc(p->stats.nr_forced_migrations); |
3a7053b3 | 8399 | } |
1e3c88bd PZ |
8400 | return 1; |
8401 | } | |
8402 | ||
ceeadb83 | 8403 | schedstat_inc(p->stats.nr_failed_migrations_hot); |
4e2dcb73 | 8404 | return 0; |
1e3c88bd PZ |
8405 | } |
8406 | ||
897c395f | 8407 | /* |
163122b7 KT |
8408 | * detach_task() -- detach the task for the migration specified in env |
8409 | */ | |
8410 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
8411 | { | |
5cb9eaa3 | 8412 | lockdep_assert_rq_held(env->src_rq); |
163122b7 | 8413 | |
5704ac0a | 8414 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
8415 | set_task_cpu(p, env->dst_cpu); |
8416 | } | |
8417 | ||
897c395f | 8418 | /* |
e5673f28 | 8419 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 8420 | * part of active balancing operations within "domain". |
897c395f | 8421 | * |
e5673f28 | 8422 | * Returns a task if successful and NULL otherwise. |
897c395f | 8423 | */ |
e5673f28 | 8424 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 8425 | { |
93824900 | 8426 | struct task_struct *p; |
897c395f | 8427 | |
5cb9eaa3 | 8428 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 8429 | |
93824900 UR |
8430 | list_for_each_entry_reverse(p, |
8431 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
8432 | if (!can_migrate_task(p, env)) |
8433 | continue; | |
897c395f | 8434 | |
163122b7 | 8435 | detach_task(p, env); |
e5673f28 | 8436 | |
367456c7 | 8437 | /* |
e5673f28 | 8438 | * Right now, this is only the second place where |
163122b7 | 8439 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 8440 | * so we can safely collect stats here rather than |
163122b7 | 8441 | * inside detach_tasks(). |
367456c7 | 8442 | */ |
ae92882e | 8443 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 8444 | return p; |
897c395f | 8445 | } |
e5673f28 | 8446 | return NULL; |
897c395f PZ |
8447 | } |
8448 | ||
5d6523eb | 8449 | /* |
0b0695f2 | 8450 | * detach_tasks() -- tries to detach up to imbalance load/util/tasks from |
163122b7 | 8451 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 8452 | * |
163122b7 | 8453 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 8454 | */ |
163122b7 | 8455 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 8456 | { |
5d6523eb | 8457 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
0b0695f2 | 8458 | unsigned long util, load; |
5d6523eb | 8459 | struct task_struct *p; |
163122b7 KT |
8460 | int detached = 0; |
8461 | ||
5cb9eaa3 | 8462 | lockdep_assert_rq_held(env->src_rq); |
1e3c88bd | 8463 | |
acb4decc AL |
8464 | /* |
8465 | * Source run queue has been emptied by another CPU, clear | |
8466 | * LBF_ALL_PINNED flag as we will not test any task. | |
8467 | */ | |
8468 | if (env->src_rq->nr_running <= 1) { | |
8469 | env->flags &= ~LBF_ALL_PINNED; | |
8470 | return 0; | |
8471 | } | |
8472 | ||
bd939f45 | 8473 | if (env->imbalance <= 0) |
5d6523eb | 8474 | return 0; |
1e3c88bd | 8475 | |
5d6523eb | 8476 | while (!list_empty(tasks)) { |
985d3a4c YD |
8477 | /* |
8478 | * We don't want to steal all, otherwise we may be treated likewise, | |
8479 | * which could at worst lead to a livelock crash. | |
8480 | */ | |
8481 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
8482 | break; | |
8483 | ||
367456c7 | 8484 | env->loop++; |
b0defa7a VG |
8485 | /* |
8486 | * We've more or less seen every task there is, call it quits | |
8487 | * unless we haven't found any movable task yet. | |
8488 | */ | |
8489 | if (env->loop > env->loop_max && | |
8490 | !(env->flags & LBF_ALL_PINNED)) | |
367456c7 | 8491 | break; |
5d6523eb PZ |
8492 | |
8493 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 8494 | if (env->loop > env->loop_break) { |
c59862f8 | 8495 | env->loop_break += SCHED_NR_MIGRATE_BREAK; |
8e45cb54 | 8496 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 8497 | break; |
a195f004 | 8498 | } |
1e3c88bd | 8499 | |
7e9518ba VG |
8500 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
8501 | ||
d3198084 | 8502 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
8503 | goto next; |
8504 | ||
0b0695f2 VG |
8505 | switch (env->migration_type) { |
8506 | case migrate_load: | |
01cfcde9 VG |
8507 | /* |
8508 | * Depending of the number of CPUs and tasks and the | |
8509 | * cgroup hierarchy, task_h_load() can return a null | |
8510 | * value. Make sure that env->imbalance decreases | |
8511 | * otherwise detach_tasks() will stop only after | |
8512 | * detaching up to loop_max tasks. | |
8513 | */ | |
8514 | load = max_t(unsigned long, task_h_load(p), 1); | |
5d6523eb | 8515 | |
0b0695f2 VG |
8516 | if (sched_feat(LB_MIN) && |
8517 | load < 16 && !env->sd->nr_balance_failed) | |
8518 | goto next; | |
367456c7 | 8519 | |
6cf82d55 VG |
8520 | /* |
8521 | * Make sure that we don't migrate too much load. | |
8522 | * Nevertheless, let relax the constraint if | |
8523 | * scheduler fails to find a good waiting task to | |
8524 | * migrate. | |
8525 | */ | |
39a2a6eb | 8526 | if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) |
0b0695f2 VG |
8527 | goto next; |
8528 | ||
8529 | env->imbalance -= load; | |
8530 | break; | |
8531 | ||
8532 | case migrate_util: | |
8533 | util = task_util_est(p); | |
8534 | ||
8535 | if (util > env->imbalance) | |
8536 | goto next; | |
8537 | ||
8538 | env->imbalance -= util; | |
8539 | break; | |
8540 | ||
8541 | case migrate_task: | |
8542 | env->imbalance--; | |
8543 | break; | |
8544 | ||
8545 | case migrate_misfit: | |
c63be7be | 8546 | /* This is not a misfit task */ |
b48e16a6 | 8547 | if (task_fits_cpu(p, env->src_cpu)) |
0b0695f2 VG |
8548 | goto next; |
8549 | ||
8550 | env->imbalance = 0; | |
8551 | break; | |
8552 | } | |
1e3c88bd | 8553 | |
163122b7 KT |
8554 | detach_task(p, env); |
8555 | list_add(&p->se.group_node, &env->tasks); | |
8556 | ||
8557 | detached++; | |
1e3c88bd | 8558 | |
c1a280b6 | 8559 | #ifdef CONFIG_PREEMPTION |
ee00e66f PZ |
8560 | /* |
8561 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 8562 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
8563 | * the critical section. |
8564 | */ | |
5d6523eb | 8565 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 8566 | break; |
1e3c88bd PZ |
8567 | #endif |
8568 | ||
ee00e66f PZ |
8569 | /* |
8570 | * We only want to steal up to the prescribed amount of | |
0b0695f2 | 8571 | * load/util/tasks. |
ee00e66f | 8572 | */ |
bd939f45 | 8573 | if (env->imbalance <= 0) |
ee00e66f | 8574 | break; |
367456c7 PZ |
8575 | |
8576 | continue; | |
8577 | next: | |
93824900 | 8578 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 8579 | } |
5d6523eb | 8580 | |
1e3c88bd | 8581 | /* |
163122b7 KT |
8582 | * Right now, this is one of only two places we collect this stat |
8583 | * so we can safely collect detach_one_task() stats here rather | |
8584 | * than inside detach_one_task(). | |
1e3c88bd | 8585 | */ |
ae92882e | 8586 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 8587 | |
163122b7 KT |
8588 | return detached; |
8589 | } | |
8590 | ||
8591 | /* | |
8592 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
8593 | */ | |
8594 | static void attach_task(struct rq *rq, struct task_struct *p) | |
8595 | { | |
5cb9eaa3 | 8596 | lockdep_assert_rq_held(rq); |
163122b7 | 8597 | |
09348d75 | 8598 | WARN_ON_ONCE(task_rq(p) != rq); |
5704ac0a | 8599 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
163122b7 KT |
8600 | check_preempt_curr(rq, p, 0); |
8601 | } | |
8602 | ||
8603 | /* | |
8604 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
8605 | * its new rq. | |
8606 | */ | |
8607 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
8608 | { | |
8a8c69c3 PZ |
8609 | struct rq_flags rf; |
8610 | ||
8611 | rq_lock(rq, &rf); | |
5704ac0a | 8612 | update_rq_clock(rq); |
163122b7 | 8613 | attach_task(rq, p); |
8a8c69c3 | 8614 | rq_unlock(rq, &rf); |
163122b7 KT |
8615 | } |
8616 | ||
8617 | /* | |
8618 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
8619 | * new rq. | |
8620 | */ | |
8621 | static void attach_tasks(struct lb_env *env) | |
8622 | { | |
8623 | struct list_head *tasks = &env->tasks; | |
8624 | struct task_struct *p; | |
8a8c69c3 | 8625 | struct rq_flags rf; |
163122b7 | 8626 | |
8a8c69c3 | 8627 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 8628 | update_rq_clock(env->dst_rq); |
163122b7 KT |
8629 | |
8630 | while (!list_empty(tasks)) { | |
8631 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
8632 | list_del_init(&p->se.group_node); | |
1e3c88bd | 8633 | |
163122b7 KT |
8634 | attach_task(env->dst_rq, p); |
8635 | } | |
8636 | ||
8a8c69c3 | 8637 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
8638 | } |
8639 | ||
b0c79224 | 8640 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
8641 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
8642 | { | |
8643 | if (cfs_rq->avg.load_avg) | |
8644 | return true; | |
8645 | ||
8646 | if (cfs_rq->avg.util_avg) | |
8647 | return true; | |
8648 | ||
8649 | return false; | |
8650 | } | |
8651 | ||
91c27493 | 8652 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
8653 | { |
8654 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
8655 | return true; | |
8656 | ||
3727e0e1 VG |
8657 | if (READ_ONCE(rq->avg_dl.util_avg)) |
8658 | return true; | |
8659 | ||
b4eccf5f TG |
8660 | if (thermal_load_avg(rq)) |
8661 | return true; | |
8662 | ||
11d4afd4 | 8663 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
8664 | if (READ_ONCE(rq->avg_irq.util_avg)) |
8665 | return true; | |
8666 | #endif | |
8667 | ||
371bf427 VG |
8668 | return false; |
8669 | } | |
8670 | ||
39b6a429 | 8671 | static inline void update_blocked_load_tick(struct rq *rq) |
b0c79224 | 8672 | { |
39b6a429 VG |
8673 | WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); |
8674 | } | |
b0c79224 | 8675 | |
39b6a429 VG |
8676 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
8677 | { | |
b0c79224 VS |
8678 | if (!has_blocked) |
8679 | rq->has_blocked_load = 0; | |
8680 | } | |
8681 | #else | |
8682 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
8683 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
39b6a429 | 8684 | static inline void update_blocked_load_tick(struct rq *rq) {} |
b0c79224 VS |
8685 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} |
8686 | #endif | |
8687 | ||
bef69dd8 VG |
8688 | static bool __update_blocked_others(struct rq *rq, bool *done) |
8689 | { | |
8690 | const struct sched_class *curr_class; | |
8691 | u64 now = rq_clock_pelt(rq); | |
b4eccf5f | 8692 | unsigned long thermal_pressure; |
bef69dd8 VG |
8693 | bool decayed; |
8694 | ||
8695 | /* | |
8696 | * update_load_avg() can call cpufreq_update_util(). Make sure that RT, | |
8697 | * DL and IRQ signals have been updated before updating CFS. | |
8698 | */ | |
8699 | curr_class = rq->curr->sched_class; | |
8700 | ||
b4eccf5f TG |
8701 | thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); |
8702 | ||
bef69dd8 VG |
8703 | decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | |
8704 | update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | | |
05289b90 | 8705 | update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | |
bef69dd8 VG |
8706 | update_irq_load_avg(rq, 0); |
8707 | ||
8708 | if (others_have_blocked(rq)) | |
8709 | *done = false; | |
8710 | ||
8711 | return decayed; | |
8712 | } | |
8713 | ||
1936c53c VG |
8714 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8715 | ||
bef69dd8 | 8716 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8717 | { |
039ae8bc | 8718 | struct cfs_rq *cfs_rq, *pos; |
bef69dd8 VG |
8719 | bool decayed = false; |
8720 | int cpu = cpu_of(rq); | |
b90f7c9d | 8721 | |
9763b67f PZ |
8722 | /* |
8723 | * Iterates the task_group tree in a bottom up fashion, see | |
8724 | * list_add_leaf_cfs_rq() for details. | |
8725 | */ | |
039ae8bc | 8726 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
8727 | struct sched_entity *se; |
8728 | ||
bef69dd8 | 8729 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { |
fe749158 | 8730 | update_tg_load_avg(cfs_rq); |
4e516076 | 8731 | |
e2f3e35f VD |
8732 | if (cfs_rq->nr_running == 0) |
8733 | update_idle_cfs_rq_clock_pelt(cfs_rq); | |
8734 | ||
bef69dd8 VG |
8735 | if (cfs_rq == &rq->cfs) |
8736 | decayed = true; | |
8737 | } | |
8738 | ||
bc427898 VG |
8739 | /* Propagate pending load changes to the parent, if any: */ |
8740 | se = cfs_rq->tg->se[cpu]; | |
8741 | if (se && !skip_blocked_update(se)) | |
02da26ad | 8742 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
a9e7f654 | 8743 | |
039ae8bc VG |
8744 | /* |
8745 | * There can be a lot of idle CPU cgroups. Don't let fully | |
8746 | * decayed cfs_rqs linger on the list. | |
8747 | */ | |
8748 | if (cfs_rq_is_decayed(cfs_rq)) | |
8749 | list_del_leaf_cfs_rq(cfs_rq); | |
8750 | ||
1936c53c VG |
8751 | /* Don't need periodic decay once load/util_avg are null */ |
8752 | if (cfs_rq_has_blocked(cfs_rq)) | |
bef69dd8 | 8753 | *done = false; |
9d89c257 | 8754 | } |
12b04875 | 8755 | |
bef69dd8 | 8756 | return decayed; |
9e3081ca PZ |
8757 | } |
8758 | ||
9763b67f | 8759 | /* |
68520796 | 8760 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
8761 | * This needs to be done in a top-down fashion because the load of a child |
8762 | * group is a fraction of its parents load. | |
8763 | */ | |
68520796 | 8764 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 8765 | { |
68520796 VD |
8766 | struct rq *rq = rq_of(cfs_rq); |
8767 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 8768 | unsigned long now = jiffies; |
68520796 | 8769 | unsigned long load; |
a35b6466 | 8770 | |
68520796 | 8771 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
8772 | return; |
8773 | ||
0e9f0245 | 8774 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
8775 | for_each_sched_entity(se) { |
8776 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 8777 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
8778 | if (cfs_rq->last_h_load_update == now) |
8779 | break; | |
8780 | } | |
a35b6466 | 8781 | |
68520796 | 8782 | if (!se) { |
7ea241af | 8783 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
8784 | cfs_rq->last_h_load_update = now; |
8785 | } | |
8786 | ||
0e9f0245 | 8787 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 8788 | load = cfs_rq->h_load; |
7ea241af YD |
8789 | load = div64_ul(load * se->avg.load_avg, |
8790 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
8791 | cfs_rq = group_cfs_rq(se); |
8792 | cfs_rq->h_load = load; | |
8793 | cfs_rq->last_h_load_update = now; | |
8794 | } | |
9763b67f PZ |
8795 | } |
8796 | ||
367456c7 | 8797 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 8798 | { |
367456c7 | 8799 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 8800 | |
68520796 | 8801 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 8802 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 8803 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
8804 | } |
8805 | #else | |
bef69dd8 | 8806 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8807 | { |
6c1d47c0 | 8808 | struct cfs_rq *cfs_rq = &rq->cfs; |
bef69dd8 | 8809 | bool decayed; |
b90f7c9d | 8810 | |
bef69dd8 VG |
8811 | decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
8812 | if (cfs_rq_has_blocked(cfs_rq)) | |
8813 | *done = false; | |
b90f7c9d | 8814 | |
bef69dd8 | 8815 | return decayed; |
9e3081ca PZ |
8816 | } |
8817 | ||
367456c7 | 8818 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 8819 | { |
9d89c257 | 8820 | return p->se.avg.load_avg; |
1e3c88bd | 8821 | } |
230059de | 8822 | #endif |
1e3c88bd | 8823 | |
bef69dd8 VG |
8824 | static void update_blocked_averages(int cpu) |
8825 | { | |
8826 | bool decayed = false, done = true; | |
8827 | struct rq *rq = cpu_rq(cpu); | |
8828 | struct rq_flags rf; | |
8829 | ||
8830 | rq_lock_irqsave(rq, &rf); | |
39b6a429 | 8831 | update_blocked_load_tick(rq); |
bef69dd8 VG |
8832 | update_rq_clock(rq); |
8833 | ||
8834 | decayed |= __update_blocked_others(rq, &done); | |
8835 | decayed |= __update_blocked_fair(rq, &done); | |
8836 | ||
8837 | update_blocked_load_status(rq, !done); | |
8838 | if (decayed) | |
8839 | cpufreq_update_util(rq, 0); | |
8840 | rq_unlock_irqrestore(rq, &rf); | |
8841 | } | |
8842 | ||
1e3c88bd | 8843 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 8844 | |
1e3c88bd PZ |
8845 | /* |
8846 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
8847 | */ | |
8848 | struct sg_lb_stats { | |
8849 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
8850 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
63b2ca30 | 8851 | unsigned long group_capacity; |
070f5e86 VG |
8852 | unsigned long group_util; /* Total utilization over the CPUs of the group */ |
8853 | unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ | |
5e23e474 | 8854 | unsigned int sum_nr_running; /* Nr of tasks running in the group */ |
a3498347 | 8855 | unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ |
147c5fc2 PZ |
8856 | unsigned int idle_cpus; |
8857 | unsigned int group_weight; | |
caeb178c | 8858 | enum group_type group_type; |
490ba971 | 8859 | unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ |
3b1baa64 | 8860 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
8861 | #ifdef CONFIG_NUMA_BALANCING |
8862 | unsigned int nr_numa_running; | |
8863 | unsigned int nr_preferred_running; | |
8864 | #endif | |
1e3c88bd PZ |
8865 | }; |
8866 | ||
56cf515b JK |
8867 | /* |
8868 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
8869 | * during load balancing. | |
8870 | */ | |
8871 | struct sd_lb_stats { | |
8872 | struct sched_group *busiest; /* Busiest group in this sd */ | |
8873 | struct sched_group *local; /* Local group in this sd */ | |
8874 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 8875 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b | 8876 | unsigned long avg_load; /* Average load across all groups in sd */ |
0b0695f2 | 8877 | unsigned int prefer_sibling; /* tasks should go to sibling first */ |
56cf515b | 8878 | |
56cf515b | 8879 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 8880 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
8881 | }; |
8882 | ||
147c5fc2 PZ |
8883 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
8884 | { | |
8885 | /* | |
8886 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
8887 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
0b0695f2 VG |
8888 | * We must however set busiest_stat::group_type and |
8889 | * busiest_stat::idle_cpus to the worst busiest group because | |
8890 | * update_sd_pick_busiest() reads these before assignment. | |
147c5fc2 PZ |
8891 | */ |
8892 | *sds = (struct sd_lb_stats){ | |
8893 | .busiest = NULL, | |
8894 | .local = NULL, | |
8895 | .total_load = 0UL, | |
63b2ca30 | 8896 | .total_capacity = 0UL, |
147c5fc2 | 8897 | .busiest_stat = { |
0b0695f2 VG |
8898 | .idle_cpus = UINT_MAX, |
8899 | .group_type = group_has_spare, | |
147c5fc2 PZ |
8900 | }, |
8901 | }; | |
8902 | } | |
8903 | ||
1ca2034e | 8904 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
8905 | { |
8906 | struct rq *rq = cpu_rq(cpu); | |
8ec59c0f | 8907 | unsigned long max = arch_scale_cpu_capacity(cpu); |
523e979d | 8908 | unsigned long used, free; |
523e979d | 8909 | unsigned long irq; |
b654f7de | 8910 | |
2e62c474 | 8911 | irq = cpu_util_irq(rq); |
cadefd3d | 8912 | |
523e979d VG |
8913 | if (unlikely(irq >= max)) |
8914 | return 1; | |
aa483808 | 8915 | |
467b7d01 TG |
8916 | /* |
8917 | * avg_rt.util_avg and avg_dl.util_avg track binary signals | |
8918 | * (running and not running) with weights 0 and 1024 respectively. | |
8919 | * avg_thermal.load_avg tracks thermal pressure and the weighted | |
8920 | * average uses the actual delta max capacity(load). | |
8921 | */ | |
523e979d VG |
8922 | used = READ_ONCE(rq->avg_rt.util_avg); |
8923 | used += READ_ONCE(rq->avg_dl.util_avg); | |
467b7d01 | 8924 | used += thermal_load_avg(rq); |
1e3c88bd | 8925 | |
523e979d VG |
8926 | if (unlikely(used >= max)) |
8927 | return 1; | |
1e3c88bd | 8928 | |
523e979d | 8929 | free = max - used; |
2e62c474 VG |
8930 | |
8931 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
8932 | } |
8933 | ||
ced549fa | 8934 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 8935 | { |
1ca2034e | 8936 | unsigned long capacity = scale_rt_capacity(cpu); |
1e3c88bd PZ |
8937 | struct sched_group *sdg = sd->groups; |
8938 | ||
b3740796 | 8939 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); |
1e3c88bd | 8940 | |
ced549fa NP |
8941 | if (!capacity) |
8942 | capacity = 1; | |
1e3c88bd | 8943 | |
b3740796 VG |
8944 | cpu_rq(cpu)->cpu_capacity = capacity; |
8945 | trace_sched_cpu_capacity_tp(cpu_rq(cpu)); | |
51cf18c9 | 8946 | |
ced549fa | 8947 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8948 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 8949 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
8950 | } |
8951 | ||
63b2ca30 | 8952 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
8953 | { |
8954 | struct sched_domain *child = sd->child; | |
8955 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 8956 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
8957 | unsigned long interval; |
8958 | ||
8959 | interval = msecs_to_jiffies(sd->balance_interval); | |
8960 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 8961 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
8962 | |
8963 | if (!child) { | |
ced549fa | 8964 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
8965 | return; |
8966 | } | |
8967 | ||
dc7ff76e | 8968 | capacity = 0; |
bf475ce0 | 8969 | min_capacity = ULONG_MAX; |
e3d6d0cb | 8970 | max_capacity = 0; |
1e3c88bd | 8971 | |
74a5ce20 PZ |
8972 | if (child->flags & SD_OVERLAP) { |
8973 | /* | |
8974 | * SD_OVERLAP domains cannot assume that child groups | |
8975 | * span the current group. | |
8976 | */ | |
8977 | ||
ae4df9d6 | 8978 | for_each_cpu(cpu, sched_group_span(sdg)) { |
4c58f57f | 8979 | unsigned long cpu_cap = capacity_of(cpu); |
863bffc8 | 8980 | |
4c58f57f PL |
8981 | capacity += cpu_cap; |
8982 | min_capacity = min(cpu_cap, min_capacity); | |
8983 | max_capacity = max(cpu_cap, max_capacity); | |
863bffc8 | 8984 | } |
74a5ce20 PZ |
8985 | } else { |
8986 | /* | |
8987 | * !SD_OVERLAP domains can assume that child groups | |
8988 | * span the current group. | |
97a7142f | 8989 | */ |
74a5ce20 PZ |
8990 | |
8991 | group = child->groups; | |
8992 | do { | |
bf475ce0 MR |
8993 | struct sched_group_capacity *sgc = group->sgc; |
8994 | ||
8995 | capacity += sgc->capacity; | |
8996 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 8997 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
8998 | group = group->next; |
8999 | } while (group != child->groups); | |
9000 | } | |
1e3c88bd | 9001 | |
63b2ca30 | 9002 | sdg->sgc->capacity = capacity; |
bf475ce0 | 9003 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 9004 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
9005 | } |
9006 | ||
9d5efe05 | 9007 | /* |
ea67821b VG |
9008 | * Check whether the capacity of the rq has been noticeably reduced by side |
9009 | * activity. The imbalance_pct is used for the threshold. | |
9010 | * Return true is the capacity is reduced | |
9d5efe05 SV |
9011 | */ |
9012 | static inline int | |
ea67821b | 9013 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 9014 | { |
ea67821b VG |
9015 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
9016 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
9017 | } |
9018 | ||
a0fe2cf0 VS |
9019 | /* |
9020 | * Check whether a rq has a misfit task and if it looks like we can actually | |
9021 | * help that task: we can migrate the task to a CPU of higher capacity, or | |
9022 | * the task's current CPU is heavily pressured. | |
9023 | */ | |
9024 | static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) | |
9025 | { | |
9026 | return rq->misfit_task_load && | |
9027 | (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || | |
9028 | check_cpu_capacity(rq, sd)); | |
9029 | } | |
9030 | ||
30ce5dab PZ |
9031 | /* |
9032 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 9033 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 9034 | * |
97fb7a0a IM |
9035 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
9036 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
9037 | * Something like: |
9038 | * | |
2b4d5b25 IM |
9039 | * { 0 1 2 3 } { 4 5 6 7 } |
9040 | * * * * * | |
30ce5dab PZ |
9041 | * |
9042 | * If we were to balance group-wise we'd place two tasks in the first group and | |
9043 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 9044 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
9045 | * |
9046 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
9047 | * by noticing the lower domain failed to reach balance and had difficulty |
9048 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
9049 | * |
9050 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 9051 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 9052 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
9053 | * to create an effective group imbalance. |
9054 | * | |
9055 | * This is a somewhat tricky proposition since the next run might not find the | |
9056 | * group imbalance and decide the groups need to be balanced again. A most | |
9057 | * subtle and fragile situation. | |
9058 | */ | |
9059 | ||
6263322c | 9060 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 9061 | { |
63b2ca30 | 9062 | return group->sgc->imbalance; |
30ce5dab PZ |
9063 | } |
9064 | ||
b37d9316 | 9065 | /* |
ea67821b VG |
9066 | * group_has_capacity returns true if the group has spare capacity that could |
9067 | * be used by some tasks. | |
fb95a5a0 | 9068 | * We consider that a group has spare capacity if the number of task is |
9e91d61d DE |
9069 | * smaller than the number of CPUs or if the utilization is lower than the |
9070 | * available capacity for CFS tasks. | |
ea67821b VG |
9071 | * For the latter, we use a threshold to stabilize the state, to take into |
9072 | * account the variance of the tasks' load and to return true if the available | |
9073 | * capacity in meaningful for the load balancer. | |
9074 | * As an example, an available capacity of 1% can appear but it doesn't make | |
9075 | * any benefit for the load balance. | |
b37d9316 | 9076 | */ |
ea67821b | 9077 | static inline bool |
57abff06 | 9078 | group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
b37d9316 | 9079 | { |
5e23e474 | 9080 | if (sgs->sum_nr_running < sgs->group_weight) |
ea67821b | 9081 | return true; |
c61037e9 | 9082 | |
070f5e86 VG |
9083 | if ((sgs->group_capacity * imbalance_pct) < |
9084 | (sgs->group_runnable * 100)) | |
9085 | return false; | |
9086 | ||
ea67821b | 9087 | if ((sgs->group_capacity * 100) > |
57abff06 | 9088 | (sgs->group_util * imbalance_pct)) |
ea67821b | 9089 | return true; |
b37d9316 | 9090 | |
ea67821b VG |
9091 | return false; |
9092 | } | |
9093 | ||
9094 | /* | |
9095 | * group_is_overloaded returns true if the group has more tasks than it can | |
9096 | * handle. | |
9097 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
9098 | * with the exact right number of tasks, has no more spare capacity but is not | |
9099 | * overloaded so both group_has_capacity and group_is_overloaded return | |
9100 | * false. | |
9101 | */ | |
9102 | static inline bool | |
57abff06 | 9103 | group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
ea67821b | 9104 | { |
5e23e474 | 9105 | if (sgs->sum_nr_running <= sgs->group_weight) |
ea67821b | 9106 | return false; |
b37d9316 | 9107 | |
ea67821b | 9108 | if ((sgs->group_capacity * 100) < |
57abff06 | 9109 | (sgs->group_util * imbalance_pct)) |
ea67821b | 9110 | return true; |
b37d9316 | 9111 | |
070f5e86 VG |
9112 | if ((sgs->group_capacity * imbalance_pct) < |
9113 | (sgs->group_runnable * 100)) | |
9114 | return true; | |
9115 | ||
ea67821b | 9116 | return false; |
b37d9316 PZ |
9117 | } |
9118 | ||
79a89f92 | 9119 | static inline enum |
57abff06 | 9120 | group_type group_classify(unsigned int imbalance_pct, |
0b0695f2 | 9121 | struct sched_group *group, |
79a89f92 | 9122 | struct sg_lb_stats *sgs) |
caeb178c | 9123 | { |
57abff06 | 9124 | if (group_is_overloaded(imbalance_pct, sgs)) |
caeb178c RR |
9125 | return group_overloaded; |
9126 | ||
9127 | if (sg_imbalanced(group)) | |
9128 | return group_imbalanced; | |
9129 | ||
0b0695f2 VG |
9130 | if (sgs->group_asym_packing) |
9131 | return group_asym_packing; | |
9132 | ||
3b1baa64 MR |
9133 | if (sgs->group_misfit_task_load) |
9134 | return group_misfit_task; | |
9135 | ||
57abff06 | 9136 | if (!group_has_capacity(imbalance_pct, sgs)) |
0b0695f2 VG |
9137 | return group_fully_busy; |
9138 | ||
9139 | return group_has_spare; | |
caeb178c RR |
9140 | } |
9141 | ||
4006a72b RN |
9142 | /** |
9143 | * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks | |
9144 | * @dst_cpu: Destination CPU of the load balancing | |
9145 | * @sds: Load-balancing data with statistics of the local group | |
9146 | * @sgs: Load-balancing statistics of the candidate busiest group | |
9147 | * @sg: The candidate busiest group | |
9148 | * | |
9149 | * Check the state of the SMT siblings of both @sds::local and @sg and decide | |
9150 | * if @dst_cpu can pull tasks. | |
9151 | * | |
9152 | * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of | |
9153 | * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks | |
9154 | * only if @dst_cpu has higher priority. | |
9155 | * | |
9156 | * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more | |
9157 | * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority. | |
9158 | * Bigger imbalances in the number of busy CPUs will be dealt with in | |
9159 | * update_sd_pick_busiest(). | |
9160 | * | |
9161 | * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings | |
9162 | * of @dst_cpu are idle and @sg has lower priority. | |
a315da5e RD |
9163 | * |
9164 | * Return: true if @dst_cpu can pull tasks, false otherwise. | |
4006a72b RN |
9165 | */ |
9166 | static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds, | |
9167 | struct sg_lb_stats *sgs, | |
9168 | struct sched_group *sg) | |
9169 | { | |
9170 | #ifdef CONFIG_SCHED_SMT | |
9171 | bool local_is_smt, sg_is_smt; | |
9172 | int sg_busy_cpus; | |
9173 | ||
9174 | local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY; | |
9175 | sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY; | |
9176 | ||
9177 | sg_busy_cpus = sgs->group_weight - sgs->idle_cpus; | |
9178 | ||
9179 | if (!local_is_smt) { | |
9180 | /* | |
9181 | * If we are here, @dst_cpu is idle and does not have SMT | |
9182 | * siblings. Pull tasks if candidate group has two or more | |
9183 | * busy CPUs. | |
9184 | */ | |
9185 | if (sg_busy_cpus >= 2) /* implies sg_is_smt */ | |
9186 | return true; | |
9187 | ||
9188 | /* | |
9189 | * @dst_cpu does not have SMT siblings. @sg may have SMT | |
9190 | * siblings and only one is busy. In such case, @dst_cpu | |
9191 | * can help if it has higher priority and is idle (i.e., | |
9192 | * it has no running tasks). | |
9193 | */ | |
9194 | return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu); | |
9195 | } | |
9196 | ||
9197 | /* @dst_cpu has SMT siblings. */ | |
9198 | ||
9199 | if (sg_is_smt) { | |
9200 | int local_busy_cpus = sds->local->group_weight - | |
9201 | sds->local_stat.idle_cpus; | |
9202 | int busy_cpus_delta = sg_busy_cpus - local_busy_cpus; | |
9203 | ||
9204 | if (busy_cpus_delta == 1) | |
9205 | return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu); | |
9206 | ||
9207 | return false; | |
9208 | } | |
9209 | ||
9210 | /* | |
9211 | * @sg does not have SMT siblings. Ensure that @sds::local does not end | |
9212 | * up with more than one busy SMT sibling and only pull tasks if there | |
9213 | * are not busy CPUs (i.e., no CPU has running tasks). | |
9214 | */ | |
9215 | if (!sds->local_stat.sum_nr_running) | |
9216 | return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu); | |
9217 | ||
9218 | return false; | |
9219 | #else | |
9220 | /* Always return false so that callers deal with non-SMT cases. */ | |
9221 | return false; | |
9222 | #endif | |
9223 | } | |
9224 | ||
aafc917a RN |
9225 | static inline bool |
9226 | sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs, | |
9227 | struct sched_group *group) | |
9228 | { | |
4006a72b RN |
9229 | /* Only do SMT checks if either local or candidate have SMT siblings */ |
9230 | if ((sds->local->flags & SD_SHARE_CPUCAPACITY) || | |
9231 | (group->flags & SD_SHARE_CPUCAPACITY)) | |
9232 | return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group); | |
9233 | ||
aafc917a RN |
9234 | return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu); |
9235 | } | |
9236 | ||
c82a6962 VG |
9237 | static inline bool |
9238 | sched_reduced_capacity(struct rq *rq, struct sched_domain *sd) | |
9239 | { | |
9240 | /* | |
9241 | * When there is more than 1 task, the group_overloaded case already | |
9242 | * takes care of cpu with reduced capacity | |
9243 | */ | |
9244 | if (rq->cfs.h_nr_running != 1) | |
9245 | return false; | |
9246 | ||
9247 | return check_cpu_capacity(rq, sd); | |
9248 | } | |
9249 | ||
1e3c88bd PZ |
9250 | /** |
9251 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 9252 | * @env: The load balancing environment. |
a315da5e | 9253 | * @sds: Load-balancing data with statistics of the local group. |
1e3c88bd | 9254 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 9255 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 9256 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 9257 | */ |
bd939f45 | 9258 | static inline void update_sg_lb_stats(struct lb_env *env, |
c0d14b57 | 9259 | struct sd_lb_stats *sds, |
630246a0 QP |
9260 | struct sched_group *group, |
9261 | struct sg_lb_stats *sgs, | |
9262 | int *sg_status) | |
1e3c88bd | 9263 | { |
0b0695f2 | 9264 | int i, nr_running, local_group; |
1e3c88bd | 9265 | |
b72ff13c PZ |
9266 | memset(sgs, 0, sizeof(*sgs)); |
9267 | ||
c0d14b57 | 9268 | local_group = group == sds->local; |
0b0695f2 | 9269 | |
ae4df9d6 | 9270 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd | 9271 | struct rq *rq = cpu_rq(i); |
c82a6962 | 9272 | unsigned long load = cpu_load(rq); |
1e3c88bd | 9273 | |
c82a6962 | 9274 | sgs->group_load += load; |
82762d2a | 9275 | sgs->group_util += cpu_util_cfs(i); |
070f5e86 | 9276 | sgs->group_runnable += cpu_runnable(rq); |
a3498347 | 9277 | sgs->sum_h_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 9278 | |
a426f99c | 9279 | nr_running = rq->nr_running; |
5e23e474 VG |
9280 | sgs->sum_nr_running += nr_running; |
9281 | ||
a426f99c | 9282 | if (nr_running > 1) |
630246a0 | 9283 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 9284 | |
2802bf3c MR |
9285 | if (cpu_overutilized(i)) |
9286 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 9287 | |
0ec8aa00 PZ |
9288 | #ifdef CONFIG_NUMA_BALANCING |
9289 | sgs->nr_numa_running += rq->nr_numa_running; | |
9290 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
9291 | #endif | |
a426f99c WL |
9292 | /* |
9293 | * No need to call idle_cpu() if nr_running is not 0 | |
9294 | */ | |
0b0695f2 | 9295 | if (!nr_running && idle_cpu(i)) { |
aae6d3dd | 9296 | sgs->idle_cpus++; |
0b0695f2 VG |
9297 | /* Idle cpu can't have misfit task */ |
9298 | continue; | |
9299 | } | |
9300 | ||
9301 | if (local_group) | |
9302 | continue; | |
3b1baa64 | 9303 | |
c82a6962 VG |
9304 | if (env->sd->flags & SD_ASYM_CPUCAPACITY) { |
9305 | /* Check for a misfit task on the cpu */ | |
9306 | if (sgs->group_misfit_task_load < rq->misfit_task_load) { | |
9307 | sgs->group_misfit_task_load = rq->misfit_task_load; | |
9308 | *sg_status |= SG_OVERLOAD; | |
9309 | } | |
9310 | } else if ((env->idle != CPU_NOT_IDLE) && | |
9311 | sched_reduced_capacity(rq, env->sd)) { | |
9312 | /* Check for a task running on a CPU with reduced capacity */ | |
9313 | if (sgs->group_misfit_task_load < load) | |
9314 | sgs->group_misfit_task_load = load; | |
757ffdd7 | 9315 | } |
1e3c88bd PZ |
9316 | } |
9317 | ||
aafc917a RN |
9318 | sgs->group_capacity = group->sgc->capacity; |
9319 | ||
9320 | sgs->group_weight = group->group_weight; | |
9321 | ||
0b0695f2 | 9322 | /* Check if dst CPU is idle and preferred to this group */ |
60256435 | 9323 | if (!local_group && env->sd->flags & SD_ASYM_PACKING && |
aafc917a RN |
9324 | env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running && |
9325 | sched_asym(env, sds, sgs, group)) { | |
0b0695f2 VG |
9326 | sgs->group_asym_packing = 1; |
9327 | } | |
9328 | ||
57abff06 | 9329 | sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); |
0b0695f2 VG |
9330 | |
9331 | /* Computing avg_load makes sense only when group is overloaded */ | |
9332 | if (sgs->group_type == group_overloaded) | |
9333 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / | |
9334 | sgs->group_capacity; | |
1e3c88bd PZ |
9335 | } |
9336 | ||
532cb4c4 MN |
9337 | /** |
9338 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 9339 | * @env: The load balancing environment. |
532cb4c4 MN |
9340 | * @sds: sched_domain statistics |
9341 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 9342 | * @sgs: sched_group statistics |
532cb4c4 MN |
9343 | * |
9344 | * Determine if @sg is a busier group than the previously selected | |
9345 | * busiest group. | |
e69f6186 YB |
9346 | * |
9347 | * Return: %true if @sg is a busier group than the previously selected | |
9348 | * busiest group. %false otherwise. | |
532cb4c4 | 9349 | */ |
bd939f45 | 9350 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
9351 | struct sd_lb_stats *sds, |
9352 | struct sched_group *sg, | |
bd939f45 | 9353 | struct sg_lb_stats *sgs) |
532cb4c4 | 9354 | { |
caeb178c | 9355 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 9356 | |
0b0695f2 VG |
9357 | /* Make sure that there is at least one task to pull */ |
9358 | if (!sgs->sum_h_nr_running) | |
9359 | return false; | |
9360 | ||
cad68e55 MR |
9361 | /* |
9362 | * Don't try to pull misfit tasks we can't help. | |
9363 | * We can use max_capacity here as reduction in capacity on some | |
9364 | * CPUs in the group should either be possible to resolve | |
9365 | * internally or be covered by avg_load imbalance (eventually). | |
9366 | */ | |
c82a6962 VG |
9367 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && |
9368 | (sgs->group_type == group_misfit_task) && | |
4aed8aa4 | 9369 | (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || |
0b0695f2 | 9370 | sds->local_stat.group_type != group_has_spare)) |
cad68e55 MR |
9371 | return false; |
9372 | ||
caeb178c | 9373 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
9374 | return true; |
9375 | ||
caeb178c RR |
9376 | if (sgs->group_type < busiest->group_type) |
9377 | return false; | |
9378 | ||
9e0994c0 | 9379 | /* |
0b0695f2 VG |
9380 | * The candidate and the current busiest group are the same type of |
9381 | * group. Let check which one is the busiest according to the type. | |
9e0994c0 | 9382 | */ |
9e0994c0 | 9383 | |
0b0695f2 VG |
9384 | switch (sgs->group_type) { |
9385 | case group_overloaded: | |
9386 | /* Select the overloaded group with highest avg_load. */ | |
9387 | if (sgs->avg_load <= busiest->avg_load) | |
9388 | return false; | |
9389 | break; | |
9390 | ||
9391 | case group_imbalanced: | |
9392 | /* | |
9393 | * Select the 1st imbalanced group as we don't have any way to | |
9394 | * choose one more than another. | |
9395 | */ | |
9e0994c0 MR |
9396 | return false; |
9397 | ||
0b0695f2 VG |
9398 | case group_asym_packing: |
9399 | /* Prefer to move from lowest priority CPU's work */ | |
9400 | if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) | |
9401 | return false; | |
9402 | break; | |
532cb4c4 | 9403 | |
0b0695f2 VG |
9404 | case group_misfit_task: |
9405 | /* | |
9406 | * If we have more than one misfit sg go with the biggest | |
9407 | * misfit. | |
9408 | */ | |
9409 | if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
9410 | return false; | |
9411 | break; | |
532cb4c4 | 9412 | |
0b0695f2 VG |
9413 | case group_fully_busy: |
9414 | /* | |
9415 | * Select the fully busy group with highest avg_load. In | |
9416 | * theory, there is no need to pull task from such kind of | |
9417 | * group because tasks have all compute capacity that they need | |
9418 | * but we can still improve the overall throughput by reducing | |
9419 | * contention when accessing shared HW resources. | |
9420 | * | |
9421 | * XXX for now avg_load is not computed and always 0 so we | |
9422 | * select the 1st one. | |
9423 | */ | |
9424 | if (sgs->avg_load <= busiest->avg_load) | |
9425 | return false; | |
9426 | break; | |
9427 | ||
9428 | case group_has_spare: | |
9429 | /* | |
5f68eb19 VG |
9430 | * Select not overloaded group with lowest number of idle cpus |
9431 | * and highest number of running tasks. We could also compare | |
9432 | * the spare capacity which is more stable but it can end up | |
9433 | * that the group has less spare capacity but finally more idle | |
0b0695f2 VG |
9434 | * CPUs which means less opportunity to pull tasks. |
9435 | */ | |
5f68eb19 | 9436 | if (sgs->idle_cpus > busiest->idle_cpus) |
0b0695f2 | 9437 | return false; |
5f68eb19 VG |
9438 | else if ((sgs->idle_cpus == busiest->idle_cpus) && |
9439 | (sgs->sum_nr_running <= busiest->sum_nr_running)) | |
9440 | return false; | |
9441 | ||
0b0695f2 | 9442 | break; |
532cb4c4 MN |
9443 | } |
9444 | ||
0b0695f2 VG |
9445 | /* |
9446 | * Candidate sg has no more than one task per CPU and has higher | |
9447 | * per-CPU capacity. Migrating tasks to less capable CPUs may harm | |
9448 | * throughput. Maximize throughput, power/energy consequences are not | |
9449 | * considered. | |
9450 | */ | |
9451 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && | |
9452 | (sgs->group_type <= group_fully_busy) && | |
4aed8aa4 | 9453 | (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) |
0b0695f2 VG |
9454 | return false; |
9455 | ||
9456 | return true; | |
532cb4c4 MN |
9457 | } |
9458 | ||
0ec8aa00 PZ |
9459 | #ifdef CONFIG_NUMA_BALANCING |
9460 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
9461 | { | |
a3498347 | 9462 | if (sgs->sum_h_nr_running > sgs->nr_numa_running) |
0ec8aa00 | 9463 | return regular; |
a3498347 | 9464 | if (sgs->sum_h_nr_running > sgs->nr_preferred_running) |
0ec8aa00 PZ |
9465 | return remote; |
9466 | return all; | |
9467 | } | |
9468 | ||
9469 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
9470 | { | |
9471 | if (rq->nr_running > rq->nr_numa_running) | |
9472 | return regular; | |
9473 | if (rq->nr_running > rq->nr_preferred_running) | |
9474 | return remote; | |
9475 | return all; | |
9476 | } | |
9477 | #else | |
9478 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
9479 | { | |
9480 | return all; | |
9481 | } | |
9482 | ||
9483 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
9484 | { | |
9485 | return regular; | |
9486 | } | |
9487 | #endif /* CONFIG_NUMA_BALANCING */ | |
9488 | ||
57abff06 VG |
9489 | |
9490 | struct sg_lb_stats; | |
9491 | ||
3318544b VG |
9492 | /* |
9493 | * task_running_on_cpu - return 1 if @p is running on @cpu. | |
9494 | */ | |
9495 | ||
9496 | static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) | |
9497 | { | |
9498 | /* Task has no contribution or is new */ | |
9499 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
9500 | return 0; | |
9501 | ||
9502 | if (task_on_rq_queued(p)) | |
9503 | return 1; | |
9504 | ||
9505 | return 0; | |
9506 | } | |
9507 | ||
9508 | /** | |
9509 | * idle_cpu_without - would a given CPU be idle without p ? | |
9510 | * @cpu: the processor on which idleness is tested. | |
9511 | * @p: task which should be ignored. | |
9512 | * | |
9513 | * Return: 1 if the CPU would be idle. 0 otherwise. | |
9514 | */ | |
9515 | static int idle_cpu_without(int cpu, struct task_struct *p) | |
9516 | { | |
9517 | struct rq *rq = cpu_rq(cpu); | |
9518 | ||
9519 | if (rq->curr != rq->idle && rq->curr != p) | |
9520 | return 0; | |
9521 | ||
9522 | /* | |
9523 | * rq->nr_running can't be used but an updated version without the | |
9524 | * impact of p on cpu must be used instead. The updated nr_running | |
9525 | * be computed and tested before calling idle_cpu_without(). | |
9526 | */ | |
9527 | ||
9528 | #ifdef CONFIG_SMP | |
126c2092 | 9529 | if (rq->ttwu_pending) |
3318544b VG |
9530 | return 0; |
9531 | #endif | |
9532 | ||
9533 | return 1; | |
9534 | } | |
9535 | ||
57abff06 VG |
9536 | /* |
9537 | * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. | |
3318544b | 9538 | * @sd: The sched_domain level to look for idlest group. |
57abff06 VG |
9539 | * @group: sched_group whose statistics are to be updated. |
9540 | * @sgs: variable to hold the statistics for this group. | |
3318544b | 9541 | * @p: The task for which we look for the idlest group/CPU. |
57abff06 VG |
9542 | */ |
9543 | static inline void update_sg_wakeup_stats(struct sched_domain *sd, | |
9544 | struct sched_group *group, | |
9545 | struct sg_lb_stats *sgs, | |
9546 | struct task_struct *p) | |
9547 | { | |
9548 | int i, nr_running; | |
9549 | ||
9550 | memset(sgs, 0, sizeof(*sgs)); | |
9551 | ||
b48e16a6 QY |
9552 | /* Assume that task can't fit any CPU of the group */ |
9553 | if (sd->flags & SD_ASYM_CPUCAPACITY) | |
9554 | sgs->group_misfit_task_load = 1; | |
9555 | ||
57abff06 VG |
9556 | for_each_cpu(i, sched_group_span(group)) { |
9557 | struct rq *rq = cpu_rq(i); | |
3318544b | 9558 | unsigned int local; |
57abff06 | 9559 | |
3318544b | 9560 | sgs->group_load += cpu_load_without(rq, p); |
57abff06 | 9561 | sgs->group_util += cpu_util_without(i, p); |
070f5e86 | 9562 | sgs->group_runnable += cpu_runnable_without(rq, p); |
3318544b VG |
9563 | local = task_running_on_cpu(i, p); |
9564 | sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; | |
57abff06 | 9565 | |
3318544b | 9566 | nr_running = rq->nr_running - local; |
57abff06 VG |
9567 | sgs->sum_nr_running += nr_running; |
9568 | ||
9569 | /* | |
3318544b | 9570 | * No need to call idle_cpu_without() if nr_running is not 0 |
57abff06 | 9571 | */ |
3318544b | 9572 | if (!nr_running && idle_cpu_without(i, p)) |
57abff06 VG |
9573 | sgs->idle_cpus++; |
9574 | ||
b48e16a6 QY |
9575 | /* Check if task fits in the CPU */ |
9576 | if (sd->flags & SD_ASYM_CPUCAPACITY && | |
9577 | sgs->group_misfit_task_load && | |
9578 | task_fits_cpu(p, i)) | |
9579 | sgs->group_misfit_task_load = 0; | |
57abff06 | 9580 | |
57abff06 VG |
9581 | } |
9582 | ||
9583 | sgs->group_capacity = group->sgc->capacity; | |
9584 | ||
289de359 VG |
9585 | sgs->group_weight = group->group_weight; |
9586 | ||
57abff06 VG |
9587 | sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); |
9588 | ||
9589 | /* | |
9590 | * Computing avg_load makes sense only when group is fully busy or | |
9591 | * overloaded | |
9592 | */ | |
6c8116c9 TZ |
9593 | if (sgs->group_type == group_fully_busy || |
9594 | sgs->group_type == group_overloaded) | |
57abff06 VG |
9595 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / |
9596 | sgs->group_capacity; | |
9597 | } | |
9598 | ||
9599 | static bool update_pick_idlest(struct sched_group *idlest, | |
9600 | struct sg_lb_stats *idlest_sgs, | |
9601 | struct sched_group *group, | |
9602 | struct sg_lb_stats *sgs) | |
9603 | { | |
9604 | if (sgs->group_type < idlest_sgs->group_type) | |
9605 | return true; | |
9606 | ||
9607 | if (sgs->group_type > idlest_sgs->group_type) | |
9608 | return false; | |
9609 | ||
9610 | /* | |
9611 | * The candidate and the current idlest group are the same type of | |
9612 | * group. Let check which one is the idlest according to the type. | |
9613 | */ | |
9614 | ||
9615 | switch (sgs->group_type) { | |
9616 | case group_overloaded: | |
9617 | case group_fully_busy: | |
9618 | /* Select the group with lowest avg_load. */ | |
9619 | if (idlest_sgs->avg_load <= sgs->avg_load) | |
9620 | return false; | |
9621 | break; | |
9622 | ||
9623 | case group_imbalanced: | |
9624 | case group_asym_packing: | |
9625 | /* Those types are not used in the slow wakeup path */ | |
9626 | return false; | |
9627 | ||
9628 | case group_misfit_task: | |
9629 | /* Select group with the highest max capacity */ | |
9630 | if (idlest->sgc->max_capacity >= group->sgc->max_capacity) | |
9631 | return false; | |
9632 | break; | |
9633 | ||
9634 | case group_has_spare: | |
9635 | /* Select group with most idle CPUs */ | |
3edecfef | 9636 | if (idlest_sgs->idle_cpus > sgs->idle_cpus) |
57abff06 | 9637 | return false; |
3edecfef PP |
9638 | |
9639 | /* Select group with lowest group_util */ | |
9640 | if (idlest_sgs->idle_cpus == sgs->idle_cpus && | |
9641 | idlest_sgs->group_util <= sgs->group_util) | |
9642 | return false; | |
9643 | ||
57abff06 VG |
9644 | break; |
9645 | } | |
9646 | ||
9647 | return true; | |
9648 | } | |
9649 | ||
9650 | /* | |
9651 | * find_idlest_group() finds and returns the least busy CPU group within the | |
9652 | * domain. | |
9653 | * | |
9654 | * Assumes p is allowed on at least one CPU in sd. | |
9655 | */ | |
9656 | static struct sched_group * | |
45da2773 | 9657 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
57abff06 VG |
9658 | { |
9659 | struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; | |
9660 | struct sg_lb_stats local_sgs, tmp_sgs; | |
9661 | struct sg_lb_stats *sgs; | |
9662 | unsigned long imbalance; | |
9663 | struct sg_lb_stats idlest_sgs = { | |
9664 | .avg_load = UINT_MAX, | |
9665 | .group_type = group_overloaded, | |
9666 | }; | |
9667 | ||
57abff06 VG |
9668 | do { |
9669 | int local_group; | |
9670 | ||
9671 | /* Skip over this group if it has no CPUs allowed */ | |
9672 | if (!cpumask_intersects(sched_group_span(group), | |
9673 | p->cpus_ptr)) | |
9674 | continue; | |
9675 | ||
97886d9d AL |
9676 | /* Skip over this group if no cookie matched */ |
9677 | if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group)) | |
9678 | continue; | |
9679 | ||
57abff06 VG |
9680 | local_group = cpumask_test_cpu(this_cpu, |
9681 | sched_group_span(group)); | |
9682 | ||
9683 | if (local_group) { | |
9684 | sgs = &local_sgs; | |
9685 | local = group; | |
9686 | } else { | |
9687 | sgs = &tmp_sgs; | |
9688 | } | |
9689 | ||
9690 | update_sg_wakeup_stats(sd, group, sgs, p); | |
9691 | ||
9692 | if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { | |
9693 | idlest = group; | |
9694 | idlest_sgs = *sgs; | |
9695 | } | |
9696 | ||
9697 | } while (group = group->next, group != sd->groups); | |
9698 | ||
9699 | ||
9700 | /* There is no idlest group to push tasks to */ | |
9701 | if (!idlest) | |
9702 | return NULL; | |
9703 | ||
7ed735c3 VG |
9704 | /* The local group has been skipped because of CPU affinity */ |
9705 | if (!local) | |
9706 | return idlest; | |
9707 | ||
57abff06 VG |
9708 | /* |
9709 | * If the local group is idler than the selected idlest group | |
9710 | * don't try and push the task. | |
9711 | */ | |
9712 | if (local_sgs.group_type < idlest_sgs.group_type) | |
9713 | return NULL; | |
9714 | ||
9715 | /* | |
9716 | * If the local group is busier than the selected idlest group | |
9717 | * try and push the task. | |
9718 | */ | |
9719 | if (local_sgs.group_type > idlest_sgs.group_type) | |
9720 | return idlest; | |
9721 | ||
9722 | switch (local_sgs.group_type) { | |
9723 | case group_overloaded: | |
9724 | case group_fully_busy: | |
5c339005 MG |
9725 | |
9726 | /* Calculate allowed imbalance based on load */ | |
9727 | imbalance = scale_load_down(NICE_0_LOAD) * | |
9728 | (sd->imbalance_pct-100) / 100; | |
9729 | ||
57abff06 VG |
9730 | /* |
9731 | * When comparing groups across NUMA domains, it's possible for | |
9732 | * the local domain to be very lightly loaded relative to the | |
9733 | * remote domains but "imbalance" skews the comparison making | |
9734 | * remote CPUs look much more favourable. When considering | |
9735 | * cross-domain, add imbalance to the load on the remote node | |
9736 | * and consider staying local. | |
9737 | */ | |
9738 | ||
9739 | if ((sd->flags & SD_NUMA) && | |
9740 | ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) | |
9741 | return NULL; | |
9742 | ||
9743 | /* | |
9744 | * If the local group is less loaded than the selected | |
9745 | * idlest group don't try and push any tasks. | |
9746 | */ | |
9747 | if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) | |
9748 | return NULL; | |
9749 | ||
9750 | if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) | |
9751 | return NULL; | |
9752 | break; | |
9753 | ||
9754 | case group_imbalanced: | |
9755 | case group_asym_packing: | |
9756 | /* Those type are not used in the slow wakeup path */ | |
9757 | return NULL; | |
9758 | ||
9759 | case group_misfit_task: | |
9760 | /* Select group with the highest max capacity */ | |
9761 | if (local->sgc->max_capacity >= idlest->sgc->max_capacity) | |
9762 | return NULL; | |
9763 | break; | |
9764 | ||
9765 | case group_has_spare: | |
cb29a5c1 | 9766 | #ifdef CONFIG_NUMA |
57abff06 | 9767 | if (sd->flags & SD_NUMA) { |
f5b2eeb4 | 9768 | int imb_numa_nr = sd->imb_numa_nr; |
57abff06 VG |
9769 | #ifdef CONFIG_NUMA_BALANCING |
9770 | int idlest_cpu; | |
9771 | /* | |
9772 | * If there is spare capacity at NUMA, try to select | |
9773 | * the preferred node | |
9774 | */ | |
9775 | if (cpu_to_node(this_cpu) == p->numa_preferred_nid) | |
9776 | return NULL; | |
9777 | ||
9778 | idlest_cpu = cpumask_first(sched_group_span(idlest)); | |
9779 | if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) | |
9780 | return idlest; | |
cb29a5c1 | 9781 | #endif /* CONFIG_NUMA_BALANCING */ |
57abff06 | 9782 | /* |
2cfb7a1b MG |
9783 | * Otherwise, keep the task close to the wakeup source |
9784 | * and improve locality if the number of running tasks | |
9785 | * would remain below threshold where an imbalance is | |
f5b2eeb4 PN |
9786 | * allowed while accounting for the possibility the |
9787 | * task is pinned to a subset of CPUs. If there is a | |
9788 | * real need of migration, periodic load balance will | |
9789 | * take care of it. | |
57abff06 | 9790 | */ |
f5b2eeb4 | 9791 | if (p->nr_cpus_allowed != NR_CPUS) { |
ec4fc801 | 9792 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
f5b2eeb4 PN |
9793 | |
9794 | cpumask_and(cpus, sched_group_span(local), p->cpus_ptr); | |
9795 | imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr); | |
9796 | } | |
9797 | ||
cb29a5c1 MG |
9798 | imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus); |
9799 | if (!adjust_numa_imbalance(imbalance, | |
9800 | local_sgs.sum_nr_running + 1, | |
f5b2eeb4 | 9801 | imb_numa_nr)) { |
57abff06 | 9802 | return NULL; |
cb29a5c1 | 9803 | } |
57abff06 | 9804 | } |
cb29a5c1 | 9805 | #endif /* CONFIG_NUMA */ |
57abff06 VG |
9806 | |
9807 | /* | |
9808 | * Select group with highest number of idle CPUs. We could also | |
9809 | * compare the utilization which is more stable but it can end | |
9810 | * up that the group has less spare capacity but finally more | |
9811 | * idle CPUs which means more opportunity to run task. | |
9812 | */ | |
9813 | if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) | |
9814 | return NULL; | |
9815 | break; | |
9816 | } | |
9817 | ||
9818 | return idlest; | |
9819 | } | |
9820 | ||
70fb5ccf CY |
9821 | static void update_idle_cpu_scan(struct lb_env *env, |
9822 | unsigned long sum_util) | |
9823 | { | |
9824 | struct sched_domain_shared *sd_share; | |
9825 | int llc_weight, pct; | |
9826 | u64 x, y, tmp; | |
9827 | /* | |
9828 | * Update the number of CPUs to scan in LLC domain, which could | |
9829 | * be used as a hint in select_idle_cpu(). The update of sd_share | |
9830 | * could be expensive because it is within a shared cache line. | |
9831 | * So the write of this hint only occurs during periodic load | |
9832 | * balancing, rather than CPU_NEWLY_IDLE, because the latter | |
9833 | * can fire way more frequently than the former. | |
9834 | */ | |
9835 | if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE) | |
9836 | return; | |
9837 | ||
9838 | llc_weight = per_cpu(sd_llc_size, env->dst_cpu); | |
9839 | if (env->sd->span_weight != llc_weight) | |
9840 | return; | |
9841 | ||
9842 | sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu)); | |
9843 | if (!sd_share) | |
9844 | return; | |
9845 | ||
9846 | /* | |
9847 | * The number of CPUs to search drops as sum_util increases, when | |
9848 | * sum_util hits 85% or above, the scan stops. | |
9849 | * The reason to choose 85% as the threshold is because this is the | |
9850 | * imbalance_pct(117) when a LLC sched group is overloaded. | |
9851 | * | |
9852 | * let y = SCHED_CAPACITY_SCALE - p * x^2 [1] | |
9853 | * and y'= y / SCHED_CAPACITY_SCALE | |
9854 | * | |
9855 | * x is the ratio of sum_util compared to the CPU capacity: | |
9856 | * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE) | |
9857 | * y' is the ratio of CPUs to be scanned in the LLC domain, | |
9858 | * and the number of CPUs to scan is calculated by: | |
9859 | * | |
9860 | * nr_scan = llc_weight * y' [2] | |
9861 | * | |
9862 | * When x hits the threshold of overloaded, AKA, when | |
9863 | * x = 100 / pct, y drops to 0. According to [1], | |
9864 | * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000 | |
9865 | * | |
9866 | * Scale x by SCHED_CAPACITY_SCALE: | |
9867 | * x' = sum_util / llc_weight; [3] | |
9868 | * | |
9869 | * and finally [1] becomes: | |
9870 | * y = SCHED_CAPACITY_SCALE - | |
9871 | * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4] | |
9872 | * | |
9873 | */ | |
9874 | /* equation [3] */ | |
9875 | x = sum_util; | |
9876 | do_div(x, llc_weight); | |
9877 | ||
9878 | /* equation [4] */ | |
9879 | pct = env->sd->imbalance_pct; | |
9880 | tmp = x * x * pct * pct; | |
9881 | do_div(tmp, 10000 * SCHED_CAPACITY_SCALE); | |
9882 | tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE); | |
9883 | y = SCHED_CAPACITY_SCALE - tmp; | |
9884 | ||
9885 | /* equation [2] */ | |
9886 | y *= llc_weight; | |
9887 | do_div(y, SCHED_CAPACITY_SCALE); | |
9888 | if ((int)y != sd_share->nr_idle_scan) | |
9889 | WRITE_ONCE(sd_share->nr_idle_scan, (int)y); | |
9890 | } | |
9891 | ||
1e3c88bd | 9892 | /** |
461819ac | 9893 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 9894 | * @env: The load balancing environment. |
1e3c88bd PZ |
9895 | * @sds: variable to hold the statistics for this sched_domain. |
9896 | */ | |
0b0695f2 | 9897 | |
0ec8aa00 | 9898 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9899 | { |
bd939f45 PZ |
9900 | struct sched_domain *child = env->sd->child; |
9901 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 9902 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 9903 | struct sg_lb_stats tmp_sgs; |
70fb5ccf | 9904 | unsigned long sum_util = 0; |
630246a0 | 9905 | int sg_status = 0; |
1e3c88bd | 9906 | |
1e3c88bd | 9907 | do { |
56cf515b | 9908 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
9909 | int local_group; |
9910 | ||
ae4df9d6 | 9911 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
9912 | if (local_group) { |
9913 | sds->local = sg; | |
05b40e05 | 9914 | sgs = local; |
b72ff13c PZ |
9915 | |
9916 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
9917 | time_after_eq(jiffies, sg->sgc->next_update)) |
9918 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 9919 | } |
1e3c88bd | 9920 | |
c0d14b57 | 9921 | update_sg_lb_stats(env, sds, sg, sgs, &sg_status); |
1e3c88bd | 9922 | |
b72ff13c PZ |
9923 | if (local_group) |
9924 | goto next_group; | |
9925 | ||
1e3c88bd | 9926 | |
b72ff13c | 9927 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 9928 | sds->busiest = sg; |
56cf515b | 9929 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
9930 | } |
9931 | ||
b72ff13c PZ |
9932 | next_group: |
9933 | /* Now, start updating sd_lb_stats */ | |
9934 | sds->total_load += sgs->group_load; | |
63b2ca30 | 9935 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 9936 | |
70fb5ccf | 9937 | sum_util += sgs->group_util; |
532cb4c4 | 9938 | sg = sg->next; |
bd939f45 | 9939 | } while (sg != env->sd->groups); |
0ec8aa00 | 9940 | |
0b0695f2 VG |
9941 | /* Tag domain that child domain prefers tasks go to siblings first */ |
9942 | sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; | |
9943 | ||
f643ea22 | 9944 | |
0ec8aa00 PZ |
9945 | if (env->sd->flags & SD_NUMA) |
9946 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
9947 | |
9948 | if (!env->sd->parent) { | |
2802bf3c MR |
9949 | struct root_domain *rd = env->dst_rq->rd; |
9950 | ||
4486edd1 | 9951 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
9952 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
9953 | ||
9954 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
9955 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
f9f240f9 | 9956 | trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); |
2802bf3c | 9957 | } else if (sg_status & SG_OVERUTILIZED) { |
f9f240f9 QY |
9958 | struct root_domain *rd = env->dst_rq->rd; |
9959 | ||
9960 | WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); | |
9961 | trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); | |
4486edd1 | 9962 | } |
70fb5ccf CY |
9963 | |
9964 | update_idle_cpu_scan(env, sum_util); | |
532cb4c4 MN |
9965 | } |
9966 | ||
1e3c88bd PZ |
9967 | /** |
9968 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
9969 | * groups of a given sched_domain during load balance. | |
bd939f45 | 9970 | * @env: load balance environment |
1e3c88bd | 9971 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 9972 | */ |
bd939f45 | 9973 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9974 | { |
56cf515b JK |
9975 | struct sg_lb_stats *local, *busiest; |
9976 | ||
9977 | local = &sds->local_stat; | |
56cf515b | 9978 | busiest = &sds->busiest_stat; |
dd5feea1 | 9979 | |
0b0695f2 | 9980 | if (busiest->group_type == group_misfit_task) { |
c82a6962 VG |
9981 | if (env->sd->flags & SD_ASYM_CPUCAPACITY) { |
9982 | /* Set imbalance to allow misfit tasks to be balanced. */ | |
9983 | env->migration_type = migrate_misfit; | |
9984 | env->imbalance = 1; | |
9985 | } else { | |
9986 | /* | |
9987 | * Set load imbalance to allow moving task from cpu | |
9988 | * with reduced capacity. | |
9989 | */ | |
9990 | env->migration_type = migrate_load; | |
9991 | env->imbalance = busiest->group_misfit_task_load; | |
9992 | } | |
0b0695f2 VG |
9993 | return; |
9994 | } | |
9995 | ||
9996 | if (busiest->group_type == group_asym_packing) { | |
9997 | /* | |
9998 | * In case of asym capacity, we will try to migrate all load to | |
9999 | * the preferred CPU. | |
10000 | */ | |
10001 | env->migration_type = migrate_task; | |
10002 | env->imbalance = busiest->sum_h_nr_running; | |
10003 | return; | |
10004 | } | |
10005 | ||
10006 | if (busiest->group_type == group_imbalanced) { | |
10007 | /* | |
10008 | * In the group_imb case we cannot rely on group-wide averages | |
10009 | * to ensure CPU-load equilibrium, try to move any task to fix | |
10010 | * the imbalance. The next load balance will take care of | |
10011 | * balancing back the system. | |
10012 | */ | |
10013 | env->migration_type = migrate_task; | |
10014 | env->imbalance = 1; | |
490ba971 VG |
10015 | return; |
10016 | } | |
10017 | ||
1e3c88bd | 10018 | /* |
0b0695f2 | 10019 | * Try to use spare capacity of local group without overloading it or |
a9723389 | 10020 | * emptying busiest. |
1e3c88bd | 10021 | */ |
0b0695f2 | 10022 | if (local->group_type == group_has_spare) { |
16b0a7a1 VG |
10023 | if ((busiest->group_type > group_fully_busy) && |
10024 | !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { | |
0b0695f2 VG |
10025 | /* |
10026 | * If busiest is overloaded, try to fill spare | |
10027 | * capacity. This might end up creating spare capacity | |
10028 | * in busiest or busiest still being overloaded but | |
10029 | * there is no simple way to directly compute the | |
10030 | * amount of load to migrate in order to balance the | |
10031 | * system. | |
10032 | */ | |
10033 | env->migration_type = migrate_util; | |
10034 | env->imbalance = max(local->group_capacity, local->group_util) - | |
10035 | local->group_util; | |
10036 | ||
10037 | /* | |
10038 | * In some cases, the group's utilization is max or even | |
10039 | * higher than capacity because of migrations but the | |
10040 | * local CPU is (newly) idle. There is at least one | |
10041 | * waiting task in this overloaded busiest group. Let's | |
10042 | * try to pull it. | |
10043 | */ | |
10044 | if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { | |
10045 | env->migration_type = migrate_task; | |
10046 | env->imbalance = 1; | |
10047 | } | |
10048 | ||
10049 | return; | |
10050 | } | |
10051 | ||
10052 | if (busiest->group_weight == 1 || sds->prefer_sibling) { | |
5e23e474 | 10053 | unsigned int nr_diff = busiest->sum_nr_running; |
0b0695f2 VG |
10054 | /* |
10055 | * When prefer sibling, evenly spread running tasks on | |
10056 | * groups. | |
10057 | */ | |
10058 | env->migration_type = migrate_task; | |
5e23e474 | 10059 | lsub_positive(&nr_diff, local->sum_nr_running); |
cb29a5c1 | 10060 | env->imbalance = nr_diff; |
b396f523 | 10061 | } else { |
0b0695f2 | 10062 | |
b396f523 MG |
10063 | /* |
10064 | * If there is no overload, we just want to even the number of | |
10065 | * idle cpus. | |
10066 | */ | |
10067 | env->migration_type = migrate_task; | |
cb29a5c1 MG |
10068 | env->imbalance = max_t(long, 0, |
10069 | (local->idle_cpus - busiest->idle_cpus)); | |
b396f523 MG |
10070 | } |
10071 | ||
cb29a5c1 | 10072 | #ifdef CONFIG_NUMA |
b396f523 | 10073 | /* Consider allowing a small imbalance between NUMA groups */ |
7d2b5dd0 | 10074 | if (env->sd->flags & SD_NUMA) { |
fb86f5b2 | 10075 | env->imbalance = adjust_numa_imbalance(env->imbalance, |
cb29a5c1 MG |
10076 | local->sum_nr_running + 1, |
10077 | env->sd->imb_numa_nr); | |
7d2b5dd0 | 10078 | } |
cb29a5c1 MG |
10079 | #endif |
10080 | ||
10081 | /* Number of tasks to move to restore balance */ | |
10082 | env->imbalance >>= 1; | |
b396f523 | 10083 | |
fcf0553d | 10084 | return; |
1e3c88bd PZ |
10085 | } |
10086 | ||
9a5d9ba6 | 10087 | /* |
0b0695f2 VG |
10088 | * Local is fully busy but has to take more load to relieve the |
10089 | * busiest group | |
9a5d9ba6 | 10090 | */ |
0b0695f2 VG |
10091 | if (local->group_type < group_overloaded) { |
10092 | /* | |
10093 | * Local will become overloaded so the avg_load metrics are | |
10094 | * finally needed. | |
10095 | */ | |
10096 | ||
10097 | local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / | |
10098 | local->group_capacity; | |
10099 | ||
111688ca AL |
10100 | /* |
10101 | * If the local group is more loaded than the selected | |
10102 | * busiest group don't try to pull any tasks. | |
10103 | */ | |
10104 | if (local->avg_load >= busiest->avg_load) { | |
10105 | env->imbalance = 0; | |
10106 | return; | |
10107 | } | |
06354900 | 10108 | |
10109 | sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / | |
10110 | sds->total_capacity; | |
b04099e7 VG |
10111 | |
10112 | /* | |
10113 | * If the local group is more loaded than the average system | |
10114 | * load, don't try to pull any tasks. | |
10115 | */ | |
10116 | if (local->avg_load >= sds->avg_load) { | |
10117 | env->imbalance = 0; | |
10118 | return; | |
10119 | } | |
10120 | ||
dd5feea1 SS |
10121 | } |
10122 | ||
10123 | /* | |
0b0695f2 VG |
10124 | * Both group are or will become overloaded and we're trying to get all |
10125 | * the CPUs to the average_load, so we don't want to push ourselves | |
10126 | * above the average load, nor do we wish to reduce the max loaded CPU | |
10127 | * below the average load. At the same time, we also don't want to | |
10128 | * reduce the group load below the group capacity. Thus we look for | |
10129 | * the minimum possible imbalance. | |
dd5feea1 | 10130 | */ |
0b0695f2 | 10131 | env->migration_type = migrate_load; |
56cf515b | 10132 | env->imbalance = min( |
0b0695f2 | 10133 | (busiest->avg_load - sds->avg_load) * busiest->group_capacity, |
63b2ca30 | 10134 | (sds->avg_load - local->avg_load) * local->group_capacity |
ca8ce3d0 | 10135 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 10136 | } |
fab47622 | 10137 | |
1e3c88bd PZ |
10138 | /******* find_busiest_group() helpers end here *********************/ |
10139 | ||
0b0695f2 VG |
10140 | /* |
10141 | * Decision matrix according to the local and busiest group type: | |
10142 | * | |
10143 | * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded | |
10144 | * has_spare nr_idle balanced N/A N/A balanced balanced | |
10145 | * fully_busy nr_idle nr_idle N/A N/A balanced balanced | |
a6583531 | 10146 | * misfit_task force N/A N/A N/A N/A N/A |
0b0695f2 VG |
10147 | * asym_packing force force N/A N/A force force |
10148 | * imbalanced force force N/A N/A force force | |
10149 | * overloaded force force N/A N/A force avg_load | |
10150 | * | |
10151 | * N/A : Not Applicable because already filtered while updating | |
10152 | * statistics. | |
10153 | * balanced : The system is balanced for these 2 groups. | |
10154 | * force : Calculate the imbalance as load migration is probably needed. | |
10155 | * avg_load : Only if imbalance is significant enough. | |
10156 | * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite | |
10157 | * different in groups. | |
10158 | */ | |
10159 | ||
1e3c88bd PZ |
10160 | /** |
10161 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 10162 | * if there is an imbalance. |
a315da5e | 10163 | * @env: The load balancing environment. |
1e3c88bd | 10164 | * |
a3df0679 | 10165 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
10166 | * to restore balance. |
10167 | * | |
e69f6186 | 10168 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 10169 | */ |
56cf515b | 10170 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 10171 | { |
56cf515b | 10172 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
10173 | struct sd_lb_stats sds; |
10174 | ||
147c5fc2 | 10175 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
10176 | |
10177 | /* | |
b0fb1eb4 | 10178 | * Compute the various statistics relevant for load balancing at |
1e3c88bd PZ |
10179 | * this level. |
10180 | */ | |
23f0d209 | 10181 | update_sd_lb_stats(env, &sds); |
2802bf3c | 10182 | |
cc57aa8f | 10183 | /* There is no busy sibling group to pull tasks from */ |
0b0695f2 | 10184 | if (!sds.busiest) |
1e3c88bd PZ |
10185 | goto out_balanced; |
10186 | ||
b40e128f VG |
10187 | busiest = &sds.busiest_stat; |
10188 | ||
0b0695f2 VG |
10189 | /* Misfit tasks should be dealt with regardless of the avg load */ |
10190 | if (busiest->group_type == group_misfit_task) | |
10191 | goto force_balance; | |
10192 | ||
b40e128f VG |
10193 | if (sched_energy_enabled()) { |
10194 | struct root_domain *rd = env->dst_rq->rd; | |
10195 | ||
10196 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
10197 | goto out_balanced; | |
10198 | } | |
10199 | ||
0b0695f2 VG |
10200 | /* ASYM feature bypasses nice load balance check */ |
10201 | if (busiest->group_type == group_asym_packing) | |
10202 | goto force_balance; | |
b0432d8f | 10203 | |
866ab43e PZ |
10204 | /* |
10205 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 10206 | * work because they assume all things are equal, which typically |
3bd37062 | 10207 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 10208 | */ |
caeb178c | 10209 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
10210 | goto force_balance; |
10211 | ||
b40e128f | 10212 | local = &sds.local_stat; |
cc57aa8f | 10213 | /* |
9c58c79a | 10214 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
10215 | * don't try and pull any tasks. |
10216 | */ | |
0b0695f2 | 10217 | if (local->group_type > busiest->group_type) |
1e3c88bd PZ |
10218 | goto out_balanced; |
10219 | ||
cc57aa8f | 10220 | /* |
0b0695f2 VG |
10221 | * When groups are overloaded, use the avg_load to ensure fairness |
10222 | * between tasks. | |
cc57aa8f | 10223 | */ |
0b0695f2 VG |
10224 | if (local->group_type == group_overloaded) { |
10225 | /* | |
10226 | * If the local group is more loaded than the selected | |
10227 | * busiest group don't try to pull any tasks. | |
10228 | */ | |
10229 | if (local->avg_load >= busiest->avg_load) | |
10230 | goto out_balanced; | |
10231 | ||
10232 | /* XXX broken for overlapping NUMA groups */ | |
10233 | sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / | |
10234 | sds.total_capacity; | |
1e3c88bd | 10235 | |
aae6d3dd | 10236 | /* |
0b0695f2 VG |
10237 | * Don't pull any tasks if this group is already above the |
10238 | * domain average load. | |
aae6d3dd | 10239 | */ |
0b0695f2 | 10240 | if (local->avg_load >= sds.avg_load) |
aae6d3dd | 10241 | goto out_balanced; |
0b0695f2 | 10242 | |
c186fafe | 10243 | /* |
0b0695f2 VG |
10244 | * If the busiest group is more loaded, use imbalance_pct to be |
10245 | * conservative. | |
c186fafe | 10246 | */ |
56cf515b JK |
10247 | if (100 * busiest->avg_load <= |
10248 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 10249 | goto out_balanced; |
aae6d3dd | 10250 | } |
1e3c88bd | 10251 | |
0b0695f2 VG |
10252 | /* Try to move all excess tasks to child's sibling domain */ |
10253 | if (sds.prefer_sibling && local->group_type == group_has_spare && | |
5e23e474 | 10254 | busiest->sum_nr_running > local->sum_nr_running + 1) |
0b0695f2 VG |
10255 | goto force_balance; |
10256 | ||
2ab4092f VG |
10257 | if (busiest->group_type != group_overloaded) { |
10258 | if (env->idle == CPU_NOT_IDLE) | |
10259 | /* | |
10260 | * If the busiest group is not overloaded (and as a | |
10261 | * result the local one too) but this CPU is already | |
10262 | * busy, let another idle CPU try to pull task. | |
10263 | */ | |
10264 | goto out_balanced; | |
10265 | ||
10266 | if (busiest->group_weight > 1 && | |
10267 | local->idle_cpus <= (busiest->idle_cpus + 1)) | |
10268 | /* | |
10269 | * If the busiest group is not overloaded | |
10270 | * and there is no imbalance between this and busiest | |
10271 | * group wrt idle CPUs, it is balanced. The imbalance | |
10272 | * becomes significant if the diff is greater than 1 | |
10273 | * otherwise we might end up to just move the imbalance | |
10274 | * on another group. Of course this applies only if | |
10275 | * there is more than 1 CPU per group. | |
10276 | */ | |
10277 | goto out_balanced; | |
10278 | ||
10279 | if (busiest->sum_h_nr_running == 1) | |
10280 | /* | |
10281 | * busiest doesn't have any tasks waiting to run | |
10282 | */ | |
10283 | goto out_balanced; | |
10284 | } | |
0b0695f2 | 10285 | |
fab47622 | 10286 | force_balance: |
1e3c88bd | 10287 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 10288 | calculate_imbalance(env, &sds); |
bb3485c8 | 10289 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
10290 | |
10291 | out_balanced: | |
bd939f45 | 10292 | env->imbalance = 0; |
1e3c88bd PZ |
10293 | return NULL; |
10294 | } | |
10295 | ||
10296 | /* | |
97fb7a0a | 10297 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 10298 | */ |
bd939f45 | 10299 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 10300 | struct sched_group *group) |
1e3c88bd PZ |
10301 | { |
10302 | struct rq *busiest = NULL, *rq; | |
0b0695f2 VG |
10303 | unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; |
10304 | unsigned int busiest_nr = 0; | |
1e3c88bd PZ |
10305 | int i; |
10306 | ||
ae4df9d6 | 10307 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
0b0695f2 VG |
10308 | unsigned long capacity, load, util; |
10309 | unsigned int nr_running; | |
0ec8aa00 PZ |
10310 | enum fbq_type rt; |
10311 | ||
10312 | rq = cpu_rq(i); | |
10313 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 10314 | |
0ec8aa00 PZ |
10315 | /* |
10316 | * We classify groups/runqueues into three groups: | |
10317 | * - regular: there are !numa tasks | |
10318 | * - remote: there are numa tasks that run on the 'wrong' node | |
10319 | * - all: there is no distinction | |
10320 | * | |
10321 | * In order to avoid migrating ideally placed numa tasks, | |
10322 | * ignore those when there's better options. | |
10323 | * | |
10324 | * If we ignore the actual busiest queue to migrate another | |
10325 | * task, the next balance pass can still reduce the busiest | |
10326 | * queue by moving tasks around inside the node. | |
10327 | * | |
10328 | * If we cannot move enough load due to this classification | |
10329 | * the next pass will adjust the group classification and | |
10330 | * allow migration of more tasks. | |
10331 | * | |
10332 | * Both cases only affect the total convergence complexity. | |
10333 | */ | |
10334 | if (rt > env->fbq_type) | |
10335 | continue; | |
10336 | ||
0b0695f2 | 10337 | nr_running = rq->cfs.h_nr_running; |
fc488ffd VG |
10338 | if (!nr_running) |
10339 | continue; | |
10340 | ||
10341 | capacity = capacity_of(i); | |
9d5efe05 | 10342 | |
4ad3831a CR |
10343 | /* |
10344 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
10345 | * eventually lead to active_balancing high->low capacity. | |
10346 | * Higher per-CPU capacity is considered better than balancing | |
10347 | * average load. | |
10348 | */ | |
10349 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
4aed8aa4 | 10350 | !capacity_greater(capacity_of(env->dst_cpu), capacity) && |
0b0695f2 | 10351 | nr_running == 1) |
4ad3831a CR |
10352 | continue; |
10353 | ||
4006a72b RN |
10354 | /* Make sure we only pull tasks from a CPU of lower priority */ |
10355 | if ((env->sd->flags & SD_ASYM_PACKING) && | |
10356 | sched_asym_prefer(i, env->dst_cpu) && | |
10357 | nr_running == 1) | |
10358 | continue; | |
10359 | ||
0b0695f2 VG |
10360 | switch (env->migration_type) { |
10361 | case migrate_load: | |
10362 | /* | |
b0fb1eb4 VG |
10363 | * When comparing with load imbalance, use cpu_load() |
10364 | * which is not scaled with the CPU capacity. | |
0b0695f2 | 10365 | */ |
b0fb1eb4 | 10366 | load = cpu_load(rq); |
1e3c88bd | 10367 | |
0b0695f2 VG |
10368 | if (nr_running == 1 && load > env->imbalance && |
10369 | !check_cpu_capacity(rq, env->sd)) | |
10370 | break; | |
ea67821b | 10371 | |
0b0695f2 VG |
10372 | /* |
10373 | * For the load comparisons with the other CPUs, | |
b0fb1eb4 VG |
10374 | * consider the cpu_load() scaled with the CPU |
10375 | * capacity, so that the load can be moved away | |
10376 | * from the CPU that is potentially running at a | |
10377 | * lower capacity. | |
0b0695f2 VG |
10378 | * |
10379 | * Thus we're looking for max(load_i / capacity_i), | |
10380 | * crosswise multiplication to rid ourselves of the | |
10381 | * division works out to: | |
10382 | * load_i * capacity_j > load_j * capacity_i; | |
10383 | * where j is our previous maximum. | |
10384 | */ | |
10385 | if (load * busiest_capacity > busiest_load * capacity) { | |
10386 | busiest_load = load; | |
10387 | busiest_capacity = capacity; | |
10388 | busiest = rq; | |
10389 | } | |
10390 | break; | |
10391 | ||
10392 | case migrate_util: | |
82762d2a | 10393 | util = cpu_util_cfs(i); |
0b0695f2 | 10394 | |
c32b4308 VG |
10395 | /* |
10396 | * Don't try to pull utilization from a CPU with one | |
10397 | * running task. Whatever its utilization, we will fail | |
10398 | * detach the task. | |
10399 | */ | |
10400 | if (nr_running <= 1) | |
10401 | continue; | |
10402 | ||
0b0695f2 VG |
10403 | if (busiest_util < util) { |
10404 | busiest_util = util; | |
10405 | busiest = rq; | |
10406 | } | |
10407 | break; | |
10408 | ||
10409 | case migrate_task: | |
10410 | if (busiest_nr < nr_running) { | |
10411 | busiest_nr = nr_running; | |
10412 | busiest = rq; | |
10413 | } | |
10414 | break; | |
10415 | ||
10416 | case migrate_misfit: | |
10417 | /* | |
10418 | * For ASYM_CPUCAPACITY domains with misfit tasks we | |
10419 | * simply seek the "biggest" misfit task. | |
10420 | */ | |
10421 | if (rq->misfit_task_load > busiest_load) { | |
10422 | busiest_load = rq->misfit_task_load; | |
10423 | busiest = rq; | |
10424 | } | |
10425 | ||
10426 | break; | |
1e3c88bd | 10427 | |
1e3c88bd PZ |
10428 | } |
10429 | } | |
10430 | ||
10431 | return busiest; | |
10432 | } | |
10433 | ||
10434 | /* | |
10435 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
10436 | * so long as it is large enough. | |
10437 | */ | |
10438 | #define MAX_PINNED_INTERVAL 512 | |
10439 | ||
46a745d9 VG |
10440 | static inline bool |
10441 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 10442 | { |
46a745d9 VG |
10443 | /* |
10444 | * ASYM_PACKING needs to force migrate tasks from busy but | |
10445 | * lower priority CPUs in order to pack all tasks in the | |
10446 | * highest priority CPUs. | |
10447 | */ | |
10448 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
10449 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
10450 | } | |
bd939f45 | 10451 | |
46a745d9 | 10452 | static inline bool |
e9b9734b VG |
10453 | imbalanced_active_balance(struct lb_env *env) |
10454 | { | |
10455 | struct sched_domain *sd = env->sd; | |
10456 | ||
10457 | /* | |
10458 | * The imbalanced case includes the case of pinned tasks preventing a fair | |
10459 | * distribution of the load on the system but also the even distribution of the | |
10460 | * threads on a system with spare capacity | |
10461 | */ | |
10462 | if ((env->migration_type == migrate_task) && | |
10463 | (sd->nr_balance_failed > sd->cache_nice_tries+2)) | |
10464 | return 1; | |
10465 | ||
10466 | return 0; | |
10467 | } | |
10468 | ||
10469 | static int need_active_balance(struct lb_env *env) | |
46a745d9 VG |
10470 | { |
10471 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 10472 | |
46a745d9 VG |
10473 | if (asym_active_balance(env)) |
10474 | return 1; | |
1af3ed3d | 10475 | |
e9b9734b VG |
10476 | if (imbalanced_active_balance(env)) |
10477 | return 1; | |
10478 | ||
1aaf90a4 VG |
10479 | /* |
10480 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
10481 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
10482 | * because of other sched_class or IRQs if more capacity stays | |
10483 | * available on dst_cpu. | |
10484 | */ | |
10485 | if ((env->idle != CPU_NOT_IDLE) && | |
10486 | (env->src_rq->cfs.h_nr_running == 1)) { | |
10487 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
10488 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
10489 | return 1; | |
10490 | } | |
10491 | ||
0b0695f2 | 10492 | if (env->migration_type == migrate_misfit) |
cad68e55 MR |
10493 | return 1; |
10494 | ||
46a745d9 VG |
10495 | return 0; |
10496 | } | |
10497 | ||
969c7921 TH |
10498 | static int active_load_balance_cpu_stop(void *data); |
10499 | ||
23f0d209 JK |
10500 | static int should_we_balance(struct lb_env *env) |
10501 | { | |
10502 | struct sched_group *sg = env->sd->groups; | |
64297f2b | 10503 | int cpu; |
23f0d209 | 10504 | |
024c9d2f PZ |
10505 | /* |
10506 | * Ensure the balancing environment is consistent; can happen | |
10507 | * when the softirq triggers 'during' hotplug. | |
10508 | */ | |
10509 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
10510 | return 0; | |
10511 | ||
23f0d209 | 10512 | /* |
97fb7a0a | 10513 | * In the newly idle case, we will allow all the CPUs |
23f0d209 | 10514 | * to do the newly idle load balance. |
792b9f65 JD |
10515 | * |
10516 | * However, we bail out if we already have tasks or a wakeup pending, | |
10517 | * to optimize wakeup latency. | |
23f0d209 | 10518 | */ |
792b9f65 JD |
10519 | if (env->idle == CPU_NEWLY_IDLE) { |
10520 | if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending) | |
10521 | return 0; | |
23f0d209 | 10522 | return 1; |
792b9f65 | 10523 | } |
23f0d209 | 10524 | |
97fb7a0a | 10525 | /* Try to find first idle CPU */ |
e5c14b1f | 10526 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 10527 | if (!idle_cpu(cpu)) |
23f0d209 JK |
10528 | continue; |
10529 | ||
64297f2b PW |
10530 | /* Are we the first idle CPU? */ |
10531 | return cpu == env->dst_cpu; | |
23f0d209 JK |
10532 | } |
10533 | ||
64297f2b PW |
10534 | /* Are we the first CPU of this group ? */ |
10535 | return group_balance_cpu(sg) == env->dst_cpu; | |
23f0d209 JK |
10536 | } |
10537 | ||
1e3c88bd PZ |
10538 | /* |
10539 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
10540 | * tasks if there is an imbalance. | |
10541 | */ | |
10542 | static int load_balance(int this_cpu, struct rq *this_rq, | |
10543 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 10544 | int *continue_balancing) |
1e3c88bd | 10545 | { |
88b8dac0 | 10546 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 10547 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 10548 | struct sched_group *group; |
1e3c88bd | 10549 | struct rq *busiest; |
8a8c69c3 | 10550 | struct rq_flags rf; |
4ba29684 | 10551 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
8e45cb54 PZ |
10552 | struct lb_env env = { |
10553 | .sd = sd, | |
ddcdf6e7 PZ |
10554 | .dst_cpu = this_cpu, |
10555 | .dst_rq = this_rq, | |
178dad9e | 10556 | .dst_grpmask = group_balance_mask(sd->groups), |
8e45cb54 | 10557 | .idle = idle, |
c59862f8 | 10558 | .loop_break = SCHED_NR_MIGRATE_BREAK, |
b9403130 | 10559 | .cpus = cpus, |
0ec8aa00 | 10560 | .fbq_type = all, |
163122b7 | 10561 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
10562 | }; |
10563 | ||
65a4433a | 10564 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 10565 | |
ae92882e | 10566 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
10567 | |
10568 | redo: | |
23f0d209 JK |
10569 | if (!should_we_balance(&env)) { |
10570 | *continue_balancing = 0; | |
1e3c88bd | 10571 | goto out_balanced; |
23f0d209 | 10572 | } |
1e3c88bd | 10573 | |
23f0d209 | 10574 | group = find_busiest_group(&env); |
1e3c88bd | 10575 | if (!group) { |
ae92882e | 10576 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
10577 | goto out_balanced; |
10578 | } | |
10579 | ||
b9403130 | 10580 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 10581 | if (!busiest) { |
ae92882e | 10582 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
10583 | goto out_balanced; |
10584 | } | |
10585 | ||
09348d75 | 10586 | WARN_ON_ONCE(busiest == env.dst_rq); |
1e3c88bd | 10587 | |
ae92882e | 10588 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 10589 | |
1aaf90a4 VG |
10590 | env.src_cpu = busiest->cpu; |
10591 | env.src_rq = busiest; | |
10592 | ||
1e3c88bd | 10593 | ld_moved = 0; |
8a41dfcd VG |
10594 | /* Clear this flag as soon as we find a pullable task */ |
10595 | env.flags |= LBF_ALL_PINNED; | |
1e3c88bd PZ |
10596 | if (busiest->nr_running > 1) { |
10597 | /* | |
10598 | * Attempt to move tasks. If find_busiest_group has found | |
10599 | * an imbalance but busiest->nr_running <= 1, the group is | |
10600 | * still unbalanced. ld_moved simply stays zero, so it is | |
10601 | * correctly treated as an imbalance. | |
10602 | */ | |
c82513e5 | 10603 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 10604 | |
5d6523eb | 10605 | more_balance: |
8a8c69c3 | 10606 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 10607 | update_rq_clock(busiest); |
88b8dac0 SV |
10608 | |
10609 | /* | |
10610 | * cur_ld_moved - load moved in current iteration | |
10611 | * ld_moved - cumulative load moved across iterations | |
10612 | */ | |
163122b7 | 10613 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
10614 | |
10615 | /* | |
163122b7 KT |
10616 | * We've detached some tasks from busiest_rq. Every |
10617 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
10618 | * unlock busiest->lock, and we are able to be sure | |
10619 | * that nobody can manipulate the tasks in parallel. | |
10620 | * See task_rq_lock() family for the details. | |
1e3c88bd | 10621 | */ |
163122b7 | 10622 | |
8a8c69c3 | 10623 | rq_unlock(busiest, &rf); |
163122b7 KT |
10624 | |
10625 | if (cur_ld_moved) { | |
10626 | attach_tasks(&env); | |
10627 | ld_moved += cur_ld_moved; | |
10628 | } | |
10629 | ||
8a8c69c3 | 10630 | local_irq_restore(rf.flags); |
88b8dac0 | 10631 | |
f1cd0858 JK |
10632 | if (env.flags & LBF_NEED_BREAK) { |
10633 | env.flags &= ~LBF_NEED_BREAK; | |
b0defa7a VG |
10634 | /* Stop if we tried all running tasks */ |
10635 | if (env.loop < busiest->nr_running) | |
10636 | goto more_balance; | |
f1cd0858 JK |
10637 | } |
10638 | ||
88b8dac0 SV |
10639 | /* |
10640 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
10641 | * us and move them to an alternate dst_cpu in our sched_group | |
10642 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 10643 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
10644 | * sched_group. |
10645 | * | |
10646 | * This changes load balance semantics a bit on who can move | |
10647 | * load to a given_cpu. In addition to the given_cpu itself | |
10648 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
10649 | * nohz-idle), we now have balance_cpu in a position to move | |
10650 | * load to given_cpu. In rare situations, this may cause | |
10651 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
10652 | * _independently_ and at _same_ time to move some load to | |
3b03706f | 10653 | * given_cpu) causing excess load to be moved to given_cpu. |
88b8dac0 SV |
10654 | * This however should not happen so much in practice and |
10655 | * moreover subsequent load balance cycles should correct the | |
10656 | * excess load moved. | |
10657 | */ | |
6263322c | 10658 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 10659 | |
97fb7a0a | 10660 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 10661 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 10662 | |
78feefc5 | 10663 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 10664 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 10665 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 | 10666 | env.loop = 0; |
c59862f8 | 10667 | env.loop_break = SCHED_NR_MIGRATE_BREAK; |
e02e60c1 | 10668 | |
88b8dac0 SV |
10669 | /* |
10670 | * Go back to "more_balance" rather than "redo" since we | |
10671 | * need to continue with same src_cpu. | |
10672 | */ | |
10673 | goto more_balance; | |
10674 | } | |
1e3c88bd | 10675 | |
6263322c PZ |
10676 | /* |
10677 | * We failed to reach balance because of affinity. | |
10678 | */ | |
10679 | if (sd_parent) { | |
63b2ca30 | 10680 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 10681 | |
afdeee05 | 10682 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 10683 | *group_imbalance = 1; |
6263322c PZ |
10684 | } |
10685 | ||
1e3c88bd | 10686 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 10687 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 10688 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
10689 | /* |
10690 | * Attempting to continue load balancing at the current | |
10691 | * sched_domain level only makes sense if there are | |
10692 | * active CPUs remaining as possible busiest CPUs to | |
10693 | * pull load from which are not contained within the | |
10694 | * destination group that is receiving any migrated | |
10695 | * load. | |
10696 | */ | |
10697 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 | 10698 | env.loop = 0; |
c59862f8 | 10699 | env.loop_break = SCHED_NR_MIGRATE_BREAK; |
1e3c88bd | 10700 | goto redo; |
bbf18b19 | 10701 | } |
afdeee05 | 10702 | goto out_all_pinned; |
1e3c88bd PZ |
10703 | } |
10704 | } | |
10705 | ||
10706 | if (!ld_moved) { | |
ae92882e | 10707 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
10708 | /* |
10709 | * Increment the failure counter only on periodic balance. | |
10710 | * We do not want newidle balance, which can be very | |
10711 | * frequent, pollute the failure counter causing | |
10712 | * excessive cache_hot migrations and active balances. | |
10713 | */ | |
10714 | if (idle != CPU_NEWLY_IDLE) | |
10715 | sd->nr_balance_failed++; | |
1e3c88bd | 10716 | |
bd939f45 | 10717 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
10718 | unsigned long flags; |
10719 | ||
5cb9eaa3 | 10720 | raw_spin_rq_lock_irqsave(busiest, flags); |
1e3c88bd | 10721 | |
97fb7a0a IM |
10722 | /* |
10723 | * Don't kick the active_load_balance_cpu_stop, | |
10724 | * if the curr task on busiest CPU can't be | |
10725 | * moved to this_cpu: | |
1e3c88bd | 10726 | */ |
3bd37062 | 10727 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
5cb9eaa3 | 10728 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
1e3c88bd PZ |
10729 | goto out_one_pinned; |
10730 | } | |
10731 | ||
8a41dfcd VG |
10732 | /* Record that we found at least one task that could run on this_cpu */ |
10733 | env.flags &= ~LBF_ALL_PINNED; | |
10734 | ||
969c7921 TH |
10735 | /* |
10736 | * ->active_balance synchronizes accesses to | |
10737 | * ->active_balance_work. Once set, it's cleared | |
10738 | * only after active load balance is finished. | |
10739 | */ | |
1e3c88bd PZ |
10740 | if (!busiest->active_balance) { |
10741 | busiest->active_balance = 1; | |
10742 | busiest->push_cpu = this_cpu; | |
10743 | active_balance = 1; | |
10744 | } | |
5cb9eaa3 | 10745 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
969c7921 | 10746 | |
bd939f45 | 10747 | if (active_balance) { |
969c7921 TH |
10748 | stop_one_cpu_nowait(cpu_of(busiest), |
10749 | active_load_balance_cpu_stop, busiest, | |
10750 | &busiest->active_balance_work); | |
bd939f45 | 10751 | } |
1e3c88bd | 10752 | } |
e9b9734b | 10753 | } else { |
1e3c88bd | 10754 | sd->nr_balance_failed = 0; |
e9b9734b | 10755 | } |
1e3c88bd | 10756 | |
e9b9734b | 10757 | if (likely(!active_balance) || need_active_balance(&env)) { |
1e3c88bd PZ |
10758 | /* We were unbalanced, so reset the balancing interval */ |
10759 | sd->balance_interval = sd->min_interval; | |
1e3c88bd PZ |
10760 | } |
10761 | ||
1e3c88bd PZ |
10762 | goto out; |
10763 | ||
10764 | out_balanced: | |
afdeee05 VG |
10765 | /* |
10766 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
10767 | * constraints. Clear the imbalance flag only if other tasks got |
10768 | * a chance to move and fix the imbalance. | |
afdeee05 | 10769 | */ |
f6cad8df | 10770 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
10771 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
10772 | ||
10773 | if (*group_imbalance) | |
10774 | *group_imbalance = 0; | |
10775 | } | |
10776 | ||
10777 | out_all_pinned: | |
10778 | /* | |
10779 | * We reach balance because all tasks are pinned at this level so | |
10780 | * we can't migrate them. Let the imbalance flag set so parent level | |
10781 | * can try to migrate them. | |
10782 | */ | |
ae92882e | 10783 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
10784 | |
10785 | sd->nr_balance_failed = 0; | |
10786 | ||
10787 | out_one_pinned: | |
3f130a37 VS |
10788 | ld_moved = 0; |
10789 | ||
10790 | /* | |
5ba553ef PZ |
10791 | * newidle_balance() disregards balance intervals, so we could |
10792 | * repeatedly reach this code, which would lead to balance_interval | |
3b03706f | 10793 | * skyrocketing in a short amount of time. Skip the balance_interval |
5ba553ef | 10794 | * increase logic to avoid that. |
3f130a37 VS |
10795 | */ |
10796 | if (env.idle == CPU_NEWLY_IDLE) | |
10797 | goto out; | |
10798 | ||
1e3c88bd | 10799 | /* tune up the balancing interval */ |
47b7aee1 VS |
10800 | if ((env.flags & LBF_ALL_PINNED && |
10801 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
10802 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 10803 | sd->balance_interval *= 2; |
1e3c88bd | 10804 | out: |
1e3c88bd PZ |
10805 | return ld_moved; |
10806 | } | |
10807 | ||
52a08ef1 JL |
10808 | static inline unsigned long |
10809 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
10810 | { | |
10811 | unsigned long interval = sd->balance_interval; | |
10812 | ||
10813 | if (cpu_busy) | |
10814 | interval *= sd->busy_factor; | |
10815 | ||
10816 | /* scale ms to jiffies */ | |
10817 | interval = msecs_to_jiffies(interval); | |
e4d32e4d VG |
10818 | |
10819 | /* | |
10820 | * Reduce likelihood of busy balancing at higher domains racing with | |
10821 | * balancing at lower domains by preventing their balancing periods | |
10822 | * from being multiples of each other. | |
10823 | */ | |
10824 | if (cpu_busy) | |
10825 | interval -= 1; | |
10826 | ||
52a08ef1 JL |
10827 | interval = clamp(interval, 1UL, max_load_balance_interval); |
10828 | ||
10829 | return interval; | |
10830 | } | |
10831 | ||
10832 | static inline void | |
31851a98 | 10833 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
10834 | { |
10835 | unsigned long interval, next; | |
10836 | ||
31851a98 LY |
10837 | /* used by idle balance, so cpu_busy = 0 */ |
10838 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
10839 | next = sd->last_balance + interval; |
10840 | ||
10841 | if (time_after(*next_balance, next)) | |
10842 | *next_balance = next; | |
10843 | } | |
10844 | ||
1e3c88bd | 10845 | /* |
97fb7a0a | 10846 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
10847 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
10848 | * least 1 task to be running on each physical CPU where possible, and | |
10849 | * avoids physical / logical imbalances. | |
1e3c88bd | 10850 | */ |
969c7921 | 10851 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 10852 | { |
969c7921 TH |
10853 | struct rq *busiest_rq = data; |
10854 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 10855 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 10856 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 10857 | struct sched_domain *sd; |
e5673f28 | 10858 | struct task_struct *p = NULL; |
8a8c69c3 | 10859 | struct rq_flags rf; |
969c7921 | 10860 | |
8a8c69c3 | 10861 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
10862 | /* |
10863 | * Between queueing the stop-work and running it is a hole in which | |
10864 | * CPUs can become inactive. We should not move tasks from or to | |
10865 | * inactive CPUs. | |
10866 | */ | |
10867 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
10868 | goto out_unlock; | |
969c7921 | 10869 | |
97fb7a0a | 10870 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
10871 | if (unlikely(busiest_cpu != smp_processor_id() || |
10872 | !busiest_rq->active_balance)) | |
10873 | goto out_unlock; | |
1e3c88bd PZ |
10874 | |
10875 | /* Is there any task to move? */ | |
10876 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 10877 | goto out_unlock; |
1e3c88bd PZ |
10878 | |
10879 | /* | |
10880 | * This condition is "impossible", if it occurs | |
10881 | * we need to fix it. Originally reported by | |
97fb7a0a | 10882 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd | 10883 | */ |
09348d75 | 10884 | WARN_ON_ONCE(busiest_rq == target_rq); |
1e3c88bd | 10885 | |
1e3c88bd | 10886 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 10887 | rcu_read_lock(); |
1e3c88bd | 10888 | for_each_domain(target_cpu, sd) { |
e669ac8a VS |
10889 | if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
10890 | break; | |
1e3c88bd PZ |
10891 | } |
10892 | ||
10893 | if (likely(sd)) { | |
8e45cb54 PZ |
10894 | struct lb_env env = { |
10895 | .sd = sd, | |
ddcdf6e7 PZ |
10896 | .dst_cpu = target_cpu, |
10897 | .dst_rq = target_rq, | |
10898 | .src_cpu = busiest_rq->cpu, | |
10899 | .src_rq = busiest_rq, | |
8e45cb54 | 10900 | .idle = CPU_IDLE, |
23fb06d9 | 10901 | .flags = LBF_ACTIVE_LB, |
8e45cb54 PZ |
10902 | }; |
10903 | ||
ae92882e | 10904 | schedstat_inc(sd->alb_count); |
3bed5e21 | 10905 | update_rq_clock(busiest_rq); |
1e3c88bd | 10906 | |
e5673f28 | 10907 | p = detach_one_task(&env); |
d02c0711 | 10908 | if (p) { |
ae92882e | 10909 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
10910 | /* Active balancing done, reset the failure counter. */ |
10911 | sd->nr_balance_failed = 0; | |
10912 | } else { | |
ae92882e | 10913 | schedstat_inc(sd->alb_failed); |
d02c0711 | 10914 | } |
1e3c88bd | 10915 | } |
dce840a0 | 10916 | rcu_read_unlock(); |
969c7921 TH |
10917 | out_unlock: |
10918 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 10919 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
10920 | |
10921 | if (p) | |
10922 | attach_one_task(target_rq, p); | |
10923 | ||
10924 | local_irq_enable(); | |
10925 | ||
969c7921 | 10926 | return 0; |
1e3c88bd PZ |
10927 | } |
10928 | ||
af3fe03c PZ |
10929 | static DEFINE_SPINLOCK(balancing); |
10930 | ||
10931 | /* | |
10932 | * Scale the max load_balance interval with the number of CPUs in the system. | |
10933 | * This trades load-balance latency on larger machines for less cross talk. | |
10934 | */ | |
10935 | void update_max_interval(void) | |
10936 | { | |
10937 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
10938 | } | |
10939 | ||
e60b56e4 VG |
10940 | static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost) |
10941 | { | |
10942 | if (cost > sd->max_newidle_lb_cost) { | |
10943 | /* | |
10944 | * Track max cost of a domain to make sure to not delay the | |
10945 | * next wakeup on the CPU. | |
10946 | */ | |
10947 | sd->max_newidle_lb_cost = cost; | |
10948 | sd->last_decay_max_lb_cost = jiffies; | |
10949 | } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) { | |
10950 | /* | |
10951 | * Decay the newidle max times by ~1% per second to ensure that | |
10952 | * it is not outdated and the current max cost is actually | |
10953 | * shorter. | |
10954 | */ | |
10955 | sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256; | |
10956 | sd->last_decay_max_lb_cost = jiffies; | |
10957 | ||
10958 | return true; | |
10959 | } | |
10960 | ||
10961 | return false; | |
10962 | } | |
10963 | ||
af3fe03c PZ |
10964 | /* |
10965 | * It checks each scheduling domain to see if it is due to be balanced, | |
10966 | * and initiates a balancing operation if so. | |
10967 | * | |
10968 | * Balancing parameters are set up in init_sched_domains. | |
10969 | */ | |
10970 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
10971 | { | |
10972 | int continue_balancing = 1; | |
10973 | int cpu = rq->cpu; | |
323af6de | 10974 | int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10975 | unsigned long interval; |
10976 | struct sched_domain *sd; | |
10977 | /* Earliest time when we have to do rebalance again */ | |
10978 | unsigned long next_balance = jiffies + 60*HZ; | |
10979 | int update_next_balance = 0; | |
10980 | int need_serialize, need_decay = 0; | |
10981 | u64 max_cost = 0; | |
10982 | ||
10983 | rcu_read_lock(); | |
10984 | for_each_domain(cpu, sd) { | |
10985 | /* | |
10986 | * Decay the newidle max times here because this is a regular | |
e60b56e4 | 10987 | * visit to all the domains. |
af3fe03c | 10988 | */ |
e60b56e4 | 10989 | need_decay = update_newidle_cost(sd, 0); |
af3fe03c PZ |
10990 | max_cost += sd->max_newidle_lb_cost; |
10991 | ||
af3fe03c PZ |
10992 | /* |
10993 | * Stop the load balance at this level. There is another | |
10994 | * CPU in our sched group which is doing load balancing more | |
10995 | * actively. | |
10996 | */ | |
10997 | if (!continue_balancing) { | |
10998 | if (need_decay) | |
10999 | continue; | |
11000 | break; | |
11001 | } | |
11002 | ||
323af6de | 11003 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
11004 | |
11005 | need_serialize = sd->flags & SD_SERIALIZE; | |
11006 | if (need_serialize) { | |
11007 | if (!spin_trylock(&balancing)) | |
11008 | goto out; | |
11009 | } | |
11010 | ||
11011 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
11012 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
11013 | /* | |
11014 | * The LBF_DST_PINNED logic could have changed | |
11015 | * env->dst_cpu, so we can't know our idle | |
11016 | * state even if we migrated tasks. Update it. | |
11017 | */ | |
11018 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
323af6de | 11019 | busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
11020 | } |
11021 | sd->last_balance = jiffies; | |
323af6de | 11022 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
11023 | } |
11024 | if (need_serialize) | |
11025 | spin_unlock(&balancing); | |
11026 | out: | |
11027 | if (time_after(next_balance, sd->last_balance + interval)) { | |
11028 | next_balance = sd->last_balance + interval; | |
11029 | update_next_balance = 1; | |
11030 | } | |
11031 | } | |
11032 | if (need_decay) { | |
11033 | /* | |
11034 | * Ensure the rq-wide value also decays but keep it at a | |
11035 | * reasonable floor to avoid funnies with rq->avg_idle. | |
11036 | */ | |
11037 | rq->max_idle_balance_cost = | |
11038 | max((u64)sysctl_sched_migration_cost, max_cost); | |
11039 | } | |
11040 | rcu_read_unlock(); | |
11041 | ||
11042 | /* | |
11043 | * next_balance will be updated only when there is a need. | |
11044 | * When the cpu is attached to null domain for ex, it will not be | |
11045 | * updated. | |
11046 | */ | |
7a82e5f5 | 11047 | if (likely(update_next_balance)) |
af3fe03c PZ |
11048 | rq->next_balance = next_balance; |
11049 | ||
af3fe03c PZ |
11050 | } |
11051 | ||
d987fc7f MG |
11052 | static inline int on_null_domain(struct rq *rq) |
11053 | { | |
11054 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
11055 | } | |
11056 | ||
3451d024 | 11057 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
11058 | /* |
11059 | * idle load balancing details | |
83cd4fe2 VP |
11060 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
11061 | * needed, they will kick the idle load balancer, which then does idle | |
11062 | * load balancing for all the idle CPUs. | |
04d4e665 | 11063 | * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set |
9b019acb | 11064 | * anywhere yet. |
83cd4fe2 | 11065 | */ |
1e3c88bd | 11066 | |
3dd0337d | 11067 | static inline int find_new_ilb(void) |
1e3c88bd | 11068 | { |
9b019acb | 11069 | int ilb; |
031e3bd8 | 11070 | const struct cpumask *hk_mask; |
1e3c88bd | 11071 | |
04d4e665 | 11072 | hk_mask = housekeeping_cpumask(HK_TYPE_MISC); |
1e3c88bd | 11073 | |
031e3bd8 | 11074 | for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) { |
45da7a2b PZ |
11075 | |
11076 | if (ilb == smp_processor_id()) | |
11077 | continue; | |
11078 | ||
9b019acb NP |
11079 | if (idle_cpu(ilb)) |
11080 | return ilb; | |
11081 | } | |
786d6dc7 SS |
11082 | |
11083 | return nr_cpu_ids; | |
1e3c88bd | 11084 | } |
1e3c88bd | 11085 | |
83cd4fe2 | 11086 | /* |
9b019acb | 11087 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick any |
04d4e665 | 11088 | * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one). |
83cd4fe2 | 11089 | */ |
a4064fb6 | 11090 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
11091 | { |
11092 | int ilb_cpu; | |
11093 | ||
3ea2f097 VG |
11094 | /* |
11095 | * Increase nohz.next_balance only when if full ilb is triggered but | |
11096 | * not if we only update stats. | |
11097 | */ | |
11098 | if (flags & NOHZ_BALANCE_KICK) | |
11099 | nohz.next_balance = jiffies+1; | |
83cd4fe2 | 11100 | |
3dd0337d | 11101 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 11102 | |
0b005cf5 SS |
11103 | if (ilb_cpu >= nr_cpu_ids) |
11104 | return; | |
83cd4fe2 | 11105 | |
19a1f5ec PZ |
11106 | /* |
11107 | * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets | |
11108 | * the first flag owns it; cleared by nohz_csd_func(). | |
11109 | */ | |
a4064fb6 | 11110 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 11111 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 11112 | return; |
4550487a | 11113 | |
1c792db7 | 11114 | /* |
90b5363a | 11115 | * This way we generate an IPI on the target CPU which |
1c792db7 SS |
11116 | * is idle. And the softirq performing nohz idle load balance |
11117 | * will be run before returning from the IPI. | |
11118 | */ | |
90b5363a | 11119 | smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); |
4550487a PZ |
11120 | } |
11121 | ||
11122 | /* | |
9f132742 VS |
11123 | * Current decision point for kicking the idle load balancer in the presence |
11124 | * of idle CPUs in the system. | |
4550487a PZ |
11125 | */ |
11126 | static void nohz_balancer_kick(struct rq *rq) | |
11127 | { | |
11128 | unsigned long now = jiffies; | |
11129 | struct sched_domain_shared *sds; | |
11130 | struct sched_domain *sd; | |
11131 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 11132 | unsigned int flags = 0; |
4550487a PZ |
11133 | |
11134 | if (unlikely(rq->idle_balance)) | |
11135 | return; | |
11136 | ||
11137 | /* | |
11138 | * We may be recently in ticked or tickless idle mode. At the first | |
11139 | * busy tick after returning from idle, we will update the busy stats. | |
11140 | */ | |
00357f5e | 11141 | nohz_balance_exit_idle(rq); |
4550487a PZ |
11142 | |
11143 | /* | |
11144 | * None are in tickless mode and hence no need for NOHZ idle load | |
11145 | * balancing. | |
11146 | */ | |
11147 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
11148 | return; | |
11149 | ||
f643ea22 VG |
11150 | if (READ_ONCE(nohz.has_blocked) && |
11151 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
11152 | flags = NOHZ_STATS_KICK; |
11153 | ||
4550487a | 11154 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 11155 | goto out; |
4550487a | 11156 | |
a0fe2cf0 | 11157 | if (rq->nr_running >= 2) { |
efd984c4 | 11158 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
11159 | goto out; |
11160 | } | |
11161 | ||
11162 | rcu_read_lock(); | |
4550487a PZ |
11163 | |
11164 | sd = rcu_dereference(rq->sd); | |
11165 | if (sd) { | |
e25a7a94 VS |
11166 | /* |
11167 | * If there's a CFS task and the current CPU has reduced | |
11168 | * capacity; kick the ILB to see if there's a better CPU to run | |
11169 | * on. | |
11170 | */ | |
11171 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
efd984c4 | 11172 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
11173 | goto unlock; |
11174 | } | |
11175 | } | |
11176 | ||
011b27bb | 11177 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 11178 | if (sd) { |
b9a7b883 VS |
11179 | /* |
11180 | * When ASYM_PACKING; see if there's a more preferred CPU | |
11181 | * currently idle; in which case, kick the ILB to move tasks | |
11182 | * around. | |
11183 | */ | |
7edab78d | 11184 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
4550487a | 11185 | if (sched_asym_prefer(i, cpu)) { |
efd984c4 | 11186 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
11187 | goto unlock; |
11188 | } | |
11189 | } | |
11190 | } | |
b9a7b883 | 11191 | |
a0fe2cf0 VS |
11192 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
11193 | if (sd) { | |
11194 | /* | |
11195 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
11196 | * to run the misfit task on. | |
11197 | */ | |
11198 | if (check_misfit_status(rq, sd)) { | |
efd984c4 | 11199 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
a0fe2cf0 VS |
11200 | goto unlock; |
11201 | } | |
b9a7b883 VS |
11202 | |
11203 | /* | |
11204 | * For asymmetric systems, we do not want to nicely balance | |
11205 | * cache use, instead we want to embrace asymmetry and only | |
11206 | * ensure tasks have enough CPU capacity. | |
11207 | * | |
11208 | * Skip the LLC logic because it's not relevant in that case. | |
11209 | */ | |
11210 | goto unlock; | |
a0fe2cf0 VS |
11211 | } |
11212 | ||
b9a7b883 VS |
11213 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
11214 | if (sds) { | |
e25a7a94 | 11215 | /* |
b9a7b883 VS |
11216 | * If there is an imbalance between LLC domains (IOW we could |
11217 | * increase the overall cache use), we need some less-loaded LLC | |
11218 | * domain to pull some load. Likewise, we may need to spread | |
11219 | * load within the current LLC domain (e.g. packed SMT cores but | |
11220 | * other CPUs are idle). We can't really know from here how busy | |
11221 | * the others are - so just get a nohz balance going if it looks | |
11222 | * like this LLC domain has tasks we could move. | |
e25a7a94 | 11223 | */ |
b9a7b883 VS |
11224 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
11225 | if (nr_busy > 1) { | |
efd984c4 | 11226 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
b9a7b883 | 11227 | goto unlock; |
4550487a PZ |
11228 | } |
11229 | } | |
11230 | unlock: | |
11231 | rcu_read_unlock(); | |
11232 | out: | |
7fd7a9e0 VS |
11233 | if (READ_ONCE(nohz.needs_update)) |
11234 | flags |= NOHZ_NEXT_KICK; | |
11235 | ||
a4064fb6 PZ |
11236 | if (flags) |
11237 | kick_ilb(flags); | |
83cd4fe2 VP |
11238 | } |
11239 | ||
00357f5e | 11240 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 11241 | { |
00357f5e | 11242 | struct sched_domain *sd; |
a22e47a4 | 11243 | |
00357f5e PZ |
11244 | rcu_read_lock(); |
11245 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 11246 | |
00357f5e PZ |
11247 | if (!sd || !sd->nohz_idle) |
11248 | goto unlock; | |
11249 | sd->nohz_idle = 0; | |
11250 | ||
11251 | atomic_inc(&sd->shared->nr_busy_cpus); | |
11252 | unlock: | |
11253 | rcu_read_unlock(); | |
71325960 SS |
11254 | } |
11255 | ||
00357f5e PZ |
11256 | void nohz_balance_exit_idle(struct rq *rq) |
11257 | { | |
11258 | SCHED_WARN_ON(rq != this_rq()); | |
11259 | ||
11260 | if (likely(!rq->nohz_tick_stopped)) | |
11261 | return; | |
11262 | ||
11263 | rq->nohz_tick_stopped = 0; | |
11264 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
11265 | atomic_dec(&nohz.nr_cpus); | |
11266 | ||
11267 | set_cpu_sd_state_busy(rq->cpu); | |
11268 | } | |
11269 | ||
11270 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
11271 | { |
11272 | struct sched_domain *sd; | |
69e1e811 | 11273 | |
69e1e811 | 11274 | rcu_read_lock(); |
0e369d75 | 11275 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
11276 | |
11277 | if (!sd || sd->nohz_idle) | |
11278 | goto unlock; | |
11279 | sd->nohz_idle = 1; | |
11280 | ||
0e369d75 | 11281 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 11282 | unlock: |
69e1e811 SS |
11283 | rcu_read_unlock(); |
11284 | } | |
11285 | ||
1e3c88bd | 11286 | /* |
97fb7a0a | 11287 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 11288 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 11289 | */ |
c1cc017c | 11290 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 11291 | { |
00357f5e PZ |
11292 | struct rq *rq = cpu_rq(cpu); |
11293 | ||
11294 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
11295 | ||
97fb7a0a | 11296 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
11297 | if (!cpu_active(cpu)) |
11298 | return; | |
11299 | ||
387bc8b5 | 11300 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
04d4e665 | 11301 | if (!housekeeping_cpu(cpu, HK_TYPE_SCHED)) |
387bc8b5 FW |
11302 | return; |
11303 | ||
f643ea22 VG |
11304 | /* |
11305 | * Can be set safely without rq->lock held | |
11306 | * If a clear happens, it will have evaluated last additions because | |
11307 | * rq->lock is held during the check and the clear | |
11308 | */ | |
11309 | rq->has_blocked_load = 1; | |
11310 | ||
11311 | /* | |
11312 | * The tick is still stopped but load could have been added in the | |
11313 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
11314 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
11315 | * of nohz.has_blocked can only happen after checking the new load | |
11316 | */ | |
00357f5e | 11317 | if (rq->nohz_tick_stopped) |
f643ea22 | 11318 | goto out; |
1e3c88bd | 11319 | |
97fb7a0a | 11320 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 11321 | if (on_null_domain(rq)) |
d987fc7f MG |
11322 | return; |
11323 | ||
00357f5e PZ |
11324 | rq->nohz_tick_stopped = 1; |
11325 | ||
c1cc017c AS |
11326 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
11327 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 11328 | |
f643ea22 VG |
11329 | /* |
11330 | * Ensures that if nohz_idle_balance() fails to observe our | |
11331 | * @idle_cpus_mask store, it must observe the @has_blocked | |
7fd7a9e0 | 11332 | * and @needs_update stores. |
f643ea22 VG |
11333 | */ |
11334 | smp_mb__after_atomic(); | |
11335 | ||
00357f5e | 11336 | set_cpu_sd_state_idle(cpu); |
f643ea22 | 11337 | |
7fd7a9e0 | 11338 | WRITE_ONCE(nohz.needs_update, 1); |
f643ea22 VG |
11339 | out: |
11340 | /* | |
11341 | * Each time a cpu enter idle, we assume that it has blocked load and | |
11342 | * enable the periodic update of the load of idle cpus | |
11343 | */ | |
11344 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 11345 | } |
1e3c88bd | 11346 | |
3f5ad914 Y |
11347 | static bool update_nohz_stats(struct rq *rq) |
11348 | { | |
11349 | unsigned int cpu = rq->cpu; | |
11350 | ||
11351 | if (!rq->has_blocked_load) | |
11352 | return false; | |
11353 | ||
11354 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) | |
11355 | return false; | |
11356 | ||
11357 | if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) | |
11358 | return true; | |
11359 | ||
11360 | update_blocked_averages(cpu); | |
11361 | ||
11362 | return rq->has_blocked_load; | |
11363 | } | |
11364 | ||
1e3c88bd | 11365 | /* |
31e77c93 VG |
11366 | * Internal function that runs load balance for all idle cpus. The load balance |
11367 | * can be a simple update of blocked load or a complete load balance with | |
11368 | * tasks movement depending of flags. | |
1e3c88bd | 11369 | */ |
d985ee9f | 11370 | static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags) |
83cd4fe2 | 11371 | { |
c5afb6a8 | 11372 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
11373 | unsigned long now = jiffies; |
11374 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 11375 | bool has_blocked_load = false; |
c5afb6a8 | 11376 | int update_next_balance = 0; |
b7031a02 | 11377 | int this_cpu = this_rq->cpu; |
b7031a02 PZ |
11378 | int balance_cpu; |
11379 | struct rq *rq; | |
83cd4fe2 | 11380 | |
b7031a02 | 11381 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 11382 | |
f643ea22 VG |
11383 | /* |
11384 | * We assume there will be no idle load after this update and clear | |
11385 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
7fd7a9e0 | 11386 | * set the has_blocked flag and trigger another update of idle load. |
f643ea22 VG |
11387 | * Because a cpu that becomes idle, is added to idle_cpus_mask before |
11388 | * setting the flag, we are sure to not clear the state and not | |
11389 | * check the load of an idle cpu. | |
7fd7a9e0 VS |
11390 | * |
11391 | * Same applies to idle_cpus_mask vs needs_update. | |
f643ea22 | 11392 | */ |
efd984c4 VS |
11393 | if (flags & NOHZ_STATS_KICK) |
11394 | WRITE_ONCE(nohz.has_blocked, 0); | |
7fd7a9e0 VS |
11395 | if (flags & NOHZ_NEXT_KICK) |
11396 | WRITE_ONCE(nohz.needs_update, 0); | |
f643ea22 VG |
11397 | |
11398 | /* | |
11399 | * Ensures that if we miss the CPU, we must see the has_blocked | |
11400 | * store from nohz_balance_enter_idle(). | |
11401 | */ | |
11402 | smp_mb(); | |
11403 | ||
7a82e5f5 VG |
11404 | /* |
11405 | * Start with the next CPU after this_cpu so we will end with this_cpu and let a | |
11406 | * chance for other idle cpu to pull load. | |
11407 | */ | |
11408 | for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) { | |
11409 | if (!idle_cpu(balance_cpu)) | |
83cd4fe2 VP |
11410 | continue; |
11411 | ||
11412 | /* | |
97fb7a0a IM |
11413 | * If this CPU gets work to do, stop the load balancing |
11414 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
11415 | * balancing owner will pick it up. |
11416 | */ | |
f643ea22 | 11417 | if (need_resched()) { |
efd984c4 VS |
11418 | if (flags & NOHZ_STATS_KICK) |
11419 | has_blocked_load = true; | |
7fd7a9e0 VS |
11420 | if (flags & NOHZ_NEXT_KICK) |
11421 | WRITE_ONCE(nohz.needs_update, 1); | |
f643ea22 VG |
11422 | goto abort; |
11423 | } | |
83cd4fe2 | 11424 | |
5ed4f1d9 VG |
11425 | rq = cpu_rq(balance_cpu); |
11426 | ||
efd984c4 VS |
11427 | if (flags & NOHZ_STATS_KICK) |
11428 | has_blocked_load |= update_nohz_stats(rq); | |
f643ea22 | 11429 | |
ed61bbc6 TC |
11430 | /* |
11431 | * If time for next balance is due, | |
11432 | * do the balance. | |
11433 | */ | |
11434 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
11435 | struct rq_flags rf; |
11436 | ||
31e77c93 | 11437 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 11438 | update_rq_clock(rq); |
31e77c93 | 11439 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 11440 | |
b7031a02 PZ |
11441 | if (flags & NOHZ_BALANCE_KICK) |
11442 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 11443 | } |
83cd4fe2 | 11444 | |
c5afb6a8 VG |
11445 | if (time_after(next_balance, rq->next_balance)) { |
11446 | next_balance = rq->next_balance; | |
11447 | update_next_balance = 1; | |
11448 | } | |
83cd4fe2 | 11449 | } |
c5afb6a8 | 11450 | |
3ea2f097 VG |
11451 | /* |
11452 | * next_balance will be updated only when there is a need. | |
11453 | * When the CPU is attached to null domain for ex, it will not be | |
11454 | * updated. | |
11455 | */ | |
11456 | if (likely(update_next_balance)) | |
11457 | nohz.next_balance = next_balance; | |
11458 | ||
efd984c4 VS |
11459 | if (flags & NOHZ_STATS_KICK) |
11460 | WRITE_ONCE(nohz.next_blocked, | |
11461 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
f643ea22 VG |
11462 | |
11463 | abort: | |
11464 | /* There is still blocked load, enable periodic update */ | |
11465 | if (has_blocked_load) | |
11466 | WRITE_ONCE(nohz.has_blocked, 1); | |
31e77c93 VG |
11467 | } |
11468 | ||
11469 | /* | |
11470 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
11471 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
11472 | */ | |
11473 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
11474 | { | |
19a1f5ec | 11475 | unsigned int flags = this_rq->nohz_idle_balance; |
31e77c93 | 11476 | |
19a1f5ec | 11477 | if (!flags) |
31e77c93 VG |
11478 | return false; |
11479 | ||
19a1f5ec | 11480 | this_rq->nohz_idle_balance = 0; |
31e77c93 | 11481 | |
19a1f5ec | 11482 | if (idle != CPU_IDLE) |
31e77c93 VG |
11483 | return false; |
11484 | ||
d985ee9f | 11485 | _nohz_idle_balance(this_rq, flags); |
31e77c93 | 11486 | |
b7031a02 | 11487 | return true; |
83cd4fe2 | 11488 | } |
31e77c93 | 11489 | |
c6f88654 VG |
11490 | /* |
11491 | * Check if we need to run the ILB for updating blocked load before entering | |
11492 | * idle state. | |
11493 | */ | |
11494 | void nohz_run_idle_balance(int cpu) | |
11495 | { | |
11496 | unsigned int flags; | |
11497 | ||
11498 | flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu)); | |
11499 | ||
11500 | /* | |
11501 | * Update the blocked load only if no SCHED_SOFTIRQ is about to happen | |
11502 | * (ie NOHZ_STATS_KICK set) and will do the same. | |
11503 | */ | |
11504 | if ((flags == NOHZ_NEWILB_KICK) && !need_resched()) | |
d985ee9f | 11505 | _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK); |
c6f88654 VG |
11506 | } |
11507 | ||
31e77c93 VG |
11508 | static void nohz_newidle_balance(struct rq *this_rq) |
11509 | { | |
11510 | int this_cpu = this_rq->cpu; | |
11511 | ||
11512 | /* | |
11513 | * This CPU doesn't want to be disturbed by scheduler | |
11514 | * housekeeping | |
11515 | */ | |
04d4e665 | 11516 | if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED)) |
31e77c93 VG |
11517 | return; |
11518 | ||
11519 | /* Will wake up very soon. No time for doing anything else*/ | |
11520 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
11521 | return; | |
11522 | ||
11523 | /* Don't need to update blocked load of idle CPUs*/ | |
11524 | if (!READ_ONCE(nohz.has_blocked) || | |
11525 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
11526 | return; | |
11527 | ||
31e77c93 | 11528 | /* |
c6f88654 VG |
11529 | * Set the need to trigger ILB in order to update blocked load |
11530 | * before entering idle state. | |
31e77c93 | 11531 | */ |
c6f88654 | 11532 | atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu)); |
31e77c93 VG |
11533 | } |
11534 | ||
dd707247 PZ |
11535 | #else /* !CONFIG_NO_HZ_COMMON */ |
11536 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
11537 | ||
31e77c93 | 11538 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
11539 | { |
11540 | return false; | |
11541 | } | |
31e77c93 VG |
11542 | |
11543 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 11544 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 11545 | |
47ea5412 | 11546 | /* |
5b78f2dc | 11547 | * newidle_balance is called by schedule() if this_cpu is about to become |
47ea5412 | 11548 | * idle. Attempts to pull tasks from other CPUs. |
7277a34c PZ |
11549 | * |
11550 | * Returns: | |
11551 | * < 0 - we released the lock and there are !fair tasks present | |
11552 | * 0 - failed, no new tasks | |
11553 | * > 0 - success, new (fair) tasks present | |
47ea5412 | 11554 | */ |
d91cecc1 | 11555 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) |
47ea5412 PZ |
11556 | { |
11557 | unsigned long next_balance = jiffies + HZ; | |
11558 | int this_cpu = this_rq->cpu; | |
9e9af819 | 11559 | u64 t0, t1, curr_cost = 0; |
47ea5412 PZ |
11560 | struct sched_domain *sd; |
11561 | int pulled_task = 0; | |
47ea5412 | 11562 | |
5ba553ef | 11563 | update_misfit_status(NULL, this_rq); |
e5e678e4 RR |
11564 | |
11565 | /* | |
11566 | * There is a task waiting to run. No need to search for one. | |
11567 | * Return 0; the task will be enqueued when switching to idle. | |
11568 | */ | |
11569 | if (this_rq->ttwu_pending) | |
11570 | return 0; | |
11571 | ||
47ea5412 PZ |
11572 | /* |
11573 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
11574 | * measure the duration of idle_balance() as idle time. | |
11575 | */ | |
11576 | this_rq->idle_stamp = rq_clock(this_rq); | |
11577 | ||
11578 | /* | |
11579 | * Do not pull tasks towards !active CPUs... | |
11580 | */ | |
11581 | if (!cpu_active(this_cpu)) | |
11582 | return 0; | |
11583 | ||
11584 | /* | |
11585 | * This is OK, because current is on_cpu, which avoids it being picked | |
11586 | * for load-balance and preemption/IRQs are still disabled avoiding | |
11587 | * further scheduler activity on it and we're being very careful to | |
11588 | * re-start the picking loop. | |
11589 | */ | |
11590 | rq_unpin_lock(this_rq, rf); | |
11591 | ||
9d783c8d VG |
11592 | rcu_read_lock(); |
11593 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
11594 | ||
c5b0a7ee | 11595 | if (!READ_ONCE(this_rq->rd->overload) || |
9d783c8d | 11596 | (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) { |
31e77c93 | 11597 | |
47ea5412 PZ |
11598 | if (sd) |
11599 | update_next_balance(sd, &next_balance); | |
11600 | rcu_read_unlock(); | |
11601 | ||
11602 | goto out; | |
11603 | } | |
9d783c8d | 11604 | rcu_read_unlock(); |
47ea5412 | 11605 | |
5cb9eaa3 | 11606 | raw_spin_rq_unlock(this_rq); |
47ea5412 | 11607 | |
9e9af819 | 11608 | t0 = sched_clock_cpu(this_cpu); |
47ea5412 | 11609 | update_blocked_averages(this_cpu); |
9e9af819 | 11610 | |
47ea5412 PZ |
11611 | rcu_read_lock(); |
11612 | for_each_domain(this_cpu, sd) { | |
11613 | int continue_balancing = 1; | |
9e9af819 | 11614 | u64 domain_cost; |
47ea5412 | 11615 | |
8ea9183d VG |
11616 | update_next_balance(sd, &next_balance); |
11617 | ||
11618 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) | |
47ea5412 | 11619 | break; |
47ea5412 PZ |
11620 | |
11621 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
47ea5412 PZ |
11622 | |
11623 | pulled_task = load_balance(this_cpu, this_rq, | |
11624 | sd, CPU_NEWLY_IDLE, | |
11625 | &continue_balancing); | |
11626 | ||
9e9af819 VG |
11627 | t1 = sched_clock_cpu(this_cpu); |
11628 | domain_cost = t1 - t0; | |
e60b56e4 | 11629 | update_newidle_cost(sd, domain_cost); |
47ea5412 PZ |
11630 | |
11631 | curr_cost += domain_cost; | |
9e9af819 | 11632 | t0 = t1; |
47ea5412 PZ |
11633 | } |
11634 | ||
47ea5412 PZ |
11635 | /* |
11636 | * Stop searching for tasks to pull if there are | |
11637 | * now runnable tasks on this rq. | |
11638 | */ | |
e5e678e4 RR |
11639 | if (pulled_task || this_rq->nr_running > 0 || |
11640 | this_rq->ttwu_pending) | |
47ea5412 PZ |
11641 | break; |
11642 | } | |
11643 | rcu_read_unlock(); | |
11644 | ||
5cb9eaa3 | 11645 | raw_spin_rq_lock(this_rq); |
47ea5412 PZ |
11646 | |
11647 | if (curr_cost > this_rq->max_idle_balance_cost) | |
11648 | this_rq->max_idle_balance_cost = curr_cost; | |
11649 | ||
11650 | /* | |
11651 | * While browsing the domains, we released the rq lock, a task could | |
11652 | * have been enqueued in the meantime. Since we're not going idle, | |
11653 | * pretend we pulled a task. | |
11654 | */ | |
11655 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
11656 | pulled_task = 1; | |
11657 | ||
47ea5412 PZ |
11658 | /* Is there a task of a high priority class? */ |
11659 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
11660 | pulled_task = -1; | |
11661 | ||
6553fc18 VG |
11662 | out: |
11663 | /* Move the next balance forward */ | |
11664 | if (time_after(this_rq->next_balance, next_balance)) | |
11665 | this_rq->next_balance = next_balance; | |
11666 | ||
47ea5412 PZ |
11667 | if (pulled_task) |
11668 | this_rq->idle_stamp = 0; | |
0826530d VG |
11669 | else |
11670 | nohz_newidle_balance(this_rq); | |
47ea5412 PZ |
11671 | |
11672 | rq_repin_lock(this_rq, rf); | |
11673 | ||
11674 | return pulled_task; | |
11675 | } | |
11676 | ||
83cd4fe2 VP |
11677 | /* |
11678 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
11679 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
11680 | */ | |
0766f788 | 11681 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 11682 | { |
208cb16b | 11683 | struct rq *this_rq = this_rq(); |
6eb57e0d | 11684 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
11685 | CPU_IDLE : CPU_NOT_IDLE; |
11686 | ||
1e3c88bd | 11687 | /* |
97fb7a0a IM |
11688 | * If this CPU has a pending nohz_balance_kick, then do the |
11689 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 11690 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 11691 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
11692 | * load balance only within the local sched_domain hierarchy |
11693 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 11694 | */ |
b7031a02 PZ |
11695 | if (nohz_idle_balance(this_rq, idle)) |
11696 | return; | |
11697 | ||
11698 | /* normal load balance */ | |
11699 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 11700 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
11701 | } |
11702 | ||
1e3c88bd PZ |
11703 | /* |
11704 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 11705 | */ |
7caff66f | 11706 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 11707 | { |
e0b257c3 AMB |
11708 | /* |
11709 | * Don't need to rebalance while attached to NULL domain or | |
11710 | * runqueue CPU is not active | |
11711 | */ | |
11712 | if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq)))) | |
c726099e DL |
11713 | return; |
11714 | ||
11715 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 11716 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
11717 | |
11718 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
11719 | } |
11720 | ||
0bcdcf28 CE |
11721 | static void rq_online_fair(struct rq *rq) |
11722 | { | |
11723 | update_sysctl(); | |
0e59bdae KT |
11724 | |
11725 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
11726 | } |
11727 | ||
11728 | static void rq_offline_fair(struct rq *rq) | |
11729 | { | |
11730 | update_sysctl(); | |
a4c96ae3 PB |
11731 | |
11732 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
11733 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
11734 | } |
11735 | ||
55e12e5e | 11736 | #endif /* CONFIG_SMP */ |
e1d1484f | 11737 | |
8039e96f VP |
11738 | #ifdef CONFIG_SCHED_CORE |
11739 | static inline bool | |
11740 | __entity_slice_used(struct sched_entity *se, int min_nr_tasks) | |
11741 | { | |
11742 | u64 slice = sched_slice(cfs_rq_of(se), se); | |
11743 | u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
11744 | ||
11745 | return (rtime * min_nr_tasks > slice); | |
11746 | } | |
11747 | ||
11748 | #define MIN_NR_TASKS_DURING_FORCEIDLE 2 | |
11749 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) | |
11750 | { | |
11751 | if (!sched_core_enabled(rq)) | |
11752 | return; | |
11753 | ||
11754 | /* | |
11755 | * If runqueue has only one task which used up its slice and | |
11756 | * if the sibling is forced idle, then trigger schedule to | |
11757 | * give forced idle task a chance. | |
11758 | * | |
11759 | * sched_slice() considers only this active rq and it gets the | |
11760 | * whole slice. But during force idle, we have siblings acting | |
11761 | * like a single runqueue and hence we need to consider runnable | |
cc00c198 | 11762 | * tasks on this CPU and the forced idle CPU. Ideally, we should |
8039e96f | 11763 | * go through the forced idle rq, but that would be a perf hit. |
cc00c198 | 11764 | * We can assume that the forced idle CPU has at least |
8039e96f | 11765 | * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check |
cc00c198 | 11766 | * if we need to give up the CPU. |
8039e96f | 11767 | */ |
4feee7d1 | 11768 | if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 && |
8039e96f VP |
11769 | __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE)) |
11770 | resched_curr(rq); | |
11771 | } | |
c6047c2e JFG |
11772 | |
11773 | /* | |
11774 | * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed. | |
11775 | */ | |
11776 | static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle) | |
11777 | { | |
11778 | for_each_sched_entity(se) { | |
11779 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11780 | ||
11781 | if (forceidle) { | |
11782 | if (cfs_rq->forceidle_seq == fi_seq) | |
11783 | break; | |
11784 | cfs_rq->forceidle_seq = fi_seq; | |
11785 | } | |
11786 | ||
11787 | cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime; | |
11788 | } | |
11789 | } | |
11790 | ||
11791 | void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi) | |
11792 | { | |
11793 | struct sched_entity *se = &p->se; | |
11794 | ||
11795 | if (p->sched_class != &fair_sched_class) | |
11796 | return; | |
11797 | ||
11798 | se_fi_update(se, rq->core->core_forceidle_seq, in_fi); | |
11799 | } | |
11800 | ||
11801 | bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi) | |
11802 | { | |
11803 | struct rq *rq = task_rq(a); | |
11804 | struct sched_entity *sea = &a->se; | |
11805 | struct sched_entity *seb = &b->se; | |
11806 | struct cfs_rq *cfs_rqa; | |
11807 | struct cfs_rq *cfs_rqb; | |
11808 | s64 delta; | |
11809 | ||
11810 | SCHED_WARN_ON(task_rq(b)->core != rq->core); | |
11811 | ||
11812 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
11813 | /* | |
11814 | * Find an se in the hierarchy for tasks a and b, such that the se's | |
11815 | * are immediate siblings. | |
11816 | */ | |
11817 | while (sea->cfs_rq->tg != seb->cfs_rq->tg) { | |
11818 | int sea_depth = sea->depth; | |
11819 | int seb_depth = seb->depth; | |
11820 | ||
11821 | if (sea_depth >= seb_depth) | |
11822 | sea = parent_entity(sea); | |
11823 | if (sea_depth <= seb_depth) | |
11824 | seb = parent_entity(seb); | |
11825 | } | |
11826 | ||
11827 | se_fi_update(sea, rq->core->core_forceidle_seq, in_fi); | |
11828 | se_fi_update(seb, rq->core->core_forceidle_seq, in_fi); | |
11829 | ||
11830 | cfs_rqa = sea->cfs_rq; | |
11831 | cfs_rqb = seb->cfs_rq; | |
11832 | #else | |
11833 | cfs_rqa = &task_rq(a)->cfs; | |
11834 | cfs_rqb = &task_rq(b)->cfs; | |
11835 | #endif | |
11836 | ||
11837 | /* | |
11838 | * Find delta after normalizing se's vruntime with its cfs_rq's | |
11839 | * min_vruntime_fi, which would have been updated in prior calls | |
11840 | * to se_fi_update(). | |
11841 | */ | |
11842 | delta = (s64)(sea->vruntime - seb->vruntime) + | |
11843 | (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi); | |
11844 | ||
11845 | return delta > 0; | |
11846 | } | |
8039e96f VP |
11847 | #else |
11848 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {} | |
11849 | #endif | |
11850 | ||
bf0f6f24 | 11851 | /* |
d84b3131 FW |
11852 | * scheduler tick hitting a task of our scheduling class. |
11853 | * | |
11854 | * NOTE: This function can be called remotely by the tick offload that | |
11855 | * goes along full dynticks. Therefore no local assumption can be made | |
11856 | * and everything must be accessed through the @rq and @curr passed in | |
11857 | * parameters. | |
bf0f6f24 | 11858 | */ |
8f4d37ec | 11859 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
11860 | { |
11861 | struct cfs_rq *cfs_rq; | |
11862 | struct sched_entity *se = &curr->se; | |
11863 | ||
11864 | for_each_sched_entity(se) { | |
11865 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 11866 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 11867 | } |
18bf2805 | 11868 | |
b52da86e | 11869 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 11870 | task_tick_numa(rq, curr); |
3b1baa64 MR |
11871 | |
11872 | update_misfit_status(curr, rq); | |
2802bf3c | 11873 | update_overutilized_status(task_rq(curr)); |
8039e96f VP |
11874 | |
11875 | task_tick_core(rq, curr); | |
bf0f6f24 IM |
11876 | } |
11877 | ||
11878 | /* | |
cd29fe6f PZ |
11879 | * called on fork with the child task as argument from the parent's context |
11880 | * - child not yet on the tasklist | |
11881 | * - preemption disabled | |
bf0f6f24 | 11882 | */ |
cd29fe6f | 11883 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 11884 | { |
4fc420c9 DN |
11885 | struct cfs_rq *cfs_rq; |
11886 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 11887 | struct rq *rq = this_rq(); |
8a8c69c3 | 11888 | struct rq_flags rf; |
bf0f6f24 | 11889 | |
8a8c69c3 | 11890 | rq_lock(rq, &rf); |
861d034e PZ |
11891 | update_rq_clock(rq); |
11892 | ||
4fc420c9 DN |
11893 | cfs_rq = task_cfs_rq(current); |
11894 | curr = cfs_rq->curr; | |
e210bffd PZ |
11895 | if (curr) { |
11896 | update_curr(cfs_rq); | |
b5d9d734 | 11897 | se->vruntime = curr->vruntime; |
e210bffd | 11898 | } |
aeb73b04 | 11899 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 11900 | |
cd29fe6f | 11901 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 11902 | /* |
edcb60a3 IM |
11903 | * Upon rescheduling, sched_class::put_prev_task() will place |
11904 | * 'current' within the tree based on its new key value. | |
11905 | */ | |
4d78e7b6 | 11906 | swap(curr->vruntime, se->vruntime); |
8875125e | 11907 | resched_curr(rq); |
4d78e7b6 | 11908 | } |
bf0f6f24 | 11909 | |
88ec22d3 | 11910 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 11911 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
11912 | } |
11913 | ||
cb469845 SR |
11914 | /* |
11915 | * Priority of the task has changed. Check to see if we preempt | |
11916 | * the current task. | |
11917 | */ | |
da7a735e PZ |
11918 | static void |
11919 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 11920 | { |
da0c1e65 | 11921 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
11922 | return; |
11923 | ||
7c2e8bbd FW |
11924 | if (rq->cfs.nr_running == 1) |
11925 | return; | |
11926 | ||
cb469845 SR |
11927 | /* |
11928 | * Reschedule if we are currently running on this runqueue and | |
11929 | * our priority decreased, or if we are not currently running on | |
11930 | * this runqueue and our priority is higher than the current's | |
11931 | */ | |
65bcf072 | 11932 | if (task_current(rq, p)) { |
cb469845 | 11933 | if (p->prio > oldprio) |
8875125e | 11934 | resched_curr(rq); |
cb469845 | 11935 | } else |
15afe09b | 11936 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
11937 | } |
11938 | ||
daa59407 | 11939 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
11940 | { |
11941 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
11942 | |
11943 | /* | |
daa59407 BP |
11944 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
11945 | * the dequeue_entity(.flags=0) will already have normalized the | |
11946 | * vruntime. | |
11947 | */ | |
11948 | if (p->on_rq) | |
11949 | return true; | |
11950 | ||
11951 | /* | |
11952 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
11953 | * But there are some cases where it has already been normalized: | |
da7a735e | 11954 | * |
daa59407 BP |
11955 | * - A forked child which is waiting for being woken up by |
11956 | * wake_up_new_task(). | |
11957 | * - A task which has been woken up by try_to_wake_up() and | |
11958 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 11959 | */ |
d0cdb3ce | 11960 | if (!se->sum_exec_runtime || |
2f064a59 | 11961 | (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup)) |
daa59407 BP |
11962 | return true; |
11963 | ||
11964 | return false; | |
11965 | } | |
11966 | ||
09a43ace VG |
11967 | #ifdef CONFIG_FAIR_GROUP_SCHED |
11968 | /* | |
11969 | * Propagate the changes of the sched_entity across the tg tree to make it | |
11970 | * visible to the root | |
11971 | */ | |
11972 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
11973 | { | |
51bf903b CZ |
11974 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
11975 | ||
11976 | if (cfs_rq_throttled(cfs_rq)) | |
11977 | return; | |
09a43ace | 11978 | |
51bf903b CZ |
11979 | if (!throttled_hierarchy(cfs_rq)) |
11980 | list_add_leaf_cfs_rq(cfs_rq); | |
0258bdfa | 11981 | |
09a43ace VG |
11982 | /* Start to propagate at parent */ |
11983 | se = se->parent; | |
11984 | ||
11985 | for_each_sched_entity(se) { | |
11986 | cfs_rq = cfs_rq_of(se); | |
11987 | ||
51bf903b | 11988 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace | 11989 | |
51bf903b | 11990 | if (cfs_rq_throttled(cfs_rq)) |
0258bdfa | 11991 | break; |
51bf903b CZ |
11992 | |
11993 | if (!throttled_hierarchy(cfs_rq)) | |
11994 | list_add_leaf_cfs_rq(cfs_rq); | |
09a43ace VG |
11995 | } |
11996 | } | |
11997 | #else | |
11998 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
11999 | #endif | |
12000 | ||
df217913 | 12001 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 12002 | { |
daa59407 BP |
12003 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
12004 | ||
7e2edaf6 CZ |
12005 | #ifdef CONFIG_SMP |
12006 | /* | |
12007 | * In case the task sched_avg hasn't been attached: | |
12008 | * - A forked task which hasn't been woken up by wake_up_new_task(). | |
12009 | * - A task which has been woken up by try_to_wake_up() but is | |
12010 | * waiting for actually being woken up by sched_ttwu_pending(). | |
12011 | */ | |
12012 | if (!se->avg.last_update_time) | |
12013 | return; | |
12014 | #endif | |
12015 | ||
9d89c257 | 12016 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 12017 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 12018 | detach_entity_load_avg(cfs_rq, se); |
fe749158 | 12019 | update_tg_load_avg(cfs_rq); |
09a43ace | 12020 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
12021 | } |
12022 | ||
df217913 | 12023 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 12024 | { |
daa59407 | 12025 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a | 12026 | |
df217913 | 12027 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 12028 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
a4f9a0e5 | 12029 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 12030 | update_tg_load_avg(cfs_rq); |
09a43ace | 12031 | propagate_entity_cfs_rq(se); |
df217913 VG |
12032 | } |
12033 | ||
12034 | static void detach_task_cfs_rq(struct task_struct *p) | |
12035 | { | |
12036 | struct sched_entity *se = &p->se; | |
12037 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
12038 | ||
12039 | if (!vruntime_normalized(p)) { | |
12040 | /* | |
12041 | * Fix up our vruntime so that the current sleep doesn't | |
12042 | * cause 'unlimited' sleep bonus. | |
12043 | */ | |
12044 | place_entity(cfs_rq, se, 0); | |
12045 | se->vruntime -= cfs_rq->min_vruntime; | |
12046 | } | |
12047 | ||
12048 | detach_entity_cfs_rq(se); | |
12049 | } | |
12050 | ||
12051 | static void attach_task_cfs_rq(struct task_struct *p) | |
12052 | { | |
12053 | struct sched_entity *se = &p->se; | |
12054 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
12055 | ||
12056 | attach_entity_cfs_rq(se); | |
daa59407 BP |
12057 | |
12058 | if (!vruntime_normalized(p)) | |
12059 | se->vruntime += cfs_rq->min_vruntime; | |
12060 | } | |
6efdb105 | 12061 | |
daa59407 BP |
12062 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
12063 | { | |
12064 | detach_task_cfs_rq(p); | |
12065 | } | |
12066 | ||
12067 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
12068 | { | |
12069 | attach_task_cfs_rq(p); | |
7855a35a | 12070 | |
daa59407 | 12071 | if (task_on_rq_queued(p)) { |
7855a35a | 12072 | /* |
daa59407 BP |
12073 | * We were most likely switched from sched_rt, so |
12074 | * kick off the schedule if running, otherwise just see | |
12075 | * if we can still preempt the current task. | |
7855a35a | 12076 | */ |
65bcf072 | 12077 | if (task_current(rq, p)) |
daa59407 BP |
12078 | resched_curr(rq); |
12079 | else | |
12080 | check_preempt_curr(rq, p, 0); | |
7855a35a | 12081 | } |
cb469845 SR |
12082 | } |
12083 | ||
83b699ed SV |
12084 | /* Account for a task changing its policy or group. |
12085 | * | |
12086 | * This routine is mostly called to set cfs_rq->curr field when a task | |
12087 | * migrates between groups/classes. | |
12088 | */ | |
a0e813f2 | 12089 | static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) |
83b699ed | 12090 | { |
03b7fad1 PZ |
12091 | struct sched_entity *se = &p->se; |
12092 | ||
12093 | #ifdef CONFIG_SMP | |
12094 | if (task_on_rq_queued(p)) { | |
12095 | /* | |
12096 | * Move the next running task to the front of the list, so our | |
12097 | * cfs_tasks list becomes MRU one. | |
12098 | */ | |
12099 | list_move(&se->group_node, &rq->cfs_tasks); | |
12100 | } | |
12101 | #endif | |
83b699ed | 12102 | |
ec12cb7f PT |
12103 | for_each_sched_entity(se) { |
12104 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
12105 | ||
12106 | set_next_entity(cfs_rq, se); | |
12107 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
12108 | account_cfs_rq_runtime(cfs_rq, 0); | |
12109 | } | |
83b699ed SV |
12110 | } |
12111 | ||
029632fb PZ |
12112 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
12113 | { | |
bfb06889 | 12114 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
d05b4305 | 12115 | u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20))); |
141965c7 | 12116 | #ifdef CONFIG_SMP |
2a2f5d4e | 12117 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 12118 | #endif |
029632fb PZ |
12119 | } |
12120 | ||
810b3817 | 12121 | #ifdef CONFIG_FAIR_GROUP_SCHED |
39c42611 | 12122 | static void task_change_group_fair(struct task_struct *p) |
810b3817 | 12123 | { |
df16b71c CZ |
12124 | /* |
12125 | * We couldn't detach or attach a forked task which | |
12126 | * hasn't been woken up by wake_up_new_task(). | |
12127 | */ | |
12128 | if (READ_ONCE(p->__state) == TASK_NEW) | |
12129 | return; | |
12130 | ||
daa59407 | 12131 | detach_task_cfs_rq(p); |
6efdb105 BP |
12132 | |
12133 | #ifdef CONFIG_SMP | |
12134 | /* Tell se's cfs_rq has been changed -- migrated */ | |
12135 | p->se.avg.last_update_time = 0; | |
12136 | #endif | |
5d6da83c | 12137 | set_task_rq(p, task_cpu(p)); |
daa59407 | 12138 | attach_task_cfs_rq(p); |
810b3817 | 12139 | } |
029632fb PZ |
12140 | |
12141 | void free_fair_sched_group(struct task_group *tg) | |
12142 | { | |
12143 | int i; | |
12144 | ||
029632fb PZ |
12145 | for_each_possible_cpu(i) { |
12146 | if (tg->cfs_rq) | |
12147 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 12148 | if (tg->se) |
029632fb PZ |
12149 | kfree(tg->se[i]); |
12150 | } | |
12151 | ||
12152 | kfree(tg->cfs_rq); | |
12153 | kfree(tg->se); | |
12154 | } | |
12155 | ||
12156 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
12157 | { | |
029632fb | 12158 | struct sched_entity *se; |
b7fa30c9 | 12159 | struct cfs_rq *cfs_rq; |
029632fb PZ |
12160 | int i; |
12161 | ||
6396bb22 | 12162 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
12163 | if (!tg->cfs_rq) |
12164 | goto err; | |
6396bb22 | 12165 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
12166 | if (!tg->se) |
12167 | goto err; | |
12168 | ||
12169 | tg->shares = NICE_0_LOAD; | |
12170 | ||
12171 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
12172 | ||
12173 | for_each_possible_cpu(i) { | |
12174 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
12175 | GFP_KERNEL, cpu_to_node(i)); | |
12176 | if (!cfs_rq) | |
12177 | goto err; | |
12178 | ||
ceeadb83 | 12179 | se = kzalloc_node(sizeof(struct sched_entity_stats), |
029632fb PZ |
12180 | GFP_KERNEL, cpu_to_node(i)); |
12181 | if (!se) | |
12182 | goto err_free_rq; | |
12183 | ||
12184 | init_cfs_rq(cfs_rq); | |
12185 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 12186 | init_entity_runnable_average(se); |
029632fb PZ |
12187 | } |
12188 | ||
12189 | return 1; | |
12190 | ||
12191 | err_free_rq: | |
12192 | kfree(cfs_rq); | |
12193 | err: | |
12194 | return 0; | |
12195 | } | |
12196 | ||
8663e24d PZ |
12197 | void online_fair_sched_group(struct task_group *tg) |
12198 | { | |
12199 | struct sched_entity *se; | |
a46d14ec | 12200 | struct rq_flags rf; |
8663e24d PZ |
12201 | struct rq *rq; |
12202 | int i; | |
12203 | ||
12204 | for_each_possible_cpu(i) { | |
12205 | rq = cpu_rq(i); | |
12206 | se = tg->se[i]; | |
a46d14ec | 12207 | rq_lock_irq(rq, &rf); |
4126bad6 | 12208 | update_rq_clock(rq); |
d0326691 | 12209 | attach_entity_cfs_rq(se); |
55e16d30 | 12210 | sync_throttle(tg, i); |
a46d14ec | 12211 | rq_unlock_irq(rq, &rf); |
8663e24d PZ |
12212 | } |
12213 | } | |
12214 | ||
6fe1f348 | 12215 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 12216 | { |
029632fb | 12217 | unsigned long flags; |
6fe1f348 PZ |
12218 | struct rq *rq; |
12219 | int cpu; | |
029632fb | 12220 | |
b027789e MK |
12221 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); |
12222 | ||
6fe1f348 PZ |
12223 | for_each_possible_cpu(cpu) { |
12224 | if (tg->se[cpu]) | |
12225 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 12226 | |
6fe1f348 PZ |
12227 | /* |
12228 | * Only empty task groups can be destroyed; so we can speculatively | |
12229 | * check on_list without danger of it being re-added. | |
12230 | */ | |
12231 | if (!tg->cfs_rq[cpu]->on_list) | |
12232 | continue; | |
12233 | ||
12234 | rq = cpu_rq(cpu); | |
12235 | ||
5cb9eaa3 | 12236 | raw_spin_rq_lock_irqsave(rq, flags); |
6fe1f348 | 12237 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); |
5cb9eaa3 | 12238 | raw_spin_rq_unlock_irqrestore(rq, flags); |
6fe1f348 | 12239 | } |
029632fb PZ |
12240 | } |
12241 | ||
12242 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
12243 | struct sched_entity *se, int cpu, | |
12244 | struct sched_entity *parent) | |
12245 | { | |
12246 | struct rq *rq = cpu_rq(cpu); | |
12247 | ||
12248 | cfs_rq->tg = tg; | |
12249 | cfs_rq->rq = rq; | |
029632fb PZ |
12250 | init_cfs_rq_runtime(cfs_rq); |
12251 | ||
12252 | tg->cfs_rq[cpu] = cfs_rq; | |
12253 | tg->se[cpu] = se; | |
12254 | ||
12255 | /* se could be NULL for root_task_group */ | |
12256 | if (!se) | |
12257 | return; | |
12258 | ||
fed14d45 | 12259 | if (!parent) { |
029632fb | 12260 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
12261 | se->depth = 0; |
12262 | } else { | |
029632fb | 12263 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
12264 | se->depth = parent->depth + 1; |
12265 | } | |
029632fb PZ |
12266 | |
12267 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
12268 | /* guarantee group entities always have weight */ |
12269 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
12270 | se->parent = parent; |
12271 | } | |
12272 | ||
12273 | static DEFINE_MUTEX(shares_mutex); | |
12274 | ||
30400039 | 12275 | static int __sched_group_set_shares(struct task_group *tg, unsigned long shares) |
029632fb PZ |
12276 | { |
12277 | int i; | |
029632fb | 12278 | |
30400039 JD |
12279 | lockdep_assert_held(&shares_mutex); |
12280 | ||
029632fb PZ |
12281 | /* |
12282 | * We can't change the weight of the root cgroup. | |
12283 | */ | |
12284 | if (!tg->se[0]) | |
12285 | return -EINVAL; | |
12286 | ||
12287 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
12288 | ||
029632fb | 12289 | if (tg->shares == shares) |
30400039 | 12290 | return 0; |
029632fb PZ |
12291 | |
12292 | tg->shares = shares; | |
12293 | for_each_possible_cpu(i) { | |
12294 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
12295 | struct sched_entity *se = tg->se[i]; |
12296 | struct rq_flags rf; | |
029632fb | 12297 | |
029632fb | 12298 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 12299 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 12300 | update_rq_clock(rq); |
89ee048f | 12301 | for_each_sched_entity(se) { |
88c0616e | 12302 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 12303 | update_cfs_group(se); |
89ee048f | 12304 | } |
8a8c69c3 | 12305 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
12306 | } |
12307 | ||
30400039 JD |
12308 | return 0; |
12309 | } | |
12310 | ||
12311 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
12312 | { | |
12313 | int ret; | |
12314 | ||
12315 | mutex_lock(&shares_mutex); | |
12316 | if (tg_is_idle(tg)) | |
12317 | ret = -EINVAL; | |
12318 | else | |
12319 | ret = __sched_group_set_shares(tg, shares); | |
12320 | mutex_unlock(&shares_mutex); | |
12321 | ||
12322 | return ret; | |
12323 | } | |
12324 | ||
12325 | int sched_group_set_idle(struct task_group *tg, long idle) | |
12326 | { | |
12327 | int i; | |
12328 | ||
12329 | if (tg == &root_task_group) | |
12330 | return -EINVAL; | |
12331 | ||
12332 | if (idle < 0 || idle > 1) | |
12333 | return -EINVAL; | |
12334 | ||
12335 | mutex_lock(&shares_mutex); | |
12336 | ||
12337 | if (tg->idle == idle) { | |
12338 | mutex_unlock(&shares_mutex); | |
12339 | return 0; | |
12340 | } | |
12341 | ||
12342 | tg->idle = idle; | |
12343 | ||
12344 | for_each_possible_cpu(i) { | |
12345 | struct rq *rq = cpu_rq(i); | |
12346 | struct sched_entity *se = tg->se[i]; | |
a480adde | 12347 | struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i]; |
30400039 JD |
12348 | bool was_idle = cfs_rq_is_idle(grp_cfs_rq); |
12349 | long idle_task_delta; | |
12350 | struct rq_flags rf; | |
12351 | ||
12352 | rq_lock_irqsave(rq, &rf); | |
12353 | ||
12354 | grp_cfs_rq->idle = idle; | |
12355 | if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq))) | |
12356 | goto next_cpu; | |
12357 | ||
a480adde JD |
12358 | if (se->on_rq) { |
12359 | parent_cfs_rq = cfs_rq_of(se); | |
12360 | if (cfs_rq_is_idle(grp_cfs_rq)) | |
12361 | parent_cfs_rq->idle_nr_running++; | |
12362 | else | |
12363 | parent_cfs_rq->idle_nr_running--; | |
12364 | } | |
12365 | ||
30400039 JD |
12366 | idle_task_delta = grp_cfs_rq->h_nr_running - |
12367 | grp_cfs_rq->idle_h_nr_running; | |
12368 | if (!cfs_rq_is_idle(grp_cfs_rq)) | |
12369 | idle_task_delta *= -1; | |
12370 | ||
12371 | for_each_sched_entity(se) { | |
12372 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
12373 | ||
12374 | if (!se->on_rq) | |
12375 | break; | |
12376 | ||
12377 | cfs_rq->idle_h_nr_running += idle_task_delta; | |
12378 | ||
12379 | /* Already accounted at parent level and above. */ | |
12380 | if (cfs_rq_is_idle(cfs_rq)) | |
12381 | break; | |
12382 | } | |
12383 | ||
12384 | next_cpu: | |
12385 | rq_unlock_irqrestore(rq, &rf); | |
12386 | } | |
12387 | ||
12388 | /* Idle groups have minimum weight. */ | |
12389 | if (tg_is_idle(tg)) | |
12390 | __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO)); | |
12391 | else | |
12392 | __sched_group_set_shares(tg, NICE_0_LOAD); | |
12393 | ||
029632fb PZ |
12394 | mutex_unlock(&shares_mutex); |
12395 | return 0; | |
12396 | } | |
30400039 | 12397 | |
029632fb PZ |
12398 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
12399 | ||
12400 | void free_fair_sched_group(struct task_group *tg) { } | |
12401 | ||
12402 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
12403 | { | |
12404 | return 1; | |
12405 | } | |
12406 | ||
8663e24d PZ |
12407 | void online_fair_sched_group(struct task_group *tg) { } |
12408 | ||
6fe1f348 | 12409 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
12410 | |
12411 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
12412 | ||
810b3817 | 12413 | |
6d686f45 | 12414 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
12415 | { |
12416 | struct sched_entity *se = &task->se; | |
0d721cea PW |
12417 | unsigned int rr_interval = 0; |
12418 | ||
12419 | /* | |
12420 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
12421 | * idle runqueue: | |
12422 | */ | |
0d721cea | 12423 | if (rq->cfs.load.weight) |
a59f4e07 | 12424 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
12425 | |
12426 | return rr_interval; | |
12427 | } | |
12428 | ||
bf0f6f24 IM |
12429 | /* |
12430 | * All the scheduling class methods: | |
12431 | */ | |
43c31ac0 PZ |
12432 | DEFINE_SCHED_CLASS(fair) = { |
12433 | ||
bf0f6f24 IM |
12434 | .enqueue_task = enqueue_task_fair, |
12435 | .dequeue_task = dequeue_task_fair, | |
12436 | .yield_task = yield_task_fair, | |
d95f4122 | 12437 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 12438 | |
2e09bf55 | 12439 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 | 12440 | |
98c2f700 | 12441 | .pick_next_task = __pick_next_task_fair, |
bf0f6f24 | 12442 | .put_prev_task = put_prev_task_fair, |
03b7fad1 | 12443 | .set_next_task = set_next_task_fair, |
bf0f6f24 | 12444 | |
681f3e68 | 12445 | #ifdef CONFIG_SMP |
6e2df058 | 12446 | .balance = balance_fair, |
21f56ffe | 12447 | .pick_task = pick_task_fair, |
4ce72a2c | 12448 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 12449 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 12450 | |
0bcdcf28 CE |
12451 | .rq_online = rq_online_fair, |
12452 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 12453 | |
12695578 | 12454 | .task_dead = task_dead_fair, |
c5b28038 | 12455 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 12456 | #endif |
bf0f6f24 | 12457 | |
bf0f6f24 | 12458 | .task_tick = task_tick_fair, |
cd29fe6f | 12459 | .task_fork = task_fork_fair, |
cb469845 SR |
12460 | |
12461 | .prio_changed = prio_changed_fair, | |
da7a735e | 12462 | .switched_from = switched_from_fair, |
cb469845 | 12463 | .switched_to = switched_to_fair, |
810b3817 | 12464 | |
0d721cea PW |
12465 | .get_rr_interval = get_rr_interval_fair, |
12466 | ||
6e998916 SG |
12467 | .update_curr = update_curr_fair, |
12468 | ||
810b3817 | 12469 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 12470 | .task_change_group = task_change_group_fair, |
810b3817 | 12471 | #endif |
982d9cdc PB |
12472 | |
12473 | #ifdef CONFIG_UCLAMP_TASK | |
12474 | .uclamp_enabled = 1, | |
12475 | #endif | |
bf0f6f24 IM |
12476 | }; |
12477 | ||
12478 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 12479 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 12480 | { |
039ae8bc | 12481 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 12482 | |
5973e5b9 | 12483 | rcu_read_lock(); |
039ae8bc | 12484 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 12485 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 12486 | rcu_read_unlock(); |
bf0f6f24 | 12487 | } |
397f2378 SD |
12488 | |
12489 | #ifdef CONFIG_NUMA_BALANCING | |
12490 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
12491 | { | |
12492 | int node; | |
12493 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 12494 | struct numa_group *ng; |
397f2378 | 12495 | |
cb361d8c JH |
12496 | rcu_read_lock(); |
12497 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
12498 | for_each_online_node(node) { |
12499 | if (p->numa_faults) { | |
12500 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
12501 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
12502 | } | |
cb361d8c JH |
12503 | if (ng) { |
12504 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
12505 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
12506 | } |
12507 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
12508 | } | |
cb361d8c | 12509 | rcu_read_unlock(); |
397f2378 SD |
12510 | } |
12511 | #endif /* CONFIG_NUMA_BALANCING */ | |
12512 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
12513 | |
12514 | __init void init_sched_fair_class(void) | |
12515 | { | |
12516 | #ifdef CONFIG_SMP | |
18c31c97 BH |
12517 | int i; |
12518 | ||
12519 | for_each_possible_cpu(i) { | |
12520 | zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i)); | |
12521 | zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i)); | |
12522 | } | |
12523 | ||
029632fb PZ |
12524 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); |
12525 | ||
3451d024 | 12526 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 12527 | nohz.next_balance = jiffies; |
f643ea22 | 12528 | nohz.next_blocked = jiffies; |
029632fb | 12529 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
12530 | #endif |
12531 | #endif /* SMP */ | |
12532 | ||
12533 | } |