Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0
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>
19 : *
20 : * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 : * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
22 : */
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>
39 : #include <linux/sched/nohz.h>
40 :
41 : #include <linux/cpuidle.h>
42 : #include <linux/interrupt.h>
43 : #include <linux/memory-tiers.h>
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>
49 : #include <linux/task_work.h>
50 :
51 : #include <asm/switch_to.h>
52 :
53 : #include <linux/sched/cond_resched.h>
54 :
55 : #include "sched.h"
56 : #include "stats.h"
57 : #include "autogroup.h"
58 :
59 : /*
60 : * Targeted preemption latency for CPU-bound tasks:
61 : *
62 : * NOTE: this latency value is not the same as the concept of
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.
66 : *
67 : * (to see the precise effective timeslice length of your workload,
68 : * run vmstat and monitor the context-switches (cs) field)
69 : *
70 : * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
71 : */
72 : unsigned int sysctl_sched_latency = 6000000ULL;
73 : static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
74 :
75 : /*
76 : * The initial- and re-scaling of tunables is configurable
77 : *
78 : * Options are:
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))
85 : */
86 : unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
87 :
88 : /*
89 : * Minimal preemption granularity for CPU-bound tasks:
90 : *
91 : * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
92 : */
93 : unsigned int sysctl_sched_min_granularity = 750000ULL;
94 : static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
95 :
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 :
104 : /*
105 : * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
106 : */
107 : static unsigned int sched_nr_latency = 8;
108 :
109 : /*
110 : * After fork, child runs first. If set to 0 (default) then
111 : * parent will (try to) run first.
112 : */
113 : unsigned int sysctl_sched_child_runs_first __read_mostly;
114 :
115 : /*
116 : * SCHED_OTHER wake-up granularity.
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.
121 : *
122 : * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
123 : */
124 : unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
125 : static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
126 :
127 : const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
128 :
129 : int sched_thermal_decay_shift;
130 0 : static int __init setup_sched_thermal_decay_shift(char *str)
131 : {
132 0 : int _shift = 0;
133 :
134 0 : if (kstrtoint(str, 0, &_shift))
135 0 : pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
136 :
137 0 : sched_thermal_decay_shift = clamp(_shift, 0, 10);
138 0 : return 1;
139 : }
140 : __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
141 :
142 : #ifdef CONFIG_SMP
143 : /*
144 : * For asym packing, by default the lower numbered CPU has higher priority.
145 : */
146 : int __weak arch_asym_cpu_priority(int cpu)
147 : {
148 : return -cpu;
149 : }
150 :
151 : /*
152 : * The margin used when comparing utilization with CPU capacity.
153 : *
154 : * (default: ~20%)
155 : */
156 : #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
157 :
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)
165 : #endif
166 :
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 : *
176 : * (default: 5 msec, units: microseconds)
177 : */
178 : static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
179 : #endif
180 :
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 :
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
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 */
215 : {}
216 : };
217 :
218 1 : static int __init sched_fair_sysctl_init(void)
219 : {
220 1 : register_sysctl_init("kernel", sched_fair_sysctls);
221 1 : return 0;
222 : }
223 : late_initcall(sched_fair_sysctl_init);
224 : #endif
225 :
226 : static inline void update_load_add(struct load_weight *lw, unsigned long inc)
227 : {
228 2828 : lw->weight += inc;
229 2828 : lw->inv_weight = 0;
230 : }
231 :
232 : static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
233 : {
234 2444 : lw->weight -= dec;
235 2444 : lw->inv_weight = 0;
236 : }
237 :
238 : static inline void update_load_set(struct load_weight *lw, unsigned long w)
239 : {
240 5 : lw->weight = w;
241 5 : lw->inv_weight = 0;
242 : }
243 :
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 : */
253 : static unsigned int get_update_sysctl_factor(void)
254 : {
255 1 : unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
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 1 : unsigned int factor = get_update_sysctl_factor();
277 :
278 : #define SET_SYSCTL(name) \
279 : (sysctl_##name = (factor) * normalized_sysctl_##name)
280 1 : SET_SYSCTL(sched_min_granularity);
281 1 : SET_SYSCTL(sched_latency);
282 1 : SET_SYSCTL(sched_wakeup_granularity);
283 : #undef SET_SYSCTL
284 : }
285 :
286 1 : void __init sched_init_granularity(void)
287 : {
288 : update_sysctl();
289 1 : }
290 :
291 : #define WMULT_CONST (~0U)
292 : #define WMULT_SHIFT 32
293 :
294 : static void __update_inv_weight(struct load_weight *lw)
295 : {
296 : unsigned long w;
297 :
298 617 : if (likely(lw->inv_weight))
299 : return;
300 :
301 433 : w = scale_load_down(lw->weight);
302 :
303 433 : if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
304 0 : lw->inv_weight = 1;
305 433 : else if (unlikely(!w))
306 0 : lw->inv_weight = WMULT_CONST;
307 : else
308 433 : lw->inv_weight = WMULT_CONST / w;
309 : }
310 :
311 : /*
312 : * delta_exec * weight / lw.weight
313 : * OR
314 : * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
315 : *
316 : * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
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.
322 : */
323 617 : static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
324 : {
325 617 : u64 fact = scale_load_down(weight);
326 617 : u32 fact_hi = (u32)(fact >> 32);
327 617 : int shift = WMULT_SHIFT;
328 : int fs;
329 :
330 617 : __update_inv_weight(lw);
331 :
332 617 : if (unlikely(fact_hi)) {
333 0 : fs = fls(fact_hi);
334 0 : shift -= fs;
335 0 : fact >>= fs;
336 : }
337 :
338 1234 : fact = mul_u32_u32(fact, lw->inv_weight);
339 :
340 617 : fact_hi = (u32)(fact >> 32);
341 617 : if (fact_hi) {
342 0 : fs = fls(fact_hi);
343 0 : shift -= fs;
344 0 : fact >>= fs;
345 : }
346 :
347 1234 : return mul_u64_u32_shr(delta_exec, fact, shift);
348 : }
349 :
350 :
351 : const struct sched_class fair_sched_class;
352 :
353 : /**************************************************************
354 : * CFS operations on generic schedulable entities:
355 : */
356 :
357 : #ifdef CONFIG_FAIR_GROUP_SCHED
358 :
359 : /* Walk up scheduling entities hierarchy */
360 : #define for_each_sched_entity(se) \
361 : for (; se; se = se->parent)
362 :
363 : static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
364 : {
365 : struct rq *rq = rq_of(cfs_rq);
366 : int cpu = cpu_of(rq);
367 :
368 : if (cfs_rq->on_list)
369 : return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
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) {
384 : /*
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.
389 : */
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;
398 : return true;
399 : }
400 :
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;
413 : return true;
414 : }
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;
428 : return false;
429 : }
430 :
431 : static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
432 : {
433 : if (cfs_rq->on_list) {
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 :
446 : list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
447 : cfs_rq->on_list = 0;
448 : }
449 : }
450 :
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 :
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)
460 :
461 : /* Do the two (enqueued) entities belong to the same group ? */
462 : static inline struct cfs_rq *
463 : is_same_group(struct sched_entity *se, struct sched_entity *pse)
464 : {
465 : if (se->cfs_rq == pse->cfs_rq)
466 : return se->cfs_rq;
467 :
468 : return NULL;
469 : }
470 :
471 : static inline struct sched_entity *parent_entity(const struct sched_entity *se)
472 : {
473 : return se->parent;
474 : }
475 :
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 */
489 : se_depth = (*se)->depth;
490 : pse_depth = (*pse)->depth;
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 :
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 :
525 : #else /* !CONFIG_FAIR_GROUP_SCHED */
526 :
527 : #define for_each_sched_entity(se) \
528 : for (; se; se = NULL)
529 :
530 : static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
531 : {
532 : return true;
533 : }
534 :
535 : static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
536 : {
537 : }
538 :
539 : static inline void assert_list_leaf_cfs_rq(struct rq *rq)
540 : {
541 : }
542 :
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)
545 :
546 : static inline struct sched_entity *parent_entity(struct sched_entity *se)
547 : {
548 : return NULL;
549 : }
550 :
551 : static inline void
552 : find_matching_se(struct sched_entity **se, struct sched_entity **pse)
553 : {
554 : }
555 :
556 : static inline int tg_is_idle(struct task_group *tg)
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 :
571 : #endif /* CONFIG_FAIR_GROUP_SCHED */
572 :
573 : static __always_inline
574 : void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
575 :
576 : /**************************************************************
577 : * Scheduling class tree data structure manipulation methods:
578 : */
579 :
580 : static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
581 : {
582 7812 : s64 delta = (s64)(vruntime - max_vruntime);
583 7812 : if (delta > 0)
584 4759 : max_vruntime = vruntime;
585 :
586 : return max_vruntime;
587 : }
588 :
589 : static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
590 : {
591 235 : s64 delta = (s64)(vruntime - min_vruntime);
592 235 : if (delta < 0)
593 235 : min_vruntime = vruntime;
594 :
595 : return min_vruntime;
596 : }
597 :
598 : static inline bool entity_before(const struct sched_entity *a,
599 : const struct sched_entity *b)
600 : {
601 981 : return (s64)(a->vruntime - b->vruntime) < 0;
602 : }
603 :
604 : #define __node_2_se(node) \
605 : rb_entry((node), struct sched_entity, run_node)
606 :
607 5370 : static void update_min_vruntime(struct cfs_rq *cfs_rq)
608 : {
609 5370 : struct sched_entity *curr = cfs_rq->curr;
610 5370 : struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
611 :
612 5370 : u64 vruntime = cfs_rq->min_vruntime;
613 :
614 5370 : if (curr) {
615 5370 : if (curr->on_rq)
616 2930 : vruntime = curr->vruntime;
617 : else
618 : curr = NULL;
619 : }
620 :
621 5370 : if (leftmost) { /* non-empty tree */
622 2674 : struct sched_entity *se = __node_2_se(leftmost);
623 :
624 2674 : if (!curr)
625 2439 : vruntime = se->vruntime;
626 : else
627 235 : vruntime = min_vruntime(vruntime, se->vruntime);
628 : }
629 :
630 : /* ensure we never gain time by being placed backwards. */
631 10740 : u64_u32_store(cfs_rq->min_vruntime,
632 : max_vruntime(cfs_rq->min_vruntime, vruntime));
633 5370 : }
634 :
635 : static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
636 : {
637 981 : return entity_before(__node_2_se(a), __node_2_se(b));
638 : }
639 :
640 : /*
641 : * Enqueue an entity into the rb-tree:
642 : */
643 2517 : static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
644 : {
645 5034 : rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
646 2517 : }
647 :
648 : static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
649 : {
650 2516 : rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
651 : }
652 :
653 0 : struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
654 : {
655 2512 : struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
656 :
657 2512 : if (!left)
658 : return NULL;
659 :
660 2512 : return __node_2_se(left);
661 : }
662 :
663 : static struct sched_entity *__pick_next_entity(struct sched_entity *se)
664 : {
665 0 : struct rb_node *next = rb_next(&se->run_node);
666 :
667 0 : if (!next)
668 : return NULL;
669 :
670 0 : return __node_2_se(next);
671 : }
672 :
673 : #ifdef CONFIG_SCHED_DEBUG
674 : struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
675 : {
676 : struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
677 :
678 : if (!last)
679 : return NULL;
680 :
681 : return __node_2_se(last);
682 : }
683 :
684 : /**************************************************************
685 : * Scheduling class statistics methods:
686 : */
687 :
688 : int sched_update_scaling(void)
689 : {
690 : unsigned int factor = get_update_sysctl_factor();
691 :
692 : sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
693 : sysctl_sched_min_granularity);
694 :
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);
700 : #undef WRT_SYSCTL
701 :
702 : return 0;
703 : }
704 : #endif
705 :
706 : /*
707 : * delta /= w
708 : */
709 : static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
710 : {
711 4150 : if (unlikely(se->load.weight != NICE_0_LOAD))
712 0 : delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
713 :
714 : return delta;
715 : }
716 :
717 : /*
718 : * The idea is to set a period in which each task runs once.
719 : *
720 : * When there are too many tasks (sched_nr_latency) we have to stretch
721 : * this period because otherwise the slices get too small.
722 : *
723 : * p = (nr <= nl) ? l : l*nr/nl
724 : */
725 : static u64 __sched_period(unsigned long nr_running)
726 : {
727 617 : if (unlikely(nr_running > sched_nr_latency))
728 0 : return nr_running * sysctl_sched_min_granularity;
729 : else
730 617 : return sysctl_sched_latency;
731 : }
732 :
733 : static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq);
734 :
735 : /*
736 : * We calculate the wall-time slice from the period by taking a part
737 : * proportional to the weight.
738 : *
739 : * s = p*P[w/rw]
740 : */
741 617 : static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 : {
743 617 : unsigned int nr_running = cfs_rq->nr_running;
744 617 : struct sched_entity *init_se = se;
745 : unsigned int min_gran;
746 : u64 slice;
747 :
748 : if (sched_feat(ALT_PERIOD))
749 617 : nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
750 :
751 617 : slice = __sched_period(nr_running + !se->on_rq);
752 :
753 617 : for_each_sched_entity(se) {
754 : struct load_weight *load;
755 : struct load_weight lw;
756 : struct cfs_rq *qcfs_rq;
757 :
758 1234 : qcfs_rq = cfs_rq_of(se);
759 617 : load = &qcfs_rq->load;
760 :
761 617 : if (unlikely(!se->on_rq)) {
762 382 : lw = qcfs_rq->load;
763 :
764 764 : update_load_add(&lw, se->load.weight);
765 382 : load = &lw;
766 : }
767 617 : slice = __calc_delta(slice, se->load.weight, load);
768 : }
769 :
770 : if (sched_feat(BASE_SLICE)) {
771 617 : if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq))
772 : min_gran = sysctl_sched_idle_min_granularity;
773 : else
774 617 : min_gran = sysctl_sched_min_granularity;
775 :
776 617 : slice = max_t(u64, slice, min_gran);
777 : }
778 :
779 617 : return slice;
780 : }
781 :
782 : /*
783 : * We calculate the vruntime slice of a to-be-inserted task.
784 : *
785 : * vs = s/w
786 : */
787 382 : static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
788 : {
789 764 : return calc_delta_fair(sched_slice(cfs_rq, se), se);
790 : }
791 :
792 : #include "pelt.h"
793 : #ifdef CONFIG_SMP
794 :
795 : static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
796 : static unsigned long task_h_load(struct task_struct *p);
797 : static unsigned long capacity_of(int cpu);
798 :
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)
801 : {
802 : struct sched_avg *sa = &se->avg;
803 :
804 : memset(sa, 0, sizeof(*sa));
805 :
806 : /*
807 : * Tasks are initialized with full load to be seen as heavy tasks until
808 : * they get a chance to stabilize to their real load level.
809 : * Group entities are initialized with zero load to reflect the fact that
810 : * nothing has been attached to the task group yet.
811 : */
812 : if (entity_is_task(se))
813 : sa->load_avg = scale_load_down(se->load.weight);
814 :
815 : /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
816 : }
817 :
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 : *
831 : * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
832 : *
833 : * where n denotes the nth task and cpu_scale the CPU capacity.
834 : *
835 : * For example, for a CPU with 1024 of capacity, a simplest series from
836 : * the beginning would be like:
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 : */
844 : void post_init_entity_util_avg(struct task_struct *p)
845 : {
846 : struct sched_entity *se = &p->se;
847 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
848 : struct sched_avg *sa = &se->avg;
849 : long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
850 : long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
851 :
852 : if (p->sched_class != &fair_sched_class) {
853 : /*
854 : * For !fair tasks do:
855 : *
856 : update_cfs_rq_load_avg(now, cfs_rq);
857 : attach_entity_load_avg(cfs_rq, se);
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;
865 : }
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;
880 : }
881 :
882 : #else /* !CONFIG_SMP */
883 382 : void init_entity_runnable_average(struct sched_entity *se)
884 : {
885 382 : }
886 382 : void post_init_entity_util_avg(struct task_struct *p)
887 : {
888 382 : }
889 : static void update_tg_load_avg(struct cfs_rq *cfs_rq)
890 : {
891 : }
892 : #endif /* CONFIG_SMP */
893 :
894 : /*
895 : * Update the current task's runtime statistics.
896 : */
897 10298 : static void update_curr(struct cfs_rq *cfs_rq)
898 : {
899 10298 : struct sched_entity *curr = cfs_rq->curr;
900 20596 : u64 now = rq_clock_task(rq_of(cfs_rq));
901 : u64 delta_exec;
902 :
903 10298 : if (unlikely(!curr))
904 : return;
905 :
906 10291 : delta_exec = now - curr->exec_start;
907 10291 : if (unlikely((s64)delta_exec <= 0))
908 : return;
909 :
910 2930 : curr->exec_start = now;
911 :
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 : }
919 :
920 2930 : curr->sum_exec_runtime += delta_exec;
921 : schedstat_add(cfs_rq->exec_clock, delta_exec);
922 :
923 2930 : curr->vruntime += calc_delta_fair(delta_exec, curr);
924 2930 : update_min_vruntime(cfs_rq);
925 :
926 : if (entity_is_task(curr)) {
927 2930 : struct task_struct *curtask = task_of(curr);
928 :
929 2930 : trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
930 2930 : cgroup_account_cputime(curtask, delta_exec);
931 : account_group_exec_runtime(curtask, delta_exec);
932 : }
933 :
934 : account_cfs_rq_runtime(cfs_rq, delta_exec);
935 : }
936 :
937 0 : static void update_curr_fair(struct rq *rq)
938 : {
939 0 : update_curr(cfs_rq_of(&rq->curr->se));
940 0 : }
941 :
942 : static inline void
943 : update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
944 : {
945 : struct sched_statistics *stats;
946 71 : struct task_struct *p = NULL;
947 :
948 : if (!schedstat_enabled())
949 : return;
950 :
951 : stats = __schedstats_from_se(se);
952 :
953 : if (entity_is_task(se))
954 : p = task_of(se);
955 :
956 : __update_stats_wait_start(rq_of(cfs_rq), p, stats);
957 : }
958 :
959 : static inline void
960 : update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
961 : {
962 : struct sched_statistics *stats;
963 2516 : struct task_struct *p = NULL;
964 :
965 : if (!schedstat_enabled())
966 : return;
967 :
968 : stats = __schedstats_from_se(se);
969 :
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 : */
976 : if (unlikely(!schedstat_val(stats->wait_start)))
977 : return;
978 :
979 : if (entity_is_task(se))
980 : p = task_of(se);
981 :
982 : __update_stats_wait_end(rq_of(cfs_rq), p, stats);
983 : }
984 :
985 : static inline void
986 : update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
987 : {
988 : struct sched_statistics *stats;
989 : struct task_struct *tsk = NULL;
990 :
991 : if (!schedstat_enabled())
992 : return;
993 :
994 : stats = __schedstats_from_se(se);
995 :
996 : if (entity_is_task(se))
997 : tsk = task_of(se);
998 :
999 : __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1000 : }
1001 :
1002 : /*
1003 : * Task is being enqueued - update stats:
1004 : */
1005 : static inline void
1006 : update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1007 : {
1008 : if (!schedstat_enabled())
1009 : return;
1010 :
1011 : /*
1012 : * Are we enqueueing a waiting task? (for current tasks
1013 : * a dequeue/enqueue event is a NOP)
1014 : */
1015 : if (se != cfs_rq->curr)
1016 : update_stats_wait_start_fair(cfs_rq, se);
1017 :
1018 : if (flags & ENQUEUE_WAKEUP)
1019 : update_stats_enqueue_sleeper_fair(cfs_rq, se);
1020 : }
1021 :
1022 : static inline void
1023 : update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1024 : {
1025 :
1026 : if (!schedstat_enabled())
1027 : return;
1028 :
1029 : /*
1030 : * Mark the end of the wait period if dequeueing a
1031 : * waiting task:
1032 : */
1033 : if (se != cfs_rq->curr)
1034 : update_stats_wait_end_fair(cfs_rq, se);
1035 :
1036 : if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1037 : struct task_struct *tsk = task_of(se);
1038 : unsigned int state;
1039 :
1040 : /* XXX racy against TTWU */
1041 : state = READ_ONCE(tsk->__state);
1042 : if (state & TASK_INTERRUPTIBLE)
1043 : __schedstat_set(tsk->stats.sleep_start,
1044 : rq_clock(rq_of(cfs_rq)));
1045 : if (state & TASK_UNINTERRUPTIBLE)
1046 : __schedstat_set(tsk->stats.block_start,
1047 : rq_clock(rq_of(cfs_rq)));
1048 : }
1049 : }
1050 :
1051 : /*
1052 : * We are picking a new current task - update its stats:
1053 : */
1054 : static inline void
1055 : update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1056 : {
1057 : /*
1058 : * We are starting a new run period:
1059 : */
1060 5032 : se->exec_start = rq_clock_task(rq_of(cfs_rq));
1061 : }
1062 :
1063 : /**************************************************
1064 : * Scheduling class queueing methods:
1065 : */
1066 :
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 :
1094 : #ifdef CONFIG_NUMA_BALANCING
1095 : /*
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.
1099 : */
1100 : unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1101 : unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1102 :
1103 : /* Portion of address space to scan in MB */
1104 : unsigned int sysctl_numa_balancing_scan_size = 256;
1105 :
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 :
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 :
1112 : struct numa_group {
1113 : refcount_t refcount;
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 : /*
1124 : * faults[] array is split into two regions: faults_mem and faults_cpu.
1125 : *
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 : */
1130 : unsigned long faults[];
1131 : };
1132 :
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 ||
1140 : (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
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 :
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 :
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 :
1170 : /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1171 : #define MAX_SCAN_WINDOW 2560
1172 :
1173 : static unsigned int task_scan_min(struct task_struct *p)
1174 : {
1175 : unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1176 : unsigned int scan, floor;
1177 : unsigned int windows = 1;
1178 :
1179 : if (scan_size < MAX_SCAN_WINDOW)
1180 : windows = MAX_SCAN_WINDOW / scan_size;
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 :
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;
1191 : struct numa_group *ng;
1192 :
1193 : /* Scale the maximum scan period with the amount of shared memory. */
1194 : rcu_read_lock();
1195 : ng = rcu_dereference(p->numa_group);
1196 : if (ng) {
1197 : unsigned long shared = group_faults_shared(ng);
1198 : unsigned long private = group_faults_priv(ng);
1199 :
1200 : period *= refcount_read(&ng->refcount);
1201 : period *= shared + 1;
1202 : period /= private + shared + 1;
1203 : }
1204 : rcu_read_unlock();
1205 :
1206 : return max(smin, period);
1207 : }
1208 :
1209 : static unsigned int task_scan_max(struct task_struct *p)
1210 : {
1211 : unsigned long smin = task_scan_min(p);
1212 : unsigned long smax;
1213 : struct numa_group *ng;
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);
1217 :
1218 : /* Scale the maximum scan period with the amount of shared memory. */
1219 : ng = deref_curr_numa_group(p);
1220 : if (ng) {
1221 : unsigned long shared = group_faults_shared(ng);
1222 : unsigned long private = group_faults_priv(ng);
1223 : unsigned long period = smax;
1224 :
1225 : period *= refcount_read(&ng->refcount);
1226 : period *= shared + 1;
1227 : period /= private + shared + 1;
1228 :
1229 : smax = max(smax, period);
1230 : }
1231 :
1232 : return max(smin, smax);
1233 : }
1234 :
1235 : static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1236 : {
1237 : rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
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 : {
1243 : rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1244 : rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1245 : }
1246 :
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 :
1256 : pid_t task_numa_group_id(struct task_struct *p)
1257 : {
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;
1268 : }
1269 :
1270 : /*
1271 : * The averaged statistics, shared & private, memory & CPU,
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)
1277 : {
1278 : return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1279 : }
1280 :
1281 : static inline unsigned long task_faults(struct task_struct *p, int nid)
1282 : {
1283 : if (!p->numa_faults)
1284 : return 0;
1285 :
1286 : return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1287 : p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1288 : }
1289 :
1290 : static inline unsigned long group_faults(struct task_struct *p, int nid)
1291 : {
1292 : struct numa_group *ng = deref_task_numa_group(p);
1293 :
1294 : if (!ng)
1295 : return 0;
1296 :
1297 : return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1298 : ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1299 : }
1300 :
1301 : static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1302 : {
1303 : return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1304 : group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1305 : }
1306 :
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 :
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 :
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,
1345 : int lim_dist, bool task)
1346 : {
1347 : unsigned long score = 0;
1348 : int node, max_dist;
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 :
1357 : /* sched_max_numa_distance may be changed in parallel. */
1358 : max_dist = READ_ONCE(sched_max_numa_distance);
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 : */
1371 : if (dist >= max_dist || node == nid)
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 : */
1381 : if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
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) {
1399 : faults *= (max_dist - dist);
1400 : faults /= (max_dist - LOCAL_DISTANCE);
1401 : }
1402 :
1403 : score += faults;
1404 : }
1405 :
1406 : return score;
1407 : }
1408 :
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 : */
1415 : static inline unsigned long task_weight(struct task_struct *p, int nid,
1416 : int dist)
1417 : {
1418 : unsigned long faults, total_faults;
1419 :
1420 : if (!p->numa_faults)
1421 : return 0;
1422 :
1423 : total_faults = p->total_numa_faults;
1424 :
1425 : if (!total_faults)
1426 : return 0;
1427 :
1428 : faults = task_faults(p, nid);
1429 : faults += score_nearby_nodes(p, nid, dist, true);
1430 :
1431 : return 1000 * faults / total_faults;
1432 : }
1433 :
1434 : static inline unsigned long group_weight(struct task_struct *p, int nid,
1435 : int dist)
1436 : {
1437 : struct numa_group *ng = deref_task_numa_group(p);
1438 : unsigned long faults, total_faults;
1439 :
1440 : if (!ng)
1441 : return 0;
1442 :
1443 : total_faults = ng->total_faults;
1444 :
1445 : if (!total_faults)
1446 : return 0;
1447 :
1448 : faults = group_faults(p, nid);
1449 : faults += score_nearby_nodes(p, nid, dist, false);
1450 :
1451 : return 1000 * faults / total_faults;
1452 : }
1453 :
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 :
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 :
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 :
1568 : bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1569 : int src_nid, int dst_cpu)
1570 : {
1571 : struct numa_group *ng = deref_curr_numa_group(p);
1572 : int dst_nid = cpu_to_node(dst_cpu);
1573 : int last_cpupid, this_cpupid;
1574 :
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;
1582 : unsigned long rate_limit;
1583 : unsigned int latency, th, def_th;
1584 :
1585 : pgdat = NODE_DATA(dst_nid);
1586 : if (pgdat_free_space_enough(pgdat)) {
1587 : /* workload changed, reset hot threshold */
1588 : pgdat->nbp_threshold = 0;
1589 : return true;
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);
1596 :
1597 : th = pgdat->nbp_threshold ? : def_th;
1598 : latency = numa_hint_fault_latency(page);
1599 : if (latency >= th)
1600 : return false;
1601 :
1602 : return !numa_promotion_rate_limit(pgdat, rate_limit,
1603 : thp_nr_pages(page));
1604 : }
1605 :
1606 : this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1607 : last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1608 :
1609 : if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1610 : !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1611 : return false;
1612 :
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 : */
1619 : if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1620 : (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1621 : return true;
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 : */
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 : /*
1653 : * Destination node is much more heavily used than the source
1654 : * node? Allow migration.
1655 : */
1656 : if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1657 : ACTIVE_NODE_FRACTION)
1658 : return true;
1659 :
1660 : /*
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)
1667 : */
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;
1670 : }
1671 :
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 : };
1689 :
1690 : /* Cached statistics for all CPUs within a node */
1691 : struct numa_stats {
1692 : unsigned long load;
1693 : unsigned long runnable;
1694 : unsigned long util;
1695 : /* Total compute capacity of CPUs on a node */
1696 : unsigned long compute_capacity;
1697 : unsigned int nr_running;
1698 : unsigned int weight;
1699 : enum numa_type node_type;
1700 : int idle_cpu;
1701 : };
1702 :
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 :
1712 : if (!idle_cpu(sibling))
1713 : return false;
1714 : }
1715 : #endif
1716 :
1717 : return true;
1718 : }
1719 :
1720 : struct task_numa_env {
1721 : struct task_struct *p;
1722 :
1723 : int src_cpu, src_nid;
1724 : int dst_cpu, dst_nid;
1725 : int imb_numa_nr;
1726 :
1727 : struct numa_stats src_stats, dst_stats;
1728 :
1729 : int imbalance_pct;
1730 : int dist;
1731 :
1732 : struct task_struct *best_task;
1733 : long best_imp;
1734 : int best_cpu;
1735 : };
1736 :
1737 : static unsigned long cpu_load(struct rq *rq);
1738 : static unsigned long cpu_runnable(struct rq *rq);
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) &&
1745 : (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1746 : ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1747 : return node_overloaded;
1748 :
1749 : if ((ns->nr_running < ns->weight) ||
1750 : (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1751 : ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1752 : return node_has_spare;
1753 :
1754 : return node_fully_busy;
1755 : }
1756 :
1757 : #ifdef CONFIG_SCHED_SMT
1758 : /* Forward declarations of select_idle_sibling helpers */
1759 : static inline bool test_idle_cores(int cpu);
1760 : static inline int numa_idle_core(int idle_core, int cpu)
1761 : {
1762 : if (!static_branch_likely(&sched_smt_present) ||
1763 : idle_core >= 0 || !test_idle_cores(cpu))
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;
1772 :
1773 : return idle_core;
1774 : }
1775 : #else
1776 : static inline int numa_idle_core(int idle_core, int cpu)
1777 : {
1778 : return idle_core;
1779 : }
1780 : #endif
1781 :
1782 : /*
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.
1787 : */
1788 : static void update_numa_stats(struct task_numa_env *env,
1789 : struct numa_stats *ns, int nid,
1790 : bool find_idle)
1791 : {
1792 : int cpu, idle_core = -1;
1793 :
1794 : memset(ns, 0, sizeof(*ns));
1795 : ns->idle_cpu = -1;
1796 :
1797 : rcu_read_lock();
1798 : for_each_cpu(cpu, cpumask_of_node(nid)) {
1799 : struct rq *rq = cpu_rq(cpu);
1800 :
1801 : ns->load += cpu_load(rq);
1802 : ns->runnable += cpu_runnable(rq);
1803 : ns->util += cpu_util_cfs(cpu);
1804 : ns->nr_running += rq->cfs.h_nr_running;
1805 : ns->compute_capacity += capacity_of(cpu);
1806 :
1807 : if (find_idle && idle_core < 0 && !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 : }
1817 : }
1818 : rcu_read_unlock();
1819 :
1820 : ns->weight = cpumask_weight(cpumask_of_node(nid));
1821 :
1822 : ns->node_type = numa_classify(env->imbalance_pct, ns);
1823 :
1824 : if (idle_core >= 0)
1825 : ns->idle_cpu = idle_core;
1826 : }
1827 :
1828 : static void task_numa_assign(struct task_numa_env *env,
1829 : struct task_struct *p, long imp)
1830 : {
1831 : struct rq *rq = cpu_rq(env->dst_cpu);
1832 :
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 + 1) {
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 */
1852 : return;
1853 : }
1854 :
1855 : assign:
1856 : /*
1857 : * Clear previous best_cpu/rq numa-migrate flag, since task now
1858 : * found a better CPU to move/swap.
1859 : */
1860 : if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1861 : rq = cpu_rq(env->best_cpu);
1862 : WRITE_ONCE(rq->numa_migrate_on, 0);
1863 : }
1864 :
1865 : if (env->best_task)
1866 : put_task_struct(env->best_task);
1867 : if (p)
1868 : get_task_struct(p);
1869 :
1870 : env->best_task = p;
1871 : env->best_imp = imp;
1872 : env->best_cpu = env->dst_cpu;
1873 : }
1874 :
1875 : static bool load_too_imbalanced(long src_load, long dst_load,
1876 : struct task_numa_env *env)
1877 : {
1878 : long imb, old_imb;
1879 : long orig_src_load, orig_dst_load;
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;
1891 :
1892 : imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1893 :
1894 : orig_src_load = env->src_stats.load;
1895 : orig_dst_load = env->dst_stats.load;
1896 :
1897 : old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1898 :
1899 : /* Would this change make things worse? */
1900 : return (imb > old_imb);
1901 : }
1902 :
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 :
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 : */
1916 : static bool task_numa_compare(struct task_numa_env *env,
1917 : long taskimp, long groupimp, bool maymove)
1918 : {
1919 : struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1920 : struct rq *dst_rq = cpu_rq(env->dst_cpu);
1921 : long imp = p_ng ? groupimp : taskimp;
1922 : struct task_struct *cur;
1923 : long src_load, dst_load;
1924 : int dist = env->dist;
1925 : long moveimp = imp;
1926 : long load;
1927 : bool stopsearch = false;
1928 :
1929 : if (READ_ONCE(dst_rq->numa_migrate_on))
1930 : return false;
1931 :
1932 : rcu_read_lock();
1933 : cur = rcu_dereference(dst_rq->curr);
1934 : if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1935 : cur = NULL;
1936 :
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 : */
1941 : if (cur == env->p) {
1942 : stopsearch = true;
1943 : goto unlock;
1944 : }
1945 :
1946 : if (!cur) {
1947 : if (maymove && moveimp >= env->best_imp)
1948 : goto assign;
1949 : else
1950 : goto unlock;
1951 : }
1952 :
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 :
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
1971 : * the value is, the more remote accesses that would be expected to
1972 : * be incurred if the tasks were swapped.
1973 : *
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 : */
1977 : cur_ng = rcu_dereference(cur->numa_group);
1978 : if (cur_ng == p_ng) {
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 :
1988 : imp = taskimp + task_weight(cur, env->src_nid, dist) -
1989 : task_weight(cur, env->dst_nid, dist);
1990 : /*
1991 : * Add some hysteresis to prevent swapping the
1992 : * tasks within a group over tiny differences.
1993 : */
1994 : if (cur_ng)
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 : */
2001 : if (cur_ng && p_ng)
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);
2007 : }
2008 :
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 :
2022 : if (maymove && moveimp > imp && moveimp > env->best_imp) {
2023 : imp = moveimp;
2024 : cur = NULL;
2025 : goto assign;
2026 : }
2027 :
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 :
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 :
2046 : /*
2047 : * In the overloaded case, try and keep the load balanced.
2048 : */
2049 : load = task_h_load(env->p) - task_h_load(cur);
2050 : if (!load)
2051 : goto assign;
2052 :
2053 : dst_load = env->dst_stats.load + load;
2054 : src_load = env->src_stats.load - load;
2055 :
2056 : if (load_too_imbalanced(src_load, dst_load, env))
2057 : goto unlock;
2058 :
2059 : assign:
2060 : /* Evaluate an idle CPU for a task numa move. */
2061 : if (!cur) {
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 :
2068 : /*
2069 : * If the CPU is no longer truly idle and the previous best CPU
2070 : * is, keep using it.
2071 : */
2072 : if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2073 : idle_cpu(env->best_cpu)) {
2074 : cpu = env->best_cpu;
2075 : }
2076 :
2077 : env->dst_cpu = cpu;
2078 : }
2079 :
2080 : task_numa_assign(env, cur, imp);
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 : }
2098 : unlock:
2099 : rcu_read_unlock();
2100 :
2101 : return stopsearch;
2102 : }
2103 :
2104 : static void task_numa_find_cpu(struct task_numa_env *env,
2105 : long taskimp, long groupimp)
2106 : {
2107 : bool maymove = false;
2108 : int cpu;
2109 :
2110 : /*
2111 : * If dst node has spare capacity, then check if there is an
2112 : * imbalance that would be overruled by the load balancer.
2113 : */
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);
2127 : imbalance = adjust_numa_imbalance(imbalance, dst_running,
2128 : env->imb_numa_nr);
2129 :
2130 : /* Use idle CPU if there is no imbalance */
2131 : if (!imbalance) {
2132 : maymove = true;
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 : }
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 : }
2150 :
2151 : for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2152 : /* Skip this CPU if the source task cannot migrate */
2153 : if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2154 : continue;
2155 :
2156 : env->dst_cpu = cpu;
2157 : if (task_numa_compare(env, taskimp, groupimp, maymove))
2158 : break;
2159 : }
2160 : }
2161 :
2162 : static int task_numa_migrate(struct task_struct *p)
2163 : {
2164 : struct task_numa_env env = {
2165 : .p = p,
2166 :
2167 : .src_cpu = task_cpu(p),
2168 : .src_nid = task_node(p),
2169 :
2170 : .imbalance_pct = 112,
2171 :
2172 : .best_task = NULL,
2173 : .best_imp = 0,
2174 : .best_cpu = -1,
2175 : };
2176 : unsigned long taskweight, groupweight;
2177 : struct sched_domain *sd;
2178 : long taskimp, groupimp;
2179 : struct numa_group *ng;
2180 : struct rq *best_rq;
2181 : int nid, ret, dist;
2182 :
2183 : /*
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.
2190 : */
2191 : rcu_read_lock();
2192 : sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2193 : if (sd) {
2194 : env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2195 : env.imb_numa_nr = sd->imb_numa_nr;
2196 : }
2197 : rcu_read_unlock();
2198 :
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)) {
2206 : sched_setnuma(p, task_node(p));
2207 : return -EINVAL;
2208 : }
2209 :
2210 : env.dst_nid = p->numa_preferred_nid;
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);
2214 : update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2215 : taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2216 : groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2217 : update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2218 :
2219 : /* Try to find a spot on the preferred nid. */
2220 : task_numa_find_cpu(&env, taskimp, groupimp);
2221 :
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 : */
2229 : ng = deref_curr_numa_group(p);
2230 : if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2231 : for_each_node_state(nid, N_CPU) {
2232 : if (nid == env.src_nid || nid == p->numa_preferred_nid)
2233 : continue;
2234 :
2235 : dist = node_distance(env.src_nid, env.dst_nid);
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 : }
2241 :
2242 : /* Only consider nodes where both task and groups benefit */
2243 : taskimp = task_weight(p, nid, dist) - taskweight;
2244 : groupimp = group_weight(p, nid, dist) - groupweight;
2245 : if (taskimp < 0 && groupimp < 0)
2246 : continue;
2247 :
2248 : env.dist = dist;
2249 : env.dst_nid = nid;
2250 : update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2251 : task_numa_find_cpu(&env, taskimp, groupimp);
2252 : }
2253 : }
2254 :
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 : */
2263 : if (ng) {
2264 : if (env.best_cpu == -1)
2265 : nid = env.src_nid;
2266 : else
2267 : nid = cpu_to_node(env.best_cpu);
2268 :
2269 : if (nid != p->numa_preferred_nid)
2270 : sched_setnuma(p, nid);
2271 : }
2272 :
2273 : /* No better CPU than the current one was found. */
2274 : if (env.best_cpu == -1) {
2275 : trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2276 : return -EAGAIN;
2277 : }
2278 :
2279 : best_rq = cpu_rq(env.best_cpu);
2280 : if (env.best_task == NULL) {
2281 : ret = migrate_task_to(p, env.best_cpu);
2282 : WRITE_ONCE(best_rq->numa_migrate_on, 0);
2283 : if (ret != 0)
2284 : trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2285 : return ret;
2286 : }
2287 :
2288 : ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2289 : WRITE_ONCE(best_rq->numa_migrate_on, 0);
2290 :
2291 : if (ret != 0)
2292 : trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2293 : put_task_struct(env.best_task);
2294 : return ret;
2295 : }
2296 :
2297 : /* Attempt to migrate a task to a CPU on the preferred node. */
2298 : static void numa_migrate_preferred(struct task_struct *p)
2299 : {
2300 : unsigned long interval = HZ;
2301 :
2302 : /* This task has no NUMA fault statistics yet */
2303 : if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2304 : return;
2305 :
2306 : /* Periodically retry migrating the task to the preferred node */
2307 : interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2308 : p->numa_migrate_retry = jiffies + interval;
2309 :
2310 : /* Success if task is already running on preferred CPU */
2311 : if (task_node(p) == p->numa_preferred_nid)
2312 : return;
2313 :
2314 : /* Otherwise, try migrate to a CPU on the preferred node */
2315 : task_numa_migrate(p);
2316 : }
2317 :
2318 : /*
2319 : * Find out how many nodes the workload is actively running on. Do this by
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.
2323 : */
2324 : static void numa_group_count_active_nodes(struct numa_group *numa_group)
2325 : {
2326 : unsigned long faults, max_faults = 0;
2327 : int nid, active_nodes = 0;
2328 :
2329 : for_each_node_state(nid, N_CPU) {
2330 : faults = group_faults_cpu(numa_group, nid);
2331 : if (faults > max_faults)
2332 : max_faults = faults;
2333 : }
2334 :
2335 : for_each_node_state(nid, N_CPU) {
2336 : faults = group_faults_cpu(numa_group, nid);
2337 : if (faults * ACTIVE_NODE_FRACTION > max_faults)
2338 : active_nodes++;
2339 : }
2340 :
2341 : numa_group->max_faults_cpu = max_faults;
2342 : numa_group->active_nodes = active_nodes;
2343 : }
2344 :
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
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.
2351 : */
2352 : #define NUMA_PERIOD_SLOTS 10
2353 : #define NUMA_PERIOD_THRESHOLD 7
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;
2365 : int lr_ratio, ps_ratio;
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
2373 : * completely idle or all activity is in areas that are not of interest
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
2377 : */
2378 : if (local + shared == 0 || p->numa_faults_locality[2]) {
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);
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;
2414 : if (!slot)
2415 : slot = 1;
2416 : diff = slot * period_slot;
2417 : } else {
2418 : /*
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.
2422 : */
2423 : int ratio = max(lr_ratio, ps_ratio);
2424 : diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
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 :
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;
2449 :
2450 : /* Avoid time going backwards, prevent potential divide error: */
2451 : if (unlikely((s64)*period < 0))
2452 : *period = 0;
2453 : } else {
2454 : delta = p->se.avg.load_sum;
2455 : *period = LOAD_AVG_MAX;
2456 : }
2457 :
2458 : p->last_sum_exec_runtime = runtime;
2459 : p->last_task_numa_placement = now;
2460 :
2461 : return delta;
2462 : }
2463 :
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 :
2489 : for_each_node_state(node, N_CPU) {
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 : */
2508 : nodes = node_states[N_CPU];
2509 : for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2510 : unsigned long max_faults = 0;
2511 : nodemask_t max_group = NODE_MASK_NONE;
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. */
2545 : if (!max_faults)
2546 : break;
2547 : nodes = max_group;
2548 : }
2549 : return nid;
2550 : }
2551 :
2552 : static void task_numa_placement(struct task_struct *p)
2553 : {
2554 : int seq, nid, max_nid = NUMA_NO_NODE;
2555 : unsigned long max_faults = 0;
2556 : unsigned long fault_types[2] = { 0, 0 };
2557 : unsigned long total_faults;
2558 : u64 runtime, period;
2559 : spinlock_t *group_lock = NULL;
2560 : struct numa_group *ng;
2561 :
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 : */
2567 : seq = READ_ONCE(p->mm->numa_scan_seq);
2568 : if (p->numa_scan_seq == seq)
2569 : return;
2570 : p->numa_scan_seq = seq;
2571 : p->numa_scan_period_max = task_scan_max(p);
2572 :
2573 : total_faults = p->numa_faults_locality[0] +
2574 : p->numa_faults_locality[1];
2575 : runtime = numa_get_avg_runtime(p, &period);
2576 :
2577 : /* If the task is part of a group prevent parallel updates to group stats */
2578 : ng = deref_curr_numa_group(p);
2579 : if (ng) {
2580 : group_lock = &ng->lock;
2581 : spin_lock_irq(group_lock);
2582 : }
2583 :
2584 : /* Find the node with the highest number of faults */
2585 : for_each_online_node(nid) {
2586 : /* Keep track of the offsets in numa_faults array */
2587 : int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2588 : unsigned long faults = 0, group_faults = 0;
2589 : int priv;
2590 :
2591 : for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2592 : long diff, f_diff, f_weight;
2593 :
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);
2598 :
2599 : /* Decay existing window, copy faults since last scan */
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;
2603 :
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);
2612 : f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2613 : (total_faults + 1);
2614 : f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2615 : p->numa_faults[cpubuf_idx] = 0;
2616 :
2617 : p->numa_faults[mem_idx] += diff;
2618 : p->numa_faults[cpu_idx] += f_diff;
2619 : faults += p->numa_faults[mem_idx];
2620 : p->total_numa_faults += diff;
2621 : if (ng) {
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 : */
2629 : ng->faults[mem_idx] += diff;
2630 : ng->faults[cpu_idx] += f_diff;
2631 : ng->total_faults += diff;
2632 : group_faults += ng->faults[mem_idx];
2633 : }
2634 : }
2635 :
2636 : if (!ng) {
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;
2643 : max_nid = nid;
2644 : }
2645 : }
2646 :
2647 : /* Cannot migrate task to CPU-less node */
2648 : if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
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 :
2662 : if (ng) {
2663 : numa_group_count_active_nodes(ng);
2664 : spin_unlock_irq(group_lock);
2665 : max_nid = preferred_group_nid(p, max_nid);
2666 : }
2667 :
2668 : if (max_faults) {
2669 : /* Set the new preferred node */
2670 : if (max_nid != p->numa_preferred_nid)
2671 : sched_setnuma(p, max_nid);
2672 : }
2673 :
2674 : update_task_scan_period(p, fault_types[0], fault_types[1]);
2675 : }
2676 :
2677 : static inline int get_numa_group(struct numa_group *grp)
2678 : {
2679 : return refcount_inc_not_zero(&grp->refcount);
2680 : }
2681 :
2682 : static inline void put_numa_group(struct numa_group *grp)
2683 : {
2684 : if (refcount_dec_and_test(&grp->refcount))
2685 : kfree_rcu(grp, rcu);
2686 : }
2687 :
2688 : static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2689 : int *priv)
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 :
2697 : if (unlikely(!deref_curr_numa_group(p))) {
2698 : unsigned int size = sizeof(struct numa_group) +
2699 : NR_NUMA_HINT_FAULT_STATS *
2700 : nr_node_ids * sizeof(unsigned long);
2701 :
2702 : grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2703 : if (!grp)
2704 : return;
2705 :
2706 : refcount_set(&grp->refcount, 1);
2707 : grp->active_nodes = 1;
2708 : grp->max_faults_cpu = 0;
2709 : spin_lock_init(&grp->lock);
2710 : grp->gid = p->pid;
2711 :
2712 : for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2713 : grp->faults[i] = p->numa_faults[i];
2714 :
2715 : grp->total_faults = p->total_numa_faults;
2716 :
2717 : grp->nr_tasks++;
2718 : rcu_assign_pointer(p->numa_group, grp);
2719 : }
2720 :
2721 : rcu_read_lock();
2722 : tsk = READ_ONCE(cpu_rq(cpu)->curr);
2723 :
2724 : if (!cpupid_match_pid(tsk, cpupid))
2725 : goto no_join;
2726 :
2727 : grp = rcu_dereference(tsk->numa_group);
2728 : if (!grp)
2729 : goto no_join;
2730 :
2731 : my_grp = deref_curr_numa_group(p);
2732 : if (grp == my_grp)
2733 : goto no_join;
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)
2740 : goto no_join;
2741 :
2742 : /*
2743 : * Tie-break on the grp address.
2744 : */
2745 : if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2746 : goto no_join;
2747 :
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;
2755 :
2756 : /* Update priv based on whether false sharing was detected */
2757 : *priv = !join;
2758 :
2759 : if (join && !get_numa_group(grp))
2760 : goto no_join;
2761 :
2762 : rcu_read_unlock();
2763 :
2764 : if (!join)
2765 : return;
2766 :
2767 : WARN_ON_ONCE(irqs_disabled());
2768 : double_lock_irq(&my_grp->lock, &grp->lock);
2769 :
2770 : for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2771 : my_grp->faults[i] -= p->numa_faults[i];
2772 : grp->faults[i] += p->numa_faults[i];
2773 : }
2774 : my_grp->total_faults -= p->total_numa_faults;
2775 : grp->total_faults += p->total_numa_faults;
2776 :
2777 : my_grp->nr_tasks--;
2778 : grp->nr_tasks++;
2779 :
2780 : spin_unlock(&my_grp->lock);
2781 : spin_unlock_irq(&grp->lock);
2782 :
2783 : rcu_assign_pointer(p->numa_group, grp);
2784 :
2785 : put_numa_group(my_grp);
2786 : return;
2787 :
2788 : no_join:
2789 : rcu_read_unlock();
2790 : return;
2791 : }
2792 :
2793 : /*
2794 : * Get rid of NUMA statistics associated with a task (either current or dead).
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)
2801 : {
2802 : /* safe: p either is current or is being freed by current */
2803 : struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2804 : unsigned long *numa_faults = p->numa_faults;
2805 : unsigned long flags;
2806 : int i;
2807 :
2808 : if (!numa_faults)
2809 : return;
2810 :
2811 : if (grp) {
2812 : spin_lock_irqsave(&grp->lock, flags);
2813 : for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2814 : grp->faults[i] -= p->numa_faults[i];
2815 : grp->total_faults -= p->total_numa_faults;
2816 :
2817 : grp->nr_tasks--;
2818 : spin_unlock_irqrestore(&grp->lock, flags);
2819 : RCU_INIT_POINTER(p->numa_group, NULL);
2820 : put_numa_group(grp);
2821 : }
2822 :
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 : }
2831 : }
2832 :
2833 : /*
2834 : * Got a PROT_NONE fault for a page on @node.
2835 : */
2836 : void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2837 : {
2838 : struct task_struct *p = current;
2839 : bool migrated = flags & TNF_MIGRATED;
2840 : int cpu_node = task_node(current);
2841 : int local = !!(flags & TNF_FAULT_LOCAL);
2842 : struct numa_group *ng;
2843 : int priv;
2844 :
2845 : if (!static_branch_likely(&sched_numa_balancing))
2846 : return;
2847 :
2848 : /* for example, ksmd faulting in a user's mm */
2849 : if (!p->mm)
2850 : return;
2851 :
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 :
2861 : /* Allocate buffer to track faults on a per-node basis */
2862 : if (unlikely(!p->numa_faults)) {
2863 : int size = sizeof(*p->numa_faults) *
2864 : NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2865 :
2866 : p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2867 : if (!p->numa_faults)
2868 : return;
2869 :
2870 : p->total_numa_faults = 0;
2871 : memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2872 : }
2873 :
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);
2882 : if (!priv && !(flags & TNF_NO_GROUP))
2883 : task_numa_group(p, last_cpupid, flags, &priv);
2884 : }
2885 :
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 : */
2892 : ng = deref_curr_numa_group(p);
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))
2896 : local = 1;
2897 :
2898 : /*
2899 : * Retry to migrate task to preferred node periodically, in case it
2900 : * previously failed, or the scheduler moved us.
2901 : */
2902 : if (time_after(jiffies, p->numa_migrate_retry)) {
2903 : task_numa_placement(p);
2904 : numa_migrate_preferred(p);
2905 : }
2906 :
2907 : if (migrated)
2908 : p->numa_pages_migrated += pages;
2909 : if (flags & TNF_MIGRATE_FAIL)
2910 : p->numa_faults_locality[2] += pages;
2911 :
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;
2914 : p->numa_faults_locality[local] += pages;
2915 : }
2916 :
2917 : static void reset_ptenuma_scan(struct task_struct *p)
2918 : {
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 : */
2927 : WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2928 : p->mm->numa_scan_offset = 0;
2929 : }
2930 :
2931 : static bool vma_is_accessed(struct vm_area_struct *vma)
2932 : {
2933 : unsigned long pids;
2934 : /*
2935 : * Allow unconditional access first two times, so that all the (pages)
2936 : * of VMAs get prot_none fault introduced irrespective of accesses.
2937 : * This is also done to avoid any side effect of task scanning
2938 : * amplifying the unfairness of disjoint set of VMAs' access.
2939 : */
2940 : if (READ_ONCE(current->mm->numa_scan_seq) < 2)
2941 : return true;
2942 :
2943 : pids = vma->numab_state->access_pids[0] | vma->numab_state->access_pids[1];
2944 : return test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids);
2945 : }
2946 :
2947 : #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
2948 :
2949 : /*
2950 : * The expensive part of numa migration is done from task_work context.
2951 : * Triggered from task_tick_numa().
2952 : */
2953 : static void task_numa_work(struct callback_head *work)
2954 : {
2955 : unsigned long migrate, next_scan, now = jiffies;
2956 : struct task_struct *p = current;
2957 : struct mm_struct *mm = p->mm;
2958 : u64 runtime = p->se.sum_exec_runtime;
2959 : struct vm_area_struct *vma;
2960 : unsigned long start, end;
2961 : unsigned long nr_pte_updates = 0;
2962 : long pages, virtpages;
2963 : struct vma_iterator vmi;
2964 :
2965 : SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2966 :
2967 : work->next = work;
2968 : /*
2969 : * Who cares about NUMA placement when they're dying.
2970 : *
2971 : * NOTE: make sure not to dereference p->mm before this check,
2972 : * exit_task_work() happens _after_ exit_mm() so we could be called
2973 : * without p->mm even though we still had it when we enqueued this
2974 : * work.
2975 : */
2976 : if (p->flags & PF_EXITING)
2977 : return;
2978 :
2979 : if (!mm->numa_next_scan) {
2980 : mm->numa_next_scan = now +
2981 : msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2982 : }
2983 :
2984 : /*
2985 : * Enforce maximal scan/migration frequency..
2986 : */
2987 : migrate = mm->numa_next_scan;
2988 : if (time_before(now, migrate))
2989 : return;
2990 :
2991 : if (p->numa_scan_period == 0) {
2992 : p->numa_scan_period_max = task_scan_max(p);
2993 : p->numa_scan_period = task_scan_start(p);
2994 : }
2995 :
2996 : next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2997 : if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
2998 : return;
2999 :
3000 : /*
3001 : * Delay this task enough that another task of this mm will likely win
3002 : * the next time around.
3003 : */
3004 : p->node_stamp += 2 * TICK_NSEC;
3005 :
3006 : start = mm->numa_scan_offset;
3007 : pages = sysctl_numa_balancing_scan_size;
3008 : pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3009 : virtpages = pages * 8; /* Scan up to this much virtual space */
3010 : if (!pages)
3011 : return;
3012 :
3013 :
3014 : if (!mmap_read_trylock(mm))
3015 : return;
3016 : vma_iter_init(&vmi, mm, start);
3017 : vma = vma_next(&vmi);
3018 : if (!vma) {
3019 : reset_ptenuma_scan(p);
3020 : start = 0;
3021 : vma_iter_set(&vmi, start);
3022 : vma = vma_next(&vmi);
3023 : }
3024 :
3025 : do {
3026 : if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3027 : is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3028 : continue;
3029 : }
3030 :
3031 : /*
3032 : * Shared library pages mapped by multiple processes are not
3033 : * migrated as it is expected they are cache replicated. Avoid
3034 : * hinting faults in read-only file-backed mappings or the vdso
3035 : * as migrating the pages will be of marginal benefit.
3036 : */
3037 : if (!vma->vm_mm ||
3038 : (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
3039 : continue;
3040 :
3041 : /*
3042 : * Skip inaccessible VMAs to avoid any confusion between
3043 : * PROT_NONE and NUMA hinting ptes
3044 : */
3045 : if (!vma_is_accessible(vma))
3046 : continue;
3047 :
3048 : /* Initialise new per-VMA NUMAB state. */
3049 : if (!vma->numab_state) {
3050 : vma->numab_state = kzalloc(sizeof(struct vma_numab_state),
3051 : GFP_KERNEL);
3052 : if (!vma->numab_state)
3053 : continue;
3054 :
3055 : vma->numab_state->next_scan = now +
3056 : msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3057 :
3058 : /* Reset happens after 4 times scan delay of scan start */
3059 : vma->numab_state->next_pid_reset = vma->numab_state->next_scan +
3060 : msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3061 : }
3062 :
3063 : /*
3064 : * Scanning the VMA's of short lived tasks add more overhead. So
3065 : * delay the scan for new VMAs.
3066 : */
3067 : if (mm->numa_scan_seq && time_before(jiffies,
3068 : vma->numab_state->next_scan))
3069 : continue;
3070 :
3071 : /* Do not scan the VMA if task has not accessed */
3072 : if (!vma_is_accessed(vma))
3073 : continue;
3074 :
3075 : /*
3076 : * RESET access PIDs regularly for old VMAs. Resetting after checking
3077 : * vma for recent access to avoid clearing PID info before access..
3078 : */
3079 : if (mm->numa_scan_seq &&
3080 : time_after(jiffies, vma->numab_state->next_pid_reset)) {
3081 : vma->numab_state->next_pid_reset = vma->numab_state->next_pid_reset +
3082 : msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3083 : vma->numab_state->access_pids[0] = READ_ONCE(vma->numab_state->access_pids[1]);
3084 : vma->numab_state->access_pids[1] = 0;
3085 : }
3086 :
3087 : do {
3088 : start = max(start, vma->vm_start);
3089 : end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3090 : end = min(end, vma->vm_end);
3091 : nr_pte_updates = change_prot_numa(vma, start, end);
3092 :
3093 : /*
3094 : * Try to scan sysctl_numa_balancing_size worth of
3095 : * hpages that have at least one present PTE that
3096 : * is not already pte-numa. If the VMA contains
3097 : * areas that are unused or already full of prot_numa
3098 : * PTEs, scan up to virtpages, to skip through those
3099 : * areas faster.
3100 : */
3101 : if (nr_pte_updates)
3102 : pages -= (end - start) >> PAGE_SHIFT;
3103 : virtpages -= (end - start) >> PAGE_SHIFT;
3104 :
3105 : start = end;
3106 : if (pages <= 0 || virtpages <= 0)
3107 : goto out;
3108 :
3109 : cond_resched();
3110 : } while (end != vma->vm_end);
3111 : } for_each_vma(vmi, vma);
3112 :
3113 : out:
3114 : /*
3115 : * It is possible to reach the end of the VMA list but the last few
3116 : * VMAs are not guaranteed to the vma_migratable. If they are not, we
3117 : * would find the !migratable VMA on the next scan but not reset the
3118 : * scanner to the start so check it now.
3119 : */
3120 : if (vma)
3121 : mm->numa_scan_offset = start;
3122 : else
3123 : reset_ptenuma_scan(p);
3124 : mmap_read_unlock(mm);
3125 :
3126 : /*
3127 : * Make sure tasks use at least 32x as much time to run other code
3128 : * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3129 : * Usually update_task_scan_period slows down scanning enough; on an
3130 : * overloaded system we need to limit overhead on a per task basis.
3131 : */
3132 : if (unlikely(p->se.sum_exec_runtime != runtime)) {
3133 : u64 diff = p->se.sum_exec_runtime - runtime;
3134 : p->node_stamp += 32 * diff;
3135 : }
3136 : }
3137 :
3138 : void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3139 : {
3140 : int mm_users = 0;
3141 : struct mm_struct *mm = p->mm;
3142 :
3143 : if (mm) {
3144 : mm_users = atomic_read(&mm->mm_users);
3145 : if (mm_users == 1) {
3146 : mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3147 : mm->numa_scan_seq = 0;
3148 : }
3149 : }
3150 : p->node_stamp = 0;
3151 : p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3152 : p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3153 : p->numa_migrate_retry = 0;
3154 : /* Protect against double add, see task_tick_numa and task_numa_work */
3155 : p->numa_work.next = &p->numa_work;
3156 : p->numa_faults = NULL;
3157 : p->numa_pages_migrated = 0;
3158 : p->total_numa_faults = 0;
3159 : RCU_INIT_POINTER(p->numa_group, NULL);
3160 : p->last_task_numa_placement = 0;
3161 : p->last_sum_exec_runtime = 0;
3162 :
3163 : init_task_work(&p->numa_work, task_numa_work);
3164 :
3165 : /* New address space, reset the preferred nid */
3166 : if (!(clone_flags & CLONE_VM)) {
3167 : p->numa_preferred_nid = NUMA_NO_NODE;
3168 : return;
3169 : }
3170 :
3171 : /*
3172 : * New thread, keep existing numa_preferred_nid which should be copied
3173 : * already by arch_dup_task_struct but stagger when scans start.
3174 : */
3175 : if (mm) {
3176 : unsigned int delay;
3177 :
3178 : delay = min_t(unsigned int, task_scan_max(current),
3179 : current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3180 : delay += 2 * TICK_NSEC;
3181 : p->node_stamp = delay;
3182 : }
3183 : }
3184 :
3185 : /*
3186 : * Drive the periodic memory faults..
3187 : */
3188 : static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3189 : {
3190 : struct callback_head *work = &curr->numa_work;
3191 : u64 period, now;
3192 :
3193 : /*
3194 : * We don't care about NUMA placement if we don't have memory.
3195 : */
3196 : if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3197 : return;
3198 :
3199 : /*
3200 : * Using runtime rather than walltime has the dual advantage that
3201 : * we (mostly) drive the selection from busy threads and that the
3202 : * task needs to have done some actual work before we bother with
3203 : * NUMA placement.
3204 : */
3205 : now = curr->se.sum_exec_runtime;
3206 : period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3207 :
3208 : if (now > curr->node_stamp + period) {
3209 : if (!curr->node_stamp)
3210 : curr->numa_scan_period = task_scan_start(curr);
3211 : curr->node_stamp += period;
3212 :
3213 : if (!time_before(jiffies, curr->mm->numa_next_scan))
3214 : task_work_add(curr, work, TWA_RESUME);
3215 : }
3216 : }
3217 :
3218 : static void update_scan_period(struct task_struct *p, int new_cpu)
3219 : {
3220 : int src_nid = cpu_to_node(task_cpu(p));
3221 : int dst_nid = cpu_to_node(new_cpu);
3222 :
3223 : if (!static_branch_likely(&sched_numa_balancing))
3224 : return;
3225 :
3226 : if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3227 : return;
3228 :
3229 : if (src_nid == dst_nid)
3230 : return;
3231 :
3232 : /*
3233 : * Allow resets if faults have been trapped before one scan
3234 : * has completed. This is most likely due to a new task that
3235 : * is pulled cross-node due to wakeups or load balancing.
3236 : */
3237 : if (p->numa_scan_seq) {
3238 : /*
3239 : * Avoid scan adjustments if moving to the preferred
3240 : * node or if the task was not previously running on
3241 : * the preferred node.
3242 : */
3243 : if (dst_nid == p->numa_preferred_nid ||
3244 : (p->numa_preferred_nid != NUMA_NO_NODE &&
3245 : src_nid != p->numa_preferred_nid))
3246 : return;
3247 : }
3248 :
3249 : p->numa_scan_period = task_scan_start(p);
3250 : }
3251 :
3252 : #else
3253 : static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3254 : {
3255 : }
3256 :
3257 : static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3258 : {
3259 : }
3260 :
3261 : static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3262 : {
3263 : }
3264 :
3265 : static inline void update_scan_period(struct task_struct *p, int new_cpu)
3266 : {
3267 : }
3268 :
3269 : #endif /* CONFIG_NUMA_BALANCING */
3270 :
3271 : static void
3272 : account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3273 : {
3274 4892 : update_load_add(&cfs_rq->load, se->load.weight);
3275 : #ifdef CONFIG_SMP
3276 : if (entity_is_task(se)) {
3277 : struct rq *rq = rq_of(cfs_rq);
3278 :
3279 : account_numa_enqueue(rq, task_of(se));
3280 : list_add(&se->group_node, &rq->cfs_tasks);
3281 : }
3282 : #endif
3283 2446 : cfs_rq->nr_running++;
3284 2446 : if (se_is_idle(se))
3285 : cfs_rq->idle_nr_running++;
3286 : }
3287 :
3288 : static void
3289 : account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3290 : {
3291 4888 : update_load_sub(&cfs_rq->load, se->load.weight);
3292 : #ifdef CONFIG_SMP
3293 : if (entity_is_task(se)) {
3294 : account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3295 : list_del_init(&se->group_node);
3296 : }
3297 : #endif
3298 2444 : cfs_rq->nr_running--;
3299 2444 : if (se_is_idle(se))
3300 : cfs_rq->idle_nr_running--;
3301 : }
3302 :
3303 : /*
3304 : * Signed add and clamp on underflow.
3305 : *
3306 : * Explicitly do a load-store to ensure the intermediate value never hits
3307 : * memory. This allows lockless observations without ever seeing the negative
3308 : * values.
3309 : */
3310 : #define add_positive(_ptr, _val) do { \
3311 : typeof(_ptr) ptr = (_ptr); \
3312 : typeof(_val) val = (_val); \
3313 : typeof(*ptr) res, var = READ_ONCE(*ptr); \
3314 : \
3315 : res = var + val; \
3316 : \
3317 : if (val < 0 && res > var) \
3318 : res = 0; \
3319 : \
3320 : WRITE_ONCE(*ptr, res); \
3321 : } while (0)
3322 :
3323 : /*
3324 : * Unsigned subtract and clamp on underflow.
3325 : *
3326 : * Explicitly do a load-store to ensure the intermediate value never hits
3327 : * memory. This allows lockless observations without ever seeing the negative
3328 : * values.
3329 : */
3330 : #define sub_positive(_ptr, _val) do { \
3331 : typeof(_ptr) ptr = (_ptr); \
3332 : typeof(*ptr) val = (_val); \
3333 : typeof(*ptr) res, var = READ_ONCE(*ptr); \
3334 : res = var - val; \
3335 : if (res > var) \
3336 : res = 0; \
3337 : WRITE_ONCE(*ptr, res); \
3338 : } while (0)
3339 :
3340 : /*
3341 : * Remove and clamp on negative, from a local variable.
3342 : *
3343 : * A variant of sub_positive(), which does not use explicit load-store
3344 : * and is thus optimized for local variable updates.
3345 : */
3346 : #define lsub_positive(_ptr, _val) do { \
3347 : typeof(_ptr) ptr = (_ptr); \
3348 : *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3349 : } while (0)
3350 :
3351 : #ifdef CONFIG_SMP
3352 : static inline void
3353 : enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3354 : {
3355 : cfs_rq->avg.load_avg += se->avg.load_avg;
3356 : cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3357 : }
3358 :
3359 : static inline void
3360 : dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3361 : {
3362 : sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3363 : sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3364 : /* See update_cfs_rq_load_avg() */
3365 : cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3366 : cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3367 : }
3368 : #else
3369 : static inline void
3370 : enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3371 : static inline void
3372 : dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3373 : #endif
3374 :
3375 5 : static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3376 : unsigned long weight)
3377 : {
3378 5 : if (se->on_rq) {
3379 : /* commit outstanding execution time */
3380 0 : if (cfs_rq->curr == se)
3381 0 : update_curr(cfs_rq);
3382 0 : update_load_sub(&cfs_rq->load, se->load.weight);
3383 : }
3384 5 : dequeue_load_avg(cfs_rq, se);
3385 :
3386 10 : update_load_set(&se->load, weight);
3387 :
3388 : #ifdef CONFIG_SMP
3389 : do {
3390 : u32 divider = get_pelt_divider(&se->avg);
3391 :
3392 : se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3393 : } while (0);
3394 : #endif
3395 :
3396 5 : enqueue_load_avg(cfs_rq, se);
3397 5 : if (se->on_rq)
3398 0 : update_load_add(&cfs_rq->load, se->load.weight);
3399 :
3400 5 : }
3401 :
3402 5 : void reweight_task(struct task_struct *p, int prio)
3403 : {
3404 5 : struct sched_entity *se = &p->se;
3405 10 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
3406 5 : struct load_weight *load = &se->load;
3407 5 : unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3408 :
3409 5 : reweight_entity(cfs_rq, se, weight);
3410 5 : load->inv_weight = sched_prio_to_wmult[prio];
3411 5 : }
3412 :
3413 : static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3414 :
3415 : #ifdef CONFIG_FAIR_GROUP_SCHED
3416 : #ifdef CONFIG_SMP
3417 : /*
3418 : * All this does is approximate the hierarchical proportion which includes that
3419 : * global sum we all love to hate.
3420 : *
3421 : * That is, the weight of a group entity, is the proportional share of the
3422 : * group weight based on the group runqueue weights. That is:
3423 : *
3424 : * tg->weight * grq->load.weight
3425 : * ge->load.weight = ----------------------------- (1)
3426 : * \Sum grq->load.weight
3427 : *
3428 : * Now, because computing that sum is prohibitively expensive to compute (been
3429 : * there, done that) we approximate it with this average stuff. The average
3430 : * moves slower and therefore the approximation is cheaper and more stable.
3431 : *
3432 : * So instead of the above, we substitute:
3433 : *
3434 : * grq->load.weight -> grq->avg.load_avg (2)
3435 : *
3436 : * which yields the following:
3437 : *
3438 : * tg->weight * grq->avg.load_avg
3439 : * ge->load.weight = ------------------------------ (3)
3440 : * tg->load_avg
3441 : *
3442 : * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3443 : *
3444 : * That is shares_avg, and it is right (given the approximation (2)).
3445 : *
3446 : * The problem with it is that because the average is slow -- it was designed
3447 : * to be exactly that of course -- this leads to transients in boundary
3448 : * conditions. In specific, the case where the group was idle and we start the
3449 : * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3450 : * yielding bad latency etc..
3451 : *
3452 : * Now, in that special case (1) reduces to:
3453 : *
3454 : * tg->weight * grq->load.weight
3455 : * ge->load.weight = ----------------------------- = tg->weight (4)
3456 : * grp->load.weight
3457 : *
3458 : * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3459 : *
3460 : * So what we do is modify our approximation (3) to approach (4) in the (near)
3461 : * UP case, like:
3462 : *
3463 : * ge->load.weight =
3464 : *
3465 : * tg->weight * grq->load.weight
3466 : * --------------------------------------------------- (5)
3467 : * tg->load_avg - grq->avg.load_avg + grq->load.weight
3468 : *
3469 : * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3470 : * we need to use grq->avg.load_avg as its lower bound, which then gives:
3471 : *
3472 : *
3473 : * tg->weight * grq->load.weight
3474 : * ge->load.weight = ----------------------------- (6)
3475 : * tg_load_avg'
3476 : *
3477 : * Where:
3478 : *
3479 : * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3480 : * max(grq->load.weight, grq->avg.load_avg)
3481 : *
3482 : * And that is shares_weight and is icky. In the (near) UP case it approaches
3483 : * (4) while in the normal case it approaches (3). It consistently
3484 : * overestimates the ge->load.weight and therefore:
3485 : *
3486 : * \Sum ge->load.weight >= tg->weight
3487 : *
3488 : * hence icky!
3489 : */
3490 : static long calc_group_shares(struct cfs_rq *cfs_rq)
3491 : {
3492 : long tg_weight, tg_shares, load, shares;
3493 : struct task_group *tg = cfs_rq->tg;
3494 :
3495 : tg_shares = READ_ONCE(tg->shares);
3496 :
3497 : load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3498 :
3499 : tg_weight = atomic_long_read(&tg->load_avg);
3500 :
3501 : /* Ensure tg_weight >= load */
3502 : tg_weight -= cfs_rq->tg_load_avg_contrib;
3503 : tg_weight += load;
3504 :
3505 : shares = (tg_shares * load);
3506 : if (tg_weight)
3507 : shares /= tg_weight;
3508 :
3509 : /*
3510 : * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3511 : * of a group with small tg->shares value. It is a floor value which is
3512 : * assigned as a minimum load.weight to the sched_entity representing
3513 : * the group on a CPU.
3514 : *
3515 : * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3516 : * on an 8-core system with 8 tasks each runnable on one CPU shares has
3517 : * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3518 : * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3519 : * instead of 0.
3520 : */
3521 : return clamp_t(long, shares, MIN_SHARES, tg_shares);
3522 : }
3523 : #endif /* CONFIG_SMP */
3524 :
3525 : /*
3526 : * Recomputes the group entity based on the current state of its group
3527 : * runqueue.
3528 : */
3529 : static void update_cfs_group(struct sched_entity *se)
3530 : {
3531 : struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3532 : long shares;
3533 :
3534 : if (!gcfs_rq)
3535 : return;
3536 :
3537 : if (throttled_hierarchy(gcfs_rq))
3538 : return;
3539 :
3540 : #ifndef CONFIG_SMP
3541 : shares = READ_ONCE(gcfs_rq->tg->shares);
3542 :
3543 : if (likely(se->load.weight == shares))
3544 : return;
3545 : #else
3546 : shares = calc_group_shares(gcfs_rq);
3547 : #endif
3548 :
3549 : reweight_entity(cfs_rq_of(se), se, shares);
3550 : }
3551 :
3552 : #else /* CONFIG_FAIR_GROUP_SCHED */
3553 : static inline void update_cfs_group(struct sched_entity *se)
3554 : {
3555 : }
3556 : #endif /* CONFIG_FAIR_GROUP_SCHED */
3557 :
3558 : static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3559 : {
3560 7820 : struct rq *rq = rq_of(cfs_rq);
3561 :
3562 : if (&rq->cfs == cfs_rq) {
3563 : /*
3564 : * There are a few boundary cases this might miss but it should
3565 : * get called often enough that that should (hopefully) not be
3566 : * a real problem.
3567 : *
3568 : * It will not get called when we go idle, because the idle
3569 : * thread is a different class (!fair), nor will the utilization
3570 : * number include things like RT tasks.
3571 : *
3572 : * As is, the util number is not freq-invariant (we'd have to
3573 : * implement arch_scale_freq_capacity() for that).
3574 : *
3575 : * See cpu_util_cfs().
3576 : */
3577 : cpufreq_update_util(rq, flags);
3578 : }
3579 : }
3580 :
3581 : #ifdef CONFIG_SMP
3582 : static inline bool load_avg_is_decayed(struct sched_avg *sa)
3583 : {
3584 : if (sa->load_sum)
3585 : return false;
3586 :
3587 : if (sa->util_sum)
3588 : return false;
3589 :
3590 : if (sa->runnable_sum)
3591 : return false;
3592 :
3593 : /*
3594 : * _avg must be null when _sum are null because _avg = _sum / divider
3595 : * Make sure that rounding and/or propagation of PELT values never
3596 : * break this.
3597 : */
3598 : SCHED_WARN_ON(sa->load_avg ||
3599 : sa->util_avg ||
3600 : sa->runnable_avg);
3601 :
3602 : return true;
3603 : }
3604 :
3605 : static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3606 : {
3607 : return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3608 : cfs_rq->last_update_time_copy);
3609 : }
3610 : #ifdef CONFIG_FAIR_GROUP_SCHED
3611 : /*
3612 : * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3613 : * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3614 : * bottom-up, we only have to test whether the cfs_rq before us on the list
3615 : * is our child.
3616 : * If cfs_rq is not on the list, test whether a child needs its to be added to
3617 : * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3618 : */
3619 : static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3620 : {
3621 : struct cfs_rq *prev_cfs_rq;
3622 : struct list_head *prev;
3623 :
3624 : if (cfs_rq->on_list) {
3625 : prev = cfs_rq->leaf_cfs_rq_list.prev;
3626 : } else {
3627 : struct rq *rq = rq_of(cfs_rq);
3628 :
3629 : prev = rq->tmp_alone_branch;
3630 : }
3631 :
3632 : prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3633 :
3634 : return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3635 : }
3636 :
3637 : static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3638 : {
3639 : if (cfs_rq->load.weight)
3640 : return false;
3641 :
3642 : if (!load_avg_is_decayed(&cfs_rq->avg))
3643 : return false;
3644 :
3645 : if (child_cfs_rq_on_list(cfs_rq))
3646 : return false;
3647 :
3648 : return true;
3649 : }
3650 :
3651 : /**
3652 : * update_tg_load_avg - update the tg's load avg
3653 : * @cfs_rq: the cfs_rq whose avg changed
3654 : *
3655 : * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3656 : * However, because tg->load_avg is a global value there are performance
3657 : * considerations.
3658 : *
3659 : * In order to avoid having to look at the other cfs_rq's, we use a
3660 : * differential update where we store the last value we propagated. This in
3661 : * turn allows skipping updates if the differential is 'small'.
3662 : *
3663 : * Updating tg's load_avg is necessary before update_cfs_share().
3664 : */
3665 : static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3666 : {
3667 : long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3668 :
3669 : /*
3670 : * No need to update load_avg for root_task_group as it is not used.
3671 : */
3672 : if (cfs_rq->tg == &root_task_group)
3673 : return;
3674 :
3675 : if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3676 : atomic_long_add(delta, &cfs_rq->tg->load_avg);
3677 : cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3678 : }
3679 : }
3680 :
3681 : /*
3682 : * Called within set_task_rq() right before setting a task's CPU. The
3683 : * caller only guarantees p->pi_lock is held; no other assumptions,
3684 : * including the state of rq->lock, should be made.
3685 : */
3686 : void set_task_rq_fair(struct sched_entity *se,
3687 : struct cfs_rq *prev, struct cfs_rq *next)
3688 : {
3689 : u64 p_last_update_time;
3690 : u64 n_last_update_time;
3691 :
3692 : if (!sched_feat(ATTACH_AGE_LOAD))
3693 : return;
3694 :
3695 : /*
3696 : * We are supposed to update the task to "current" time, then its up to
3697 : * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3698 : * getting what current time is, so simply throw away the out-of-date
3699 : * time. This will result in the wakee task is less decayed, but giving
3700 : * the wakee more load sounds not bad.
3701 : */
3702 : if (!(se->avg.last_update_time && prev))
3703 : return;
3704 :
3705 : p_last_update_time = cfs_rq_last_update_time(prev);
3706 : n_last_update_time = cfs_rq_last_update_time(next);
3707 :
3708 : __update_load_avg_blocked_se(p_last_update_time, se);
3709 : se->avg.last_update_time = n_last_update_time;
3710 : }
3711 :
3712 : /*
3713 : * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3714 : * propagate its contribution. The key to this propagation is the invariant
3715 : * that for each group:
3716 : *
3717 : * ge->avg == grq->avg (1)
3718 : *
3719 : * _IFF_ we look at the pure running and runnable sums. Because they
3720 : * represent the very same entity, just at different points in the hierarchy.
3721 : *
3722 : * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3723 : * and simply copies the running/runnable sum over (but still wrong, because
3724 : * the group entity and group rq do not have their PELT windows aligned).
3725 : *
3726 : * However, update_tg_cfs_load() is more complex. So we have:
3727 : *
3728 : * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3729 : *
3730 : * And since, like util, the runnable part should be directly transferable,
3731 : * the following would _appear_ to be the straight forward approach:
3732 : *
3733 : * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3734 : *
3735 : * And per (1) we have:
3736 : *
3737 : * ge->avg.runnable_avg == grq->avg.runnable_avg
3738 : *
3739 : * Which gives:
3740 : *
3741 : * ge->load.weight * grq->avg.load_avg
3742 : * ge->avg.load_avg = ----------------------------------- (4)
3743 : * grq->load.weight
3744 : *
3745 : * Except that is wrong!
3746 : *
3747 : * Because while for entities historical weight is not important and we
3748 : * really only care about our future and therefore can consider a pure
3749 : * runnable sum, runqueues can NOT do this.
3750 : *
3751 : * We specifically want runqueues to have a load_avg that includes
3752 : * historical weights. Those represent the blocked load, the load we expect
3753 : * to (shortly) return to us. This only works by keeping the weights as
3754 : * integral part of the sum. We therefore cannot decompose as per (3).
3755 : *
3756 : * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3757 : * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3758 : * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3759 : * runnable section of these tasks overlap (or not). If they were to perfectly
3760 : * align the rq as a whole would be runnable 2/3 of the time. If however we
3761 : * always have at least 1 runnable task, the rq as a whole is always runnable.
3762 : *
3763 : * So we'll have to approximate.. :/
3764 : *
3765 : * Given the constraint:
3766 : *
3767 : * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3768 : *
3769 : * We can construct a rule that adds runnable to a rq by assuming minimal
3770 : * overlap.
3771 : *
3772 : * On removal, we'll assume each task is equally runnable; which yields:
3773 : *
3774 : * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3775 : *
3776 : * XXX: only do this for the part of runnable > running ?
3777 : *
3778 : */
3779 : static inline void
3780 : update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3781 : {
3782 : long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3783 : u32 new_sum, divider;
3784 :
3785 : /* Nothing to update */
3786 : if (!delta_avg)
3787 : return;
3788 :
3789 : /*
3790 : * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3791 : * See ___update_load_avg() for details.
3792 : */
3793 : divider = get_pelt_divider(&cfs_rq->avg);
3794 :
3795 :
3796 : /* Set new sched_entity's utilization */
3797 : se->avg.util_avg = gcfs_rq->avg.util_avg;
3798 : new_sum = se->avg.util_avg * divider;
3799 : delta_sum = (long)new_sum - (long)se->avg.util_sum;
3800 : se->avg.util_sum = new_sum;
3801 :
3802 : /* Update parent cfs_rq utilization */
3803 : add_positive(&cfs_rq->avg.util_avg, delta_avg);
3804 : add_positive(&cfs_rq->avg.util_sum, delta_sum);
3805 :
3806 : /* See update_cfs_rq_load_avg() */
3807 : cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3808 : cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3809 : }
3810 :
3811 : static inline void
3812 : update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3813 : {
3814 : long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3815 : u32 new_sum, divider;
3816 :
3817 : /* Nothing to update */
3818 : if (!delta_avg)
3819 : return;
3820 :
3821 : /*
3822 : * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3823 : * See ___update_load_avg() for details.
3824 : */
3825 : divider = get_pelt_divider(&cfs_rq->avg);
3826 :
3827 : /* Set new sched_entity's runnable */
3828 : se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3829 : new_sum = se->avg.runnable_avg * divider;
3830 : delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3831 : se->avg.runnable_sum = new_sum;
3832 :
3833 : /* Update parent cfs_rq runnable */
3834 : add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3835 : add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3836 : /* See update_cfs_rq_load_avg() */
3837 : cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3838 : cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3839 : }
3840 :
3841 : static inline void
3842 : update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3843 : {
3844 : long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3845 : unsigned long load_avg;
3846 : u64 load_sum = 0;
3847 : s64 delta_sum;
3848 : u32 divider;
3849 :
3850 : if (!runnable_sum)
3851 : return;
3852 :
3853 : gcfs_rq->prop_runnable_sum = 0;
3854 :
3855 : /*
3856 : * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3857 : * See ___update_load_avg() for details.
3858 : */
3859 : divider = get_pelt_divider(&cfs_rq->avg);
3860 :
3861 : if (runnable_sum >= 0) {
3862 : /*
3863 : * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3864 : * the CPU is saturated running == runnable.
3865 : */
3866 : runnable_sum += se->avg.load_sum;
3867 : runnable_sum = min_t(long, runnable_sum, divider);
3868 : } else {
3869 : /*
3870 : * Estimate the new unweighted runnable_sum of the gcfs_rq by
3871 : * assuming all tasks are equally runnable.
3872 : */
3873 : if (scale_load_down(gcfs_rq->load.weight)) {
3874 : load_sum = div_u64(gcfs_rq->avg.load_sum,
3875 : scale_load_down(gcfs_rq->load.weight));
3876 : }
3877 :
3878 : /* But make sure to not inflate se's runnable */
3879 : runnable_sum = min(se->avg.load_sum, load_sum);
3880 : }
3881 :
3882 : /*
3883 : * runnable_sum can't be lower than running_sum
3884 : * Rescale running sum to be in the same range as runnable sum
3885 : * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3886 : * runnable_sum is in [0 : LOAD_AVG_MAX]
3887 : */
3888 : running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3889 : runnable_sum = max(runnable_sum, running_sum);
3890 :
3891 : load_sum = se_weight(se) * runnable_sum;
3892 : load_avg = div_u64(load_sum, divider);
3893 :
3894 : delta_avg = load_avg - se->avg.load_avg;
3895 : if (!delta_avg)
3896 : return;
3897 :
3898 : delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3899 :
3900 : se->avg.load_sum = runnable_sum;
3901 : se->avg.load_avg = load_avg;
3902 : add_positive(&cfs_rq->avg.load_avg, delta_avg);
3903 : add_positive(&cfs_rq->avg.load_sum, delta_sum);
3904 : /* See update_cfs_rq_load_avg() */
3905 : cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3906 : cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3907 : }
3908 :
3909 : static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3910 : {
3911 : cfs_rq->propagate = 1;
3912 : cfs_rq->prop_runnable_sum += runnable_sum;
3913 : }
3914 :
3915 : /* Update task and its cfs_rq load average */
3916 : static inline int propagate_entity_load_avg(struct sched_entity *se)
3917 : {
3918 : struct cfs_rq *cfs_rq, *gcfs_rq;
3919 :
3920 : if (entity_is_task(se))
3921 : return 0;
3922 :
3923 : gcfs_rq = group_cfs_rq(se);
3924 : if (!gcfs_rq->propagate)
3925 : return 0;
3926 :
3927 : gcfs_rq->propagate = 0;
3928 :
3929 : cfs_rq = cfs_rq_of(se);
3930 :
3931 : add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3932 :
3933 : update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3934 : update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3935 : update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3936 :
3937 : trace_pelt_cfs_tp(cfs_rq);
3938 : trace_pelt_se_tp(se);
3939 :
3940 : return 1;
3941 : }
3942 :
3943 : /*
3944 : * Check if we need to update the load and the utilization of a blocked
3945 : * group_entity:
3946 : */
3947 : static inline bool skip_blocked_update(struct sched_entity *se)
3948 : {
3949 : struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3950 :
3951 : /*
3952 : * If sched_entity still have not zero load or utilization, we have to
3953 : * decay it:
3954 : */
3955 : if (se->avg.load_avg || se->avg.util_avg)
3956 : return false;
3957 :
3958 : /*
3959 : * If there is a pending propagation, we have to update the load and
3960 : * the utilization of the sched_entity:
3961 : */
3962 : if (gcfs_rq->propagate)
3963 : return false;
3964 :
3965 : /*
3966 : * Otherwise, the load and the utilization of the sched_entity is
3967 : * already zero and there is no pending propagation, so it will be a
3968 : * waste of time to try to decay it:
3969 : */
3970 : return true;
3971 : }
3972 :
3973 : #else /* CONFIG_FAIR_GROUP_SCHED */
3974 :
3975 : static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3976 :
3977 : static inline int propagate_entity_load_avg(struct sched_entity *se)
3978 : {
3979 : return 0;
3980 : }
3981 :
3982 : static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3983 :
3984 : #endif /* CONFIG_FAIR_GROUP_SCHED */
3985 :
3986 : #ifdef CONFIG_NO_HZ_COMMON
3987 : static inline void migrate_se_pelt_lag(struct sched_entity *se)
3988 : {
3989 : u64 throttled = 0, now, lut;
3990 : struct cfs_rq *cfs_rq;
3991 : struct rq *rq;
3992 : bool is_idle;
3993 :
3994 : if (load_avg_is_decayed(&se->avg))
3995 : return;
3996 :
3997 : cfs_rq = cfs_rq_of(se);
3998 : rq = rq_of(cfs_rq);
3999 :
4000 : rcu_read_lock();
4001 : is_idle = is_idle_task(rcu_dereference(rq->curr));
4002 : rcu_read_unlock();
4003 :
4004 : /*
4005 : * The lag estimation comes with a cost we don't want to pay all the
4006 : * time. Hence, limiting to the case where the source CPU is idle and
4007 : * we know we are at the greatest risk to have an outdated clock.
4008 : */
4009 : if (!is_idle)
4010 : return;
4011 :
4012 : /*
4013 : * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4014 : *
4015 : * last_update_time (the cfs_rq's last_update_time)
4016 : * = cfs_rq_clock_pelt()@cfs_rq_idle
4017 : * = rq_clock_pelt()@cfs_rq_idle
4018 : * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4019 : *
4020 : * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4021 : * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4022 : *
4023 : * rq_idle_lag (delta between now and rq's update)
4024 : * = sched_clock_cpu() - rq_clock()@rq_idle
4025 : *
4026 : * We can then write:
4027 : *
4028 : * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4029 : * sched_clock_cpu() - rq_clock()@rq_idle
4030 : * Where:
4031 : * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4032 : * rq_clock()@rq_idle is rq->clock_idle
4033 : * cfs->throttled_clock_pelt_time@cfs_rq_idle
4034 : * is cfs_rq->throttled_pelt_idle
4035 : */
4036 :
4037 : #ifdef CONFIG_CFS_BANDWIDTH
4038 : throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4039 : /* The clock has been stopped for throttling */
4040 : if (throttled == U64_MAX)
4041 : return;
4042 : #endif
4043 : now = u64_u32_load(rq->clock_pelt_idle);
4044 : /*
4045 : * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4046 : * is observed the old clock_pelt_idle value and the new clock_idle,
4047 : * which lead to an underestimation. The opposite would lead to an
4048 : * overestimation.
4049 : */
4050 : smp_rmb();
4051 : lut = cfs_rq_last_update_time(cfs_rq);
4052 :
4053 : now -= throttled;
4054 : if (now < lut)
4055 : /*
4056 : * cfs_rq->avg.last_update_time is more recent than our
4057 : * estimation, let's use it.
4058 : */
4059 : now = lut;
4060 : else
4061 : now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4062 :
4063 : __update_load_avg_blocked_se(now, se);
4064 : }
4065 : #else
4066 : static void migrate_se_pelt_lag(struct sched_entity *se) {}
4067 : #endif
4068 :
4069 : /**
4070 : * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4071 : * @now: current time, as per cfs_rq_clock_pelt()
4072 : * @cfs_rq: cfs_rq to update
4073 : *
4074 : * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4075 : * avg. The immediate corollary is that all (fair) tasks must be attached.
4076 : *
4077 : * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4078 : *
4079 : * Return: true if the load decayed or we removed load.
4080 : *
4081 : * Since both these conditions indicate a changed cfs_rq->avg.load we should
4082 : * call update_tg_load_avg() when this function returns true.
4083 : */
4084 : static inline int
4085 : update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4086 : {
4087 : unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4088 : struct sched_avg *sa = &cfs_rq->avg;
4089 : int decayed = 0;
4090 :
4091 : if (cfs_rq->removed.nr) {
4092 : unsigned long r;
4093 : u32 divider = get_pelt_divider(&cfs_rq->avg);
4094 :
4095 : raw_spin_lock(&cfs_rq->removed.lock);
4096 : swap(cfs_rq->removed.util_avg, removed_util);
4097 : swap(cfs_rq->removed.load_avg, removed_load);
4098 : swap(cfs_rq->removed.runnable_avg, removed_runnable);
4099 : cfs_rq->removed.nr = 0;
4100 : raw_spin_unlock(&cfs_rq->removed.lock);
4101 :
4102 : r = removed_load;
4103 : sub_positive(&sa->load_avg, r);
4104 : sub_positive(&sa->load_sum, r * divider);
4105 : /* See sa->util_sum below */
4106 : sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4107 :
4108 : r = removed_util;
4109 : sub_positive(&sa->util_avg, r);
4110 : sub_positive(&sa->util_sum, r * divider);
4111 : /*
4112 : * Because of rounding, se->util_sum might ends up being +1 more than
4113 : * cfs->util_sum. Although this is not a problem by itself, detaching
4114 : * a lot of tasks with the rounding problem between 2 updates of
4115 : * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4116 : * cfs_util_avg is not.
4117 : * Check that util_sum is still above its lower bound for the new
4118 : * util_avg. Given that period_contrib might have moved since the last
4119 : * sync, we are only sure that util_sum must be above or equal to
4120 : * util_avg * minimum possible divider
4121 : */
4122 : sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4123 :
4124 : r = removed_runnable;
4125 : sub_positive(&sa->runnable_avg, r);
4126 : sub_positive(&sa->runnable_sum, r * divider);
4127 : /* See sa->util_sum above */
4128 : sa->runnable_sum = max_t(u32, sa->runnable_sum,
4129 : sa->runnable_avg * PELT_MIN_DIVIDER);
4130 :
4131 : /*
4132 : * removed_runnable is the unweighted version of removed_load so we
4133 : * can use it to estimate removed_load_sum.
4134 : */
4135 : add_tg_cfs_propagate(cfs_rq,
4136 : -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4137 :
4138 : decayed = 1;
4139 : }
4140 :
4141 : decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4142 : u64_u32_store_copy(sa->last_update_time,
4143 : cfs_rq->last_update_time_copy,
4144 : sa->last_update_time);
4145 : return decayed;
4146 : }
4147 :
4148 : /**
4149 : * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4150 : * @cfs_rq: cfs_rq to attach to
4151 : * @se: sched_entity to attach
4152 : *
4153 : * Must call update_cfs_rq_load_avg() before this, since we rely on
4154 : * cfs_rq->avg.last_update_time being current.
4155 : */
4156 : static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4157 : {
4158 : /*
4159 : * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4160 : * See ___update_load_avg() for details.
4161 : */
4162 : u32 divider = get_pelt_divider(&cfs_rq->avg);
4163 :
4164 : /*
4165 : * When we attach the @se to the @cfs_rq, we must align the decay
4166 : * window because without that, really weird and wonderful things can
4167 : * happen.
4168 : *
4169 : * XXX illustrate
4170 : */
4171 : se->avg.last_update_time = cfs_rq->avg.last_update_time;
4172 : se->avg.period_contrib = cfs_rq->avg.period_contrib;
4173 :
4174 : /*
4175 : * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4176 : * period_contrib. This isn't strictly correct, but since we're
4177 : * entirely outside of the PELT hierarchy, nobody cares if we truncate
4178 : * _sum a little.
4179 : */
4180 : se->avg.util_sum = se->avg.util_avg * divider;
4181 :
4182 : se->avg.runnable_sum = se->avg.runnable_avg * divider;
4183 :
4184 : se->avg.load_sum = se->avg.load_avg * divider;
4185 : if (se_weight(se) < se->avg.load_sum)
4186 : se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4187 : else
4188 : se->avg.load_sum = 1;
4189 :
4190 : enqueue_load_avg(cfs_rq, se);
4191 : cfs_rq->avg.util_avg += se->avg.util_avg;
4192 : cfs_rq->avg.util_sum += se->avg.util_sum;
4193 : cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4194 : cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4195 :
4196 : add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4197 :
4198 : cfs_rq_util_change(cfs_rq, 0);
4199 :
4200 : trace_pelt_cfs_tp(cfs_rq);
4201 : }
4202 :
4203 : /**
4204 : * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4205 : * @cfs_rq: cfs_rq to detach from
4206 : * @se: sched_entity to detach
4207 : *
4208 : * Must call update_cfs_rq_load_avg() before this, since we rely on
4209 : * cfs_rq->avg.last_update_time being current.
4210 : */
4211 : static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4212 : {
4213 : dequeue_load_avg(cfs_rq, se);
4214 : sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4215 : sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4216 : /* See update_cfs_rq_load_avg() */
4217 : cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4218 : cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4219 :
4220 : sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4221 : sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4222 : /* See update_cfs_rq_load_avg() */
4223 : cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4224 : cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4225 :
4226 : add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4227 :
4228 : cfs_rq_util_change(cfs_rq, 0);
4229 :
4230 : trace_pelt_cfs_tp(cfs_rq);
4231 : }
4232 :
4233 : /*
4234 : * Optional action to be done while updating the load average
4235 : */
4236 : #define UPDATE_TG 0x1
4237 : #define SKIP_AGE_LOAD 0x2
4238 : #define DO_ATTACH 0x4
4239 : #define DO_DETACH 0x8
4240 :
4241 : /* Update task and its cfs_rq load average */
4242 : static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4243 : {
4244 : u64 now = cfs_rq_clock_pelt(cfs_rq);
4245 : int decayed;
4246 :
4247 : /*
4248 : * Track task load average for carrying it to new CPU after migrated, and
4249 : * track group sched_entity load average for task_h_load calc in migration
4250 : */
4251 : if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4252 : __update_load_avg_se(now, cfs_rq, se);
4253 :
4254 : decayed = update_cfs_rq_load_avg(now, cfs_rq);
4255 : decayed |= propagate_entity_load_avg(se);
4256 :
4257 : if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4258 :
4259 : /*
4260 : * DO_ATTACH means we're here from enqueue_entity().
4261 : * !last_update_time means we've passed through
4262 : * migrate_task_rq_fair() indicating we migrated.
4263 : *
4264 : * IOW we're enqueueing a task on a new CPU.
4265 : */
4266 : attach_entity_load_avg(cfs_rq, se);
4267 : update_tg_load_avg(cfs_rq);
4268 :
4269 : } else if (flags & DO_DETACH) {
4270 : /*
4271 : * DO_DETACH means we're here from dequeue_entity()
4272 : * and we are migrating task out of the CPU.
4273 : */
4274 : detach_entity_load_avg(cfs_rq, se);
4275 : update_tg_load_avg(cfs_rq);
4276 : } else if (decayed) {
4277 : cfs_rq_util_change(cfs_rq, 0);
4278 :
4279 : if (flags & UPDATE_TG)
4280 : update_tg_load_avg(cfs_rq);
4281 : }
4282 : }
4283 :
4284 : /*
4285 : * Synchronize entity load avg of dequeued entity without locking
4286 : * the previous rq.
4287 : */
4288 : static void sync_entity_load_avg(struct sched_entity *se)
4289 : {
4290 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
4291 : u64 last_update_time;
4292 :
4293 : last_update_time = cfs_rq_last_update_time(cfs_rq);
4294 : __update_load_avg_blocked_se(last_update_time, se);
4295 : }
4296 :
4297 : /*
4298 : * Task first catches up with cfs_rq, and then subtract
4299 : * itself from the cfs_rq (task must be off the queue now).
4300 : */
4301 : static void remove_entity_load_avg(struct sched_entity *se)
4302 : {
4303 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
4304 : unsigned long flags;
4305 :
4306 : /*
4307 : * tasks cannot exit without having gone through wake_up_new_task() ->
4308 : * enqueue_task_fair() which will have added things to the cfs_rq,
4309 : * so we can remove unconditionally.
4310 : */
4311 :
4312 : sync_entity_load_avg(se);
4313 :
4314 : raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4315 : ++cfs_rq->removed.nr;
4316 : cfs_rq->removed.util_avg += se->avg.util_avg;
4317 : cfs_rq->removed.load_avg += se->avg.load_avg;
4318 : cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4319 : raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4320 : }
4321 :
4322 : static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4323 : {
4324 : return cfs_rq->avg.runnable_avg;
4325 : }
4326 :
4327 : static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4328 : {
4329 : return cfs_rq->avg.load_avg;
4330 : }
4331 :
4332 : static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4333 :
4334 : static inline unsigned long task_util(struct task_struct *p)
4335 : {
4336 : return READ_ONCE(p->se.avg.util_avg);
4337 : }
4338 :
4339 : static inline unsigned long _task_util_est(struct task_struct *p)
4340 : {
4341 : struct util_est ue = READ_ONCE(p->se.avg.util_est);
4342 :
4343 : return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4344 : }
4345 :
4346 : static inline unsigned long task_util_est(struct task_struct *p)
4347 : {
4348 : return max(task_util(p), _task_util_est(p));
4349 : }
4350 :
4351 : #ifdef CONFIG_UCLAMP_TASK
4352 : static inline unsigned long uclamp_task_util(struct task_struct *p,
4353 : unsigned long uclamp_min,
4354 : unsigned long uclamp_max)
4355 : {
4356 : return clamp(task_util_est(p), uclamp_min, uclamp_max);
4357 : }
4358 : #else
4359 : static inline unsigned long uclamp_task_util(struct task_struct *p,
4360 : unsigned long uclamp_min,
4361 : unsigned long uclamp_max)
4362 : {
4363 : return task_util_est(p);
4364 : }
4365 : #endif
4366 :
4367 : static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4368 : struct task_struct *p)
4369 : {
4370 : unsigned int enqueued;
4371 :
4372 : if (!sched_feat(UTIL_EST))
4373 : return;
4374 :
4375 : /* Update root cfs_rq's estimated utilization */
4376 : enqueued = cfs_rq->avg.util_est.enqueued;
4377 : enqueued += _task_util_est(p);
4378 : WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4379 :
4380 : trace_sched_util_est_cfs_tp(cfs_rq);
4381 : }
4382 :
4383 : static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4384 : struct task_struct *p)
4385 : {
4386 : unsigned int enqueued;
4387 :
4388 : if (!sched_feat(UTIL_EST))
4389 : return;
4390 :
4391 : /* Update root cfs_rq's estimated utilization */
4392 : enqueued = cfs_rq->avg.util_est.enqueued;
4393 : enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4394 : WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4395 :
4396 : trace_sched_util_est_cfs_tp(cfs_rq);
4397 : }
4398 :
4399 : #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4400 :
4401 : /*
4402 : * Check if a (signed) value is within a specified (unsigned) margin,
4403 : * based on the observation that:
4404 : *
4405 : * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4406 : *
4407 : * NOTE: this only works when value + margin < INT_MAX.
4408 : */
4409 : static inline bool within_margin(int value, int margin)
4410 : {
4411 : return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4412 : }
4413 :
4414 : static inline void util_est_update(struct cfs_rq *cfs_rq,
4415 : struct task_struct *p,
4416 : bool task_sleep)
4417 : {
4418 : long last_ewma_diff, last_enqueued_diff;
4419 : struct util_est ue;
4420 :
4421 : if (!sched_feat(UTIL_EST))
4422 : return;
4423 :
4424 : /*
4425 : * Skip update of task's estimated utilization when the task has not
4426 : * yet completed an activation, e.g. being migrated.
4427 : */
4428 : if (!task_sleep)
4429 : return;
4430 :
4431 : /*
4432 : * If the PELT values haven't changed since enqueue time,
4433 : * skip the util_est update.
4434 : */
4435 : ue = p->se.avg.util_est;
4436 : if (ue.enqueued & UTIL_AVG_UNCHANGED)
4437 : return;
4438 :
4439 : last_enqueued_diff = ue.enqueued;
4440 :
4441 : /*
4442 : * Reset EWMA on utilization increases, the moving average is used only
4443 : * to smooth utilization decreases.
4444 : */
4445 : ue.enqueued = task_util(p);
4446 : if (sched_feat(UTIL_EST_FASTUP)) {
4447 : if (ue.ewma < ue.enqueued) {
4448 : ue.ewma = ue.enqueued;
4449 : goto done;
4450 : }
4451 : }
4452 :
4453 : /*
4454 : * Skip update of task's estimated utilization when its members are
4455 : * already ~1% close to its last activation value.
4456 : */
4457 : last_ewma_diff = ue.enqueued - ue.ewma;
4458 : last_enqueued_diff -= ue.enqueued;
4459 : if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4460 : if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4461 : goto done;
4462 :
4463 : return;
4464 : }
4465 :
4466 : /*
4467 : * To avoid overestimation of actual task utilization, skip updates if
4468 : * we cannot grant there is idle time in this CPU.
4469 : */
4470 : if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4471 : return;
4472 :
4473 : /*
4474 : * Update Task's estimated utilization
4475 : *
4476 : * When *p completes an activation we can consolidate another sample
4477 : * of the task size. This is done by storing the current PELT value
4478 : * as ue.enqueued and by using this value to update the Exponential
4479 : * Weighted Moving Average (EWMA):
4480 : *
4481 : * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4482 : * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4483 : * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4484 : * = w * ( last_ewma_diff ) + ewma(t-1)
4485 : * = w * (last_ewma_diff + ewma(t-1) / w)
4486 : *
4487 : * Where 'w' is the weight of new samples, which is configured to be
4488 : * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4489 : */
4490 : ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4491 : ue.ewma += last_ewma_diff;
4492 : ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4493 : done:
4494 : ue.enqueued |= UTIL_AVG_UNCHANGED;
4495 : WRITE_ONCE(p->se.avg.util_est, ue);
4496 :
4497 : trace_sched_util_est_se_tp(&p->se);
4498 : }
4499 :
4500 : static inline int util_fits_cpu(unsigned long util,
4501 : unsigned long uclamp_min,
4502 : unsigned long uclamp_max,
4503 : int cpu)
4504 : {
4505 : unsigned long capacity_orig, capacity_orig_thermal;
4506 : unsigned long capacity = capacity_of(cpu);
4507 : bool fits, uclamp_max_fits;
4508 :
4509 : /*
4510 : * Check if the real util fits without any uclamp boost/cap applied.
4511 : */
4512 : fits = fits_capacity(util, capacity);
4513 :
4514 : if (!uclamp_is_used())
4515 : return fits;
4516 :
4517 : /*
4518 : * We must use capacity_orig_of() for comparing against uclamp_min and
4519 : * uclamp_max. We only care about capacity pressure (by using
4520 : * capacity_of()) for comparing against the real util.
4521 : *
4522 : * If a task is boosted to 1024 for example, we don't want a tiny
4523 : * pressure to skew the check whether it fits a CPU or not.
4524 : *
4525 : * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4526 : * should fit a little cpu even if there's some pressure.
4527 : *
4528 : * Only exception is for thermal pressure since it has a direct impact
4529 : * on available OPP of the system.
4530 : *
4531 : * We honour it for uclamp_min only as a drop in performance level
4532 : * could result in not getting the requested minimum performance level.
4533 : *
4534 : * For uclamp_max, we can tolerate a drop in performance level as the
4535 : * goal is to cap the task. So it's okay if it's getting less.
4536 : */
4537 : capacity_orig = capacity_orig_of(cpu);
4538 : capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4539 :
4540 : /*
4541 : * We want to force a task to fit a cpu as implied by uclamp_max.
4542 : * But we do have some corner cases to cater for..
4543 : *
4544 : *
4545 : * C=z
4546 : * | ___
4547 : * | C=y | |
4548 : * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4549 : * | C=x | | | |
4550 : * | ___ | | | |
4551 : * | | | | | | | (util somewhere in this region)
4552 : * | | | | | | |
4553 : * | | | | | | |
4554 : * +----------------------------------------
4555 : * cpu0 cpu1 cpu2
4556 : *
4557 : * In the above example if a task is capped to a specific performance
4558 : * point, y, then when:
4559 : *
4560 : * * util = 80% of x then it does not fit on cpu0 and should migrate
4561 : * to cpu1
4562 : * * util = 80% of y then it is forced to fit on cpu1 to honour
4563 : * uclamp_max request.
4564 : *
4565 : * which is what we're enforcing here. A task always fits if
4566 : * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4567 : * the normal upmigration rules should withhold still.
4568 : *
4569 : * Only exception is when we are on max capacity, then we need to be
4570 : * careful not to block overutilized state. This is so because:
4571 : *
4572 : * 1. There's no concept of capping at max_capacity! We can't go
4573 : * beyond this performance level anyway.
4574 : * 2. The system is being saturated when we're operating near
4575 : * max capacity, it doesn't make sense to block overutilized.
4576 : */
4577 : uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4578 : uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4579 : fits = fits || uclamp_max_fits;
4580 :
4581 : /*
4582 : *
4583 : * C=z
4584 : * | ___ (region a, capped, util >= uclamp_max)
4585 : * | C=y | |
4586 : * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4587 : * | C=x | | | |
4588 : * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
4589 : * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4590 : * | | | | | | |
4591 : * | | | | | | | (region c, boosted, util < uclamp_min)
4592 : * +----------------------------------------
4593 : * cpu0 cpu1 cpu2
4594 : *
4595 : * a) If util > uclamp_max, then we're capped, we don't care about
4596 : * actual fitness value here. We only care if uclamp_max fits
4597 : * capacity without taking margin/pressure into account.
4598 : * See comment above.
4599 : *
4600 : * b) If uclamp_min <= util <= uclamp_max, then the normal
4601 : * fits_capacity() rules apply. Except we need to ensure that we
4602 : * enforce we remain within uclamp_max, see comment above.
4603 : *
4604 : * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4605 : * need to take into account the boosted value fits the CPU without
4606 : * taking margin/pressure into account.
4607 : *
4608 : * Cases (a) and (b) are handled in the 'fits' variable already. We
4609 : * just need to consider an extra check for case (c) after ensuring we
4610 : * handle the case uclamp_min > uclamp_max.
4611 : */
4612 : uclamp_min = min(uclamp_min, uclamp_max);
4613 : if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
4614 : return -1;
4615 :
4616 : return fits;
4617 : }
4618 :
4619 : static inline int task_fits_cpu(struct task_struct *p, int cpu)
4620 : {
4621 : unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4622 : unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4623 : unsigned long util = task_util_est(p);
4624 : /*
4625 : * Return true only if the cpu fully fits the task requirements, which
4626 : * include the utilization but also the performance hints.
4627 : */
4628 : return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
4629 : }
4630 :
4631 : static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4632 : {
4633 : if (!sched_asym_cpucap_active())
4634 : return;
4635 :
4636 : if (!p || p->nr_cpus_allowed == 1) {
4637 : rq->misfit_task_load = 0;
4638 : return;
4639 : }
4640 :
4641 : if (task_fits_cpu(p, cpu_of(rq))) {
4642 : rq->misfit_task_load = 0;
4643 : return;
4644 : }
4645 :
4646 : /*
4647 : * Make sure that misfit_task_load will not be null even if
4648 : * task_h_load() returns 0.
4649 : */
4650 : rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4651 : }
4652 :
4653 : #else /* CONFIG_SMP */
4654 :
4655 : static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4656 : {
4657 : return true;
4658 : }
4659 :
4660 : #define UPDATE_TG 0x0
4661 : #define SKIP_AGE_LOAD 0x0
4662 : #define DO_ATTACH 0x0
4663 : #define DO_DETACH 0x0
4664 :
4665 : static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4666 : {
4667 7820 : cfs_rq_util_change(cfs_rq, 0);
4668 : }
4669 :
4670 : static inline void remove_entity_load_avg(struct sched_entity *se) {}
4671 :
4672 : static inline void
4673 : attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4674 : static inline void
4675 : detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4676 :
4677 : static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4678 : {
4679 : return 0;
4680 : }
4681 :
4682 : static inline void
4683 : util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4684 :
4685 : static inline void
4686 : util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4687 :
4688 : static inline void
4689 : util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4690 : bool task_sleep) {}
4691 : static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4692 :
4693 : #endif /* CONFIG_SMP */
4694 :
4695 : static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4696 : {
4697 : #ifdef CONFIG_SCHED_DEBUG
4698 : s64 d = se->vruntime - cfs_rq->min_vruntime;
4699 :
4700 : if (d < 0)
4701 : d = -d;
4702 :
4703 : if (d > 3*sysctl_sched_latency)
4704 : schedstat_inc(cfs_rq->nr_spread_over);
4705 : #endif
4706 : }
4707 :
4708 : static inline bool entity_is_long_sleeper(struct sched_entity *se)
4709 : {
4710 : struct cfs_rq *cfs_rq;
4711 : u64 sleep_time;
4712 :
4713 2442 : if (se->exec_start == 0)
4714 : return false;
4715 :
4716 4120 : cfs_rq = cfs_rq_of(se);
4717 :
4718 4120 : sleep_time = rq_clock_task(rq_of(cfs_rq));
4719 :
4720 : /* Happen while migrating because of clock task divergence */
4721 2060 : if (sleep_time <= se->exec_start)
4722 : return false;
4723 :
4724 111 : sleep_time -= se->exec_start;
4725 111 : if (sleep_time > ((1ULL << 63) / scale_load_down(NICE_0_LOAD)))
4726 : return true;
4727 :
4728 : return false;
4729 : }
4730 :
4731 : static void
4732 2442 : place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4733 : {
4734 2442 : u64 vruntime = cfs_rq->min_vruntime;
4735 :
4736 : /*
4737 : * The 'current' period is already promised to the current tasks,
4738 : * however the extra weight of the new task will slow them down a
4739 : * little, place the new task so that it fits in the slot that
4740 : * stays open at the end.
4741 : */
4742 2442 : if (initial && sched_feat(START_DEBIT))
4743 382 : vruntime += sched_vslice(cfs_rq, se);
4744 :
4745 : /* sleeps up to a single latency don't count. */
4746 2442 : if (!initial) {
4747 : unsigned long thresh;
4748 :
4749 2060 : if (se_is_idle(se))
4750 : thresh = sysctl_sched_min_granularity;
4751 : else
4752 2060 : thresh = sysctl_sched_latency;
4753 :
4754 : /*
4755 : * Halve their sleep time's effect, to allow
4756 : * for a gentler effect of sleepers:
4757 : */
4758 : if (sched_feat(GENTLE_FAIR_SLEEPERS))
4759 2060 : thresh >>= 1;
4760 :
4761 2060 : vruntime -= thresh;
4762 : }
4763 :
4764 : /*
4765 : * Pull vruntime of the entity being placed to the base level of
4766 : * cfs_rq, to prevent boosting it if placed backwards.
4767 : * However, min_vruntime can advance much faster than real time, with
4768 : * the extreme being when an entity with the minimal weight always runs
4769 : * on the cfs_rq. If the waking entity slept for a long time, its
4770 : * vruntime difference from min_vruntime may overflow s64 and their
4771 : * comparison may get inversed, so ignore the entity's original
4772 : * vruntime in that case.
4773 : * The maximal vruntime speedup is given by the ratio of normal to
4774 : * minimal weight: scale_load_down(NICE_0_LOAD) / MIN_SHARES.
4775 : * When placing a migrated waking entity, its exec_start has been set
4776 : * from a different rq. In order to take into account a possible
4777 : * divergence between new and prev rq's clocks task because of irq and
4778 : * stolen time, we take an additional margin.
4779 : * So, cutting off on the sleep time of
4780 : * 2^63 / scale_load_down(NICE_0_LOAD) ~ 104 days
4781 : * should be safe.
4782 : */
4783 4884 : if (entity_is_long_sleeper(se))
4784 0 : se->vruntime = vruntime;
4785 : else
4786 4884 : se->vruntime = max_vruntime(se->vruntime, vruntime);
4787 2442 : }
4788 :
4789 : static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4790 :
4791 : static inline bool cfs_bandwidth_used(void);
4792 :
4793 : /*
4794 : * MIGRATION
4795 : *
4796 : * dequeue
4797 : * update_curr()
4798 : * update_min_vruntime()
4799 : * vruntime -= min_vruntime
4800 : *
4801 : * enqueue
4802 : * update_curr()
4803 : * update_min_vruntime()
4804 : * vruntime += min_vruntime
4805 : *
4806 : * this way the vruntime transition between RQs is done when both
4807 : * min_vruntime are up-to-date.
4808 : *
4809 : * WAKEUP (remote)
4810 : *
4811 : * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4812 : * vruntime -= min_vruntime
4813 : *
4814 : * enqueue
4815 : * update_curr()
4816 : * update_min_vruntime()
4817 : * vruntime += min_vruntime
4818 : *
4819 : * this way we don't have the most up-to-date min_vruntime on the originating
4820 : * CPU and an up-to-date min_vruntime on the destination CPU.
4821 : */
4822 :
4823 : static void
4824 2446 : enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4825 : {
4826 2446 : bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4827 2446 : bool curr = cfs_rq->curr == se;
4828 :
4829 : /*
4830 : * If we're the current task, we must renormalise before calling
4831 : * update_curr().
4832 : */
4833 2446 : if (renorm && curr)
4834 0 : se->vruntime += cfs_rq->min_vruntime;
4835 :
4836 2446 : update_curr(cfs_rq);
4837 :
4838 : /*
4839 : * Otherwise, renormalise after, such that we're placed at the current
4840 : * moment in time, instead of some random moment in the past. Being
4841 : * placed in the past could significantly boost this task to the
4842 : * fairness detriment of existing tasks.
4843 : */
4844 2446 : if (renorm && !curr)
4845 386 : se->vruntime += cfs_rq->min_vruntime;
4846 :
4847 : /*
4848 : * When enqueuing a sched_entity, we must:
4849 : * - Update loads to have both entity and cfs_rq synced with now.
4850 : * - For group_entity, update its runnable_weight to reflect the new
4851 : * h_nr_running of its group cfs_rq.
4852 : * - For group_entity, update its weight to reflect the new share of
4853 : * its group cfs_rq
4854 : * - Add its new weight to cfs_rq->load.weight
4855 : */
4856 2446 : update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4857 2446 : se_update_runnable(se);
4858 2446 : update_cfs_group(se);
4859 4892 : account_entity_enqueue(cfs_rq, se);
4860 :
4861 2446 : if (flags & ENQUEUE_WAKEUP)
4862 2060 : place_entity(cfs_rq, se, 0);
4863 : /* Entity has migrated, no longer consider this task hot */
4864 : if (flags & ENQUEUE_MIGRATED)
4865 : se->exec_start = 0;
4866 :
4867 : check_schedstat_required();
4868 2446 : update_stats_enqueue_fair(cfs_rq, se, flags);
4869 2446 : check_spread(cfs_rq, se);
4870 2446 : if (!curr)
4871 2446 : __enqueue_entity(cfs_rq, se);
4872 2446 : se->on_rq = 1;
4873 :
4874 : if (cfs_rq->nr_running == 1) {
4875 : check_enqueue_throttle(cfs_rq);
4876 : if (!throttled_hierarchy(cfs_rq))
4877 : list_add_leaf_cfs_rq(cfs_rq);
4878 : }
4879 2446 : }
4880 :
4881 : static void __clear_buddies_last(struct sched_entity *se)
4882 : {
4883 0 : for_each_sched_entity(se) {
4884 0 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
4885 0 : if (cfs_rq->last != se)
4886 : break;
4887 :
4888 0 : cfs_rq->last = NULL;
4889 : }
4890 : }
4891 :
4892 : static void __clear_buddies_next(struct sched_entity *se)
4893 : {
4894 837 : for_each_sched_entity(se) {
4895 1674 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
4896 837 : if (cfs_rq->next != se)
4897 : break;
4898 :
4899 837 : cfs_rq->next = NULL;
4900 : }
4901 : }
4902 :
4903 : static void __clear_buddies_skip(struct sched_entity *se)
4904 : {
4905 0 : for_each_sched_entity(se) {
4906 0 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
4907 0 : if (cfs_rq->skip != se)
4908 : break;
4909 :
4910 0 : cfs_rq->skip = NULL;
4911 : }
4912 : }
4913 :
4914 5195 : static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4915 : {
4916 5195 : if (cfs_rq->last == se)
4917 : __clear_buddies_last(se);
4918 :
4919 5195 : if (cfs_rq->next == se)
4920 : __clear_buddies_next(se);
4921 :
4922 5195 : if (cfs_rq->skip == se)
4923 : __clear_buddies_skip(se);
4924 5195 : }
4925 :
4926 : static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4927 :
4928 : static void
4929 2444 : dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4930 : {
4931 2444 : int action = UPDATE_TG;
4932 :
4933 4888 : if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
4934 : action |= DO_DETACH;
4935 :
4936 : /*
4937 : * Update run-time statistics of the 'current'.
4938 : */
4939 2444 : update_curr(cfs_rq);
4940 :
4941 : /*
4942 : * When dequeuing a sched_entity, we must:
4943 : * - Update loads to have both entity and cfs_rq synced with now.
4944 : * - For group_entity, update its runnable_weight to reflect the new
4945 : * h_nr_running of its group cfs_rq.
4946 : * - Subtract its previous weight from cfs_rq->load.weight.
4947 : * - For group entity, update its weight to reflect the new share
4948 : * of its group cfs_rq.
4949 : */
4950 2444 : update_load_avg(cfs_rq, se, action);
4951 2444 : se_update_runnable(se);
4952 :
4953 2444 : update_stats_dequeue_fair(cfs_rq, se, flags);
4954 :
4955 2444 : clear_buddies(cfs_rq, se);
4956 :
4957 2444 : if (se != cfs_rq->curr)
4958 : __dequeue_entity(cfs_rq, se);
4959 2444 : se->on_rq = 0;
4960 4888 : account_entity_dequeue(cfs_rq, se);
4961 :
4962 : /*
4963 : * Normalize after update_curr(); which will also have moved
4964 : * min_vruntime if @se is the one holding it back. But before doing
4965 : * update_min_vruntime() again, which will discount @se's position and
4966 : * can move min_vruntime forward still more.
4967 : */
4968 2444 : if (!(flags & DEQUEUE_SLEEP))
4969 4 : se->vruntime -= cfs_rq->min_vruntime;
4970 :
4971 : /* return excess runtime on last dequeue */
4972 2444 : return_cfs_rq_runtime(cfs_rq);
4973 :
4974 2444 : update_cfs_group(se);
4975 :
4976 : /*
4977 : * Now advance min_vruntime if @se was the entity holding it back,
4978 : * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4979 : * put back on, and if we advance min_vruntime, we'll be placed back
4980 : * further than we started -- ie. we'll be penalized.
4981 : */
4982 2444 : if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4983 2440 : update_min_vruntime(cfs_rq);
4984 :
4985 : if (cfs_rq->nr_running == 0)
4986 : update_idle_cfs_rq_clock_pelt(cfs_rq);
4987 2444 : }
4988 :
4989 : /*
4990 : * Preempt the current task with a newly woken task if needed:
4991 : */
4992 : static void
4993 235 : check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4994 : {
4995 : unsigned long ideal_runtime, delta_exec;
4996 : struct sched_entity *se;
4997 : s64 delta;
4998 :
4999 : /*
5000 : * When many tasks blow up the sched_period; it is possible that
5001 : * sched_slice() reports unusually large results (when many tasks are
5002 : * very light for example). Therefore impose a maximum.
5003 : */
5004 235 : ideal_runtime = min_t(u64, sched_slice(cfs_rq, curr), sysctl_sched_latency);
5005 :
5006 235 : delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
5007 235 : if (delta_exec > ideal_runtime) {
5008 235 : resched_curr(rq_of(cfs_rq));
5009 : /*
5010 : * The current task ran long enough, ensure it doesn't get
5011 : * re-elected due to buddy favours.
5012 : */
5013 235 : clear_buddies(cfs_rq, curr);
5014 235 : return;
5015 : }
5016 :
5017 : /*
5018 : * Ensure that a task that missed wakeup preemption by a
5019 : * narrow margin doesn't have to wait for a full slice.
5020 : * This also mitigates buddy induced latencies under load.
5021 : */
5022 0 : if (delta_exec < sysctl_sched_min_granularity)
5023 : return;
5024 :
5025 0 : se = __pick_first_entity(cfs_rq);
5026 0 : delta = curr->vruntime - se->vruntime;
5027 :
5028 0 : if (delta < 0)
5029 : return;
5030 :
5031 0 : if (delta > ideal_runtime)
5032 0 : resched_curr(rq_of(cfs_rq));
5033 : }
5034 :
5035 : static void
5036 2516 : set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5037 : {
5038 2516 : clear_buddies(cfs_rq, se);
5039 :
5040 : /* 'current' is not kept within the tree. */
5041 2516 : if (se->on_rq) {
5042 : /*
5043 : * Any task has to be enqueued before it get to execute on
5044 : * a CPU. So account for the time it spent waiting on the
5045 : * runqueue.
5046 : */
5047 5032 : update_stats_wait_end_fair(cfs_rq, se);
5048 : __dequeue_entity(cfs_rq, se);
5049 : update_load_avg(cfs_rq, se, UPDATE_TG);
5050 : }
5051 :
5052 5032 : update_stats_curr_start(cfs_rq, se);
5053 2516 : cfs_rq->curr = se;
5054 :
5055 : /*
5056 : * Track our maximum slice length, if the CPU's load is at
5057 : * least twice that of our own weight (i.e. dont track it
5058 : * when there are only lesser-weight tasks around):
5059 : */
5060 : if (schedstat_enabled() &&
5061 : rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5062 : struct sched_statistics *stats;
5063 :
5064 : stats = __schedstats_from_se(se);
5065 : __schedstat_set(stats->slice_max,
5066 : max((u64)stats->slice_max,
5067 : se->sum_exec_runtime - se->prev_sum_exec_runtime));
5068 : }
5069 :
5070 2516 : se->prev_sum_exec_runtime = se->sum_exec_runtime;
5071 2516 : }
5072 :
5073 : static int
5074 : wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
5075 :
5076 : /*
5077 : * Pick the next process, keeping these things in mind, in this order:
5078 : * 1) keep things fair between processes/task groups
5079 : * 2) pick the "next" process, since someone really wants that to run
5080 : * 3) pick the "last" process, for cache locality
5081 : * 4) do not run the "skip" process, if something else is available
5082 : */
5083 : static struct sched_entity *
5084 2512 : pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
5085 : {
5086 2512 : struct sched_entity *left = __pick_first_entity(cfs_rq);
5087 : struct sched_entity *se;
5088 :
5089 : /*
5090 : * If curr is set we have to see if its left of the leftmost entity
5091 : * still in the tree, provided there was anything in the tree at all.
5092 : */
5093 2512 : if (!left || (curr && entity_before(curr, left)))
5094 : left = curr;
5095 :
5096 2512 : se = left; /* ideally we run the leftmost entity */
5097 :
5098 : /*
5099 : * Avoid running the skip buddy, if running something else can
5100 : * be done without getting too unfair.
5101 : */
5102 2512 : if (cfs_rq->skip && cfs_rq->skip == se) {
5103 : struct sched_entity *second;
5104 :
5105 0 : if (se == curr) {
5106 : second = __pick_first_entity(cfs_rq);
5107 : } else {
5108 0 : second = __pick_next_entity(se);
5109 0 : if (!second || (curr && entity_before(curr, second)))
5110 : second = curr;
5111 : }
5112 :
5113 0 : if (second && wakeup_preempt_entity(second, left) < 1)
5114 0 : se = second;
5115 : }
5116 :
5117 2512 : if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
5118 : /*
5119 : * Someone really wants this to run. If it's not unfair, run it.
5120 : */
5121 837 : se = cfs_rq->next;
5122 1675 : } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
5123 : /*
5124 : * Prefer last buddy, try to return the CPU to a preempted task.
5125 : */
5126 0 : se = cfs_rq->last;
5127 : }
5128 :
5129 2512 : return se;
5130 : }
5131 :
5132 : static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5133 :
5134 2515 : static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5135 : {
5136 : /*
5137 : * If still on the runqueue then deactivate_task()
5138 : * was not called and update_curr() has to be done:
5139 : */
5140 2515 : if (prev->on_rq)
5141 71 : update_curr(cfs_rq);
5142 :
5143 : /* throttle cfs_rqs exceeding runtime */
5144 2515 : check_cfs_rq_runtime(cfs_rq);
5145 :
5146 2515 : check_spread(cfs_rq, prev);
5147 :
5148 2515 : if (prev->on_rq) {
5149 71 : update_stats_wait_start_fair(cfs_rq, prev);
5150 : /* Put 'current' back into the tree. */
5151 71 : __enqueue_entity(cfs_rq, prev);
5152 : /* in !on_rq case, update occurred at dequeue */
5153 71 : update_load_avg(cfs_rq, prev, 0);
5154 : }
5155 2515 : cfs_rq->curr = NULL;
5156 2515 : }
5157 :
5158 : static void
5159 2930 : entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5160 : {
5161 : /*
5162 : * Update run-time statistics of the 'current'.
5163 : */
5164 2930 : update_curr(cfs_rq);
5165 :
5166 : /*
5167 : * Ensure that runnable average is periodically updated.
5168 : */
5169 2930 : update_load_avg(cfs_rq, curr, UPDATE_TG);
5170 2930 : update_cfs_group(curr);
5171 :
5172 : #ifdef CONFIG_SCHED_HRTICK
5173 : /*
5174 : * queued ticks are scheduled to match the slice, so don't bother
5175 : * validating it and just reschedule.
5176 : */
5177 : if (queued) {
5178 : resched_curr(rq_of(cfs_rq));
5179 : return;
5180 : }
5181 : /*
5182 : * don't let the period tick interfere with the hrtick preemption
5183 : */
5184 : if (!sched_feat(DOUBLE_TICK) &&
5185 : hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5186 : return;
5187 : #endif
5188 :
5189 2930 : if (cfs_rq->nr_running > 1)
5190 235 : check_preempt_tick(cfs_rq, curr);
5191 2930 : }
5192 :
5193 :
5194 : /**************************************************
5195 : * CFS bandwidth control machinery
5196 : */
5197 :
5198 : #ifdef CONFIG_CFS_BANDWIDTH
5199 :
5200 : #ifdef CONFIG_JUMP_LABEL
5201 : static struct static_key __cfs_bandwidth_used;
5202 :
5203 : static inline bool cfs_bandwidth_used(void)
5204 : {
5205 : return static_key_false(&__cfs_bandwidth_used);
5206 : }
5207 :
5208 : void cfs_bandwidth_usage_inc(void)
5209 : {
5210 : static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5211 : }
5212 :
5213 : void cfs_bandwidth_usage_dec(void)
5214 : {
5215 : static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5216 : }
5217 : #else /* CONFIG_JUMP_LABEL */
5218 : static bool cfs_bandwidth_used(void)
5219 : {
5220 : return true;
5221 : }
5222 :
5223 : void cfs_bandwidth_usage_inc(void) {}
5224 : void cfs_bandwidth_usage_dec(void) {}
5225 : #endif /* CONFIG_JUMP_LABEL */
5226 :
5227 : /*
5228 : * default period for cfs group bandwidth.
5229 : * default: 0.1s, units: nanoseconds
5230 : */
5231 : static inline u64 default_cfs_period(void)
5232 : {
5233 : return 100000000ULL;
5234 : }
5235 :
5236 : static inline u64 sched_cfs_bandwidth_slice(void)
5237 : {
5238 : return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5239 : }
5240 :
5241 : /*
5242 : * Replenish runtime according to assigned quota. We use sched_clock_cpu
5243 : * directly instead of rq->clock to avoid adding additional synchronization
5244 : * around rq->lock.
5245 : *
5246 : * requires cfs_b->lock
5247 : */
5248 : void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5249 : {
5250 : s64 runtime;
5251 :
5252 : if (unlikely(cfs_b->quota == RUNTIME_INF))
5253 : return;
5254 :
5255 : cfs_b->runtime += cfs_b->quota;
5256 : runtime = cfs_b->runtime_snap - cfs_b->runtime;
5257 : if (runtime > 0) {
5258 : cfs_b->burst_time += runtime;
5259 : cfs_b->nr_burst++;
5260 : }
5261 :
5262 : cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5263 : cfs_b->runtime_snap = cfs_b->runtime;
5264 : }
5265 :
5266 : static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5267 : {
5268 : return &tg->cfs_bandwidth;
5269 : }
5270 :
5271 : /* returns 0 on failure to allocate runtime */
5272 : static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5273 : struct cfs_rq *cfs_rq, u64 target_runtime)
5274 : {
5275 : u64 min_amount, amount = 0;
5276 :
5277 : lockdep_assert_held(&cfs_b->lock);
5278 :
5279 : /* note: this is a positive sum as runtime_remaining <= 0 */
5280 : min_amount = target_runtime - cfs_rq->runtime_remaining;
5281 :
5282 : if (cfs_b->quota == RUNTIME_INF)
5283 : amount = min_amount;
5284 : else {
5285 : start_cfs_bandwidth(cfs_b);
5286 :
5287 : if (cfs_b->runtime > 0) {
5288 : amount = min(cfs_b->runtime, min_amount);
5289 : cfs_b->runtime -= amount;
5290 : cfs_b->idle = 0;
5291 : }
5292 : }
5293 :
5294 : cfs_rq->runtime_remaining += amount;
5295 :
5296 : return cfs_rq->runtime_remaining > 0;
5297 : }
5298 :
5299 : /* returns 0 on failure to allocate runtime */
5300 : static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5301 : {
5302 : struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5303 : int ret;
5304 :
5305 : raw_spin_lock(&cfs_b->lock);
5306 : ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5307 : raw_spin_unlock(&cfs_b->lock);
5308 :
5309 : return ret;
5310 : }
5311 :
5312 : static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5313 : {
5314 : /* dock delta_exec before expiring quota (as it could span periods) */
5315 : cfs_rq->runtime_remaining -= delta_exec;
5316 :
5317 : if (likely(cfs_rq->runtime_remaining > 0))
5318 : return;
5319 :
5320 : if (cfs_rq->throttled)
5321 : return;
5322 : /*
5323 : * if we're unable to extend our runtime we resched so that the active
5324 : * hierarchy can be throttled
5325 : */
5326 : if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5327 : resched_curr(rq_of(cfs_rq));
5328 : }
5329 :
5330 : static __always_inline
5331 : void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5332 : {
5333 : if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5334 : return;
5335 :
5336 : __account_cfs_rq_runtime(cfs_rq, delta_exec);
5337 : }
5338 :
5339 : static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5340 : {
5341 : return cfs_bandwidth_used() && cfs_rq->throttled;
5342 : }
5343 :
5344 : /* check whether cfs_rq, or any parent, is throttled */
5345 : static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5346 : {
5347 : return cfs_bandwidth_used() && cfs_rq->throttle_count;
5348 : }
5349 :
5350 : /*
5351 : * Ensure that neither of the group entities corresponding to src_cpu or
5352 : * dest_cpu are members of a throttled hierarchy when performing group
5353 : * load-balance operations.
5354 : */
5355 : static inline int throttled_lb_pair(struct task_group *tg,
5356 : int src_cpu, int dest_cpu)
5357 : {
5358 : struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5359 :
5360 : src_cfs_rq = tg->cfs_rq[src_cpu];
5361 : dest_cfs_rq = tg->cfs_rq[dest_cpu];
5362 :
5363 : return throttled_hierarchy(src_cfs_rq) ||
5364 : throttled_hierarchy(dest_cfs_rq);
5365 : }
5366 :
5367 : static int tg_unthrottle_up(struct task_group *tg, void *data)
5368 : {
5369 : struct rq *rq = data;
5370 : struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5371 :
5372 : cfs_rq->throttle_count--;
5373 : if (!cfs_rq->throttle_count) {
5374 : cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5375 : cfs_rq->throttled_clock_pelt;
5376 :
5377 : /* Add cfs_rq with load or one or more already running entities to the list */
5378 : if (!cfs_rq_is_decayed(cfs_rq))
5379 : list_add_leaf_cfs_rq(cfs_rq);
5380 : }
5381 :
5382 : return 0;
5383 : }
5384 :
5385 : static int tg_throttle_down(struct task_group *tg, void *data)
5386 : {
5387 : struct rq *rq = data;
5388 : struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5389 :
5390 : /* group is entering throttled state, stop time */
5391 : if (!cfs_rq->throttle_count) {
5392 : cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5393 : list_del_leaf_cfs_rq(cfs_rq);
5394 : }
5395 : cfs_rq->throttle_count++;
5396 :
5397 : return 0;
5398 : }
5399 :
5400 : static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5401 : {
5402 : struct rq *rq = rq_of(cfs_rq);
5403 : struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5404 : struct sched_entity *se;
5405 : long task_delta, idle_task_delta, dequeue = 1;
5406 :
5407 : raw_spin_lock(&cfs_b->lock);
5408 : /* This will start the period timer if necessary */
5409 : if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5410 : /*
5411 : * We have raced with bandwidth becoming available, and if we
5412 : * actually throttled the timer might not unthrottle us for an
5413 : * entire period. We additionally needed to make sure that any
5414 : * subsequent check_cfs_rq_runtime calls agree not to throttle
5415 : * us, as we may commit to do cfs put_prev+pick_next, so we ask
5416 : * for 1ns of runtime rather than just check cfs_b.
5417 : */
5418 : dequeue = 0;
5419 : } else {
5420 : list_add_tail_rcu(&cfs_rq->throttled_list,
5421 : &cfs_b->throttled_cfs_rq);
5422 : }
5423 : raw_spin_unlock(&cfs_b->lock);
5424 :
5425 : if (!dequeue)
5426 : return false; /* Throttle no longer required. */
5427 :
5428 : se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5429 :
5430 : /* freeze hierarchy runnable averages while throttled */
5431 : rcu_read_lock();
5432 : walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5433 : rcu_read_unlock();
5434 :
5435 : task_delta = cfs_rq->h_nr_running;
5436 : idle_task_delta = cfs_rq->idle_h_nr_running;
5437 : for_each_sched_entity(se) {
5438 : struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5439 : /* throttled entity or throttle-on-deactivate */
5440 : if (!se->on_rq)
5441 : goto done;
5442 :
5443 : dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5444 :
5445 : if (cfs_rq_is_idle(group_cfs_rq(se)))
5446 : idle_task_delta = cfs_rq->h_nr_running;
5447 :
5448 : qcfs_rq->h_nr_running -= task_delta;
5449 : qcfs_rq->idle_h_nr_running -= idle_task_delta;
5450 :
5451 : if (qcfs_rq->load.weight) {
5452 : /* Avoid re-evaluating load for this entity: */
5453 : se = parent_entity(se);
5454 : break;
5455 : }
5456 : }
5457 :
5458 : for_each_sched_entity(se) {
5459 : struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5460 : /* throttled entity or throttle-on-deactivate */
5461 : if (!se->on_rq)
5462 : goto done;
5463 :
5464 : update_load_avg(qcfs_rq, se, 0);
5465 : se_update_runnable(se);
5466 :
5467 : if (cfs_rq_is_idle(group_cfs_rq(se)))
5468 : idle_task_delta = cfs_rq->h_nr_running;
5469 :
5470 : qcfs_rq->h_nr_running -= task_delta;
5471 : qcfs_rq->idle_h_nr_running -= idle_task_delta;
5472 : }
5473 :
5474 : /* At this point se is NULL and we are at root level*/
5475 : sub_nr_running(rq, task_delta);
5476 :
5477 : done:
5478 : /*
5479 : * Note: distribution will already see us throttled via the
5480 : * throttled-list. rq->lock protects completion.
5481 : */
5482 : cfs_rq->throttled = 1;
5483 : cfs_rq->throttled_clock = rq_clock(rq);
5484 : return true;
5485 : }
5486 :
5487 : void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5488 : {
5489 : struct rq *rq = rq_of(cfs_rq);
5490 : struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5491 : struct sched_entity *se;
5492 : long task_delta, idle_task_delta;
5493 :
5494 : se = cfs_rq->tg->se[cpu_of(rq)];
5495 :
5496 : cfs_rq->throttled = 0;
5497 :
5498 : update_rq_clock(rq);
5499 :
5500 : raw_spin_lock(&cfs_b->lock);
5501 : cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5502 : list_del_rcu(&cfs_rq->throttled_list);
5503 : raw_spin_unlock(&cfs_b->lock);
5504 :
5505 : /* update hierarchical throttle state */
5506 : walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5507 :
5508 : if (!cfs_rq->load.weight) {
5509 : if (!cfs_rq->on_list)
5510 : return;
5511 : /*
5512 : * Nothing to run but something to decay (on_list)?
5513 : * Complete the branch.
5514 : */
5515 : for_each_sched_entity(se) {
5516 : if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5517 : break;
5518 : }
5519 : goto unthrottle_throttle;
5520 : }
5521 :
5522 : task_delta = cfs_rq->h_nr_running;
5523 : idle_task_delta = cfs_rq->idle_h_nr_running;
5524 : for_each_sched_entity(se) {
5525 : struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5526 :
5527 : if (se->on_rq)
5528 : break;
5529 : enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5530 :
5531 : if (cfs_rq_is_idle(group_cfs_rq(se)))
5532 : idle_task_delta = cfs_rq->h_nr_running;
5533 :
5534 : qcfs_rq->h_nr_running += task_delta;
5535 : qcfs_rq->idle_h_nr_running += idle_task_delta;
5536 :
5537 : /* end evaluation on encountering a throttled cfs_rq */
5538 : if (cfs_rq_throttled(qcfs_rq))
5539 : goto unthrottle_throttle;
5540 : }
5541 :
5542 : for_each_sched_entity(se) {
5543 : struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5544 :
5545 : update_load_avg(qcfs_rq, se, UPDATE_TG);
5546 : se_update_runnable(se);
5547 :
5548 : if (cfs_rq_is_idle(group_cfs_rq(se)))
5549 : idle_task_delta = cfs_rq->h_nr_running;
5550 :
5551 : qcfs_rq->h_nr_running += task_delta;
5552 : qcfs_rq->idle_h_nr_running += idle_task_delta;
5553 :
5554 : /* end evaluation on encountering a throttled cfs_rq */
5555 : if (cfs_rq_throttled(qcfs_rq))
5556 : goto unthrottle_throttle;
5557 : }
5558 :
5559 : /* At this point se is NULL and we are at root level*/
5560 : add_nr_running(rq, task_delta);
5561 :
5562 : unthrottle_throttle:
5563 : assert_list_leaf_cfs_rq(rq);
5564 :
5565 : /* Determine whether we need to wake up potentially idle CPU: */
5566 : if (rq->curr == rq->idle && rq->cfs.nr_running)
5567 : resched_curr(rq);
5568 : }
5569 :
5570 : #ifdef CONFIG_SMP
5571 : static void __cfsb_csd_unthrottle(void *arg)
5572 : {
5573 : struct cfs_rq *cursor, *tmp;
5574 : struct rq *rq = arg;
5575 : struct rq_flags rf;
5576 :
5577 : rq_lock(rq, &rf);
5578 :
5579 : /*
5580 : * Since we hold rq lock we're safe from concurrent manipulation of
5581 : * the CSD list. However, this RCU critical section annotates the
5582 : * fact that we pair with sched_free_group_rcu(), so that we cannot
5583 : * race with group being freed in the window between removing it
5584 : * from the list and advancing to the next entry in the list.
5585 : */
5586 : rcu_read_lock();
5587 :
5588 : list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5589 : throttled_csd_list) {
5590 : list_del_init(&cursor->throttled_csd_list);
5591 :
5592 : if (cfs_rq_throttled(cursor))
5593 : unthrottle_cfs_rq(cursor);
5594 : }
5595 :
5596 : rcu_read_unlock();
5597 :
5598 : rq_unlock(rq, &rf);
5599 : }
5600 :
5601 : static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5602 : {
5603 : struct rq *rq = rq_of(cfs_rq);
5604 : bool first;
5605 :
5606 : if (rq == this_rq()) {
5607 : unthrottle_cfs_rq(cfs_rq);
5608 : return;
5609 : }
5610 :
5611 : /* Already enqueued */
5612 : if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5613 : return;
5614 :
5615 : first = list_empty(&rq->cfsb_csd_list);
5616 : list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5617 : if (first)
5618 : smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5619 : }
5620 : #else
5621 : static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5622 : {
5623 : unthrottle_cfs_rq(cfs_rq);
5624 : }
5625 : #endif
5626 :
5627 : static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5628 : {
5629 : lockdep_assert_rq_held(rq_of(cfs_rq));
5630 :
5631 : if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5632 : cfs_rq->runtime_remaining <= 0))
5633 : return;
5634 :
5635 : __unthrottle_cfs_rq_async(cfs_rq);
5636 : }
5637 :
5638 : static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5639 : {
5640 : struct cfs_rq *local_unthrottle = NULL;
5641 : int this_cpu = smp_processor_id();
5642 : u64 runtime, remaining = 1;
5643 : bool throttled = false;
5644 : struct cfs_rq *cfs_rq;
5645 : struct rq_flags rf;
5646 : struct rq *rq;
5647 :
5648 : rcu_read_lock();
5649 : list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5650 : throttled_list) {
5651 : rq = rq_of(cfs_rq);
5652 :
5653 : if (!remaining) {
5654 : throttled = true;
5655 : break;
5656 : }
5657 :
5658 : rq_lock_irqsave(rq, &rf);
5659 : if (!cfs_rq_throttled(cfs_rq))
5660 : goto next;
5661 :
5662 : #ifdef CONFIG_SMP
5663 : /* Already queued for async unthrottle */
5664 : if (!list_empty(&cfs_rq->throttled_csd_list))
5665 : goto next;
5666 : #endif
5667 :
5668 : /* By the above checks, this should never be true */
5669 : SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5670 :
5671 : raw_spin_lock(&cfs_b->lock);
5672 : runtime = -cfs_rq->runtime_remaining + 1;
5673 : if (runtime > cfs_b->runtime)
5674 : runtime = cfs_b->runtime;
5675 : cfs_b->runtime -= runtime;
5676 : remaining = cfs_b->runtime;
5677 : raw_spin_unlock(&cfs_b->lock);
5678 :
5679 : cfs_rq->runtime_remaining += runtime;
5680 :
5681 : /* we check whether we're throttled above */
5682 : if (cfs_rq->runtime_remaining > 0) {
5683 : if (cpu_of(rq) != this_cpu ||
5684 : SCHED_WARN_ON(local_unthrottle))
5685 : unthrottle_cfs_rq_async(cfs_rq);
5686 : else
5687 : local_unthrottle = cfs_rq;
5688 : } else {
5689 : throttled = true;
5690 : }
5691 :
5692 : next:
5693 : rq_unlock_irqrestore(rq, &rf);
5694 : }
5695 : rcu_read_unlock();
5696 :
5697 : if (local_unthrottle) {
5698 : rq = cpu_rq(this_cpu);
5699 : rq_lock_irqsave(rq, &rf);
5700 : if (cfs_rq_throttled(local_unthrottle))
5701 : unthrottle_cfs_rq(local_unthrottle);
5702 : rq_unlock_irqrestore(rq, &rf);
5703 : }
5704 :
5705 : return throttled;
5706 : }
5707 :
5708 : /*
5709 : * Responsible for refilling a task_group's bandwidth and unthrottling its
5710 : * cfs_rqs as appropriate. If there has been no activity within the last
5711 : * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5712 : * used to track this state.
5713 : */
5714 : static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5715 : {
5716 : int throttled;
5717 :
5718 : /* no need to continue the timer with no bandwidth constraint */
5719 : if (cfs_b->quota == RUNTIME_INF)
5720 : goto out_deactivate;
5721 :
5722 : throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5723 : cfs_b->nr_periods += overrun;
5724 :
5725 : /* Refill extra burst quota even if cfs_b->idle */
5726 : __refill_cfs_bandwidth_runtime(cfs_b);
5727 :
5728 : /*
5729 : * idle depends on !throttled (for the case of a large deficit), and if
5730 : * we're going inactive then everything else can be deferred
5731 : */
5732 : if (cfs_b->idle && !throttled)
5733 : goto out_deactivate;
5734 :
5735 : if (!throttled) {
5736 : /* mark as potentially idle for the upcoming period */
5737 : cfs_b->idle = 1;
5738 : return 0;
5739 : }
5740 :
5741 : /* account preceding periods in which throttling occurred */
5742 : cfs_b->nr_throttled += overrun;
5743 :
5744 : /*
5745 : * This check is repeated as we release cfs_b->lock while we unthrottle.
5746 : */
5747 : while (throttled && cfs_b->runtime > 0) {
5748 : raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5749 : /* we can't nest cfs_b->lock while distributing bandwidth */
5750 : throttled = distribute_cfs_runtime(cfs_b);
5751 : raw_spin_lock_irqsave(&cfs_b->lock, flags);
5752 : }
5753 :
5754 : /*
5755 : * While we are ensured activity in the period following an
5756 : * unthrottle, this also covers the case in which the new bandwidth is
5757 : * insufficient to cover the existing bandwidth deficit. (Forcing the
5758 : * timer to remain active while there are any throttled entities.)
5759 : */
5760 : cfs_b->idle = 0;
5761 :
5762 : return 0;
5763 :
5764 : out_deactivate:
5765 : return 1;
5766 : }
5767 :
5768 : /* a cfs_rq won't donate quota below this amount */
5769 : static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5770 : /* minimum remaining period time to redistribute slack quota */
5771 : static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5772 : /* how long we wait to gather additional slack before distributing */
5773 : static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5774 :
5775 : /*
5776 : * Are we near the end of the current quota period?
5777 : *
5778 : * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5779 : * hrtimer base being cleared by hrtimer_start. In the case of
5780 : * migrate_hrtimers, base is never cleared, so we are fine.
5781 : */
5782 : static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5783 : {
5784 : struct hrtimer *refresh_timer = &cfs_b->period_timer;
5785 : s64 remaining;
5786 :
5787 : /* if the call-back is running a quota refresh is already occurring */
5788 : if (hrtimer_callback_running(refresh_timer))
5789 : return 1;
5790 :
5791 : /* is a quota refresh about to occur? */
5792 : remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5793 : if (remaining < (s64)min_expire)
5794 : return 1;
5795 :
5796 : return 0;
5797 : }
5798 :
5799 : static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5800 : {
5801 : u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5802 :
5803 : /* if there's a quota refresh soon don't bother with slack */
5804 : if (runtime_refresh_within(cfs_b, min_left))
5805 : return;
5806 :
5807 : /* don't push forwards an existing deferred unthrottle */
5808 : if (cfs_b->slack_started)
5809 : return;
5810 : cfs_b->slack_started = true;
5811 :
5812 : hrtimer_start(&cfs_b->slack_timer,
5813 : ns_to_ktime(cfs_bandwidth_slack_period),
5814 : HRTIMER_MODE_REL);
5815 : }
5816 :
5817 : /* we know any runtime found here is valid as update_curr() precedes return */
5818 : static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5819 : {
5820 : struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5821 : s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5822 :
5823 : if (slack_runtime <= 0)
5824 : return;
5825 :
5826 : raw_spin_lock(&cfs_b->lock);
5827 : if (cfs_b->quota != RUNTIME_INF) {
5828 : cfs_b->runtime += slack_runtime;
5829 :
5830 : /* we are under rq->lock, defer unthrottling using a timer */
5831 : if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5832 : !list_empty(&cfs_b->throttled_cfs_rq))
5833 : start_cfs_slack_bandwidth(cfs_b);
5834 : }
5835 : raw_spin_unlock(&cfs_b->lock);
5836 :
5837 : /* even if it's not valid for return we don't want to try again */
5838 : cfs_rq->runtime_remaining -= slack_runtime;
5839 : }
5840 :
5841 : static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5842 : {
5843 : if (!cfs_bandwidth_used())
5844 : return;
5845 :
5846 : if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5847 : return;
5848 :
5849 : __return_cfs_rq_runtime(cfs_rq);
5850 : }
5851 :
5852 : /*
5853 : * This is done with a timer (instead of inline with bandwidth return) since
5854 : * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5855 : */
5856 : static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5857 : {
5858 : u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5859 : unsigned long flags;
5860 :
5861 : /* confirm we're still not at a refresh boundary */
5862 : raw_spin_lock_irqsave(&cfs_b->lock, flags);
5863 : cfs_b->slack_started = false;
5864 :
5865 : if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5866 : raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5867 : return;
5868 : }
5869 :
5870 : if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5871 : runtime = cfs_b->runtime;
5872 :
5873 : raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5874 :
5875 : if (!runtime)
5876 : return;
5877 :
5878 : distribute_cfs_runtime(cfs_b);
5879 : }
5880 :
5881 : /*
5882 : * When a group wakes up we want to make sure that its quota is not already
5883 : * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5884 : * runtime as update_curr() throttling can not trigger until it's on-rq.
5885 : */
5886 : static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5887 : {
5888 : if (!cfs_bandwidth_used())
5889 : return;
5890 :
5891 : /* an active group must be handled by the update_curr()->put() path */
5892 : if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5893 : return;
5894 :
5895 : /* ensure the group is not already throttled */
5896 : if (cfs_rq_throttled(cfs_rq))
5897 : return;
5898 :
5899 : /* update runtime allocation */
5900 : account_cfs_rq_runtime(cfs_rq, 0);
5901 : if (cfs_rq->runtime_remaining <= 0)
5902 : throttle_cfs_rq(cfs_rq);
5903 : }
5904 :
5905 : static void sync_throttle(struct task_group *tg, int cpu)
5906 : {
5907 : struct cfs_rq *pcfs_rq, *cfs_rq;
5908 :
5909 : if (!cfs_bandwidth_used())
5910 : return;
5911 :
5912 : if (!tg->parent)
5913 : return;
5914 :
5915 : cfs_rq = tg->cfs_rq[cpu];
5916 : pcfs_rq = tg->parent->cfs_rq[cpu];
5917 :
5918 : cfs_rq->throttle_count = pcfs_rq->throttle_count;
5919 : cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
5920 : }
5921 :
5922 : /* conditionally throttle active cfs_rq's from put_prev_entity() */
5923 : static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5924 : {
5925 : if (!cfs_bandwidth_used())
5926 : return false;
5927 :
5928 : if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5929 : return false;
5930 :
5931 : /*
5932 : * it's possible for a throttled entity to be forced into a running
5933 : * state (e.g. set_curr_task), in this case we're finished.
5934 : */
5935 : if (cfs_rq_throttled(cfs_rq))
5936 : return true;
5937 :
5938 : return throttle_cfs_rq(cfs_rq);
5939 : }
5940 :
5941 : static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5942 : {
5943 : struct cfs_bandwidth *cfs_b =
5944 : container_of(timer, struct cfs_bandwidth, slack_timer);
5945 :
5946 : do_sched_cfs_slack_timer(cfs_b);
5947 :
5948 : return HRTIMER_NORESTART;
5949 : }
5950 :
5951 : extern const u64 max_cfs_quota_period;
5952 :
5953 : static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5954 : {
5955 : struct cfs_bandwidth *cfs_b =
5956 : container_of(timer, struct cfs_bandwidth, period_timer);
5957 : unsigned long flags;
5958 : int overrun;
5959 : int idle = 0;
5960 : int count = 0;
5961 :
5962 : raw_spin_lock_irqsave(&cfs_b->lock, flags);
5963 : for (;;) {
5964 : overrun = hrtimer_forward_now(timer, cfs_b->period);
5965 : if (!overrun)
5966 : break;
5967 :
5968 : idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5969 :
5970 : if (++count > 3) {
5971 : u64 new, old = ktime_to_ns(cfs_b->period);
5972 :
5973 : /*
5974 : * Grow period by a factor of 2 to avoid losing precision.
5975 : * Precision loss in the quota/period ratio can cause __cfs_schedulable
5976 : * to fail.
5977 : */
5978 : new = old * 2;
5979 : if (new < max_cfs_quota_period) {
5980 : cfs_b->period = ns_to_ktime(new);
5981 : cfs_b->quota *= 2;
5982 : cfs_b->burst *= 2;
5983 :
5984 : pr_warn_ratelimited(
5985 : "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5986 : smp_processor_id(),
5987 : div_u64(new, NSEC_PER_USEC),
5988 : div_u64(cfs_b->quota, NSEC_PER_USEC));
5989 : } else {
5990 : pr_warn_ratelimited(
5991 : "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5992 : smp_processor_id(),
5993 : div_u64(old, NSEC_PER_USEC),
5994 : div_u64(cfs_b->quota, NSEC_PER_USEC));
5995 : }
5996 :
5997 : /* reset count so we don't come right back in here */
5998 : count = 0;
5999 : }
6000 : }
6001 : if (idle)
6002 : cfs_b->period_active = 0;
6003 : raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6004 :
6005 : return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6006 : }
6007 :
6008 : void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6009 : {
6010 : raw_spin_lock_init(&cfs_b->lock);
6011 : cfs_b->runtime = 0;
6012 : cfs_b->quota = RUNTIME_INF;
6013 : cfs_b->period = ns_to_ktime(default_cfs_period());
6014 : cfs_b->burst = 0;
6015 :
6016 : INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6017 : hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6018 : cfs_b->period_timer.function = sched_cfs_period_timer;
6019 :
6020 : /* Add a random offset so that timers interleave */
6021 : hrtimer_set_expires(&cfs_b->period_timer,
6022 : get_random_u32_below(cfs_b->period));
6023 : hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6024 : cfs_b->slack_timer.function = sched_cfs_slack_timer;
6025 : cfs_b->slack_started = false;
6026 : }
6027 :
6028 : static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6029 : {
6030 : cfs_rq->runtime_enabled = 0;
6031 : INIT_LIST_HEAD(&cfs_rq->throttled_list);
6032 : #ifdef CONFIG_SMP
6033 : INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6034 : #endif
6035 : }
6036 :
6037 : void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6038 : {
6039 : lockdep_assert_held(&cfs_b->lock);
6040 :
6041 : if (cfs_b->period_active)
6042 : return;
6043 :
6044 : cfs_b->period_active = 1;
6045 : hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6046 : hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6047 : }
6048 :
6049 : static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6050 : {
6051 : int __maybe_unused i;
6052 :
6053 : /* init_cfs_bandwidth() was not called */
6054 : if (!cfs_b->throttled_cfs_rq.next)
6055 : return;
6056 :
6057 : hrtimer_cancel(&cfs_b->period_timer);
6058 : hrtimer_cancel(&cfs_b->slack_timer);
6059 :
6060 : /*
6061 : * It is possible that we still have some cfs_rq's pending on a CSD
6062 : * list, though this race is very rare. In order for this to occur, we
6063 : * must have raced with the last task leaving the group while there
6064 : * exist throttled cfs_rq(s), and the period_timer must have queued the
6065 : * CSD item but the remote cpu has not yet processed it. To handle this,
6066 : * we can simply flush all pending CSD work inline here. We're
6067 : * guaranteed at this point that no additional cfs_rq of this group can
6068 : * join a CSD list.
6069 : */
6070 : #ifdef CONFIG_SMP
6071 : for_each_possible_cpu(i) {
6072 : struct rq *rq = cpu_rq(i);
6073 : unsigned long flags;
6074 :
6075 : if (list_empty(&rq->cfsb_csd_list))
6076 : continue;
6077 :
6078 : local_irq_save(flags);
6079 : __cfsb_csd_unthrottle(rq);
6080 : local_irq_restore(flags);
6081 : }
6082 : #endif
6083 : }
6084 :
6085 : /*
6086 : * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6087 : *
6088 : * The race is harmless, since modifying bandwidth settings of unhooked group
6089 : * bits doesn't do much.
6090 : */
6091 :
6092 : /* cpu online callback */
6093 : static void __maybe_unused update_runtime_enabled(struct rq *rq)
6094 : {
6095 : struct task_group *tg;
6096 :
6097 : lockdep_assert_rq_held(rq);
6098 :
6099 : rcu_read_lock();
6100 : list_for_each_entry_rcu(tg, &task_groups, list) {
6101 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6102 : struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6103 :
6104 : raw_spin_lock(&cfs_b->lock);
6105 : cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6106 : raw_spin_unlock(&cfs_b->lock);
6107 : }
6108 : rcu_read_unlock();
6109 : }
6110 :
6111 : /* cpu offline callback */
6112 : static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6113 : {
6114 : struct task_group *tg;
6115 :
6116 : lockdep_assert_rq_held(rq);
6117 :
6118 : rcu_read_lock();
6119 : list_for_each_entry_rcu(tg, &task_groups, list) {
6120 : struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6121 :
6122 : if (!cfs_rq->runtime_enabled)
6123 : continue;
6124 :
6125 : /*
6126 : * clock_task is not advancing so we just need to make sure
6127 : * there's some valid quota amount
6128 : */
6129 : cfs_rq->runtime_remaining = 1;
6130 : /*
6131 : * Offline rq is schedulable till CPU is completely disabled
6132 : * in take_cpu_down(), so we prevent new cfs throttling here.
6133 : */
6134 : cfs_rq->runtime_enabled = 0;
6135 :
6136 : if (cfs_rq_throttled(cfs_rq))
6137 : unthrottle_cfs_rq(cfs_rq);
6138 : }
6139 : rcu_read_unlock();
6140 : }
6141 :
6142 : #else /* CONFIG_CFS_BANDWIDTH */
6143 :
6144 : static inline bool cfs_bandwidth_used(void)
6145 : {
6146 : return false;
6147 : }
6148 :
6149 : static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6150 : static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6151 : static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6152 : static inline void sync_throttle(struct task_group *tg, int cpu) {}
6153 : static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6154 :
6155 : static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6156 : {
6157 : return 0;
6158 : }
6159 :
6160 : static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6161 : {
6162 : return 0;
6163 : }
6164 :
6165 : static inline int throttled_lb_pair(struct task_group *tg,
6166 : int src_cpu, int dest_cpu)
6167 : {
6168 : return 0;
6169 : }
6170 :
6171 0 : void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6172 :
6173 : #ifdef CONFIG_FAIR_GROUP_SCHED
6174 : static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6175 : #endif
6176 :
6177 : static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6178 : {
6179 : return NULL;
6180 : }
6181 : static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6182 : static inline void update_runtime_enabled(struct rq *rq) {}
6183 : static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6184 :
6185 : #endif /* CONFIG_CFS_BANDWIDTH */
6186 :
6187 : /**************************************************
6188 : * CFS operations on tasks:
6189 : */
6190 :
6191 : #ifdef CONFIG_SCHED_HRTICK
6192 : static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6193 : {
6194 : struct sched_entity *se = &p->se;
6195 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
6196 :
6197 : SCHED_WARN_ON(task_rq(p) != rq);
6198 :
6199 : if (rq->cfs.h_nr_running > 1) {
6200 : u64 slice = sched_slice(cfs_rq, se);
6201 : u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6202 : s64 delta = slice - ran;
6203 :
6204 : if (delta < 0) {
6205 : if (task_current(rq, p))
6206 : resched_curr(rq);
6207 : return;
6208 : }
6209 : hrtick_start(rq, delta);
6210 : }
6211 : }
6212 :
6213 : /*
6214 : * called from enqueue/dequeue and updates the hrtick when the
6215 : * current task is from our class and nr_running is low enough
6216 : * to matter.
6217 : */
6218 : static void hrtick_update(struct rq *rq)
6219 : {
6220 : struct task_struct *curr = rq->curr;
6221 :
6222 : if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6223 : return;
6224 :
6225 : if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
6226 : hrtick_start_fair(rq, curr);
6227 : }
6228 : #else /* !CONFIG_SCHED_HRTICK */
6229 : static inline void
6230 : hrtick_start_fair(struct rq *rq, struct task_struct *p)
6231 : {
6232 : }
6233 :
6234 : static inline void hrtick_update(struct rq *rq)
6235 : {
6236 : }
6237 : #endif
6238 :
6239 : #ifdef CONFIG_SMP
6240 : static inline bool cpu_overutilized(int cpu)
6241 : {
6242 : unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6243 : unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6244 :
6245 : /* Return true only if the utilization doesn't fit CPU's capacity */
6246 : return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6247 : }
6248 :
6249 : static inline void update_overutilized_status(struct rq *rq)
6250 : {
6251 : if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6252 : WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6253 : trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6254 : }
6255 : }
6256 : #else
6257 : static inline void update_overutilized_status(struct rq *rq) { }
6258 : #endif
6259 :
6260 : /* Runqueue only has SCHED_IDLE tasks enqueued */
6261 : static int sched_idle_rq(struct rq *rq)
6262 : {
6263 4888 : return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6264 : rq->nr_running);
6265 : }
6266 :
6267 : /*
6268 : * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
6269 : * of idle_nr_running, which does not consider idle descendants of normal
6270 : * entities.
6271 : */
6272 : static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
6273 : {
6274 : return cfs_rq->nr_running &&
6275 : cfs_rq->nr_running == cfs_rq->idle_nr_running;
6276 : }
6277 :
6278 : #ifdef CONFIG_SMP
6279 : static int sched_idle_cpu(int cpu)
6280 : {
6281 : return sched_idle_rq(cpu_rq(cpu));
6282 : }
6283 : #endif
6284 :
6285 : /*
6286 : * The enqueue_task method is called before nr_running is
6287 : * increased. Here we update the fair scheduling stats and
6288 : * then put the task into the rbtree:
6289 : */
6290 : static void
6291 2446 : enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6292 : {
6293 : struct cfs_rq *cfs_rq;
6294 2446 : struct sched_entity *se = &p->se;
6295 4892 : int idle_h_nr_running = task_has_idle_policy(p);
6296 2446 : int task_new = !(flags & ENQUEUE_WAKEUP);
6297 :
6298 : /*
6299 : * The code below (indirectly) updates schedutil which looks at
6300 : * the cfs_rq utilization to select a frequency.
6301 : * Let's add the task's estimated utilization to the cfs_rq's
6302 : * estimated utilization, before we update schedutil.
6303 : */
6304 2446 : util_est_enqueue(&rq->cfs, p);
6305 :
6306 : /*
6307 : * If in_iowait is set, the code below may not trigger any cpufreq
6308 : * utilization updates, so do it here explicitly with the IOWAIT flag
6309 : * passed.
6310 : */
6311 : if (p->in_iowait)
6312 : cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6313 :
6314 4892 : for_each_sched_entity(se) {
6315 2446 : if (se->on_rq)
6316 : break;
6317 4892 : cfs_rq = cfs_rq_of(se);
6318 2446 : enqueue_entity(cfs_rq, se, flags);
6319 :
6320 2446 : cfs_rq->h_nr_running++;
6321 2446 : cfs_rq->idle_h_nr_running += idle_h_nr_running;
6322 :
6323 : if (cfs_rq_is_idle(cfs_rq))
6324 : idle_h_nr_running = 1;
6325 :
6326 : /* end evaluation on encountering a throttled cfs_rq */
6327 : if (cfs_rq_throttled(cfs_rq))
6328 : goto enqueue_throttle;
6329 :
6330 : flags = ENQUEUE_WAKEUP;
6331 : }
6332 :
6333 2446 : for_each_sched_entity(se) {
6334 0 : cfs_rq = cfs_rq_of(se);
6335 :
6336 0 : update_load_avg(cfs_rq, se, UPDATE_TG);
6337 0 : se_update_runnable(se);
6338 0 : update_cfs_group(se);
6339 :
6340 0 : cfs_rq->h_nr_running++;
6341 0 : cfs_rq->idle_h_nr_running += idle_h_nr_running;
6342 :
6343 : if (cfs_rq_is_idle(cfs_rq))
6344 : idle_h_nr_running = 1;
6345 :
6346 : /* end evaluation on encountering a throttled cfs_rq */
6347 : if (cfs_rq_throttled(cfs_rq))
6348 : goto enqueue_throttle;
6349 : }
6350 :
6351 : /* At this point se is NULL and we are at root level*/
6352 4892 : add_nr_running(rq, 1);
6353 :
6354 : /*
6355 : * Since new tasks are assigned an initial util_avg equal to
6356 : * half of the spare capacity of their CPU, tiny tasks have the
6357 : * ability to cross the overutilized threshold, which will
6358 : * result in the load balancer ruining all the task placement
6359 : * done by EAS. As a way to mitigate that effect, do not account
6360 : * for the first enqueue operation of new tasks during the
6361 : * overutilized flag detection.
6362 : *
6363 : * A better way of solving this problem would be to wait for
6364 : * the PELT signals of tasks to converge before taking them
6365 : * into account, but that is not straightforward to implement,
6366 : * and the following generally works well enough in practice.
6367 : */
6368 : if (!task_new)
6369 : update_overutilized_status(rq);
6370 :
6371 : enqueue_throttle:
6372 2446 : assert_list_leaf_cfs_rq(rq);
6373 :
6374 2446 : hrtick_update(rq);
6375 2446 : }
6376 :
6377 : static void set_next_buddy(struct sched_entity *se);
6378 :
6379 : /*
6380 : * The dequeue_task method is called before nr_running is
6381 : * decreased. We remove the task from the rbtree and
6382 : * update the fair scheduling stats:
6383 : */
6384 2444 : static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6385 : {
6386 : struct cfs_rq *cfs_rq;
6387 2444 : struct sched_entity *se = &p->se;
6388 2444 : int task_sleep = flags & DEQUEUE_SLEEP;
6389 4888 : int idle_h_nr_running = task_has_idle_policy(p);
6390 4888 : bool was_sched_idle = sched_idle_rq(rq);
6391 :
6392 2444 : util_est_dequeue(&rq->cfs, p);
6393 :
6394 1 : for_each_sched_entity(se) {
6395 4888 : cfs_rq = cfs_rq_of(se);
6396 2444 : dequeue_entity(cfs_rq, se, flags);
6397 :
6398 2444 : cfs_rq->h_nr_running--;
6399 2444 : cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6400 :
6401 : if (cfs_rq_is_idle(cfs_rq))
6402 : idle_h_nr_running = 1;
6403 :
6404 : /* end evaluation on encountering a throttled cfs_rq */
6405 : if (cfs_rq_throttled(cfs_rq))
6406 : goto dequeue_throttle;
6407 :
6408 : /* Don't dequeue parent if it has other entities besides us */
6409 2444 : if (cfs_rq->load.weight) {
6410 : /* Avoid re-evaluating load for this entity: */
6411 : se = parent_entity(se);
6412 : /*
6413 : * Bias pick_next to pick a task from this cfs_rq, as
6414 : * p is sleeping when it is within its sched_slice.
6415 : */
6416 : if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6417 : set_next_buddy(se);
6418 : break;
6419 : }
6420 1 : flags |= DEQUEUE_SLEEP;
6421 : }
6422 :
6423 2444 : for_each_sched_entity(se) {
6424 0 : cfs_rq = cfs_rq_of(se);
6425 :
6426 0 : update_load_avg(cfs_rq, se, UPDATE_TG);
6427 0 : se_update_runnable(se);
6428 0 : update_cfs_group(se);
6429 :
6430 0 : cfs_rq->h_nr_running--;
6431 0 : cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6432 :
6433 : if (cfs_rq_is_idle(cfs_rq))
6434 : idle_h_nr_running = 1;
6435 :
6436 : /* end evaluation on encountering a throttled cfs_rq */
6437 : if (cfs_rq_throttled(cfs_rq))
6438 : goto dequeue_throttle;
6439 :
6440 : }
6441 :
6442 : /* At this point se is NULL and we are at root level*/
6443 4888 : sub_nr_running(rq, 1);
6444 :
6445 : /* balance early to pull high priority tasks */
6446 4888 : if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6447 0 : rq->next_balance = jiffies;
6448 :
6449 : dequeue_throttle:
6450 2444 : util_est_update(&rq->cfs, p, task_sleep);
6451 2444 : hrtick_update(rq);
6452 2444 : }
6453 :
6454 : #ifdef CONFIG_SMP
6455 :
6456 : /* Working cpumask for: load_balance, load_balance_newidle. */
6457 : static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6458 : static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6459 :
6460 : #ifdef CONFIG_NO_HZ_COMMON
6461 :
6462 : static struct {
6463 : cpumask_var_t idle_cpus_mask;
6464 : atomic_t nr_cpus;
6465 : int has_blocked; /* Idle CPUS has blocked load */
6466 : int needs_update; /* Newly idle CPUs need their next_balance collated */
6467 : unsigned long next_balance; /* in jiffy units */
6468 : unsigned long next_blocked; /* Next update of blocked load in jiffies */
6469 : } nohz ____cacheline_aligned;
6470 :
6471 : #endif /* CONFIG_NO_HZ_COMMON */
6472 :
6473 : static unsigned long cpu_load(struct rq *rq)
6474 : {
6475 : return cfs_rq_load_avg(&rq->cfs);
6476 : }
6477 :
6478 : /*
6479 : * cpu_load_without - compute CPU load without any contributions from *p
6480 : * @cpu: the CPU which load is requested
6481 : * @p: the task which load should be discounted
6482 : *
6483 : * The load of a CPU is defined by the load of tasks currently enqueued on that
6484 : * CPU as well as tasks which are currently sleeping after an execution on that
6485 : * CPU.
6486 : *
6487 : * This method returns the load of the specified CPU by discounting the load of
6488 : * the specified task, whenever the task is currently contributing to the CPU
6489 : * load.
6490 : */
6491 : static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6492 : {
6493 : struct cfs_rq *cfs_rq;
6494 : unsigned int load;
6495 :
6496 : /* Task has no contribution or is new */
6497 : if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6498 : return cpu_load(rq);
6499 :
6500 : cfs_rq = &rq->cfs;
6501 : load = READ_ONCE(cfs_rq->avg.load_avg);
6502 :
6503 : /* Discount task's util from CPU's util */
6504 : lsub_positive(&load, task_h_load(p));
6505 :
6506 : return load;
6507 : }
6508 :
6509 : static unsigned long cpu_runnable(struct rq *rq)
6510 : {
6511 : return cfs_rq_runnable_avg(&rq->cfs);
6512 : }
6513 :
6514 : static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6515 : {
6516 : struct cfs_rq *cfs_rq;
6517 : unsigned int runnable;
6518 :
6519 : /* Task has no contribution or is new */
6520 : if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6521 : return cpu_runnable(rq);
6522 :
6523 : cfs_rq = &rq->cfs;
6524 : runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6525 :
6526 : /* Discount task's runnable from CPU's runnable */
6527 : lsub_positive(&runnable, p->se.avg.runnable_avg);
6528 :
6529 : return runnable;
6530 : }
6531 :
6532 : static unsigned long capacity_of(int cpu)
6533 : {
6534 : return cpu_rq(cpu)->cpu_capacity;
6535 : }
6536 :
6537 : static void record_wakee(struct task_struct *p)
6538 : {
6539 : /*
6540 : * Only decay a single time; tasks that have less then 1 wakeup per
6541 : * jiffy will not have built up many flips.
6542 : */
6543 : if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6544 : current->wakee_flips >>= 1;
6545 : current->wakee_flip_decay_ts = jiffies;
6546 : }
6547 :
6548 : if (current->last_wakee != p) {
6549 : current->last_wakee = p;
6550 : current->wakee_flips++;
6551 : }
6552 : }
6553 :
6554 : /*
6555 : * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6556 : *
6557 : * A waker of many should wake a different task than the one last awakened
6558 : * at a frequency roughly N times higher than one of its wakees.
6559 : *
6560 : * In order to determine whether we should let the load spread vs consolidating
6561 : * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6562 : * partner, and a factor of lls_size higher frequency in the other.
6563 : *
6564 : * With both conditions met, we can be relatively sure that the relationship is
6565 : * non-monogamous, with partner count exceeding socket size.
6566 : *
6567 : * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6568 : * whatever is irrelevant, spread criteria is apparent partner count exceeds
6569 : * socket size.
6570 : */
6571 : static int wake_wide(struct task_struct *p)
6572 : {
6573 : unsigned int master = current->wakee_flips;
6574 : unsigned int slave = p->wakee_flips;
6575 : int factor = __this_cpu_read(sd_llc_size);
6576 :
6577 : if (master < slave)
6578 : swap(master, slave);
6579 : if (slave < factor || master < slave * factor)
6580 : return 0;
6581 : return 1;
6582 : }
6583 :
6584 : /*
6585 : * The purpose of wake_affine() is to quickly determine on which CPU we can run
6586 : * soonest. For the purpose of speed we only consider the waking and previous
6587 : * CPU.
6588 : *
6589 : * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6590 : * cache-affine and is (or will be) idle.
6591 : *
6592 : * wake_affine_weight() - considers the weight to reflect the average
6593 : * scheduling latency of the CPUs. This seems to work
6594 : * for the overloaded case.
6595 : */
6596 : static int
6597 : wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6598 : {
6599 : /*
6600 : * If this_cpu is idle, it implies the wakeup is from interrupt
6601 : * context. Only allow the move if cache is shared. Otherwise an
6602 : * interrupt intensive workload could force all tasks onto one
6603 : * node depending on the IO topology or IRQ affinity settings.
6604 : *
6605 : * If the prev_cpu is idle and cache affine then avoid a migration.
6606 : * There is no guarantee that the cache hot data from an interrupt
6607 : * is more important than cache hot data on the prev_cpu and from
6608 : * a cpufreq perspective, it's better to have higher utilisation
6609 : * on one CPU.
6610 : */
6611 : if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6612 : return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6613 :
6614 : if (sync && cpu_rq(this_cpu)->nr_running == 1)
6615 : return this_cpu;
6616 :
6617 : if (available_idle_cpu(prev_cpu))
6618 : return prev_cpu;
6619 :
6620 : return nr_cpumask_bits;
6621 : }
6622 :
6623 : static int
6624 : wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6625 : int this_cpu, int prev_cpu, int sync)
6626 : {
6627 : s64 this_eff_load, prev_eff_load;
6628 : unsigned long task_load;
6629 :
6630 : this_eff_load = cpu_load(cpu_rq(this_cpu));
6631 :
6632 : if (sync) {
6633 : unsigned long current_load = task_h_load(current);
6634 :
6635 : if (current_load > this_eff_load)
6636 : return this_cpu;
6637 :
6638 : this_eff_load -= current_load;
6639 : }
6640 :
6641 : task_load = task_h_load(p);
6642 :
6643 : this_eff_load += task_load;
6644 : if (sched_feat(WA_BIAS))
6645 : this_eff_load *= 100;
6646 : this_eff_load *= capacity_of(prev_cpu);
6647 :
6648 : prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6649 : prev_eff_load -= task_load;
6650 : if (sched_feat(WA_BIAS))
6651 : prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6652 : prev_eff_load *= capacity_of(this_cpu);
6653 :
6654 : /*
6655 : * If sync, adjust the weight of prev_eff_load such that if
6656 : * prev_eff == this_eff that select_idle_sibling() will consider
6657 : * stacking the wakee on top of the waker if no other CPU is
6658 : * idle.
6659 : */
6660 : if (sync)
6661 : prev_eff_load += 1;
6662 :
6663 : return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6664 : }
6665 :
6666 : static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6667 : int this_cpu, int prev_cpu, int sync)
6668 : {
6669 : int target = nr_cpumask_bits;
6670 :
6671 : if (sched_feat(WA_IDLE))
6672 : target = wake_affine_idle(this_cpu, prev_cpu, sync);
6673 :
6674 : if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6675 : target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6676 :
6677 : schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6678 : if (target != this_cpu)
6679 : return prev_cpu;
6680 :
6681 : schedstat_inc(sd->ttwu_move_affine);
6682 : schedstat_inc(p->stats.nr_wakeups_affine);
6683 : return target;
6684 : }
6685 :
6686 : static struct sched_group *
6687 : find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6688 :
6689 : /*
6690 : * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6691 : */
6692 : static int
6693 : find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6694 : {
6695 : unsigned long load, min_load = ULONG_MAX;
6696 : unsigned int min_exit_latency = UINT_MAX;
6697 : u64 latest_idle_timestamp = 0;
6698 : int least_loaded_cpu = this_cpu;
6699 : int shallowest_idle_cpu = -1;
6700 : int i;
6701 :
6702 : /* Check if we have any choice: */
6703 : if (group->group_weight == 1)
6704 : return cpumask_first(sched_group_span(group));
6705 :
6706 : /* Traverse only the allowed CPUs */
6707 : for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6708 : struct rq *rq = cpu_rq(i);
6709 :
6710 : if (!sched_core_cookie_match(rq, p))
6711 : continue;
6712 :
6713 : if (sched_idle_cpu(i))
6714 : return i;
6715 :
6716 : if (available_idle_cpu(i)) {
6717 : struct cpuidle_state *idle = idle_get_state(rq);
6718 : if (idle && idle->exit_latency < min_exit_latency) {
6719 : /*
6720 : * We give priority to a CPU whose idle state
6721 : * has the smallest exit latency irrespective
6722 : * of any idle timestamp.
6723 : */
6724 : min_exit_latency = idle->exit_latency;
6725 : latest_idle_timestamp = rq->idle_stamp;
6726 : shallowest_idle_cpu = i;
6727 : } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6728 : rq->idle_stamp > latest_idle_timestamp) {
6729 : /*
6730 : * If equal or no active idle state, then
6731 : * the most recently idled CPU might have
6732 : * a warmer cache.
6733 : */
6734 : latest_idle_timestamp = rq->idle_stamp;
6735 : shallowest_idle_cpu = i;
6736 : }
6737 : } else if (shallowest_idle_cpu == -1) {
6738 : load = cpu_load(cpu_rq(i));
6739 : if (load < min_load) {
6740 : min_load = load;
6741 : least_loaded_cpu = i;
6742 : }
6743 : }
6744 : }
6745 :
6746 : return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6747 : }
6748 :
6749 : static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6750 : int cpu, int prev_cpu, int sd_flag)
6751 : {
6752 : int new_cpu = cpu;
6753 :
6754 : if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6755 : return prev_cpu;
6756 :
6757 : /*
6758 : * We need task's util for cpu_util_without, sync it up to
6759 : * prev_cpu's last_update_time.
6760 : */
6761 : if (!(sd_flag & SD_BALANCE_FORK))
6762 : sync_entity_load_avg(&p->se);
6763 :
6764 : while (sd) {
6765 : struct sched_group *group;
6766 : struct sched_domain *tmp;
6767 : int weight;
6768 :
6769 : if (!(sd->flags & sd_flag)) {
6770 : sd = sd->child;
6771 : continue;
6772 : }
6773 :
6774 : group = find_idlest_group(sd, p, cpu);
6775 : if (!group) {
6776 : sd = sd->child;
6777 : continue;
6778 : }
6779 :
6780 : new_cpu = find_idlest_group_cpu(group, p, cpu);
6781 : if (new_cpu == cpu) {
6782 : /* Now try balancing at a lower domain level of 'cpu': */
6783 : sd = sd->child;
6784 : continue;
6785 : }
6786 :
6787 : /* Now try balancing at a lower domain level of 'new_cpu': */
6788 : cpu = new_cpu;
6789 : weight = sd->span_weight;
6790 : sd = NULL;
6791 : for_each_domain(cpu, tmp) {
6792 : if (weight <= tmp->span_weight)
6793 : break;
6794 : if (tmp->flags & sd_flag)
6795 : sd = tmp;
6796 : }
6797 : }
6798 :
6799 : return new_cpu;
6800 : }
6801 :
6802 : static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6803 : {
6804 : if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6805 : sched_cpu_cookie_match(cpu_rq(cpu), p))
6806 : return cpu;
6807 :
6808 : return -1;
6809 : }
6810 :
6811 : #ifdef CONFIG_SCHED_SMT
6812 : DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6813 : EXPORT_SYMBOL_GPL(sched_smt_present);
6814 :
6815 : static inline void set_idle_cores(int cpu, int val)
6816 : {
6817 : struct sched_domain_shared *sds;
6818 :
6819 : sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6820 : if (sds)
6821 : WRITE_ONCE(sds->has_idle_cores, val);
6822 : }
6823 :
6824 : static inline bool test_idle_cores(int cpu)
6825 : {
6826 : struct sched_domain_shared *sds;
6827 :
6828 : sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6829 : if (sds)
6830 : return READ_ONCE(sds->has_idle_cores);
6831 :
6832 : return false;
6833 : }
6834 :
6835 : /*
6836 : * Scans the local SMT mask to see if the entire core is idle, and records this
6837 : * information in sd_llc_shared->has_idle_cores.
6838 : *
6839 : * Since SMT siblings share all cache levels, inspecting this limited remote
6840 : * state should be fairly cheap.
6841 : */
6842 : void __update_idle_core(struct rq *rq)
6843 : {
6844 : int core = cpu_of(rq);
6845 : int cpu;
6846 :
6847 : rcu_read_lock();
6848 : if (test_idle_cores(core))
6849 : goto unlock;
6850 :
6851 : for_each_cpu(cpu, cpu_smt_mask(core)) {
6852 : if (cpu == core)
6853 : continue;
6854 :
6855 : if (!available_idle_cpu(cpu))
6856 : goto unlock;
6857 : }
6858 :
6859 : set_idle_cores(core, 1);
6860 : unlock:
6861 : rcu_read_unlock();
6862 : }
6863 :
6864 : /*
6865 : * Scan the entire LLC domain for idle cores; this dynamically switches off if
6866 : * there are no idle cores left in the system; tracked through
6867 : * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6868 : */
6869 : static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6870 : {
6871 : bool idle = true;
6872 : int cpu;
6873 :
6874 : for_each_cpu(cpu, cpu_smt_mask(core)) {
6875 : if (!available_idle_cpu(cpu)) {
6876 : idle = false;
6877 : if (*idle_cpu == -1) {
6878 : if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6879 : *idle_cpu = cpu;
6880 : break;
6881 : }
6882 : continue;
6883 : }
6884 : break;
6885 : }
6886 : if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6887 : *idle_cpu = cpu;
6888 : }
6889 :
6890 : if (idle)
6891 : return core;
6892 :
6893 : cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6894 : return -1;
6895 : }
6896 :
6897 : /*
6898 : * Scan the local SMT mask for idle CPUs.
6899 : */
6900 : static int select_idle_smt(struct task_struct *p, int target)
6901 : {
6902 : int cpu;
6903 :
6904 : for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
6905 : if (cpu == target)
6906 : continue;
6907 : if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6908 : return cpu;
6909 : }
6910 :
6911 : return -1;
6912 : }
6913 :
6914 : #else /* CONFIG_SCHED_SMT */
6915 :
6916 : static inline void set_idle_cores(int cpu, int val)
6917 : {
6918 : }
6919 :
6920 : static inline bool test_idle_cores(int cpu)
6921 : {
6922 : return false;
6923 : }
6924 :
6925 : static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6926 : {
6927 : return __select_idle_cpu(core, p);
6928 : }
6929 :
6930 : static inline int select_idle_smt(struct task_struct *p, int target)
6931 : {
6932 : return -1;
6933 : }
6934 :
6935 : #endif /* CONFIG_SCHED_SMT */
6936 :
6937 : /*
6938 : * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6939 : * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6940 : * average idle time for this rq (as found in rq->avg_idle).
6941 : */
6942 : static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6943 : {
6944 : struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6945 : int i, cpu, idle_cpu = -1, nr = INT_MAX;
6946 : struct sched_domain_shared *sd_share;
6947 : struct rq *this_rq = this_rq();
6948 : int this = smp_processor_id();
6949 : struct sched_domain *this_sd = NULL;
6950 : u64 time = 0;
6951 :
6952 : cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6953 :
6954 : if (sched_feat(SIS_PROP) && !has_idle_core) {
6955 : u64 avg_cost, avg_idle, span_avg;
6956 : unsigned long now = jiffies;
6957 :
6958 : this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6959 : if (!this_sd)
6960 : return -1;
6961 :
6962 : /*
6963 : * If we're busy, the assumption that the last idle period
6964 : * predicts the future is flawed; age away the remaining
6965 : * predicted idle time.
6966 : */
6967 : if (unlikely(this_rq->wake_stamp < now)) {
6968 : while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6969 : this_rq->wake_stamp++;
6970 : this_rq->wake_avg_idle >>= 1;
6971 : }
6972 : }
6973 :
6974 : avg_idle = this_rq->wake_avg_idle;
6975 : avg_cost = this_sd->avg_scan_cost + 1;
6976 :
6977 : span_avg = sd->span_weight * avg_idle;
6978 : if (span_avg > 4*avg_cost)
6979 : nr = div_u64(span_avg, avg_cost);
6980 : else
6981 : nr = 4;
6982 :
6983 : time = cpu_clock(this);
6984 : }
6985 :
6986 : if (sched_feat(SIS_UTIL)) {
6987 : sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
6988 : if (sd_share) {
6989 : /* because !--nr is the condition to stop scan */
6990 : nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
6991 : /* overloaded LLC is unlikely to have idle cpu/core */
6992 : if (nr == 1)
6993 : return -1;
6994 : }
6995 : }
6996 :
6997 : for_each_cpu_wrap(cpu, cpus, target + 1) {
6998 : if (has_idle_core) {
6999 : i = select_idle_core(p, cpu, cpus, &idle_cpu);
7000 : if ((unsigned int)i < nr_cpumask_bits)
7001 : return i;
7002 :
7003 : } else {
7004 : if (!--nr)
7005 : return -1;
7006 : idle_cpu = __select_idle_cpu(cpu, p);
7007 : if ((unsigned int)idle_cpu < nr_cpumask_bits)
7008 : break;
7009 : }
7010 : }
7011 :
7012 : if (has_idle_core)
7013 : set_idle_cores(target, false);
7014 :
7015 : if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
7016 : time = cpu_clock(this) - time;
7017 :
7018 : /*
7019 : * Account for the scan cost of wakeups against the average
7020 : * idle time.
7021 : */
7022 : this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
7023 :
7024 : update_avg(&this_sd->avg_scan_cost, time);
7025 : }
7026 :
7027 : return idle_cpu;
7028 : }
7029 :
7030 : /*
7031 : * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7032 : * the task fits. If no CPU is big enough, but there are idle ones, try to
7033 : * maximize capacity.
7034 : */
7035 : static int
7036 : select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7037 : {
7038 : unsigned long task_util, util_min, util_max, best_cap = 0;
7039 : int fits, best_fits = 0;
7040 : int cpu, best_cpu = -1;
7041 : struct cpumask *cpus;
7042 :
7043 : cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7044 : cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7045 :
7046 : task_util = task_util_est(p);
7047 : util_min = uclamp_eff_value(p, UCLAMP_MIN);
7048 : util_max = uclamp_eff_value(p, UCLAMP_MAX);
7049 :
7050 : for_each_cpu_wrap(cpu, cpus, target + 1) {
7051 : unsigned long cpu_cap = capacity_of(cpu);
7052 :
7053 : if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7054 : continue;
7055 :
7056 : fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7057 :
7058 : /* This CPU fits with all requirements */
7059 : if (fits > 0)
7060 : return cpu;
7061 : /*
7062 : * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7063 : * Look for the CPU with best capacity.
7064 : */
7065 : else if (fits < 0)
7066 : cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
7067 :
7068 : /*
7069 : * First, select CPU which fits better (-1 being better than 0).
7070 : * Then, select the one with best capacity at same level.
7071 : */
7072 : if ((fits < best_fits) ||
7073 : ((fits == best_fits) && (cpu_cap > best_cap))) {
7074 : best_cap = cpu_cap;
7075 : best_cpu = cpu;
7076 : best_fits = fits;
7077 : }
7078 : }
7079 :
7080 : return best_cpu;
7081 : }
7082 :
7083 : static inline bool asym_fits_cpu(unsigned long util,
7084 : unsigned long util_min,
7085 : unsigned long util_max,
7086 : int cpu)
7087 : {
7088 : if (sched_asym_cpucap_active())
7089 : /*
7090 : * Return true only if the cpu fully fits the task requirements
7091 : * which include the utilization and the performance hints.
7092 : */
7093 : return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7094 :
7095 : return true;
7096 : }
7097 :
7098 : /*
7099 : * Try and locate an idle core/thread in the LLC cache domain.
7100 : */
7101 : static int select_idle_sibling(struct task_struct *p, int prev, int target)
7102 : {
7103 : bool has_idle_core = false;
7104 : struct sched_domain *sd;
7105 : unsigned long task_util, util_min, util_max;
7106 : int i, recent_used_cpu;
7107 :
7108 : /*
7109 : * On asymmetric system, update task utilization because we will check
7110 : * that the task fits with cpu's capacity.
7111 : */
7112 : if (sched_asym_cpucap_active()) {
7113 : sync_entity_load_avg(&p->se);
7114 : task_util = task_util_est(p);
7115 : util_min = uclamp_eff_value(p, UCLAMP_MIN);
7116 : util_max = uclamp_eff_value(p, UCLAMP_MAX);
7117 : }
7118 :
7119 : /*
7120 : * per-cpu select_rq_mask usage
7121 : */
7122 : lockdep_assert_irqs_disabled();
7123 :
7124 : if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7125 : asym_fits_cpu(task_util, util_min, util_max, target))
7126 : return target;
7127 :
7128 : /*
7129 : * If the previous CPU is cache affine and idle, don't be stupid:
7130 : */
7131 : if (prev != target && cpus_share_cache(prev, target) &&
7132 : (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7133 : asym_fits_cpu(task_util, util_min, util_max, prev))
7134 : return prev;
7135 :
7136 : /*
7137 : * Allow a per-cpu kthread to stack with the wakee if the
7138 : * kworker thread and the tasks previous CPUs are the same.
7139 : * The assumption is that the wakee queued work for the
7140 : * per-cpu kthread that is now complete and the wakeup is
7141 : * essentially a sync wakeup. An obvious example of this
7142 : * pattern is IO completions.
7143 : */
7144 : if (is_per_cpu_kthread(current) &&
7145 : in_task() &&
7146 : prev == smp_processor_id() &&
7147 : this_rq()->nr_running <= 1 &&
7148 : asym_fits_cpu(task_util, util_min, util_max, prev)) {
7149 : return prev;
7150 : }
7151 :
7152 : /* Check a recently used CPU as a potential idle candidate: */
7153 : recent_used_cpu = p->recent_used_cpu;
7154 : p->recent_used_cpu = prev;
7155 : if (recent_used_cpu != prev &&
7156 : recent_used_cpu != target &&
7157 : cpus_share_cache(recent_used_cpu, target) &&
7158 : (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7159 : cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
7160 : asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7161 : return recent_used_cpu;
7162 : }
7163 :
7164 : /*
7165 : * For asymmetric CPU capacity systems, our domain of interest is
7166 : * sd_asym_cpucapacity rather than sd_llc.
7167 : */
7168 : if (sched_asym_cpucap_active()) {
7169 : sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7170 : /*
7171 : * On an asymmetric CPU capacity system where an exclusive
7172 : * cpuset defines a symmetric island (i.e. one unique
7173 : * capacity_orig value through the cpuset), the key will be set
7174 : * but the CPUs within that cpuset will not have a domain with
7175 : * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7176 : * capacity path.
7177 : */
7178 : if (sd) {
7179 : i = select_idle_capacity(p, sd, target);
7180 : return ((unsigned)i < nr_cpumask_bits) ? i : target;
7181 : }
7182 : }
7183 :
7184 : sd = rcu_dereference(per_cpu(sd_llc, target));
7185 : if (!sd)
7186 : return target;
7187 :
7188 : if (sched_smt_active()) {
7189 : has_idle_core = test_idle_cores(target);
7190 :
7191 : if (!has_idle_core && cpus_share_cache(prev, target)) {
7192 : i = select_idle_smt(p, prev);
7193 : if ((unsigned int)i < nr_cpumask_bits)
7194 : return i;
7195 : }
7196 : }
7197 :
7198 : i = select_idle_cpu(p, sd, has_idle_core, target);
7199 : if ((unsigned)i < nr_cpumask_bits)
7200 : return i;
7201 :
7202 : return target;
7203 : }
7204 :
7205 : /*
7206 : * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu
7207 : * (@dst_cpu = -1) or migrated to @dst_cpu.
7208 : */
7209 : static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
7210 : {
7211 : struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7212 : unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7213 :
7214 : /*
7215 : * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7216 : * contribution. If @p migrates from another CPU to @cpu add its
7217 : * contribution. In all the other cases @cpu is not impacted by the
7218 : * migration so its util_avg is already correct.
7219 : */
7220 : if (task_cpu(p) == cpu && dst_cpu != cpu)
7221 : lsub_positive(&util, task_util(p));
7222 : else if (task_cpu(p) != cpu && dst_cpu == cpu)
7223 : util += task_util(p);
7224 :
7225 : if (sched_feat(UTIL_EST)) {
7226 : unsigned long util_est;
7227 :
7228 : util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7229 :
7230 : /*
7231 : * During wake-up @p isn't enqueued yet and doesn't contribute
7232 : * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7233 : * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7234 : * has been enqueued.
7235 : *
7236 : * During exec (@dst_cpu = -1) @p is enqueued and does
7237 : * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7238 : * Remove it to "simulate" cpu_util without @p's contribution.
7239 : *
7240 : * Despite the task_on_rq_queued(@p) check there is still a
7241 : * small window for a possible race when an exec
7242 : * select_task_rq_fair() races with LB's detach_task().
7243 : *
7244 : * detach_task()
7245 : * deactivate_task()
7246 : * p->on_rq = TASK_ON_RQ_MIGRATING;
7247 : * -------------------------------- A
7248 : * dequeue_task() \
7249 : * dequeue_task_fair() + Race Time
7250 : * util_est_dequeue() /
7251 : * -------------------------------- B
7252 : *
7253 : * The additional check "current == p" is required to further
7254 : * reduce the race window.
7255 : */
7256 : if (dst_cpu == cpu)
7257 : util_est += _task_util_est(p);
7258 : else if (unlikely(task_on_rq_queued(p) || current == p))
7259 : lsub_positive(&util_est, _task_util_est(p));
7260 :
7261 : util = max(util, util_est);
7262 : }
7263 :
7264 : return min(util, capacity_orig_of(cpu));
7265 : }
7266 :
7267 : /*
7268 : * cpu_util_without: compute cpu utilization without any contributions from *p
7269 : * @cpu: the CPU which utilization is requested
7270 : * @p: the task which utilization should be discounted
7271 : *
7272 : * The utilization of a CPU is defined by the utilization of tasks currently
7273 : * enqueued on that CPU as well as tasks which are currently sleeping after an
7274 : * execution on that CPU.
7275 : *
7276 : * This method returns the utilization of the specified CPU by discounting the
7277 : * utilization of the specified task, whenever the task is currently
7278 : * contributing to the CPU utilization.
7279 : */
7280 : static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7281 : {
7282 : /* Task has no contribution or is new */
7283 : if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7284 : return cpu_util_cfs(cpu);
7285 :
7286 : return cpu_util_next(cpu, p, -1);
7287 : }
7288 :
7289 : /*
7290 : * energy_env - Utilization landscape for energy estimation.
7291 : * @task_busy_time: Utilization contribution by the task for which we test the
7292 : * placement. Given by eenv_task_busy_time().
7293 : * @pd_busy_time: Utilization of the whole perf domain without the task
7294 : * contribution. Given by eenv_pd_busy_time().
7295 : * @cpu_cap: Maximum CPU capacity for the perf domain.
7296 : * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7297 : */
7298 : struct energy_env {
7299 : unsigned long task_busy_time;
7300 : unsigned long pd_busy_time;
7301 : unsigned long cpu_cap;
7302 : unsigned long pd_cap;
7303 : };
7304 :
7305 : /*
7306 : * Compute the task busy time for compute_energy(). This time cannot be
7307 : * injected directly into effective_cpu_util() because of the IRQ scaling.
7308 : * The latter only makes sense with the most recent CPUs where the task has
7309 : * run.
7310 : */
7311 : static inline void eenv_task_busy_time(struct energy_env *eenv,
7312 : struct task_struct *p, int prev_cpu)
7313 : {
7314 : unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7315 : unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7316 :
7317 : if (unlikely(irq >= max_cap))
7318 : busy_time = max_cap;
7319 : else
7320 : busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7321 :
7322 : eenv->task_busy_time = busy_time;
7323 : }
7324 :
7325 : /*
7326 : * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7327 : * utilization for each @pd_cpus, it however doesn't take into account
7328 : * clamping since the ratio (utilization / cpu_capacity) is already enough to
7329 : * scale the EM reported power consumption at the (eventually clamped)
7330 : * cpu_capacity.
7331 : *
7332 : * The contribution of the task @p for which we want to estimate the
7333 : * energy cost is removed (by cpu_util_next()) and must be calculated
7334 : * separately (see eenv_task_busy_time). This ensures:
7335 : *
7336 : * - A stable PD utilization, no matter which CPU of that PD we want to place
7337 : * the task on.
7338 : *
7339 : * - A fair comparison between CPUs as the task contribution (task_util())
7340 : * will always be the same no matter which CPU utilization we rely on
7341 : * (util_avg or util_est).
7342 : *
7343 : * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7344 : * exceed @eenv->pd_cap.
7345 : */
7346 : static inline void eenv_pd_busy_time(struct energy_env *eenv,
7347 : struct cpumask *pd_cpus,
7348 : struct task_struct *p)
7349 : {
7350 : unsigned long busy_time = 0;
7351 : int cpu;
7352 :
7353 : for_each_cpu(cpu, pd_cpus) {
7354 : unsigned long util = cpu_util_next(cpu, p, -1);
7355 :
7356 : busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7357 : }
7358 :
7359 : eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7360 : }
7361 :
7362 : /*
7363 : * Compute the maximum utilization for compute_energy() when the task @p
7364 : * is placed on the cpu @dst_cpu.
7365 : *
7366 : * Returns the maximum utilization among @eenv->cpus. This utilization can't
7367 : * exceed @eenv->cpu_cap.
7368 : */
7369 : static inline unsigned long
7370 : eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7371 : struct task_struct *p, int dst_cpu)
7372 : {
7373 : unsigned long max_util = 0;
7374 : int cpu;
7375 :
7376 : for_each_cpu(cpu, pd_cpus) {
7377 : struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7378 : unsigned long util = cpu_util_next(cpu, p, dst_cpu);
7379 : unsigned long cpu_util;
7380 :
7381 : /*
7382 : * Performance domain frequency: utilization clamping
7383 : * must be considered since it affects the selection
7384 : * of the performance domain frequency.
7385 : * NOTE: in case RT tasks are running, by default the
7386 : * FREQUENCY_UTIL's utilization can be max OPP.
7387 : */
7388 : cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7389 : max_util = max(max_util, cpu_util);
7390 : }
7391 :
7392 : return min(max_util, eenv->cpu_cap);
7393 : }
7394 :
7395 : /*
7396 : * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7397 : * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7398 : * contribution is ignored.
7399 : */
7400 : static inline unsigned long
7401 : compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7402 : struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7403 : {
7404 : unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7405 : unsigned long busy_time = eenv->pd_busy_time;
7406 :
7407 : if (dst_cpu >= 0)
7408 : busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7409 :
7410 : return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7411 : }
7412 :
7413 : /*
7414 : * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7415 : * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7416 : * spare capacity in each performance domain and uses it as a potential
7417 : * candidate to execute the task. Then, it uses the Energy Model to figure
7418 : * out which of the CPU candidates is the most energy-efficient.
7419 : *
7420 : * The rationale for this heuristic is as follows. In a performance domain,
7421 : * all the most energy efficient CPU candidates (according to the Energy
7422 : * Model) are those for which we'll request a low frequency. When there are
7423 : * several CPUs for which the frequency request will be the same, we don't
7424 : * have enough data to break the tie between them, because the Energy Model
7425 : * only includes active power costs. With this model, if we assume that
7426 : * frequency requests follow utilization (e.g. using schedutil), the CPU with
7427 : * the maximum spare capacity in a performance domain is guaranteed to be among
7428 : * the best candidates of the performance domain.
7429 : *
7430 : * In practice, it could be preferable from an energy standpoint to pack
7431 : * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7432 : * but that could also hurt our chances to go cluster idle, and we have no
7433 : * ways to tell with the current Energy Model if this is actually a good
7434 : * idea or not. So, find_energy_efficient_cpu() basically favors
7435 : * cluster-packing, and spreading inside a cluster. That should at least be
7436 : * a good thing for latency, and this is consistent with the idea that most
7437 : * of the energy savings of EAS come from the asymmetry of the system, and
7438 : * not so much from breaking the tie between identical CPUs. That's also the
7439 : * reason why EAS is enabled in the topology code only for systems where
7440 : * SD_ASYM_CPUCAPACITY is set.
7441 : *
7442 : * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7443 : * they don't have any useful utilization data yet and it's not possible to
7444 : * forecast their impact on energy consumption. Consequently, they will be
7445 : * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7446 : * to be energy-inefficient in some use-cases. The alternative would be to
7447 : * bias new tasks towards specific types of CPUs first, or to try to infer
7448 : * their util_avg from the parent task, but those heuristics could hurt
7449 : * other use-cases too. So, until someone finds a better way to solve this,
7450 : * let's keep things simple by re-using the existing slow path.
7451 : */
7452 : static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7453 : {
7454 : struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7455 : unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7456 : unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7457 : unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7458 : struct root_domain *rd = this_rq()->rd;
7459 : int cpu, best_energy_cpu, target = -1;
7460 : int prev_fits = -1, best_fits = -1;
7461 : unsigned long best_thermal_cap = 0;
7462 : unsigned long prev_thermal_cap = 0;
7463 : struct sched_domain *sd;
7464 : struct perf_domain *pd;
7465 : struct energy_env eenv;
7466 :
7467 : rcu_read_lock();
7468 : pd = rcu_dereference(rd->pd);
7469 : if (!pd || READ_ONCE(rd->overutilized))
7470 : goto unlock;
7471 :
7472 : /*
7473 : * Energy-aware wake-up happens on the lowest sched_domain starting
7474 : * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7475 : */
7476 : sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7477 : while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7478 : sd = sd->parent;
7479 : if (!sd)
7480 : goto unlock;
7481 :
7482 : target = prev_cpu;
7483 :
7484 : sync_entity_load_avg(&p->se);
7485 : if (!uclamp_task_util(p, p_util_min, p_util_max))
7486 : goto unlock;
7487 :
7488 : eenv_task_busy_time(&eenv, p, prev_cpu);
7489 :
7490 : for (; pd; pd = pd->next) {
7491 : unsigned long util_min = p_util_min, util_max = p_util_max;
7492 : unsigned long cpu_cap, cpu_thermal_cap, util;
7493 : unsigned long cur_delta, max_spare_cap = 0;
7494 : unsigned long rq_util_min, rq_util_max;
7495 : unsigned long prev_spare_cap = 0;
7496 : int max_spare_cap_cpu = -1;
7497 : unsigned long base_energy;
7498 : int fits, max_fits = -1;
7499 :
7500 : cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7501 :
7502 : if (cpumask_empty(cpus))
7503 : continue;
7504 :
7505 : /* Account thermal pressure for the energy estimation */
7506 : cpu = cpumask_first(cpus);
7507 : cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7508 : cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7509 :
7510 : eenv.cpu_cap = cpu_thermal_cap;
7511 : eenv.pd_cap = 0;
7512 :
7513 : for_each_cpu(cpu, cpus) {
7514 : struct rq *rq = cpu_rq(cpu);
7515 :
7516 : eenv.pd_cap += cpu_thermal_cap;
7517 :
7518 : if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7519 : continue;
7520 :
7521 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7522 : continue;
7523 :
7524 : util = cpu_util_next(cpu, p, cpu);
7525 : cpu_cap = capacity_of(cpu);
7526 :
7527 : /*
7528 : * Skip CPUs that cannot satisfy the capacity request.
7529 : * IOW, placing the task there would make the CPU
7530 : * overutilized. Take uclamp into account to see how
7531 : * much capacity we can get out of the CPU; this is
7532 : * aligned with sched_cpu_util().
7533 : */
7534 : if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
7535 : /*
7536 : * Open code uclamp_rq_util_with() except for
7537 : * the clamp() part. Ie: apply max aggregation
7538 : * only. util_fits_cpu() logic requires to
7539 : * operate on non clamped util but must use the
7540 : * max-aggregated uclamp_{min, max}.
7541 : */
7542 : rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
7543 : rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
7544 :
7545 : util_min = max(rq_util_min, p_util_min);
7546 : util_max = max(rq_util_max, p_util_max);
7547 : }
7548 :
7549 : fits = util_fits_cpu(util, util_min, util_max, cpu);
7550 : if (!fits)
7551 : continue;
7552 :
7553 : lsub_positive(&cpu_cap, util);
7554 :
7555 : if (cpu == prev_cpu) {
7556 : /* Always use prev_cpu as a candidate. */
7557 : prev_spare_cap = cpu_cap;
7558 : prev_fits = fits;
7559 : } else if ((fits > max_fits) ||
7560 : ((fits == max_fits) && (cpu_cap > max_spare_cap))) {
7561 : /*
7562 : * Find the CPU with the maximum spare capacity
7563 : * among the remaining CPUs in the performance
7564 : * domain.
7565 : */
7566 : max_spare_cap = cpu_cap;
7567 : max_spare_cap_cpu = cpu;
7568 : max_fits = fits;
7569 : }
7570 : }
7571 :
7572 : if (max_spare_cap_cpu < 0 && prev_spare_cap == 0)
7573 : continue;
7574 :
7575 : eenv_pd_busy_time(&eenv, cpus, p);
7576 : /* Compute the 'base' energy of the pd, without @p */
7577 : base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7578 :
7579 : /* Evaluate the energy impact of using prev_cpu. */
7580 : if (prev_spare_cap > 0) {
7581 : prev_delta = compute_energy(&eenv, pd, cpus, p,
7582 : prev_cpu);
7583 : /* CPU utilization has changed */
7584 : if (prev_delta < base_energy)
7585 : goto unlock;
7586 : prev_delta -= base_energy;
7587 : prev_thermal_cap = cpu_thermal_cap;
7588 : best_delta = min(best_delta, prev_delta);
7589 : }
7590 :
7591 : /* Evaluate the energy impact of using max_spare_cap_cpu. */
7592 : if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7593 : /* Current best energy cpu fits better */
7594 : if (max_fits < best_fits)
7595 : continue;
7596 :
7597 : /*
7598 : * Both don't fit performance hint (i.e. uclamp_min)
7599 : * but best energy cpu has better capacity.
7600 : */
7601 : if ((max_fits < 0) &&
7602 : (cpu_thermal_cap <= best_thermal_cap))
7603 : continue;
7604 :
7605 : cur_delta = compute_energy(&eenv, pd, cpus, p,
7606 : max_spare_cap_cpu);
7607 : /* CPU utilization has changed */
7608 : if (cur_delta < base_energy)
7609 : goto unlock;
7610 : cur_delta -= base_energy;
7611 :
7612 : /*
7613 : * Both fit for the task but best energy cpu has lower
7614 : * energy impact.
7615 : */
7616 : if ((max_fits > 0) && (best_fits > 0) &&
7617 : (cur_delta >= best_delta))
7618 : continue;
7619 :
7620 : best_delta = cur_delta;
7621 : best_energy_cpu = max_spare_cap_cpu;
7622 : best_fits = max_fits;
7623 : best_thermal_cap = cpu_thermal_cap;
7624 : }
7625 : }
7626 : rcu_read_unlock();
7627 :
7628 : if ((best_fits > prev_fits) ||
7629 : ((best_fits > 0) && (best_delta < prev_delta)) ||
7630 : ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
7631 : target = best_energy_cpu;
7632 :
7633 : return target;
7634 :
7635 : unlock:
7636 : rcu_read_unlock();
7637 :
7638 : return target;
7639 : }
7640 :
7641 : /*
7642 : * select_task_rq_fair: Select target runqueue for the waking task in domains
7643 : * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7644 : * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7645 : *
7646 : * Balances load by selecting the idlest CPU in the idlest group, or under
7647 : * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7648 : *
7649 : * Returns the target CPU number.
7650 : */
7651 : static int
7652 : select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7653 : {
7654 : int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7655 : struct sched_domain *tmp, *sd = NULL;
7656 : int cpu = smp_processor_id();
7657 : int new_cpu = prev_cpu;
7658 : int want_affine = 0;
7659 : /* SD_flags and WF_flags share the first nibble */
7660 : int sd_flag = wake_flags & 0xF;
7661 :
7662 : /*
7663 : * required for stable ->cpus_allowed
7664 : */
7665 : lockdep_assert_held(&p->pi_lock);
7666 : if (wake_flags & WF_TTWU) {
7667 : record_wakee(p);
7668 :
7669 : if (sched_energy_enabled()) {
7670 : new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7671 : if (new_cpu >= 0)
7672 : return new_cpu;
7673 : new_cpu = prev_cpu;
7674 : }
7675 :
7676 : want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7677 : }
7678 :
7679 : rcu_read_lock();
7680 : for_each_domain(cpu, tmp) {
7681 : /*
7682 : * If both 'cpu' and 'prev_cpu' are part of this domain,
7683 : * cpu is a valid SD_WAKE_AFFINE target.
7684 : */
7685 : if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7686 : cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7687 : if (cpu != prev_cpu)
7688 : new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7689 :
7690 : sd = NULL; /* Prefer wake_affine over balance flags */
7691 : break;
7692 : }
7693 :
7694 : /*
7695 : * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7696 : * usually do not have SD_BALANCE_WAKE set. That means wakeup
7697 : * will usually go to the fast path.
7698 : */
7699 : if (tmp->flags & sd_flag)
7700 : sd = tmp;
7701 : else if (!want_affine)
7702 : break;
7703 : }
7704 :
7705 : if (unlikely(sd)) {
7706 : /* Slow path */
7707 : new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7708 : } else if (wake_flags & WF_TTWU) { /* XXX always ? */
7709 : /* Fast path */
7710 : new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7711 : }
7712 : rcu_read_unlock();
7713 :
7714 : return new_cpu;
7715 : }
7716 :
7717 : /*
7718 : * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7719 : * cfs_rq_of(p) references at time of call are still valid and identify the
7720 : * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7721 : */
7722 : static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
7723 : {
7724 : struct sched_entity *se = &p->se;
7725 :
7726 : /*
7727 : * As blocked tasks retain absolute vruntime the migration needs to
7728 : * deal with this by subtracting the old and adding the new
7729 : * min_vruntime -- the latter is done by enqueue_entity() when placing
7730 : * the task on the new runqueue.
7731 : */
7732 : if (READ_ONCE(p->__state) == TASK_WAKING) {
7733 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
7734 :
7735 : se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
7736 : }
7737 :
7738 : if (!task_on_rq_migrating(p)) {
7739 : remove_entity_load_avg(se);
7740 :
7741 : /*
7742 : * Here, the task's PELT values have been updated according to
7743 : * the current rq's clock. But if that clock hasn't been
7744 : * updated in a while, a substantial idle time will be missed,
7745 : * leading to an inflation after wake-up on the new rq.
7746 : *
7747 : * Estimate the missing time from the cfs_rq last_update_time
7748 : * and update sched_avg to improve the PELT continuity after
7749 : * migration.
7750 : */
7751 : migrate_se_pelt_lag(se);
7752 : }
7753 :
7754 : /* Tell new CPU we are migrated */
7755 : se->avg.last_update_time = 0;
7756 :
7757 : update_scan_period(p, new_cpu);
7758 : }
7759 :
7760 : static void task_dead_fair(struct task_struct *p)
7761 : {
7762 : remove_entity_load_avg(&p->se);
7763 : }
7764 :
7765 : static int
7766 : balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7767 : {
7768 : if (rq->nr_running)
7769 : return 1;
7770 :
7771 : return newidle_balance(rq, rf) != 0;
7772 : }
7773 : #endif /* CONFIG_SMP */
7774 :
7775 : static unsigned long wakeup_gran(struct sched_entity *se)
7776 : {
7777 838 : unsigned long gran = sysctl_sched_wakeup_granularity;
7778 :
7779 : /*
7780 : * Since its curr running now, convert the gran from real-time
7781 : * to virtual-time in his units.
7782 : *
7783 : * By using 'se' instead of 'curr' we penalize light tasks, so
7784 : * they get preempted easier. That is, if 'se' < 'curr' then
7785 : * the resulting gran will be larger, therefore penalizing the
7786 : * lighter, if otoh 'se' > 'curr' then the resulting gran will
7787 : * be smaller, again penalizing the lighter task.
7788 : *
7789 : * This is especially important for buddies when the leftmost
7790 : * task is higher priority than the buddy.
7791 : */
7792 838 : return calc_delta_fair(gran, se);
7793 : }
7794 :
7795 : /*
7796 : * Should 'se' preempt 'curr'.
7797 : *
7798 : * |s1
7799 : * |s2
7800 : * |s3
7801 : * g
7802 : * |<--->|c
7803 : *
7804 : * w(c, s1) = -1
7805 : * w(c, s2) = 0
7806 : * w(c, s3) = 1
7807 : *
7808 : */
7809 : static int
7810 2865 : wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7811 : {
7812 2865 : s64 gran, vdiff = curr->vruntime - se->vruntime;
7813 :
7814 2865 : if (vdiff <= 0)
7815 : return -1;
7816 :
7817 838 : gran = wakeup_gran(se);
7818 838 : if (vdiff > gran)
7819 : return 1;
7820 :
7821 : return 0;
7822 : }
7823 :
7824 : static void set_last_buddy(struct sched_entity *se)
7825 : {
7826 0 : for_each_sched_entity(se) {
7827 : if (SCHED_WARN_ON(!se->on_rq))
7828 : return;
7829 0 : if (se_is_idle(se))
7830 : return;
7831 0 : cfs_rq_of(se)->last = se;
7832 : }
7833 : }
7834 :
7835 : static void set_next_buddy(struct sched_entity *se)
7836 : {
7837 837 : for_each_sched_entity(se) {
7838 : if (SCHED_WARN_ON(!se->on_rq))
7839 : return;
7840 837 : if (se_is_idle(se))
7841 : return;
7842 1674 : cfs_rq_of(se)->next = se;
7843 : }
7844 : }
7845 :
7846 : static void set_skip_buddy(struct sched_entity *se)
7847 : {
7848 0 : for_each_sched_entity(se)
7849 0 : cfs_rq_of(se)->skip = se;
7850 : }
7851 :
7852 : /*
7853 : * Preempt the current task with a newly woken task if needed:
7854 : */
7855 2439 : static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7856 : {
7857 2439 : struct task_struct *curr = rq->curr;
7858 2439 : struct sched_entity *se = &curr->se, *pse = &p->se;
7859 4878 : struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7860 2439 : int scale = cfs_rq->nr_running >= sched_nr_latency;
7861 2439 : int next_buddy_marked = 0;
7862 : int cse_is_idle, pse_is_idle;
7863 :
7864 2439 : if (unlikely(se == pse))
7865 : return;
7866 :
7867 : /*
7868 : * This is possible from callers such as attach_tasks(), in which we
7869 : * unconditionally check_preempt_curr() after an enqueue (which may have
7870 : * lead to a throttle). This both saves work and prevents false
7871 : * next-buddy nomination below.
7872 : */
7873 2439 : if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7874 : return;
7875 :
7876 : if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7877 : set_next_buddy(pse);
7878 : next_buddy_marked = 1;
7879 : }
7880 :
7881 : /*
7882 : * We can come here with TIF_NEED_RESCHED already set from new task
7883 : * wake up path.
7884 : *
7885 : * Note: this also catches the edge-case of curr being in a throttled
7886 : * group (e.g. via set_curr_task), since update_curr() (in the
7887 : * enqueue of curr) will have resulted in resched being set. This
7888 : * prevents us from potentially nominating it as a false LAST_BUDDY
7889 : * below.
7890 : */
7891 2439 : if (test_tsk_need_resched(curr))
7892 : return;
7893 :
7894 : /* Idle tasks are by definition preempted by non-idle tasks. */
7895 4054 : if (unlikely(task_has_idle_policy(curr)) &&
7896 0 : likely(!task_has_idle_policy(p)))
7897 : goto preempt;
7898 :
7899 : /*
7900 : * Batch and idle tasks do not preempt non-idle tasks (their preemption
7901 : * is driven by the tick):
7902 : */
7903 2027 : if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7904 : return;
7905 :
7906 2027 : find_matching_se(&se, &pse);
7907 2027 : WARN_ON_ONCE(!pse);
7908 :
7909 2027 : cse_is_idle = se_is_idle(se);
7910 2027 : pse_is_idle = se_is_idle(pse);
7911 :
7912 : /*
7913 : * Preempt an idle group in favor of a non-idle group (and don't preempt
7914 : * in the inverse case).
7915 : */
7916 : if (cse_is_idle && !pse_is_idle)
7917 : goto preempt;
7918 : if (cse_is_idle != pse_is_idle)
7919 : return;
7920 :
7921 4054 : update_curr(cfs_rq_of(se));
7922 2027 : if (wakeup_preempt_entity(se, pse) == 1) {
7923 : /*
7924 : * Bias pick_next to pick the sched entity that is
7925 : * triggering this preemption.
7926 : */
7927 : if (!next_buddy_marked)
7928 : set_next_buddy(pse);
7929 : goto preempt;
7930 : }
7931 :
7932 : return;
7933 :
7934 : preempt:
7935 837 : resched_curr(rq);
7936 : /*
7937 : * Only set the backward buddy when the current task is still
7938 : * on the rq. This can happen when a wakeup gets interleaved
7939 : * with schedule on the ->pre_schedule() or idle_balance()
7940 : * point, either of which can * drop the rq lock.
7941 : *
7942 : * Also, during early boot the idle thread is in the fair class,
7943 : * for obvious reasons its a bad idea to schedule back to it.
7944 : */
7945 837 : if (unlikely(!se->on_rq || curr == rq->idle))
7946 : return;
7947 :
7948 837 : if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7949 : set_last_buddy(se);
7950 : }
7951 :
7952 : #ifdef CONFIG_SMP
7953 : static struct task_struct *pick_task_fair(struct rq *rq)
7954 : {
7955 : struct sched_entity *se;
7956 : struct cfs_rq *cfs_rq;
7957 :
7958 : again:
7959 : cfs_rq = &rq->cfs;
7960 : if (!cfs_rq->nr_running)
7961 : return NULL;
7962 :
7963 : do {
7964 : struct sched_entity *curr = cfs_rq->curr;
7965 :
7966 : /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7967 : if (curr) {
7968 : if (curr->on_rq)
7969 : update_curr(cfs_rq);
7970 : else
7971 : curr = NULL;
7972 :
7973 : if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7974 : goto again;
7975 : }
7976 :
7977 : se = pick_next_entity(cfs_rq, curr);
7978 : cfs_rq = group_cfs_rq(se);
7979 : } while (cfs_rq);
7980 :
7981 : return task_of(se);
7982 : }
7983 : #endif
7984 :
7985 : struct task_struct *
7986 2513 : pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7987 : {
7988 2513 : struct cfs_rq *cfs_rq = &rq->cfs;
7989 : struct sched_entity *se;
7990 : struct task_struct *p;
7991 : int new_tasks;
7992 :
7993 : again:
7994 2513 : if (!sched_fair_runnable(rq))
7995 : goto idle;
7996 :
7997 : #ifdef CONFIG_FAIR_GROUP_SCHED
7998 : if (!prev || prev->sched_class != &fair_sched_class)
7999 : goto simple;
8000 :
8001 : /*
8002 : * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8003 : * likely that a next task is from the same cgroup as the current.
8004 : *
8005 : * Therefore attempt to avoid putting and setting the entire cgroup
8006 : * hierarchy, only change the part that actually changes.
8007 : */
8008 :
8009 : do {
8010 : struct sched_entity *curr = cfs_rq->curr;
8011 :
8012 : /*
8013 : * Since we got here without doing put_prev_entity() we also
8014 : * have to consider cfs_rq->curr. If it is still a runnable
8015 : * entity, update_curr() will update its vruntime, otherwise
8016 : * forget we've ever seen it.
8017 : */
8018 : if (curr) {
8019 : if (curr->on_rq)
8020 : update_curr(cfs_rq);
8021 : else
8022 : curr = NULL;
8023 :
8024 : /*
8025 : * This call to check_cfs_rq_runtime() will do the
8026 : * throttle and dequeue its entity in the parent(s).
8027 : * Therefore the nr_running test will indeed
8028 : * be correct.
8029 : */
8030 : if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8031 : cfs_rq = &rq->cfs;
8032 :
8033 : if (!cfs_rq->nr_running)
8034 : goto idle;
8035 :
8036 : goto simple;
8037 : }
8038 : }
8039 :
8040 : se = pick_next_entity(cfs_rq, curr);
8041 : cfs_rq = group_cfs_rq(se);
8042 : } while (cfs_rq);
8043 :
8044 : p = task_of(se);
8045 :
8046 : /*
8047 : * Since we haven't yet done put_prev_entity and if the selected task
8048 : * is a different task than we started out with, try and touch the
8049 : * least amount of cfs_rqs.
8050 : */
8051 : if (prev != p) {
8052 : struct sched_entity *pse = &prev->se;
8053 :
8054 : while (!(cfs_rq = is_same_group(se, pse))) {
8055 : int se_depth = se->depth;
8056 : int pse_depth = pse->depth;
8057 :
8058 : if (se_depth <= pse_depth) {
8059 : put_prev_entity(cfs_rq_of(pse), pse);
8060 : pse = parent_entity(pse);
8061 : }
8062 : if (se_depth >= pse_depth) {
8063 : set_next_entity(cfs_rq_of(se), se);
8064 : se = parent_entity(se);
8065 : }
8066 : }
8067 :
8068 : put_prev_entity(cfs_rq, pse);
8069 : set_next_entity(cfs_rq, se);
8070 : }
8071 :
8072 : goto done;
8073 : simple:
8074 : #endif
8075 2512 : if (prev)
8076 2512 : put_prev_task(rq, prev);
8077 :
8078 : do {
8079 2512 : se = pick_next_entity(cfs_rq, NULL);
8080 2512 : set_next_entity(cfs_rq, se);
8081 2512 : cfs_rq = group_cfs_rq(se);
8082 : } while (cfs_rq);
8083 :
8084 2512 : p = task_of(se);
8085 :
8086 : done: __maybe_unused;
8087 : #ifdef CONFIG_SMP
8088 : /*
8089 : * Move the next running task to the front of
8090 : * the list, so our cfs_tasks list becomes MRU
8091 : * one.
8092 : */
8093 : list_move(&p->se.group_node, &rq->cfs_tasks);
8094 : #endif
8095 :
8096 2512 : if (hrtick_enabled_fair(rq))
8097 : hrtick_start_fair(rq, p);
8098 :
8099 2512 : update_misfit_status(p, rq);
8100 :
8101 2512 : return p;
8102 :
8103 : idle:
8104 : if (!rf)
8105 : return NULL;
8106 :
8107 : new_tasks = newidle_balance(rq, rf);
8108 :
8109 : /*
8110 : * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8111 : * possible for any higher priority task to appear. In that case we
8112 : * must re-start the pick_next_entity() loop.
8113 : */
8114 : if (new_tasks < 0)
8115 : return RETRY_TASK;
8116 :
8117 : if (new_tasks > 0)
8118 : goto again;
8119 :
8120 : /*
8121 : * rq is about to be idle, check if we need to update the
8122 : * lost_idle_time of clock_pelt
8123 : */
8124 : update_idle_rq_clock_pelt(rq);
8125 :
8126 : return NULL;
8127 : }
8128 :
8129 0 : static struct task_struct *__pick_next_task_fair(struct rq *rq)
8130 : {
8131 0 : return pick_next_task_fair(rq, NULL, NULL);
8132 : }
8133 :
8134 : /*
8135 : * Account for a descheduled task:
8136 : */
8137 2515 : static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8138 : {
8139 2515 : struct sched_entity *se = &prev->se;
8140 : struct cfs_rq *cfs_rq;
8141 :
8142 5030 : for_each_sched_entity(se) {
8143 5030 : cfs_rq = cfs_rq_of(se);
8144 2515 : put_prev_entity(cfs_rq, se);
8145 : }
8146 2515 : }
8147 :
8148 : /*
8149 : * sched_yield() is very simple
8150 : *
8151 : * The magic of dealing with the ->skip buddy is in pick_next_entity.
8152 : */
8153 0 : static void yield_task_fair(struct rq *rq)
8154 : {
8155 0 : struct task_struct *curr = rq->curr;
8156 0 : struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8157 0 : struct sched_entity *se = &curr->se;
8158 :
8159 : /*
8160 : * Are we the only task in the tree?
8161 : */
8162 0 : if (unlikely(rq->nr_running == 1))
8163 : return;
8164 :
8165 0 : clear_buddies(cfs_rq, se);
8166 :
8167 0 : if (curr->policy != SCHED_BATCH) {
8168 0 : update_rq_clock(rq);
8169 : /*
8170 : * Update run-time statistics of the 'current'.
8171 : */
8172 0 : update_curr(cfs_rq);
8173 : /*
8174 : * Tell update_rq_clock() that we've just updated,
8175 : * so we don't do microscopic update in schedule()
8176 : * and double the fastpath cost.
8177 : */
8178 0 : rq_clock_skip_update(rq);
8179 : }
8180 :
8181 : set_skip_buddy(se);
8182 : }
8183 :
8184 0 : static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8185 : {
8186 0 : struct sched_entity *se = &p->se;
8187 :
8188 : /* throttled hierarchies are not runnable */
8189 0 : if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8190 : return false;
8191 :
8192 : /* Tell the scheduler that we'd really like pse to run next. */
8193 0 : set_next_buddy(se);
8194 :
8195 0 : yield_task_fair(rq);
8196 :
8197 0 : return true;
8198 : }
8199 :
8200 : #ifdef CONFIG_SMP
8201 : /**************************************************
8202 : * Fair scheduling class load-balancing methods.
8203 : *
8204 : * BASICS
8205 : *
8206 : * The purpose of load-balancing is to achieve the same basic fairness the
8207 : * per-CPU scheduler provides, namely provide a proportional amount of compute
8208 : * time to each task. This is expressed in the following equation:
8209 : *
8210 : * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8211 : *
8212 : * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8213 : * W_i,0 is defined as:
8214 : *
8215 : * W_i,0 = \Sum_j w_i,j (2)
8216 : *
8217 : * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8218 : * is derived from the nice value as per sched_prio_to_weight[].
8219 : *
8220 : * The weight average is an exponential decay average of the instantaneous
8221 : * weight:
8222 : *
8223 : * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8224 : *
8225 : * C_i is the compute capacity of CPU i, typically it is the
8226 : * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8227 : * can also include other factors [XXX].
8228 : *
8229 : * To achieve this balance we define a measure of imbalance which follows
8230 : * directly from (1):
8231 : *
8232 : * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8233 : *
8234 : * We them move tasks around to minimize the imbalance. In the continuous
8235 : * function space it is obvious this converges, in the discrete case we get
8236 : * a few fun cases generally called infeasible weight scenarios.
8237 : *
8238 : * [XXX expand on:
8239 : * - infeasible weights;
8240 : * - local vs global optima in the discrete case. ]
8241 : *
8242 : *
8243 : * SCHED DOMAINS
8244 : *
8245 : * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8246 : * for all i,j solution, we create a tree of CPUs that follows the hardware
8247 : * topology where each level pairs two lower groups (or better). This results
8248 : * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8249 : * tree to only the first of the previous level and we decrease the frequency
8250 : * of load-balance at each level inv. proportional to the number of CPUs in
8251 : * the groups.
8252 : *
8253 : * This yields:
8254 : *
8255 : * log_2 n 1 n
8256 : * \Sum { --- * --- * 2^i } = O(n) (5)
8257 : * i = 0 2^i 2^i
8258 : * `- size of each group
8259 : * | | `- number of CPUs doing load-balance
8260 : * | `- freq
8261 : * `- sum over all levels
8262 : *
8263 : * Coupled with a limit on how many tasks we can migrate every balance pass,
8264 : * this makes (5) the runtime complexity of the balancer.
8265 : *
8266 : * An important property here is that each CPU is still (indirectly) connected
8267 : * to every other CPU in at most O(log n) steps:
8268 : *
8269 : * The adjacency matrix of the resulting graph is given by:
8270 : *
8271 : * log_2 n
8272 : * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8273 : * k = 0
8274 : *
8275 : * And you'll find that:
8276 : *
8277 : * A^(log_2 n)_i,j != 0 for all i,j (7)
8278 : *
8279 : * Showing there's indeed a path between every CPU in at most O(log n) steps.
8280 : * The task movement gives a factor of O(m), giving a convergence complexity
8281 : * of:
8282 : *
8283 : * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8284 : *
8285 : *
8286 : * WORK CONSERVING
8287 : *
8288 : * In order to avoid CPUs going idle while there's still work to do, new idle
8289 : * balancing is more aggressive and has the newly idle CPU iterate up the domain
8290 : * tree itself instead of relying on other CPUs to bring it work.
8291 : *
8292 : * This adds some complexity to both (5) and (8) but it reduces the total idle
8293 : * time.
8294 : *
8295 : * [XXX more?]
8296 : *
8297 : *
8298 : * CGROUPS
8299 : *
8300 : * Cgroups make a horror show out of (2), instead of a simple sum we get:
8301 : *
8302 : * s_k,i
8303 : * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8304 : * S_k
8305 : *
8306 : * Where
8307 : *
8308 : * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8309 : *
8310 : * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8311 : *
8312 : * The big problem is S_k, its a global sum needed to compute a local (W_i)
8313 : * property.
8314 : *
8315 : * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8316 : * rewrite all of this once again.]
8317 : */
8318 :
8319 : static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8320 :
8321 : enum fbq_type { regular, remote, all };
8322 :
8323 : /*
8324 : * 'group_type' describes the group of CPUs at the moment of load balancing.
8325 : *
8326 : * The enum is ordered by pulling priority, with the group with lowest priority
8327 : * first so the group_type can simply be compared when selecting the busiest
8328 : * group. See update_sd_pick_busiest().
8329 : */
8330 : enum group_type {
8331 : /* The group has spare capacity that can be used to run more tasks. */
8332 : group_has_spare = 0,
8333 : /*
8334 : * The group is fully used and the tasks don't compete for more CPU
8335 : * cycles. Nevertheless, some tasks might wait before running.
8336 : */
8337 : group_fully_busy,
8338 : /*
8339 : * One task doesn't fit with CPU's capacity and must be migrated to a
8340 : * more powerful CPU.
8341 : */
8342 : group_misfit_task,
8343 : /*
8344 : * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8345 : * and the task should be migrated to it instead of running on the
8346 : * current CPU.
8347 : */
8348 : group_asym_packing,
8349 : /*
8350 : * The tasks' affinity constraints previously prevented the scheduler
8351 : * from balancing the load across the system.
8352 : */
8353 : group_imbalanced,
8354 : /*
8355 : * The CPU is overloaded and can't provide expected CPU cycles to all
8356 : * tasks.
8357 : */
8358 : group_overloaded
8359 : };
8360 :
8361 : enum migration_type {
8362 : migrate_load = 0,
8363 : migrate_util,
8364 : migrate_task,
8365 : migrate_misfit
8366 : };
8367 :
8368 : #define LBF_ALL_PINNED 0x01
8369 : #define LBF_NEED_BREAK 0x02
8370 : #define LBF_DST_PINNED 0x04
8371 : #define LBF_SOME_PINNED 0x08
8372 : #define LBF_ACTIVE_LB 0x10
8373 :
8374 : struct lb_env {
8375 : struct sched_domain *sd;
8376 :
8377 : struct rq *src_rq;
8378 : int src_cpu;
8379 :
8380 : int dst_cpu;
8381 : struct rq *dst_rq;
8382 :
8383 : struct cpumask *dst_grpmask;
8384 : int new_dst_cpu;
8385 : enum cpu_idle_type idle;
8386 : long imbalance;
8387 : /* The set of CPUs under consideration for load-balancing */
8388 : struct cpumask *cpus;
8389 :
8390 : unsigned int flags;
8391 :
8392 : unsigned int loop;
8393 : unsigned int loop_break;
8394 : unsigned int loop_max;
8395 :
8396 : enum fbq_type fbq_type;
8397 : enum migration_type migration_type;
8398 : struct list_head tasks;
8399 : };
8400 :
8401 : /*
8402 : * Is this task likely cache-hot:
8403 : */
8404 : static int task_hot(struct task_struct *p, struct lb_env *env)
8405 : {
8406 : s64 delta;
8407 :
8408 : lockdep_assert_rq_held(env->src_rq);
8409 :
8410 : if (p->sched_class != &fair_sched_class)
8411 : return 0;
8412 :
8413 : if (unlikely(task_has_idle_policy(p)))
8414 : return 0;
8415 :
8416 : /* SMT siblings share cache */
8417 : if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8418 : return 0;
8419 :
8420 : /*
8421 : * Buddy candidates are cache hot:
8422 : */
8423 : if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8424 : (&p->se == cfs_rq_of(&p->se)->next ||
8425 : &p->se == cfs_rq_of(&p->se)->last))
8426 : return 1;
8427 :
8428 : if (sysctl_sched_migration_cost == -1)
8429 : return 1;
8430 :
8431 : /*
8432 : * Don't migrate task if the task's cookie does not match
8433 : * with the destination CPU's core cookie.
8434 : */
8435 : if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8436 : return 1;
8437 :
8438 : if (sysctl_sched_migration_cost == 0)
8439 : return 0;
8440 :
8441 : delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8442 :
8443 : return delta < (s64)sysctl_sched_migration_cost;
8444 : }
8445 :
8446 : #ifdef CONFIG_NUMA_BALANCING
8447 : /*
8448 : * Returns 1, if task migration degrades locality
8449 : * Returns 0, if task migration improves locality i.e migration preferred.
8450 : * Returns -1, if task migration is not affected by locality.
8451 : */
8452 : static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8453 : {
8454 : struct numa_group *numa_group = rcu_dereference(p->numa_group);
8455 : unsigned long src_weight, dst_weight;
8456 : int src_nid, dst_nid, dist;
8457 :
8458 : if (!static_branch_likely(&sched_numa_balancing))
8459 : return -1;
8460 :
8461 : if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8462 : return -1;
8463 :
8464 : src_nid = cpu_to_node(env->src_cpu);
8465 : dst_nid = cpu_to_node(env->dst_cpu);
8466 :
8467 : if (src_nid == dst_nid)
8468 : return -1;
8469 :
8470 : /* Migrating away from the preferred node is always bad. */
8471 : if (src_nid == p->numa_preferred_nid) {
8472 : if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8473 : return 1;
8474 : else
8475 : return -1;
8476 : }
8477 :
8478 : /* Encourage migration to the preferred node. */
8479 : if (dst_nid == p->numa_preferred_nid)
8480 : return 0;
8481 :
8482 : /* Leaving a core idle is often worse than degrading locality. */
8483 : if (env->idle == CPU_IDLE)
8484 : return -1;
8485 :
8486 : dist = node_distance(src_nid, dst_nid);
8487 : if (numa_group) {
8488 : src_weight = group_weight(p, src_nid, dist);
8489 : dst_weight = group_weight(p, dst_nid, dist);
8490 : } else {
8491 : src_weight = task_weight(p, src_nid, dist);
8492 : dst_weight = task_weight(p, dst_nid, dist);
8493 : }
8494 :
8495 : return dst_weight < src_weight;
8496 : }
8497 :
8498 : #else
8499 : static inline int migrate_degrades_locality(struct task_struct *p,
8500 : struct lb_env *env)
8501 : {
8502 : return -1;
8503 : }
8504 : #endif
8505 :
8506 : /*
8507 : * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8508 : */
8509 : static
8510 : int can_migrate_task(struct task_struct *p, struct lb_env *env)
8511 : {
8512 : int tsk_cache_hot;
8513 :
8514 : lockdep_assert_rq_held(env->src_rq);
8515 :
8516 : /*
8517 : * We do not migrate tasks that are:
8518 : * 1) throttled_lb_pair, or
8519 : * 2) cannot be migrated to this CPU due to cpus_ptr, or
8520 : * 3) running (obviously), or
8521 : * 4) are cache-hot on their current CPU.
8522 : */
8523 : if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8524 : return 0;
8525 :
8526 : /* Disregard pcpu kthreads; they are where they need to be. */
8527 : if (kthread_is_per_cpu(p))
8528 : return 0;
8529 :
8530 : if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8531 : int cpu;
8532 :
8533 : schedstat_inc(p->stats.nr_failed_migrations_affine);
8534 :
8535 : env->flags |= LBF_SOME_PINNED;
8536 :
8537 : /*
8538 : * Remember if this task can be migrated to any other CPU in
8539 : * our sched_group. We may want to revisit it if we couldn't
8540 : * meet load balance goals by pulling other tasks on src_cpu.
8541 : *
8542 : * Avoid computing new_dst_cpu
8543 : * - for NEWLY_IDLE
8544 : * - if we have already computed one in current iteration
8545 : * - if it's an active balance
8546 : */
8547 : if (env->idle == CPU_NEWLY_IDLE ||
8548 : env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8549 : return 0;
8550 :
8551 : /* Prevent to re-select dst_cpu via env's CPUs: */
8552 : for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8553 : if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8554 : env->flags |= LBF_DST_PINNED;
8555 : env->new_dst_cpu = cpu;
8556 : break;
8557 : }
8558 : }
8559 :
8560 : return 0;
8561 : }
8562 :
8563 : /* Record that we found at least one task that could run on dst_cpu */
8564 : env->flags &= ~LBF_ALL_PINNED;
8565 :
8566 : if (task_on_cpu(env->src_rq, p)) {
8567 : schedstat_inc(p->stats.nr_failed_migrations_running);
8568 : return 0;
8569 : }
8570 :
8571 : /*
8572 : * Aggressive migration if:
8573 : * 1) active balance
8574 : * 2) destination numa is preferred
8575 : * 3) task is cache cold, or
8576 : * 4) too many balance attempts have failed.
8577 : */
8578 : if (env->flags & LBF_ACTIVE_LB)
8579 : return 1;
8580 :
8581 : tsk_cache_hot = migrate_degrades_locality(p, env);
8582 : if (tsk_cache_hot == -1)
8583 : tsk_cache_hot = task_hot(p, env);
8584 :
8585 : if (tsk_cache_hot <= 0 ||
8586 : env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8587 : if (tsk_cache_hot == 1) {
8588 : schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8589 : schedstat_inc(p->stats.nr_forced_migrations);
8590 : }
8591 : return 1;
8592 : }
8593 :
8594 : schedstat_inc(p->stats.nr_failed_migrations_hot);
8595 : return 0;
8596 : }
8597 :
8598 : /*
8599 : * detach_task() -- detach the task for the migration specified in env
8600 : */
8601 : static void detach_task(struct task_struct *p, struct lb_env *env)
8602 : {
8603 : lockdep_assert_rq_held(env->src_rq);
8604 :
8605 : deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8606 : set_task_cpu(p, env->dst_cpu);
8607 : }
8608 :
8609 : /*
8610 : * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8611 : * part of active balancing operations within "domain".
8612 : *
8613 : * Returns a task if successful and NULL otherwise.
8614 : */
8615 : static struct task_struct *detach_one_task(struct lb_env *env)
8616 : {
8617 : struct task_struct *p;
8618 :
8619 : lockdep_assert_rq_held(env->src_rq);
8620 :
8621 : list_for_each_entry_reverse(p,
8622 : &env->src_rq->cfs_tasks, se.group_node) {
8623 : if (!can_migrate_task(p, env))
8624 : continue;
8625 :
8626 : detach_task(p, env);
8627 :
8628 : /*
8629 : * Right now, this is only the second place where
8630 : * lb_gained[env->idle] is updated (other is detach_tasks)
8631 : * so we can safely collect stats here rather than
8632 : * inside detach_tasks().
8633 : */
8634 : schedstat_inc(env->sd->lb_gained[env->idle]);
8635 : return p;
8636 : }
8637 : return NULL;
8638 : }
8639 :
8640 : /*
8641 : * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8642 : * busiest_rq, as part of a balancing operation within domain "sd".
8643 : *
8644 : * Returns number of detached tasks if successful and 0 otherwise.
8645 : */
8646 : static int detach_tasks(struct lb_env *env)
8647 : {
8648 : struct list_head *tasks = &env->src_rq->cfs_tasks;
8649 : unsigned long util, load;
8650 : struct task_struct *p;
8651 : int detached = 0;
8652 :
8653 : lockdep_assert_rq_held(env->src_rq);
8654 :
8655 : /*
8656 : * Source run queue has been emptied by another CPU, clear
8657 : * LBF_ALL_PINNED flag as we will not test any task.
8658 : */
8659 : if (env->src_rq->nr_running <= 1) {
8660 : env->flags &= ~LBF_ALL_PINNED;
8661 : return 0;
8662 : }
8663 :
8664 : if (env->imbalance <= 0)
8665 : return 0;
8666 :
8667 : while (!list_empty(tasks)) {
8668 : /*
8669 : * We don't want to steal all, otherwise we may be treated likewise,
8670 : * which could at worst lead to a livelock crash.
8671 : */
8672 : if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8673 : break;
8674 :
8675 : env->loop++;
8676 : /*
8677 : * We've more or less seen every task there is, call it quits
8678 : * unless we haven't found any movable task yet.
8679 : */
8680 : if (env->loop > env->loop_max &&
8681 : !(env->flags & LBF_ALL_PINNED))
8682 : break;
8683 :
8684 : /* take a breather every nr_migrate tasks */
8685 : if (env->loop > env->loop_break) {
8686 : env->loop_break += SCHED_NR_MIGRATE_BREAK;
8687 : env->flags |= LBF_NEED_BREAK;
8688 : break;
8689 : }
8690 :
8691 : p = list_last_entry(tasks, struct task_struct, se.group_node);
8692 :
8693 : if (!can_migrate_task(p, env))
8694 : goto next;
8695 :
8696 : switch (env->migration_type) {
8697 : case migrate_load:
8698 : /*
8699 : * Depending of the number of CPUs and tasks and the
8700 : * cgroup hierarchy, task_h_load() can return a null
8701 : * value. Make sure that env->imbalance decreases
8702 : * otherwise detach_tasks() will stop only after
8703 : * detaching up to loop_max tasks.
8704 : */
8705 : load = max_t(unsigned long, task_h_load(p), 1);
8706 :
8707 : if (sched_feat(LB_MIN) &&
8708 : load < 16 && !env->sd->nr_balance_failed)
8709 : goto next;
8710 :
8711 : /*
8712 : * Make sure that we don't migrate too much load.
8713 : * Nevertheless, let relax the constraint if
8714 : * scheduler fails to find a good waiting task to
8715 : * migrate.
8716 : */
8717 : if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8718 : goto next;
8719 :
8720 : env->imbalance -= load;
8721 : break;
8722 :
8723 : case migrate_util:
8724 : util = task_util_est(p);
8725 :
8726 : if (util > env->imbalance)
8727 : goto next;
8728 :
8729 : env->imbalance -= util;
8730 : break;
8731 :
8732 : case migrate_task:
8733 : env->imbalance--;
8734 : break;
8735 :
8736 : case migrate_misfit:
8737 : /* This is not a misfit task */
8738 : if (task_fits_cpu(p, env->src_cpu))
8739 : goto next;
8740 :
8741 : env->imbalance = 0;
8742 : break;
8743 : }
8744 :
8745 : detach_task(p, env);
8746 : list_add(&p->se.group_node, &env->tasks);
8747 :
8748 : detached++;
8749 :
8750 : #ifdef CONFIG_PREEMPTION
8751 : /*
8752 : * NEWIDLE balancing is a source of latency, so preemptible
8753 : * kernels will stop after the first task is detached to minimize
8754 : * the critical section.
8755 : */
8756 : if (env->idle == CPU_NEWLY_IDLE)
8757 : break;
8758 : #endif
8759 :
8760 : /*
8761 : * We only want to steal up to the prescribed amount of
8762 : * load/util/tasks.
8763 : */
8764 : if (env->imbalance <= 0)
8765 : break;
8766 :
8767 : continue;
8768 : next:
8769 : list_move(&p->se.group_node, tasks);
8770 : }
8771 :
8772 : /*
8773 : * Right now, this is one of only two places we collect this stat
8774 : * so we can safely collect detach_one_task() stats here rather
8775 : * than inside detach_one_task().
8776 : */
8777 : schedstat_add(env->sd->lb_gained[env->idle], detached);
8778 :
8779 : return detached;
8780 : }
8781 :
8782 : /*
8783 : * attach_task() -- attach the task detached by detach_task() to its new rq.
8784 : */
8785 : static void attach_task(struct rq *rq, struct task_struct *p)
8786 : {
8787 : lockdep_assert_rq_held(rq);
8788 :
8789 : WARN_ON_ONCE(task_rq(p) != rq);
8790 : activate_task(rq, p, ENQUEUE_NOCLOCK);
8791 : check_preempt_curr(rq, p, 0);
8792 : }
8793 :
8794 : /*
8795 : * attach_one_task() -- attaches the task returned from detach_one_task() to
8796 : * its new rq.
8797 : */
8798 : static void attach_one_task(struct rq *rq, struct task_struct *p)
8799 : {
8800 : struct rq_flags rf;
8801 :
8802 : rq_lock(rq, &rf);
8803 : update_rq_clock(rq);
8804 : attach_task(rq, p);
8805 : rq_unlock(rq, &rf);
8806 : }
8807 :
8808 : /*
8809 : * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8810 : * new rq.
8811 : */
8812 : static void attach_tasks(struct lb_env *env)
8813 : {
8814 : struct list_head *tasks = &env->tasks;
8815 : struct task_struct *p;
8816 : struct rq_flags rf;
8817 :
8818 : rq_lock(env->dst_rq, &rf);
8819 : update_rq_clock(env->dst_rq);
8820 :
8821 : while (!list_empty(tasks)) {
8822 : p = list_first_entry(tasks, struct task_struct, se.group_node);
8823 : list_del_init(&p->se.group_node);
8824 :
8825 : attach_task(env->dst_rq, p);
8826 : }
8827 :
8828 : rq_unlock(env->dst_rq, &rf);
8829 : }
8830 :
8831 : #ifdef CONFIG_NO_HZ_COMMON
8832 : static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8833 : {
8834 : if (cfs_rq->avg.load_avg)
8835 : return true;
8836 :
8837 : if (cfs_rq->avg.util_avg)
8838 : return true;
8839 :
8840 : return false;
8841 : }
8842 :
8843 : static inline bool others_have_blocked(struct rq *rq)
8844 : {
8845 : if (READ_ONCE(rq->avg_rt.util_avg))
8846 : return true;
8847 :
8848 : if (READ_ONCE(rq->avg_dl.util_avg))
8849 : return true;
8850 :
8851 : if (thermal_load_avg(rq))
8852 : return true;
8853 :
8854 : #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8855 : if (READ_ONCE(rq->avg_irq.util_avg))
8856 : return true;
8857 : #endif
8858 :
8859 : return false;
8860 : }
8861 :
8862 : static inline void update_blocked_load_tick(struct rq *rq)
8863 : {
8864 : WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8865 : }
8866 :
8867 : static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8868 : {
8869 : if (!has_blocked)
8870 : rq->has_blocked_load = 0;
8871 : }
8872 : #else
8873 : static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8874 : static inline bool others_have_blocked(struct rq *rq) { return false; }
8875 : static inline void update_blocked_load_tick(struct rq *rq) {}
8876 : static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8877 : #endif
8878 :
8879 : static bool __update_blocked_others(struct rq *rq, bool *done)
8880 : {
8881 : const struct sched_class *curr_class;
8882 : u64 now = rq_clock_pelt(rq);
8883 : unsigned long thermal_pressure;
8884 : bool decayed;
8885 :
8886 : /*
8887 : * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8888 : * DL and IRQ signals have been updated before updating CFS.
8889 : */
8890 : curr_class = rq->curr->sched_class;
8891 :
8892 : thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8893 :
8894 : decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8895 : update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8896 : update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8897 : update_irq_load_avg(rq, 0);
8898 :
8899 : if (others_have_blocked(rq))
8900 : *done = false;
8901 :
8902 : return decayed;
8903 : }
8904 :
8905 : #ifdef CONFIG_FAIR_GROUP_SCHED
8906 :
8907 : static bool __update_blocked_fair(struct rq *rq, bool *done)
8908 : {
8909 : struct cfs_rq *cfs_rq, *pos;
8910 : bool decayed = false;
8911 : int cpu = cpu_of(rq);
8912 :
8913 : /*
8914 : * Iterates the task_group tree in a bottom up fashion, see
8915 : * list_add_leaf_cfs_rq() for details.
8916 : */
8917 : for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8918 : struct sched_entity *se;
8919 :
8920 : if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8921 : update_tg_load_avg(cfs_rq);
8922 :
8923 : if (cfs_rq->nr_running == 0)
8924 : update_idle_cfs_rq_clock_pelt(cfs_rq);
8925 :
8926 : if (cfs_rq == &rq->cfs)
8927 : decayed = true;
8928 : }
8929 :
8930 : /* Propagate pending load changes to the parent, if any: */
8931 : se = cfs_rq->tg->se[cpu];
8932 : if (se && !skip_blocked_update(se))
8933 : update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8934 :
8935 : /*
8936 : * There can be a lot of idle CPU cgroups. Don't let fully
8937 : * decayed cfs_rqs linger on the list.
8938 : */
8939 : if (cfs_rq_is_decayed(cfs_rq))
8940 : list_del_leaf_cfs_rq(cfs_rq);
8941 :
8942 : /* Don't need periodic decay once load/util_avg are null */
8943 : if (cfs_rq_has_blocked(cfs_rq))
8944 : *done = false;
8945 : }
8946 :
8947 : return decayed;
8948 : }
8949 :
8950 : /*
8951 : * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8952 : * This needs to be done in a top-down fashion because the load of a child
8953 : * group is a fraction of its parents load.
8954 : */
8955 : static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8956 : {
8957 : struct rq *rq = rq_of(cfs_rq);
8958 : struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8959 : unsigned long now = jiffies;
8960 : unsigned long load;
8961 :
8962 : if (cfs_rq->last_h_load_update == now)
8963 : return;
8964 :
8965 : WRITE_ONCE(cfs_rq->h_load_next, NULL);
8966 : for_each_sched_entity(se) {
8967 : cfs_rq = cfs_rq_of(se);
8968 : WRITE_ONCE(cfs_rq->h_load_next, se);
8969 : if (cfs_rq->last_h_load_update == now)
8970 : break;
8971 : }
8972 :
8973 : if (!se) {
8974 : cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8975 : cfs_rq->last_h_load_update = now;
8976 : }
8977 :
8978 : while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8979 : load = cfs_rq->h_load;
8980 : load = div64_ul(load * se->avg.load_avg,
8981 : cfs_rq_load_avg(cfs_rq) + 1);
8982 : cfs_rq = group_cfs_rq(se);
8983 : cfs_rq->h_load = load;
8984 : cfs_rq->last_h_load_update = now;
8985 : }
8986 : }
8987 :
8988 : static unsigned long task_h_load(struct task_struct *p)
8989 : {
8990 : struct cfs_rq *cfs_rq = task_cfs_rq(p);
8991 :
8992 : update_cfs_rq_h_load(cfs_rq);
8993 : return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8994 : cfs_rq_load_avg(cfs_rq) + 1);
8995 : }
8996 : #else
8997 : static bool __update_blocked_fair(struct rq *rq, bool *done)
8998 : {
8999 : struct cfs_rq *cfs_rq = &rq->cfs;
9000 : bool decayed;
9001 :
9002 : decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9003 : if (cfs_rq_has_blocked(cfs_rq))
9004 : *done = false;
9005 :
9006 : return decayed;
9007 : }
9008 :
9009 : static unsigned long task_h_load(struct task_struct *p)
9010 : {
9011 : return p->se.avg.load_avg;
9012 : }
9013 : #endif
9014 :
9015 : static void update_blocked_averages(int cpu)
9016 : {
9017 : bool decayed = false, done = true;
9018 : struct rq *rq = cpu_rq(cpu);
9019 : struct rq_flags rf;
9020 :
9021 : rq_lock_irqsave(rq, &rf);
9022 : update_blocked_load_tick(rq);
9023 : update_rq_clock(rq);
9024 :
9025 : decayed |= __update_blocked_others(rq, &done);
9026 : decayed |= __update_blocked_fair(rq, &done);
9027 :
9028 : update_blocked_load_status(rq, !done);
9029 : if (decayed)
9030 : cpufreq_update_util(rq, 0);
9031 : rq_unlock_irqrestore(rq, &rf);
9032 : }
9033 :
9034 : /********** Helpers for find_busiest_group ************************/
9035 :
9036 : /*
9037 : * sg_lb_stats - stats of a sched_group required for load_balancing
9038 : */
9039 : struct sg_lb_stats {
9040 : unsigned long avg_load; /*Avg load across the CPUs of the group */
9041 : unsigned long group_load; /* Total load over the CPUs of the group */
9042 : unsigned long group_capacity;
9043 : unsigned long group_util; /* Total utilization over the CPUs of the group */
9044 : unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9045 : unsigned int sum_nr_running; /* Nr of tasks running in the group */
9046 : unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9047 : unsigned int idle_cpus;
9048 : unsigned int group_weight;
9049 : enum group_type group_type;
9050 : unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9051 : unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9052 : #ifdef CONFIG_NUMA_BALANCING
9053 : unsigned int nr_numa_running;
9054 : unsigned int nr_preferred_running;
9055 : #endif
9056 : };
9057 :
9058 : /*
9059 : * sd_lb_stats - Structure to store the statistics of a sched_domain
9060 : * during load balancing.
9061 : */
9062 : struct sd_lb_stats {
9063 : struct sched_group *busiest; /* Busiest group in this sd */
9064 : struct sched_group *local; /* Local group in this sd */
9065 : unsigned long total_load; /* Total load of all groups in sd */
9066 : unsigned long total_capacity; /* Total capacity of all groups in sd */
9067 : unsigned long avg_load; /* Average load across all groups in sd */
9068 : unsigned int prefer_sibling; /* tasks should go to sibling first */
9069 :
9070 : struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9071 : struct sg_lb_stats local_stat; /* Statistics of the local group */
9072 : };
9073 :
9074 : static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9075 : {
9076 : /*
9077 : * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9078 : * local_stat because update_sg_lb_stats() does a full clear/assignment.
9079 : * We must however set busiest_stat::group_type and
9080 : * busiest_stat::idle_cpus to the worst busiest group because
9081 : * update_sd_pick_busiest() reads these before assignment.
9082 : */
9083 : *sds = (struct sd_lb_stats){
9084 : .busiest = NULL,
9085 : .local = NULL,
9086 : .total_load = 0UL,
9087 : .total_capacity = 0UL,
9088 : .busiest_stat = {
9089 : .idle_cpus = UINT_MAX,
9090 : .group_type = group_has_spare,
9091 : },
9092 : };
9093 : }
9094 :
9095 : static unsigned long scale_rt_capacity(int cpu)
9096 : {
9097 : struct rq *rq = cpu_rq(cpu);
9098 : unsigned long max = arch_scale_cpu_capacity(cpu);
9099 : unsigned long used, free;
9100 : unsigned long irq;
9101 :
9102 : irq = cpu_util_irq(rq);
9103 :
9104 : if (unlikely(irq >= max))
9105 : return 1;
9106 :
9107 : /*
9108 : * avg_rt.util_avg and avg_dl.util_avg track binary signals
9109 : * (running and not running) with weights 0 and 1024 respectively.
9110 : * avg_thermal.load_avg tracks thermal pressure and the weighted
9111 : * average uses the actual delta max capacity(load).
9112 : */
9113 : used = READ_ONCE(rq->avg_rt.util_avg);
9114 : used += READ_ONCE(rq->avg_dl.util_avg);
9115 : used += thermal_load_avg(rq);
9116 :
9117 : if (unlikely(used >= max))
9118 : return 1;
9119 :
9120 : free = max - used;
9121 :
9122 : return scale_irq_capacity(free, irq, max);
9123 : }
9124 :
9125 : static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9126 : {
9127 : unsigned long capacity = scale_rt_capacity(cpu);
9128 : struct sched_group *sdg = sd->groups;
9129 :
9130 : cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9131 :
9132 : if (!capacity)
9133 : capacity = 1;
9134 :
9135 : cpu_rq(cpu)->cpu_capacity = capacity;
9136 : trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9137 :
9138 : sdg->sgc->capacity = capacity;
9139 : sdg->sgc->min_capacity = capacity;
9140 : sdg->sgc->max_capacity = capacity;
9141 : }
9142 :
9143 : void update_group_capacity(struct sched_domain *sd, int cpu)
9144 : {
9145 : struct sched_domain *child = sd->child;
9146 : struct sched_group *group, *sdg = sd->groups;
9147 : unsigned long capacity, min_capacity, max_capacity;
9148 : unsigned long interval;
9149 :
9150 : interval = msecs_to_jiffies(sd->balance_interval);
9151 : interval = clamp(interval, 1UL, max_load_balance_interval);
9152 : sdg->sgc->next_update = jiffies + interval;
9153 :
9154 : if (!child) {
9155 : update_cpu_capacity(sd, cpu);
9156 : return;
9157 : }
9158 :
9159 : capacity = 0;
9160 : min_capacity = ULONG_MAX;
9161 : max_capacity = 0;
9162 :
9163 : if (child->flags & SD_OVERLAP) {
9164 : /*
9165 : * SD_OVERLAP domains cannot assume that child groups
9166 : * span the current group.
9167 : */
9168 :
9169 : for_each_cpu(cpu, sched_group_span(sdg)) {
9170 : unsigned long cpu_cap = capacity_of(cpu);
9171 :
9172 : capacity += cpu_cap;
9173 : min_capacity = min(cpu_cap, min_capacity);
9174 : max_capacity = max(cpu_cap, max_capacity);
9175 : }
9176 : } else {
9177 : /*
9178 : * !SD_OVERLAP domains can assume that child groups
9179 : * span the current group.
9180 : */
9181 :
9182 : group = child->groups;
9183 : do {
9184 : struct sched_group_capacity *sgc = group->sgc;
9185 :
9186 : capacity += sgc->capacity;
9187 : min_capacity = min(sgc->min_capacity, min_capacity);
9188 : max_capacity = max(sgc->max_capacity, max_capacity);
9189 : group = group->next;
9190 : } while (group != child->groups);
9191 : }
9192 :
9193 : sdg->sgc->capacity = capacity;
9194 : sdg->sgc->min_capacity = min_capacity;
9195 : sdg->sgc->max_capacity = max_capacity;
9196 : }
9197 :
9198 : /*
9199 : * Check whether the capacity of the rq has been noticeably reduced by side
9200 : * activity. The imbalance_pct is used for the threshold.
9201 : * Return true is the capacity is reduced
9202 : */
9203 : static inline int
9204 : check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9205 : {
9206 : return ((rq->cpu_capacity * sd->imbalance_pct) <
9207 : (rq->cpu_capacity_orig * 100));
9208 : }
9209 :
9210 : /*
9211 : * Check whether a rq has a misfit task and if it looks like we can actually
9212 : * help that task: we can migrate the task to a CPU of higher capacity, or
9213 : * the task's current CPU is heavily pressured.
9214 : */
9215 : static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9216 : {
9217 : return rq->misfit_task_load &&
9218 : (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9219 : check_cpu_capacity(rq, sd));
9220 : }
9221 :
9222 : /*
9223 : * Group imbalance indicates (and tries to solve) the problem where balancing
9224 : * groups is inadequate due to ->cpus_ptr constraints.
9225 : *
9226 : * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9227 : * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9228 : * Something like:
9229 : *
9230 : * { 0 1 2 3 } { 4 5 6 7 }
9231 : * * * * *
9232 : *
9233 : * If we were to balance group-wise we'd place two tasks in the first group and
9234 : * two tasks in the second group. Clearly this is undesired as it will overload
9235 : * cpu 3 and leave one of the CPUs in the second group unused.
9236 : *
9237 : * The current solution to this issue is detecting the skew in the first group
9238 : * by noticing the lower domain failed to reach balance and had difficulty
9239 : * moving tasks due to affinity constraints.
9240 : *
9241 : * When this is so detected; this group becomes a candidate for busiest; see
9242 : * update_sd_pick_busiest(). And calculate_imbalance() and
9243 : * find_busiest_group() avoid some of the usual balance conditions to allow it
9244 : * to create an effective group imbalance.
9245 : *
9246 : * This is a somewhat tricky proposition since the next run might not find the
9247 : * group imbalance and decide the groups need to be balanced again. A most
9248 : * subtle and fragile situation.
9249 : */
9250 :
9251 : static inline int sg_imbalanced(struct sched_group *group)
9252 : {
9253 : return group->sgc->imbalance;
9254 : }
9255 :
9256 : /*
9257 : * group_has_capacity returns true if the group has spare capacity that could
9258 : * be used by some tasks.
9259 : * We consider that a group has spare capacity if the number of task is
9260 : * smaller than the number of CPUs or if the utilization is lower than the
9261 : * available capacity for CFS tasks.
9262 : * For the latter, we use a threshold to stabilize the state, to take into
9263 : * account the variance of the tasks' load and to return true if the available
9264 : * capacity in meaningful for the load balancer.
9265 : * As an example, an available capacity of 1% can appear but it doesn't make
9266 : * any benefit for the load balance.
9267 : */
9268 : static inline bool
9269 : group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9270 : {
9271 : if (sgs->sum_nr_running < sgs->group_weight)
9272 : return true;
9273 :
9274 : if ((sgs->group_capacity * imbalance_pct) <
9275 : (sgs->group_runnable * 100))
9276 : return false;
9277 :
9278 : if ((sgs->group_capacity * 100) >
9279 : (sgs->group_util * imbalance_pct))
9280 : return true;
9281 :
9282 : return false;
9283 : }
9284 :
9285 : /*
9286 : * group_is_overloaded returns true if the group has more tasks than it can
9287 : * handle.
9288 : * group_is_overloaded is not equals to !group_has_capacity because a group
9289 : * with the exact right number of tasks, has no more spare capacity but is not
9290 : * overloaded so both group_has_capacity and group_is_overloaded return
9291 : * false.
9292 : */
9293 : static inline bool
9294 : group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9295 : {
9296 : if (sgs->sum_nr_running <= sgs->group_weight)
9297 : return false;
9298 :
9299 : if ((sgs->group_capacity * 100) <
9300 : (sgs->group_util * imbalance_pct))
9301 : return true;
9302 :
9303 : if ((sgs->group_capacity * imbalance_pct) <
9304 : (sgs->group_runnable * 100))
9305 : return true;
9306 :
9307 : return false;
9308 : }
9309 :
9310 : static inline enum
9311 : group_type group_classify(unsigned int imbalance_pct,
9312 : struct sched_group *group,
9313 : struct sg_lb_stats *sgs)
9314 : {
9315 : if (group_is_overloaded(imbalance_pct, sgs))
9316 : return group_overloaded;
9317 :
9318 : if (sg_imbalanced(group))
9319 : return group_imbalanced;
9320 :
9321 : if (sgs->group_asym_packing)
9322 : return group_asym_packing;
9323 :
9324 : if (sgs->group_misfit_task_load)
9325 : return group_misfit_task;
9326 :
9327 : if (!group_has_capacity(imbalance_pct, sgs))
9328 : return group_fully_busy;
9329 :
9330 : return group_has_spare;
9331 : }
9332 :
9333 : /**
9334 : * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
9335 : * @dst_cpu: Destination CPU of the load balancing
9336 : * @sds: Load-balancing data with statistics of the local group
9337 : * @sgs: Load-balancing statistics of the candidate busiest group
9338 : * @sg: The candidate busiest group
9339 : *
9340 : * Check the state of the SMT siblings of both @sds::local and @sg and decide
9341 : * if @dst_cpu can pull tasks.
9342 : *
9343 : * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
9344 : * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
9345 : * only if @dst_cpu has higher priority.
9346 : *
9347 : * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
9348 : * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
9349 : * Bigger imbalances in the number of busy CPUs will be dealt with in
9350 : * update_sd_pick_busiest().
9351 : *
9352 : * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
9353 : * of @dst_cpu are idle and @sg has lower priority.
9354 : *
9355 : * Return: true if @dst_cpu can pull tasks, false otherwise.
9356 : */
9357 : static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
9358 : struct sg_lb_stats *sgs,
9359 : struct sched_group *sg)
9360 : {
9361 : #ifdef CONFIG_SCHED_SMT
9362 : bool local_is_smt, sg_is_smt;
9363 : int sg_busy_cpus;
9364 :
9365 : local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
9366 : sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
9367 :
9368 : sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
9369 :
9370 : if (!local_is_smt) {
9371 : /*
9372 : * If we are here, @dst_cpu is idle and does not have SMT
9373 : * siblings. Pull tasks if candidate group has two or more
9374 : * busy CPUs.
9375 : */
9376 : if (sg_busy_cpus >= 2) /* implies sg_is_smt */
9377 : return true;
9378 :
9379 : /*
9380 : * @dst_cpu does not have SMT siblings. @sg may have SMT
9381 : * siblings and only one is busy. In such case, @dst_cpu
9382 : * can help if it has higher priority and is idle (i.e.,
9383 : * it has no running tasks).
9384 : */
9385 : return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9386 : }
9387 :
9388 : /* @dst_cpu has SMT siblings. */
9389 :
9390 : if (sg_is_smt) {
9391 : int local_busy_cpus = sds->local->group_weight -
9392 : sds->local_stat.idle_cpus;
9393 : int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
9394 :
9395 : if (busy_cpus_delta == 1)
9396 : return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9397 :
9398 : return false;
9399 : }
9400 :
9401 : /*
9402 : * @sg does not have SMT siblings. Ensure that @sds::local does not end
9403 : * up with more than one busy SMT sibling and only pull tasks if there
9404 : * are not busy CPUs (i.e., no CPU has running tasks).
9405 : */
9406 : if (!sds->local_stat.sum_nr_running)
9407 : return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9408 :
9409 : return false;
9410 : #else
9411 : /* Always return false so that callers deal with non-SMT cases. */
9412 : return false;
9413 : #endif
9414 : }
9415 :
9416 : static inline bool
9417 : sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
9418 : struct sched_group *group)
9419 : {
9420 : /* Only do SMT checks if either local or candidate have SMT siblings */
9421 : if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
9422 : (group->flags & SD_SHARE_CPUCAPACITY))
9423 : return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
9424 :
9425 : return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9426 : }
9427 :
9428 : static inline bool
9429 : sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9430 : {
9431 : /*
9432 : * When there is more than 1 task, the group_overloaded case already
9433 : * takes care of cpu with reduced capacity
9434 : */
9435 : if (rq->cfs.h_nr_running != 1)
9436 : return false;
9437 :
9438 : return check_cpu_capacity(rq, sd);
9439 : }
9440 :
9441 : /**
9442 : * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9443 : * @env: The load balancing environment.
9444 : * @sds: Load-balancing data with statistics of the local group.
9445 : * @group: sched_group whose statistics are to be updated.
9446 : * @sgs: variable to hold the statistics for this group.
9447 : * @sg_status: Holds flag indicating the status of the sched_group
9448 : */
9449 : static inline void update_sg_lb_stats(struct lb_env *env,
9450 : struct sd_lb_stats *sds,
9451 : struct sched_group *group,
9452 : struct sg_lb_stats *sgs,
9453 : int *sg_status)
9454 : {
9455 : int i, nr_running, local_group;
9456 :
9457 : memset(sgs, 0, sizeof(*sgs));
9458 :
9459 : local_group = group == sds->local;
9460 :
9461 : for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9462 : struct rq *rq = cpu_rq(i);
9463 : unsigned long load = cpu_load(rq);
9464 :
9465 : sgs->group_load += load;
9466 : sgs->group_util += cpu_util_cfs(i);
9467 : sgs->group_runnable += cpu_runnable(rq);
9468 : sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9469 :
9470 : nr_running = rq->nr_running;
9471 : sgs->sum_nr_running += nr_running;
9472 :
9473 : if (nr_running > 1)
9474 : *sg_status |= SG_OVERLOAD;
9475 :
9476 : if (cpu_overutilized(i))
9477 : *sg_status |= SG_OVERUTILIZED;
9478 :
9479 : #ifdef CONFIG_NUMA_BALANCING
9480 : sgs->nr_numa_running += rq->nr_numa_running;
9481 : sgs->nr_preferred_running += rq->nr_preferred_running;
9482 : #endif
9483 : /*
9484 : * No need to call idle_cpu() if nr_running is not 0
9485 : */
9486 : if (!nr_running && idle_cpu(i)) {
9487 : sgs->idle_cpus++;
9488 : /* Idle cpu can't have misfit task */
9489 : continue;
9490 : }
9491 :
9492 : if (local_group)
9493 : continue;
9494 :
9495 : if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9496 : /* Check for a misfit task on the cpu */
9497 : if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9498 : sgs->group_misfit_task_load = rq->misfit_task_load;
9499 : *sg_status |= SG_OVERLOAD;
9500 : }
9501 : } else if ((env->idle != CPU_NOT_IDLE) &&
9502 : sched_reduced_capacity(rq, env->sd)) {
9503 : /* Check for a task running on a CPU with reduced capacity */
9504 : if (sgs->group_misfit_task_load < load)
9505 : sgs->group_misfit_task_load = load;
9506 : }
9507 : }
9508 :
9509 : sgs->group_capacity = group->sgc->capacity;
9510 :
9511 : sgs->group_weight = group->group_weight;
9512 :
9513 : /* Check if dst CPU is idle and preferred to this group */
9514 : if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9515 : env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9516 : sched_asym(env, sds, sgs, group)) {
9517 : sgs->group_asym_packing = 1;
9518 : }
9519 :
9520 : sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9521 :
9522 : /* Computing avg_load makes sense only when group is overloaded */
9523 : if (sgs->group_type == group_overloaded)
9524 : sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9525 : sgs->group_capacity;
9526 : }
9527 :
9528 : /**
9529 : * update_sd_pick_busiest - return 1 on busiest group
9530 : * @env: The load balancing environment.
9531 : * @sds: sched_domain statistics
9532 : * @sg: sched_group candidate to be checked for being the busiest
9533 : * @sgs: sched_group statistics
9534 : *
9535 : * Determine if @sg is a busier group than the previously selected
9536 : * busiest group.
9537 : *
9538 : * Return: %true if @sg is a busier group than the previously selected
9539 : * busiest group. %false otherwise.
9540 : */
9541 : static bool update_sd_pick_busiest(struct lb_env *env,
9542 : struct sd_lb_stats *sds,
9543 : struct sched_group *sg,
9544 : struct sg_lb_stats *sgs)
9545 : {
9546 : struct sg_lb_stats *busiest = &sds->busiest_stat;
9547 :
9548 : /* Make sure that there is at least one task to pull */
9549 : if (!sgs->sum_h_nr_running)
9550 : return false;
9551 :
9552 : /*
9553 : * Don't try to pull misfit tasks we can't help.
9554 : * We can use max_capacity here as reduction in capacity on some
9555 : * CPUs in the group should either be possible to resolve
9556 : * internally or be covered by avg_load imbalance (eventually).
9557 : */
9558 : if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9559 : (sgs->group_type == group_misfit_task) &&
9560 : (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9561 : sds->local_stat.group_type != group_has_spare))
9562 : return false;
9563 :
9564 : if (sgs->group_type > busiest->group_type)
9565 : return true;
9566 :
9567 : if (sgs->group_type < busiest->group_type)
9568 : return false;
9569 :
9570 : /*
9571 : * The candidate and the current busiest group are the same type of
9572 : * group. Let check which one is the busiest according to the type.
9573 : */
9574 :
9575 : switch (sgs->group_type) {
9576 : case group_overloaded:
9577 : /* Select the overloaded group with highest avg_load. */
9578 : if (sgs->avg_load <= busiest->avg_load)
9579 : return false;
9580 : break;
9581 :
9582 : case group_imbalanced:
9583 : /*
9584 : * Select the 1st imbalanced group as we don't have any way to
9585 : * choose one more than another.
9586 : */
9587 : return false;
9588 :
9589 : case group_asym_packing:
9590 : /* Prefer to move from lowest priority CPU's work */
9591 : if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9592 : return false;
9593 : break;
9594 :
9595 : case group_misfit_task:
9596 : /*
9597 : * If we have more than one misfit sg go with the biggest
9598 : * misfit.
9599 : */
9600 : if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9601 : return false;
9602 : break;
9603 :
9604 : case group_fully_busy:
9605 : /*
9606 : * Select the fully busy group with highest avg_load. In
9607 : * theory, there is no need to pull task from such kind of
9608 : * group because tasks have all compute capacity that they need
9609 : * but we can still improve the overall throughput by reducing
9610 : * contention when accessing shared HW resources.
9611 : *
9612 : * XXX for now avg_load is not computed and always 0 so we
9613 : * select the 1st one.
9614 : */
9615 : if (sgs->avg_load <= busiest->avg_load)
9616 : return false;
9617 : break;
9618 :
9619 : case group_has_spare:
9620 : /*
9621 : * Select not overloaded group with lowest number of idle cpus
9622 : * and highest number of running tasks. We could also compare
9623 : * the spare capacity which is more stable but it can end up
9624 : * that the group has less spare capacity but finally more idle
9625 : * CPUs which means less opportunity to pull tasks.
9626 : */
9627 : if (sgs->idle_cpus > busiest->idle_cpus)
9628 : return false;
9629 : else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9630 : (sgs->sum_nr_running <= busiest->sum_nr_running))
9631 : return false;
9632 :
9633 : break;
9634 : }
9635 :
9636 : /*
9637 : * Candidate sg has no more than one task per CPU and has higher
9638 : * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9639 : * throughput. Maximize throughput, power/energy consequences are not
9640 : * considered.
9641 : */
9642 : if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9643 : (sgs->group_type <= group_fully_busy) &&
9644 : (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9645 : return false;
9646 :
9647 : return true;
9648 : }
9649 :
9650 : #ifdef CONFIG_NUMA_BALANCING
9651 : static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9652 : {
9653 : if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9654 : return regular;
9655 : if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9656 : return remote;
9657 : return all;
9658 : }
9659 :
9660 : static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9661 : {
9662 : if (rq->nr_running > rq->nr_numa_running)
9663 : return regular;
9664 : if (rq->nr_running > rq->nr_preferred_running)
9665 : return remote;
9666 : return all;
9667 : }
9668 : #else
9669 : static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9670 : {
9671 : return all;
9672 : }
9673 :
9674 : static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9675 : {
9676 : return regular;
9677 : }
9678 : #endif /* CONFIG_NUMA_BALANCING */
9679 :
9680 :
9681 : struct sg_lb_stats;
9682 :
9683 : /*
9684 : * task_running_on_cpu - return 1 if @p is running on @cpu.
9685 : */
9686 :
9687 : static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9688 : {
9689 : /* Task has no contribution or is new */
9690 : if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9691 : return 0;
9692 :
9693 : if (task_on_rq_queued(p))
9694 : return 1;
9695 :
9696 : return 0;
9697 : }
9698 :
9699 : /**
9700 : * idle_cpu_without - would a given CPU be idle without p ?
9701 : * @cpu: the processor on which idleness is tested.
9702 : * @p: task which should be ignored.
9703 : *
9704 : * Return: 1 if the CPU would be idle. 0 otherwise.
9705 : */
9706 : static int idle_cpu_without(int cpu, struct task_struct *p)
9707 : {
9708 : struct rq *rq = cpu_rq(cpu);
9709 :
9710 : if (rq->curr != rq->idle && rq->curr != p)
9711 : return 0;
9712 :
9713 : /*
9714 : * rq->nr_running can't be used but an updated version without the
9715 : * impact of p on cpu must be used instead. The updated nr_running
9716 : * be computed and tested before calling idle_cpu_without().
9717 : */
9718 :
9719 : #ifdef CONFIG_SMP
9720 : if (rq->ttwu_pending)
9721 : return 0;
9722 : #endif
9723 :
9724 : return 1;
9725 : }
9726 :
9727 : /*
9728 : * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9729 : * @sd: The sched_domain level to look for idlest group.
9730 : * @group: sched_group whose statistics are to be updated.
9731 : * @sgs: variable to hold the statistics for this group.
9732 : * @p: The task for which we look for the idlest group/CPU.
9733 : */
9734 : static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9735 : struct sched_group *group,
9736 : struct sg_lb_stats *sgs,
9737 : struct task_struct *p)
9738 : {
9739 : int i, nr_running;
9740 :
9741 : memset(sgs, 0, sizeof(*sgs));
9742 :
9743 : /* Assume that task can't fit any CPU of the group */
9744 : if (sd->flags & SD_ASYM_CPUCAPACITY)
9745 : sgs->group_misfit_task_load = 1;
9746 :
9747 : for_each_cpu(i, sched_group_span(group)) {
9748 : struct rq *rq = cpu_rq(i);
9749 : unsigned int local;
9750 :
9751 : sgs->group_load += cpu_load_without(rq, p);
9752 : sgs->group_util += cpu_util_without(i, p);
9753 : sgs->group_runnable += cpu_runnable_without(rq, p);
9754 : local = task_running_on_cpu(i, p);
9755 : sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
9756 :
9757 : nr_running = rq->nr_running - local;
9758 : sgs->sum_nr_running += nr_running;
9759 :
9760 : /*
9761 : * No need to call idle_cpu_without() if nr_running is not 0
9762 : */
9763 : if (!nr_running && idle_cpu_without(i, p))
9764 : sgs->idle_cpus++;
9765 :
9766 : /* Check if task fits in the CPU */
9767 : if (sd->flags & SD_ASYM_CPUCAPACITY &&
9768 : sgs->group_misfit_task_load &&
9769 : task_fits_cpu(p, i))
9770 : sgs->group_misfit_task_load = 0;
9771 :
9772 : }
9773 :
9774 : sgs->group_capacity = group->sgc->capacity;
9775 :
9776 : sgs->group_weight = group->group_weight;
9777 :
9778 : sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
9779 :
9780 : /*
9781 : * Computing avg_load makes sense only when group is fully busy or
9782 : * overloaded
9783 : */
9784 : if (sgs->group_type == group_fully_busy ||
9785 : sgs->group_type == group_overloaded)
9786 : sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9787 : sgs->group_capacity;
9788 : }
9789 :
9790 : static bool update_pick_idlest(struct sched_group *idlest,
9791 : struct sg_lb_stats *idlest_sgs,
9792 : struct sched_group *group,
9793 : struct sg_lb_stats *sgs)
9794 : {
9795 : if (sgs->group_type < idlest_sgs->group_type)
9796 : return true;
9797 :
9798 : if (sgs->group_type > idlest_sgs->group_type)
9799 : return false;
9800 :
9801 : /*
9802 : * The candidate and the current idlest group are the same type of
9803 : * group. Let check which one is the idlest according to the type.
9804 : */
9805 :
9806 : switch (sgs->group_type) {
9807 : case group_overloaded:
9808 : case group_fully_busy:
9809 : /* Select the group with lowest avg_load. */
9810 : if (idlest_sgs->avg_load <= sgs->avg_load)
9811 : return false;
9812 : break;
9813 :
9814 : case group_imbalanced:
9815 : case group_asym_packing:
9816 : /* Those types are not used in the slow wakeup path */
9817 : return false;
9818 :
9819 : case group_misfit_task:
9820 : /* Select group with the highest max capacity */
9821 : if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9822 : return false;
9823 : break;
9824 :
9825 : case group_has_spare:
9826 : /* Select group with most idle CPUs */
9827 : if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9828 : return false;
9829 :
9830 : /* Select group with lowest group_util */
9831 : if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9832 : idlest_sgs->group_util <= sgs->group_util)
9833 : return false;
9834 :
9835 : break;
9836 : }
9837 :
9838 : return true;
9839 : }
9840 :
9841 : /*
9842 : * find_idlest_group() finds and returns the least busy CPU group within the
9843 : * domain.
9844 : *
9845 : * Assumes p is allowed on at least one CPU in sd.
9846 : */
9847 : static struct sched_group *
9848 : find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9849 : {
9850 : struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9851 : struct sg_lb_stats local_sgs, tmp_sgs;
9852 : struct sg_lb_stats *sgs;
9853 : unsigned long imbalance;
9854 : struct sg_lb_stats idlest_sgs = {
9855 : .avg_load = UINT_MAX,
9856 : .group_type = group_overloaded,
9857 : };
9858 :
9859 : do {
9860 : int local_group;
9861 :
9862 : /* Skip over this group if it has no CPUs allowed */
9863 : if (!cpumask_intersects(sched_group_span(group),
9864 : p->cpus_ptr))
9865 : continue;
9866 :
9867 : /* Skip over this group if no cookie matched */
9868 : if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9869 : continue;
9870 :
9871 : local_group = cpumask_test_cpu(this_cpu,
9872 : sched_group_span(group));
9873 :
9874 : if (local_group) {
9875 : sgs = &local_sgs;
9876 : local = group;
9877 : } else {
9878 : sgs = &tmp_sgs;
9879 : }
9880 :
9881 : update_sg_wakeup_stats(sd, group, sgs, p);
9882 :
9883 : if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9884 : idlest = group;
9885 : idlest_sgs = *sgs;
9886 : }
9887 :
9888 : } while (group = group->next, group != sd->groups);
9889 :
9890 :
9891 : /* There is no idlest group to push tasks to */
9892 : if (!idlest)
9893 : return NULL;
9894 :
9895 : /* The local group has been skipped because of CPU affinity */
9896 : if (!local)
9897 : return idlest;
9898 :
9899 : /*
9900 : * If the local group is idler than the selected idlest group
9901 : * don't try and push the task.
9902 : */
9903 : if (local_sgs.group_type < idlest_sgs.group_type)
9904 : return NULL;
9905 :
9906 : /*
9907 : * If the local group is busier than the selected idlest group
9908 : * try and push the task.
9909 : */
9910 : if (local_sgs.group_type > idlest_sgs.group_type)
9911 : return idlest;
9912 :
9913 : switch (local_sgs.group_type) {
9914 : case group_overloaded:
9915 : case group_fully_busy:
9916 :
9917 : /* Calculate allowed imbalance based on load */
9918 : imbalance = scale_load_down(NICE_0_LOAD) *
9919 : (sd->imbalance_pct-100) / 100;
9920 :
9921 : /*
9922 : * When comparing groups across NUMA domains, it's possible for
9923 : * the local domain to be very lightly loaded relative to the
9924 : * remote domains but "imbalance" skews the comparison making
9925 : * remote CPUs look much more favourable. When considering
9926 : * cross-domain, add imbalance to the load on the remote node
9927 : * and consider staying local.
9928 : */
9929 :
9930 : if ((sd->flags & SD_NUMA) &&
9931 : ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9932 : return NULL;
9933 :
9934 : /*
9935 : * If the local group is less loaded than the selected
9936 : * idlest group don't try and push any tasks.
9937 : */
9938 : if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9939 : return NULL;
9940 :
9941 : if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9942 : return NULL;
9943 : break;
9944 :
9945 : case group_imbalanced:
9946 : case group_asym_packing:
9947 : /* Those type are not used in the slow wakeup path */
9948 : return NULL;
9949 :
9950 : case group_misfit_task:
9951 : /* Select group with the highest max capacity */
9952 : if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9953 : return NULL;
9954 : break;
9955 :
9956 : case group_has_spare:
9957 : #ifdef CONFIG_NUMA
9958 : if (sd->flags & SD_NUMA) {
9959 : int imb_numa_nr = sd->imb_numa_nr;
9960 : #ifdef CONFIG_NUMA_BALANCING
9961 : int idlest_cpu;
9962 : /*
9963 : * If there is spare capacity at NUMA, try to select
9964 : * the preferred node
9965 : */
9966 : if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9967 : return NULL;
9968 :
9969 : idlest_cpu = cpumask_first(sched_group_span(idlest));
9970 : if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9971 : return idlest;
9972 : #endif /* CONFIG_NUMA_BALANCING */
9973 : /*
9974 : * Otherwise, keep the task close to the wakeup source
9975 : * and improve locality if the number of running tasks
9976 : * would remain below threshold where an imbalance is
9977 : * allowed while accounting for the possibility the
9978 : * task is pinned to a subset of CPUs. If there is a
9979 : * real need of migration, periodic load balance will
9980 : * take care of it.
9981 : */
9982 : if (p->nr_cpus_allowed != NR_CPUS) {
9983 : struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9984 :
9985 : cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
9986 : imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
9987 : }
9988 :
9989 : imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
9990 : if (!adjust_numa_imbalance(imbalance,
9991 : local_sgs.sum_nr_running + 1,
9992 : imb_numa_nr)) {
9993 : return NULL;
9994 : }
9995 : }
9996 : #endif /* CONFIG_NUMA */
9997 :
9998 : /*
9999 : * Select group with highest number of idle CPUs. We could also
10000 : * compare the utilization which is more stable but it can end
10001 : * up that the group has less spare capacity but finally more
10002 : * idle CPUs which means more opportunity to run task.
10003 : */
10004 : if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10005 : return NULL;
10006 : break;
10007 : }
10008 :
10009 : return idlest;
10010 : }
10011 :
10012 : static void update_idle_cpu_scan(struct lb_env *env,
10013 : unsigned long sum_util)
10014 : {
10015 : struct sched_domain_shared *sd_share;
10016 : int llc_weight, pct;
10017 : u64 x, y, tmp;
10018 : /*
10019 : * Update the number of CPUs to scan in LLC domain, which could
10020 : * be used as a hint in select_idle_cpu(). The update of sd_share
10021 : * could be expensive because it is within a shared cache line.
10022 : * So the write of this hint only occurs during periodic load
10023 : * balancing, rather than CPU_NEWLY_IDLE, because the latter
10024 : * can fire way more frequently than the former.
10025 : */
10026 : if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10027 : return;
10028 :
10029 : llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10030 : if (env->sd->span_weight != llc_weight)
10031 : return;
10032 :
10033 : sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10034 : if (!sd_share)
10035 : return;
10036 :
10037 : /*
10038 : * The number of CPUs to search drops as sum_util increases, when
10039 : * sum_util hits 85% or above, the scan stops.
10040 : * The reason to choose 85% as the threshold is because this is the
10041 : * imbalance_pct(117) when a LLC sched group is overloaded.
10042 : *
10043 : * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10044 : * and y'= y / SCHED_CAPACITY_SCALE
10045 : *
10046 : * x is the ratio of sum_util compared to the CPU capacity:
10047 : * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10048 : * y' is the ratio of CPUs to be scanned in the LLC domain,
10049 : * and the number of CPUs to scan is calculated by:
10050 : *
10051 : * nr_scan = llc_weight * y' [2]
10052 : *
10053 : * When x hits the threshold of overloaded, AKA, when
10054 : * x = 100 / pct, y drops to 0. According to [1],
10055 : * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10056 : *
10057 : * Scale x by SCHED_CAPACITY_SCALE:
10058 : * x' = sum_util / llc_weight; [3]
10059 : *
10060 : * and finally [1] becomes:
10061 : * y = SCHED_CAPACITY_SCALE -
10062 : * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10063 : *
10064 : */
10065 : /* equation [3] */
10066 : x = sum_util;
10067 : do_div(x, llc_weight);
10068 :
10069 : /* equation [4] */
10070 : pct = env->sd->imbalance_pct;
10071 : tmp = x * x * pct * pct;
10072 : do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10073 : tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10074 : y = SCHED_CAPACITY_SCALE - tmp;
10075 :
10076 : /* equation [2] */
10077 : y *= llc_weight;
10078 : do_div(y, SCHED_CAPACITY_SCALE);
10079 : if ((int)y != sd_share->nr_idle_scan)
10080 : WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10081 : }
10082 :
10083 : /**
10084 : * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10085 : * @env: The load balancing environment.
10086 : * @sds: variable to hold the statistics for this sched_domain.
10087 : */
10088 :
10089 : static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10090 : {
10091 : struct sched_domain *child = env->sd->child;
10092 : struct sched_group *sg = env->sd->groups;
10093 : struct sg_lb_stats *local = &sds->local_stat;
10094 : struct sg_lb_stats tmp_sgs;
10095 : unsigned long sum_util = 0;
10096 : int sg_status = 0;
10097 :
10098 : do {
10099 : struct sg_lb_stats *sgs = &tmp_sgs;
10100 : int local_group;
10101 :
10102 : local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10103 : if (local_group) {
10104 : sds->local = sg;
10105 : sgs = local;
10106 :
10107 : if (env->idle != CPU_NEWLY_IDLE ||
10108 : time_after_eq(jiffies, sg->sgc->next_update))
10109 : update_group_capacity(env->sd, env->dst_cpu);
10110 : }
10111 :
10112 : update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10113 :
10114 : if (local_group)
10115 : goto next_group;
10116 :
10117 :
10118 : if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10119 : sds->busiest = sg;
10120 : sds->busiest_stat = *sgs;
10121 : }
10122 :
10123 : next_group:
10124 : /* Now, start updating sd_lb_stats */
10125 : sds->total_load += sgs->group_load;
10126 : sds->total_capacity += sgs->group_capacity;
10127 :
10128 : sum_util += sgs->group_util;
10129 : sg = sg->next;
10130 : } while (sg != env->sd->groups);
10131 :
10132 : /* Tag domain that child domain prefers tasks go to siblings first */
10133 : sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
10134 :
10135 :
10136 : if (env->sd->flags & SD_NUMA)
10137 : env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10138 :
10139 : if (!env->sd->parent) {
10140 : struct root_domain *rd = env->dst_rq->rd;
10141 :
10142 : /* update overload indicator if we are at root domain */
10143 : WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10144 :
10145 : /* Update over-utilization (tipping point, U >= 0) indicator */
10146 : WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10147 : trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10148 : } else if (sg_status & SG_OVERUTILIZED) {
10149 : struct root_domain *rd = env->dst_rq->rd;
10150 :
10151 : WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10152 : trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10153 : }
10154 :
10155 : update_idle_cpu_scan(env, sum_util);
10156 : }
10157 :
10158 : /**
10159 : * calculate_imbalance - Calculate the amount of imbalance present within the
10160 : * groups of a given sched_domain during load balance.
10161 : * @env: load balance environment
10162 : * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10163 : */
10164 : static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10165 : {
10166 : struct sg_lb_stats *local, *busiest;
10167 :
10168 : local = &sds->local_stat;
10169 : busiest = &sds->busiest_stat;
10170 :
10171 : if (busiest->group_type == group_misfit_task) {
10172 : if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10173 : /* Set imbalance to allow misfit tasks to be balanced. */
10174 : env->migration_type = migrate_misfit;
10175 : env->imbalance = 1;
10176 : } else {
10177 : /*
10178 : * Set load imbalance to allow moving task from cpu
10179 : * with reduced capacity.
10180 : */
10181 : env->migration_type = migrate_load;
10182 : env->imbalance = busiest->group_misfit_task_load;
10183 : }
10184 : return;
10185 : }
10186 :
10187 : if (busiest->group_type == group_asym_packing) {
10188 : /*
10189 : * In case of asym capacity, we will try to migrate all load to
10190 : * the preferred CPU.
10191 : */
10192 : env->migration_type = migrate_task;
10193 : env->imbalance = busiest->sum_h_nr_running;
10194 : return;
10195 : }
10196 :
10197 : if (busiest->group_type == group_imbalanced) {
10198 : /*
10199 : * In the group_imb case we cannot rely on group-wide averages
10200 : * to ensure CPU-load equilibrium, try to move any task to fix
10201 : * the imbalance. The next load balance will take care of
10202 : * balancing back the system.
10203 : */
10204 : env->migration_type = migrate_task;
10205 : env->imbalance = 1;
10206 : return;
10207 : }
10208 :
10209 : /*
10210 : * Try to use spare capacity of local group without overloading it or
10211 : * emptying busiest.
10212 : */
10213 : if (local->group_type == group_has_spare) {
10214 : if ((busiest->group_type > group_fully_busy) &&
10215 : !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10216 : /*
10217 : * If busiest is overloaded, try to fill spare
10218 : * capacity. This might end up creating spare capacity
10219 : * in busiest or busiest still being overloaded but
10220 : * there is no simple way to directly compute the
10221 : * amount of load to migrate in order to balance the
10222 : * system.
10223 : */
10224 : env->migration_type = migrate_util;
10225 : env->imbalance = max(local->group_capacity, local->group_util) -
10226 : local->group_util;
10227 :
10228 : /*
10229 : * In some cases, the group's utilization is max or even
10230 : * higher than capacity because of migrations but the
10231 : * local CPU is (newly) idle. There is at least one
10232 : * waiting task in this overloaded busiest group. Let's
10233 : * try to pull it.
10234 : */
10235 : if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10236 : env->migration_type = migrate_task;
10237 : env->imbalance = 1;
10238 : }
10239 :
10240 : return;
10241 : }
10242 :
10243 : if (busiest->group_weight == 1 || sds->prefer_sibling) {
10244 : unsigned int nr_diff = busiest->sum_nr_running;
10245 : /*
10246 : * When prefer sibling, evenly spread running tasks on
10247 : * groups.
10248 : */
10249 : env->migration_type = migrate_task;
10250 : lsub_positive(&nr_diff, local->sum_nr_running);
10251 : env->imbalance = nr_diff;
10252 : } else {
10253 :
10254 : /*
10255 : * If there is no overload, we just want to even the number of
10256 : * idle cpus.
10257 : */
10258 : env->migration_type = migrate_task;
10259 : env->imbalance = max_t(long, 0,
10260 : (local->idle_cpus - busiest->idle_cpus));
10261 : }
10262 :
10263 : #ifdef CONFIG_NUMA
10264 : /* Consider allowing a small imbalance between NUMA groups */
10265 : if (env->sd->flags & SD_NUMA) {
10266 : env->imbalance = adjust_numa_imbalance(env->imbalance,
10267 : local->sum_nr_running + 1,
10268 : env->sd->imb_numa_nr);
10269 : }
10270 : #endif
10271 :
10272 : /* Number of tasks to move to restore balance */
10273 : env->imbalance >>= 1;
10274 :
10275 : return;
10276 : }
10277 :
10278 : /*
10279 : * Local is fully busy but has to take more load to relieve the
10280 : * busiest group
10281 : */
10282 : if (local->group_type < group_overloaded) {
10283 : /*
10284 : * Local will become overloaded so the avg_load metrics are
10285 : * finally needed.
10286 : */
10287 :
10288 : local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10289 : local->group_capacity;
10290 :
10291 : /*
10292 : * If the local group is more loaded than the selected
10293 : * busiest group don't try to pull any tasks.
10294 : */
10295 : if (local->avg_load >= busiest->avg_load) {
10296 : env->imbalance = 0;
10297 : return;
10298 : }
10299 :
10300 : sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10301 : sds->total_capacity;
10302 :
10303 : /*
10304 : * If the local group is more loaded than the average system
10305 : * load, don't try to pull any tasks.
10306 : */
10307 : if (local->avg_load >= sds->avg_load) {
10308 : env->imbalance = 0;
10309 : return;
10310 : }
10311 :
10312 : }
10313 :
10314 : /*
10315 : * Both group are or will become overloaded and we're trying to get all
10316 : * the CPUs to the average_load, so we don't want to push ourselves
10317 : * above the average load, nor do we wish to reduce the max loaded CPU
10318 : * below the average load. At the same time, we also don't want to
10319 : * reduce the group load below the group capacity. Thus we look for
10320 : * the minimum possible imbalance.
10321 : */
10322 : env->migration_type = migrate_load;
10323 : env->imbalance = min(
10324 : (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10325 : (sds->avg_load - local->avg_load) * local->group_capacity
10326 : ) / SCHED_CAPACITY_SCALE;
10327 : }
10328 :
10329 : /******* find_busiest_group() helpers end here *********************/
10330 :
10331 : /*
10332 : * Decision matrix according to the local and busiest group type:
10333 : *
10334 : * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10335 : * has_spare nr_idle balanced N/A N/A balanced balanced
10336 : * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10337 : * misfit_task force N/A N/A N/A N/A N/A
10338 : * asym_packing force force N/A N/A force force
10339 : * imbalanced force force N/A N/A force force
10340 : * overloaded force force N/A N/A force avg_load
10341 : *
10342 : * N/A : Not Applicable because already filtered while updating
10343 : * statistics.
10344 : * balanced : The system is balanced for these 2 groups.
10345 : * force : Calculate the imbalance as load migration is probably needed.
10346 : * avg_load : Only if imbalance is significant enough.
10347 : * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10348 : * different in groups.
10349 : */
10350 :
10351 : /**
10352 : * find_busiest_group - Returns the busiest group within the sched_domain
10353 : * if there is an imbalance.
10354 : * @env: The load balancing environment.
10355 : *
10356 : * Also calculates the amount of runnable load which should be moved
10357 : * to restore balance.
10358 : *
10359 : * Return: - The busiest group if imbalance exists.
10360 : */
10361 : static struct sched_group *find_busiest_group(struct lb_env *env)
10362 : {
10363 : struct sg_lb_stats *local, *busiest;
10364 : struct sd_lb_stats sds;
10365 :
10366 : init_sd_lb_stats(&sds);
10367 :
10368 : /*
10369 : * Compute the various statistics relevant for load balancing at
10370 : * this level.
10371 : */
10372 : update_sd_lb_stats(env, &sds);
10373 :
10374 : /* There is no busy sibling group to pull tasks from */
10375 : if (!sds.busiest)
10376 : goto out_balanced;
10377 :
10378 : busiest = &sds.busiest_stat;
10379 :
10380 : /* Misfit tasks should be dealt with regardless of the avg load */
10381 : if (busiest->group_type == group_misfit_task)
10382 : goto force_balance;
10383 :
10384 : if (sched_energy_enabled()) {
10385 : struct root_domain *rd = env->dst_rq->rd;
10386 :
10387 : if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10388 : goto out_balanced;
10389 : }
10390 :
10391 : /* ASYM feature bypasses nice load balance check */
10392 : if (busiest->group_type == group_asym_packing)
10393 : goto force_balance;
10394 :
10395 : /*
10396 : * If the busiest group is imbalanced the below checks don't
10397 : * work because they assume all things are equal, which typically
10398 : * isn't true due to cpus_ptr constraints and the like.
10399 : */
10400 : if (busiest->group_type == group_imbalanced)
10401 : goto force_balance;
10402 :
10403 : local = &sds.local_stat;
10404 : /*
10405 : * If the local group is busier than the selected busiest group
10406 : * don't try and pull any tasks.
10407 : */
10408 : if (local->group_type > busiest->group_type)
10409 : goto out_balanced;
10410 :
10411 : /*
10412 : * When groups are overloaded, use the avg_load to ensure fairness
10413 : * between tasks.
10414 : */
10415 : if (local->group_type == group_overloaded) {
10416 : /*
10417 : * If the local group is more loaded than the selected
10418 : * busiest group don't try to pull any tasks.
10419 : */
10420 : if (local->avg_load >= busiest->avg_load)
10421 : goto out_balanced;
10422 :
10423 : /* XXX broken for overlapping NUMA groups */
10424 : sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10425 : sds.total_capacity;
10426 :
10427 : /*
10428 : * Don't pull any tasks if this group is already above the
10429 : * domain average load.
10430 : */
10431 : if (local->avg_load >= sds.avg_load)
10432 : goto out_balanced;
10433 :
10434 : /*
10435 : * If the busiest group is more loaded, use imbalance_pct to be
10436 : * conservative.
10437 : */
10438 : if (100 * busiest->avg_load <=
10439 : env->sd->imbalance_pct * local->avg_load)
10440 : goto out_balanced;
10441 : }
10442 :
10443 : /* Try to move all excess tasks to child's sibling domain */
10444 : if (sds.prefer_sibling && local->group_type == group_has_spare &&
10445 : busiest->sum_nr_running > local->sum_nr_running + 1)
10446 : goto force_balance;
10447 :
10448 : if (busiest->group_type != group_overloaded) {
10449 : if (env->idle == CPU_NOT_IDLE)
10450 : /*
10451 : * If the busiest group is not overloaded (and as a
10452 : * result the local one too) but this CPU is already
10453 : * busy, let another idle CPU try to pull task.
10454 : */
10455 : goto out_balanced;
10456 :
10457 : if (busiest->group_weight > 1 &&
10458 : local->idle_cpus <= (busiest->idle_cpus + 1))
10459 : /*
10460 : * If the busiest group is not overloaded
10461 : * and there is no imbalance between this and busiest
10462 : * group wrt idle CPUs, it is balanced. The imbalance
10463 : * becomes significant if the diff is greater than 1
10464 : * otherwise we might end up to just move the imbalance
10465 : * on another group. Of course this applies only if
10466 : * there is more than 1 CPU per group.
10467 : */
10468 : goto out_balanced;
10469 :
10470 : if (busiest->sum_h_nr_running == 1)
10471 : /*
10472 : * busiest doesn't have any tasks waiting to run
10473 : */
10474 : goto out_balanced;
10475 : }
10476 :
10477 : force_balance:
10478 : /* Looks like there is an imbalance. Compute it */
10479 : calculate_imbalance(env, &sds);
10480 : return env->imbalance ? sds.busiest : NULL;
10481 :
10482 : out_balanced:
10483 : env->imbalance = 0;
10484 : return NULL;
10485 : }
10486 :
10487 : /*
10488 : * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10489 : */
10490 : static struct rq *find_busiest_queue(struct lb_env *env,
10491 : struct sched_group *group)
10492 : {
10493 : struct rq *busiest = NULL, *rq;
10494 : unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10495 : unsigned int busiest_nr = 0;
10496 : int i;
10497 :
10498 : for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10499 : unsigned long capacity, load, util;
10500 : unsigned int nr_running;
10501 : enum fbq_type rt;
10502 :
10503 : rq = cpu_rq(i);
10504 : rt = fbq_classify_rq(rq);
10505 :
10506 : /*
10507 : * We classify groups/runqueues into three groups:
10508 : * - regular: there are !numa tasks
10509 : * - remote: there are numa tasks that run on the 'wrong' node
10510 : * - all: there is no distinction
10511 : *
10512 : * In order to avoid migrating ideally placed numa tasks,
10513 : * ignore those when there's better options.
10514 : *
10515 : * If we ignore the actual busiest queue to migrate another
10516 : * task, the next balance pass can still reduce the busiest
10517 : * queue by moving tasks around inside the node.
10518 : *
10519 : * If we cannot move enough load due to this classification
10520 : * the next pass will adjust the group classification and
10521 : * allow migration of more tasks.
10522 : *
10523 : * Both cases only affect the total convergence complexity.
10524 : */
10525 : if (rt > env->fbq_type)
10526 : continue;
10527 :
10528 : nr_running = rq->cfs.h_nr_running;
10529 : if (!nr_running)
10530 : continue;
10531 :
10532 : capacity = capacity_of(i);
10533 :
10534 : /*
10535 : * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10536 : * eventually lead to active_balancing high->low capacity.
10537 : * Higher per-CPU capacity is considered better than balancing
10538 : * average load.
10539 : */
10540 : if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10541 : !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10542 : nr_running == 1)
10543 : continue;
10544 :
10545 : /* Make sure we only pull tasks from a CPU of lower priority */
10546 : if ((env->sd->flags & SD_ASYM_PACKING) &&
10547 : sched_asym_prefer(i, env->dst_cpu) &&
10548 : nr_running == 1)
10549 : continue;
10550 :
10551 : switch (env->migration_type) {
10552 : case migrate_load:
10553 : /*
10554 : * When comparing with load imbalance, use cpu_load()
10555 : * which is not scaled with the CPU capacity.
10556 : */
10557 : load = cpu_load(rq);
10558 :
10559 : if (nr_running == 1 && load > env->imbalance &&
10560 : !check_cpu_capacity(rq, env->sd))
10561 : break;
10562 :
10563 : /*
10564 : * For the load comparisons with the other CPUs,
10565 : * consider the cpu_load() scaled with the CPU
10566 : * capacity, so that the load can be moved away
10567 : * from the CPU that is potentially running at a
10568 : * lower capacity.
10569 : *
10570 : * Thus we're looking for max(load_i / capacity_i),
10571 : * crosswise multiplication to rid ourselves of the
10572 : * division works out to:
10573 : * load_i * capacity_j > load_j * capacity_i;
10574 : * where j is our previous maximum.
10575 : */
10576 : if (load * busiest_capacity > busiest_load * capacity) {
10577 : busiest_load = load;
10578 : busiest_capacity = capacity;
10579 : busiest = rq;
10580 : }
10581 : break;
10582 :
10583 : case migrate_util:
10584 : util = cpu_util_cfs(i);
10585 :
10586 : /*
10587 : * Don't try to pull utilization from a CPU with one
10588 : * running task. Whatever its utilization, we will fail
10589 : * detach the task.
10590 : */
10591 : if (nr_running <= 1)
10592 : continue;
10593 :
10594 : if (busiest_util < util) {
10595 : busiest_util = util;
10596 : busiest = rq;
10597 : }
10598 : break;
10599 :
10600 : case migrate_task:
10601 : if (busiest_nr < nr_running) {
10602 : busiest_nr = nr_running;
10603 : busiest = rq;
10604 : }
10605 : break;
10606 :
10607 : case migrate_misfit:
10608 : /*
10609 : * For ASYM_CPUCAPACITY domains with misfit tasks we
10610 : * simply seek the "biggest" misfit task.
10611 : */
10612 : if (rq->misfit_task_load > busiest_load) {
10613 : busiest_load = rq->misfit_task_load;
10614 : busiest = rq;
10615 : }
10616 :
10617 : break;
10618 :
10619 : }
10620 : }
10621 :
10622 : return busiest;
10623 : }
10624 :
10625 : /*
10626 : * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
10627 : * so long as it is large enough.
10628 : */
10629 : #define MAX_PINNED_INTERVAL 512
10630 :
10631 : static inline bool
10632 : asym_active_balance(struct lb_env *env)
10633 : {
10634 : /*
10635 : * ASYM_PACKING needs to force migrate tasks from busy but
10636 : * lower priority CPUs in order to pack all tasks in the
10637 : * highest priority CPUs.
10638 : */
10639 : return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
10640 : sched_asym_prefer(env->dst_cpu, env->src_cpu);
10641 : }
10642 :
10643 : static inline bool
10644 : imbalanced_active_balance(struct lb_env *env)
10645 : {
10646 : struct sched_domain *sd = env->sd;
10647 :
10648 : /*
10649 : * The imbalanced case includes the case of pinned tasks preventing a fair
10650 : * distribution of the load on the system but also the even distribution of the
10651 : * threads on a system with spare capacity
10652 : */
10653 : if ((env->migration_type == migrate_task) &&
10654 : (sd->nr_balance_failed > sd->cache_nice_tries+2))
10655 : return 1;
10656 :
10657 : return 0;
10658 : }
10659 :
10660 : static int need_active_balance(struct lb_env *env)
10661 : {
10662 : struct sched_domain *sd = env->sd;
10663 :
10664 : if (asym_active_balance(env))
10665 : return 1;
10666 :
10667 : if (imbalanced_active_balance(env))
10668 : return 1;
10669 :
10670 : /*
10671 : * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10672 : * It's worth migrating the task if the src_cpu's capacity is reduced
10673 : * because of other sched_class or IRQs if more capacity stays
10674 : * available on dst_cpu.
10675 : */
10676 : if ((env->idle != CPU_NOT_IDLE) &&
10677 : (env->src_rq->cfs.h_nr_running == 1)) {
10678 : if ((check_cpu_capacity(env->src_rq, sd)) &&
10679 : (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10680 : return 1;
10681 : }
10682 :
10683 : if (env->migration_type == migrate_misfit)
10684 : return 1;
10685 :
10686 : return 0;
10687 : }
10688 :
10689 : static int active_load_balance_cpu_stop(void *data);
10690 :
10691 : static int should_we_balance(struct lb_env *env)
10692 : {
10693 : struct sched_group *sg = env->sd->groups;
10694 : int cpu;
10695 :
10696 : /*
10697 : * Ensure the balancing environment is consistent; can happen
10698 : * when the softirq triggers 'during' hotplug.
10699 : */
10700 : if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10701 : return 0;
10702 :
10703 : /*
10704 : * In the newly idle case, we will allow all the CPUs
10705 : * to do the newly idle load balance.
10706 : *
10707 : * However, we bail out if we already have tasks or a wakeup pending,
10708 : * to optimize wakeup latency.
10709 : */
10710 : if (env->idle == CPU_NEWLY_IDLE) {
10711 : if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
10712 : return 0;
10713 : return 1;
10714 : }
10715 :
10716 : /* Try to find first idle CPU */
10717 : for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
10718 : if (!idle_cpu(cpu))
10719 : continue;
10720 :
10721 : /* Are we the first idle CPU? */
10722 : return cpu == env->dst_cpu;
10723 : }
10724 :
10725 : /* Are we the first CPU of this group ? */
10726 : return group_balance_cpu(sg) == env->dst_cpu;
10727 : }
10728 :
10729 : /*
10730 : * Check this_cpu to ensure it is balanced within domain. Attempt to move
10731 : * tasks if there is an imbalance.
10732 : */
10733 : static int load_balance(int this_cpu, struct rq *this_rq,
10734 : struct sched_domain *sd, enum cpu_idle_type idle,
10735 : int *continue_balancing)
10736 : {
10737 : int ld_moved, cur_ld_moved, active_balance = 0;
10738 : struct sched_domain *sd_parent = sd->parent;
10739 : struct sched_group *group;
10740 : struct rq *busiest;
10741 : struct rq_flags rf;
10742 : struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
10743 : struct lb_env env = {
10744 : .sd = sd,
10745 : .dst_cpu = this_cpu,
10746 : .dst_rq = this_rq,
10747 : .dst_grpmask = sched_group_span(sd->groups),
10748 : .idle = idle,
10749 : .loop_break = SCHED_NR_MIGRATE_BREAK,
10750 : .cpus = cpus,
10751 : .fbq_type = all,
10752 : .tasks = LIST_HEAD_INIT(env.tasks),
10753 : };
10754 :
10755 : cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
10756 :
10757 : schedstat_inc(sd->lb_count[idle]);
10758 :
10759 : redo:
10760 : if (!should_we_balance(&env)) {
10761 : *continue_balancing = 0;
10762 : goto out_balanced;
10763 : }
10764 :
10765 : group = find_busiest_group(&env);
10766 : if (!group) {
10767 : schedstat_inc(sd->lb_nobusyg[idle]);
10768 : goto out_balanced;
10769 : }
10770 :
10771 : busiest = find_busiest_queue(&env, group);
10772 : if (!busiest) {
10773 : schedstat_inc(sd->lb_nobusyq[idle]);
10774 : goto out_balanced;
10775 : }
10776 :
10777 : WARN_ON_ONCE(busiest == env.dst_rq);
10778 :
10779 : schedstat_add(sd->lb_imbalance[idle], env.imbalance);
10780 :
10781 : env.src_cpu = busiest->cpu;
10782 : env.src_rq = busiest;
10783 :
10784 : ld_moved = 0;
10785 : /* Clear this flag as soon as we find a pullable task */
10786 : env.flags |= LBF_ALL_PINNED;
10787 : if (busiest->nr_running > 1) {
10788 : /*
10789 : * Attempt to move tasks. If find_busiest_group has found
10790 : * an imbalance but busiest->nr_running <= 1, the group is
10791 : * still unbalanced. ld_moved simply stays zero, so it is
10792 : * correctly treated as an imbalance.
10793 : */
10794 : env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
10795 :
10796 : more_balance:
10797 : rq_lock_irqsave(busiest, &rf);
10798 : update_rq_clock(busiest);
10799 :
10800 : /*
10801 : * cur_ld_moved - load moved in current iteration
10802 : * ld_moved - cumulative load moved across iterations
10803 : */
10804 : cur_ld_moved = detach_tasks(&env);
10805 :
10806 : /*
10807 : * We've detached some tasks from busiest_rq. Every
10808 : * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
10809 : * unlock busiest->lock, and we are able to be sure
10810 : * that nobody can manipulate the tasks in parallel.
10811 : * See task_rq_lock() family for the details.
10812 : */
10813 :
10814 : rq_unlock(busiest, &rf);
10815 :
10816 : if (cur_ld_moved) {
10817 : attach_tasks(&env);
10818 : ld_moved += cur_ld_moved;
10819 : }
10820 :
10821 : local_irq_restore(rf.flags);
10822 :
10823 : if (env.flags & LBF_NEED_BREAK) {
10824 : env.flags &= ~LBF_NEED_BREAK;
10825 : /* Stop if we tried all running tasks */
10826 : if (env.loop < busiest->nr_running)
10827 : goto more_balance;
10828 : }
10829 :
10830 : /*
10831 : * Revisit (affine) tasks on src_cpu that couldn't be moved to
10832 : * us and move them to an alternate dst_cpu in our sched_group
10833 : * where they can run. The upper limit on how many times we
10834 : * iterate on same src_cpu is dependent on number of CPUs in our
10835 : * sched_group.
10836 : *
10837 : * This changes load balance semantics a bit on who can move
10838 : * load to a given_cpu. In addition to the given_cpu itself
10839 : * (or a ilb_cpu acting on its behalf where given_cpu is
10840 : * nohz-idle), we now have balance_cpu in a position to move
10841 : * load to given_cpu. In rare situations, this may cause
10842 : * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
10843 : * _independently_ and at _same_ time to move some load to
10844 : * given_cpu) causing excess load to be moved to given_cpu.
10845 : * This however should not happen so much in practice and
10846 : * moreover subsequent load balance cycles should correct the
10847 : * excess load moved.
10848 : */
10849 : if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
10850 :
10851 : /* Prevent to re-select dst_cpu via env's CPUs */
10852 : __cpumask_clear_cpu(env.dst_cpu, env.cpus);
10853 :
10854 : env.dst_rq = cpu_rq(env.new_dst_cpu);
10855 : env.dst_cpu = env.new_dst_cpu;
10856 : env.flags &= ~LBF_DST_PINNED;
10857 : env.loop = 0;
10858 : env.loop_break = SCHED_NR_MIGRATE_BREAK;
10859 :
10860 : /*
10861 : * Go back to "more_balance" rather than "redo" since we
10862 : * need to continue with same src_cpu.
10863 : */
10864 : goto more_balance;
10865 : }
10866 :
10867 : /*
10868 : * We failed to reach balance because of affinity.
10869 : */
10870 : if (sd_parent) {
10871 : int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10872 :
10873 : if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
10874 : *group_imbalance = 1;
10875 : }
10876 :
10877 : /* All tasks on this runqueue were pinned by CPU affinity */
10878 : if (unlikely(env.flags & LBF_ALL_PINNED)) {
10879 : __cpumask_clear_cpu(cpu_of(busiest), cpus);
10880 : /*
10881 : * Attempting to continue load balancing at the current
10882 : * sched_domain level only makes sense if there are
10883 : * active CPUs remaining as possible busiest CPUs to
10884 : * pull load from which are not contained within the
10885 : * destination group that is receiving any migrated
10886 : * load.
10887 : */
10888 : if (!cpumask_subset(cpus, env.dst_grpmask)) {
10889 : env.loop = 0;
10890 : env.loop_break = SCHED_NR_MIGRATE_BREAK;
10891 : goto redo;
10892 : }
10893 : goto out_all_pinned;
10894 : }
10895 : }
10896 :
10897 : if (!ld_moved) {
10898 : schedstat_inc(sd->lb_failed[idle]);
10899 : /*
10900 : * Increment the failure counter only on periodic balance.
10901 : * We do not want newidle balance, which can be very
10902 : * frequent, pollute the failure counter causing
10903 : * excessive cache_hot migrations and active balances.
10904 : */
10905 : if (idle != CPU_NEWLY_IDLE)
10906 : sd->nr_balance_failed++;
10907 :
10908 : if (need_active_balance(&env)) {
10909 : unsigned long flags;
10910 :
10911 : raw_spin_rq_lock_irqsave(busiest, flags);
10912 :
10913 : /*
10914 : * Don't kick the active_load_balance_cpu_stop,
10915 : * if the curr task on busiest CPU can't be
10916 : * moved to this_cpu:
10917 : */
10918 : if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10919 : raw_spin_rq_unlock_irqrestore(busiest, flags);
10920 : goto out_one_pinned;
10921 : }
10922 :
10923 : /* Record that we found at least one task that could run on this_cpu */
10924 : env.flags &= ~LBF_ALL_PINNED;
10925 :
10926 : /*
10927 : * ->active_balance synchronizes accesses to
10928 : * ->active_balance_work. Once set, it's cleared
10929 : * only after active load balance is finished.
10930 : */
10931 : if (!busiest->active_balance) {
10932 : busiest->active_balance = 1;
10933 : busiest->push_cpu = this_cpu;
10934 : active_balance = 1;
10935 : }
10936 : raw_spin_rq_unlock_irqrestore(busiest, flags);
10937 :
10938 : if (active_balance) {
10939 : stop_one_cpu_nowait(cpu_of(busiest),
10940 : active_load_balance_cpu_stop, busiest,
10941 : &busiest->active_balance_work);
10942 : }
10943 : }
10944 : } else {
10945 : sd->nr_balance_failed = 0;
10946 : }
10947 :
10948 : if (likely(!active_balance) || need_active_balance(&env)) {
10949 : /* We were unbalanced, so reset the balancing interval */
10950 : sd->balance_interval = sd->min_interval;
10951 : }
10952 :
10953 : goto out;
10954 :
10955 : out_balanced:
10956 : /*
10957 : * We reach balance although we may have faced some affinity
10958 : * constraints. Clear the imbalance flag only if other tasks got
10959 : * a chance to move and fix the imbalance.
10960 : */
10961 : if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10962 : int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10963 :
10964 : if (*group_imbalance)
10965 : *group_imbalance = 0;
10966 : }
10967 :
10968 : out_all_pinned:
10969 : /*
10970 : * We reach balance because all tasks are pinned at this level so
10971 : * we can't migrate them. Let the imbalance flag set so parent level
10972 : * can try to migrate them.
10973 : */
10974 : schedstat_inc(sd->lb_balanced[idle]);
10975 :
10976 : sd->nr_balance_failed = 0;
10977 :
10978 : out_one_pinned:
10979 : ld_moved = 0;
10980 :
10981 : /*
10982 : * newidle_balance() disregards balance intervals, so we could
10983 : * repeatedly reach this code, which would lead to balance_interval
10984 : * skyrocketing in a short amount of time. Skip the balance_interval
10985 : * increase logic to avoid that.
10986 : */
10987 : if (env.idle == CPU_NEWLY_IDLE)
10988 : goto out;
10989 :
10990 : /* tune up the balancing interval */
10991 : if ((env.flags & LBF_ALL_PINNED &&
10992 : sd->balance_interval < MAX_PINNED_INTERVAL) ||
10993 : sd->balance_interval < sd->max_interval)
10994 : sd->balance_interval *= 2;
10995 : out:
10996 : return ld_moved;
10997 : }
10998 :
10999 : static inline unsigned long
11000 : get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11001 : {
11002 : unsigned long interval = sd->balance_interval;
11003 :
11004 : if (cpu_busy)
11005 : interval *= sd->busy_factor;
11006 :
11007 : /* scale ms to jiffies */
11008 : interval = msecs_to_jiffies(interval);
11009 :
11010 : /*
11011 : * Reduce likelihood of busy balancing at higher domains racing with
11012 : * balancing at lower domains by preventing their balancing periods
11013 : * from being multiples of each other.
11014 : */
11015 : if (cpu_busy)
11016 : interval -= 1;
11017 :
11018 : interval = clamp(interval, 1UL, max_load_balance_interval);
11019 :
11020 : return interval;
11021 : }
11022 :
11023 : static inline void
11024 : update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11025 : {
11026 : unsigned long interval, next;
11027 :
11028 : /* used by idle balance, so cpu_busy = 0 */
11029 : interval = get_sd_balance_interval(sd, 0);
11030 : next = sd->last_balance + interval;
11031 :
11032 : if (time_after(*next_balance, next))
11033 : *next_balance = next;
11034 : }
11035 :
11036 : /*
11037 : * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11038 : * running tasks off the busiest CPU onto idle CPUs. It requires at
11039 : * least 1 task to be running on each physical CPU where possible, and
11040 : * avoids physical / logical imbalances.
11041 : */
11042 : static int active_load_balance_cpu_stop(void *data)
11043 : {
11044 : struct rq *busiest_rq = data;
11045 : int busiest_cpu = cpu_of(busiest_rq);
11046 : int target_cpu = busiest_rq->push_cpu;
11047 : struct rq *target_rq = cpu_rq(target_cpu);
11048 : struct sched_domain *sd;
11049 : struct task_struct *p = NULL;
11050 : struct rq_flags rf;
11051 :
11052 : rq_lock_irq(busiest_rq, &rf);
11053 : /*
11054 : * Between queueing the stop-work and running it is a hole in which
11055 : * CPUs can become inactive. We should not move tasks from or to
11056 : * inactive CPUs.
11057 : */
11058 : if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11059 : goto out_unlock;
11060 :
11061 : /* Make sure the requested CPU hasn't gone down in the meantime: */
11062 : if (unlikely(busiest_cpu != smp_processor_id() ||
11063 : !busiest_rq->active_balance))
11064 : goto out_unlock;
11065 :
11066 : /* Is there any task to move? */
11067 : if (busiest_rq->nr_running <= 1)
11068 : goto out_unlock;
11069 :
11070 : /*
11071 : * This condition is "impossible", if it occurs
11072 : * we need to fix it. Originally reported by
11073 : * Bjorn Helgaas on a 128-CPU setup.
11074 : */
11075 : WARN_ON_ONCE(busiest_rq == target_rq);
11076 :
11077 : /* Search for an sd spanning us and the target CPU. */
11078 : rcu_read_lock();
11079 : for_each_domain(target_cpu, sd) {
11080 : if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11081 : break;
11082 : }
11083 :
11084 : if (likely(sd)) {
11085 : struct lb_env env = {
11086 : .sd = sd,
11087 : .dst_cpu = target_cpu,
11088 : .dst_rq = target_rq,
11089 : .src_cpu = busiest_rq->cpu,
11090 : .src_rq = busiest_rq,
11091 : .idle = CPU_IDLE,
11092 : .flags = LBF_ACTIVE_LB,
11093 : };
11094 :
11095 : schedstat_inc(sd->alb_count);
11096 : update_rq_clock(busiest_rq);
11097 :
11098 : p = detach_one_task(&env);
11099 : if (p) {
11100 : schedstat_inc(sd->alb_pushed);
11101 : /* Active balancing done, reset the failure counter. */
11102 : sd->nr_balance_failed = 0;
11103 : } else {
11104 : schedstat_inc(sd->alb_failed);
11105 : }
11106 : }
11107 : rcu_read_unlock();
11108 : out_unlock:
11109 : busiest_rq->active_balance = 0;
11110 : rq_unlock(busiest_rq, &rf);
11111 :
11112 : if (p)
11113 : attach_one_task(target_rq, p);
11114 :
11115 : local_irq_enable();
11116 :
11117 : return 0;
11118 : }
11119 :
11120 : static DEFINE_SPINLOCK(balancing);
11121 :
11122 : /*
11123 : * Scale the max load_balance interval with the number of CPUs in the system.
11124 : * This trades load-balance latency on larger machines for less cross talk.
11125 : */
11126 : void update_max_interval(void)
11127 : {
11128 : max_load_balance_interval = HZ*num_online_cpus()/10;
11129 : }
11130 :
11131 : static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11132 : {
11133 : if (cost > sd->max_newidle_lb_cost) {
11134 : /*
11135 : * Track max cost of a domain to make sure to not delay the
11136 : * next wakeup on the CPU.
11137 : */
11138 : sd->max_newidle_lb_cost = cost;
11139 : sd->last_decay_max_lb_cost = jiffies;
11140 : } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11141 : /*
11142 : * Decay the newidle max times by ~1% per second to ensure that
11143 : * it is not outdated and the current max cost is actually
11144 : * shorter.
11145 : */
11146 : sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11147 : sd->last_decay_max_lb_cost = jiffies;
11148 :
11149 : return true;
11150 : }
11151 :
11152 : return false;
11153 : }
11154 :
11155 : /*
11156 : * It checks each scheduling domain to see if it is due to be balanced,
11157 : * and initiates a balancing operation if so.
11158 : *
11159 : * Balancing parameters are set up in init_sched_domains.
11160 : */
11161 : static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11162 : {
11163 : int continue_balancing = 1;
11164 : int cpu = rq->cpu;
11165 : int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11166 : unsigned long interval;
11167 : struct sched_domain *sd;
11168 : /* Earliest time when we have to do rebalance again */
11169 : unsigned long next_balance = jiffies + 60*HZ;
11170 : int update_next_balance = 0;
11171 : int need_serialize, need_decay = 0;
11172 : u64 max_cost = 0;
11173 :
11174 : rcu_read_lock();
11175 : for_each_domain(cpu, sd) {
11176 : /*
11177 : * Decay the newidle max times here because this is a regular
11178 : * visit to all the domains.
11179 : */
11180 : need_decay = update_newidle_cost(sd, 0);
11181 : max_cost += sd->max_newidle_lb_cost;
11182 :
11183 : /*
11184 : * Stop the load balance at this level. There is another
11185 : * CPU in our sched group which is doing load balancing more
11186 : * actively.
11187 : */
11188 : if (!continue_balancing) {
11189 : if (need_decay)
11190 : continue;
11191 : break;
11192 : }
11193 :
11194 : interval = get_sd_balance_interval(sd, busy);
11195 :
11196 : need_serialize = sd->flags & SD_SERIALIZE;
11197 : if (need_serialize) {
11198 : if (!spin_trylock(&balancing))
11199 : goto out;
11200 : }
11201 :
11202 : if (time_after_eq(jiffies, sd->last_balance + interval)) {
11203 : if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11204 : /*
11205 : * The LBF_DST_PINNED logic could have changed
11206 : * env->dst_cpu, so we can't know our idle
11207 : * state even if we migrated tasks. Update it.
11208 : */
11209 : idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11210 : busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11211 : }
11212 : sd->last_balance = jiffies;
11213 : interval = get_sd_balance_interval(sd, busy);
11214 : }
11215 : if (need_serialize)
11216 : spin_unlock(&balancing);
11217 : out:
11218 : if (time_after(next_balance, sd->last_balance + interval)) {
11219 : next_balance = sd->last_balance + interval;
11220 : update_next_balance = 1;
11221 : }
11222 : }
11223 : if (need_decay) {
11224 : /*
11225 : * Ensure the rq-wide value also decays but keep it at a
11226 : * reasonable floor to avoid funnies with rq->avg_idle.
11227 : */
11228 : rq->max_idle_balance_cost =
11229 : max((u64)sysctl_sched_migration_cost, max_cost);
11230 : }
11231 : rcu_read_unlock();
11232 :
11233 : /*
11234 : * next_balance will be updated only when there is a need.
11235 : * When the cpu is attached to null domain for ex, it will not be
11236 : * updated.
11237 : */
11238 : if (likely(update_next_balance))
11239 : rq->next_balance = next_balance;
11240 :
11241 : }
11242 :
11243 : static inline int on_null_domain(struct rq *rq)
11244 : {
11245 : return unlikely(!rcu_dereference_sched(rq->sd));
11246 : }
11247 :
11248 : #ifdef CONFIG_NO_HZ_COMMON
11249 : /*
11250 : * idle load balancing details
11251 : * - When one of the busy CPUs notice that there may be an idle rebalancing
11252 : * needed, they will kick the idle load balancer, which then does idle
11253 : * load balancing for all the idle CPUs.
11254 : * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11255 : * anywhere yet.
11256 : */
11257 :
11258 : static inline int find_new_ilb(void)
11259 : {
11260 : int ilb;
11261 : const struct cpumask *hk_mask;
11262 :
11263 : hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11264 :
11265 : for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11266 :
11267 : if (ilb == smp_processor_id())
11268 : continue;
11269 :
11270 : if (idle_cpu(ilb))
11271 : return ilb;
11272 : }
11273 :
11274 : return nr_cpu_ids;
11275 : }
11276 :
11277 : /*
11278 : * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11279 : * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11280 : */
11281 : static void kick_ilb(unsigned int flags)
11282 : {
11283 : int ilb_cpu;
11284 :
11285 : /*
11286 : * Increase nohz.next_balance only when if full ilb is triggered but
11287 : * not if we only update stats.
11288 : */
11289 : if (flags & NOHZ_BALANCE_KICK)
11290 : nohz.next_balance = jiffies+1;
11291 :
11292 : ilb_cpu = find_new_ilb();
11293 :
11294 : if (ilb_cpu >= nr_cpu_ids)
11295 : return;
11296 :
11297 : /*
11298 : * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11299 : * the first flag owns it; cleared by nohz_csd_func().
11300 : */
11301 : flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11302 : if (flags & NOHZ_KICK_MASK)
11303 : return;
11304 :
11305 : /*
11306 : * This way we generate an IPI on the target CPU which
11307 : * is idle. And the softirq performing nohz idle load balance
11308 : * will be run before returning from the IPI.
11309 : */
11310 : smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11311 : }
11312 :
11313 : /*
11314 : * Current decision point for kicking the idle load balancer in the presence
11315 : * of idle CPUs in the system.
11316 : */
11317 : static void nohz_balancer_kick(struct rq *rq)
11318 : {
11319 : unsigned long now = jiffies;
11320 : struct sched_domain_shared *sds;
11321 : struct sched_domain *sd;
11322 : int nr_busy, i, cpu = rq->cpu;
11323 : unsigned int flags = 0;
11324 :
11325 : if (unlikely(rq->idle_balance))
11326 : return;
11327 :
11328 : /*
11329 : * We may be recently in ticked or tickless idle mode. At the first
11330 : * busy tick after returning from idle, we will update the busy stats.
11331 : */
11332 : nohz_balance_exit_idle(rq);
11333 :
11334 : /*
11335 : * None are in tickless mode and hence no need for NOHZ idle load
11336 : * balancing.
11337 : */
11338 : if (likely(!atomic_read(&nohz.nr_cpus)))
11339 : return;
11340 :
11341 : if (READ_ONCE(nohz.has_blocked) &&
11342 : time_after(now, READ_ONCE(nohz.next_blocked)))
11343 : flags = NOHZ_STATS_KICK;
11344 :
11345 : if (time_before(now, nohz.next_balance))
11346 : goto out;
11347 :
11348 : if (rq->nr_running >= 2) {
11349 : flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11350 : goto out;
11351 : }
11352 :
11353 : rcu_read_lock();
11354 :
11355 : sd = rcu_dereference(rq->sd);
11356 : if (sd) {
11357 : /*
11358 : * If there's a CFS task and the current CPU has reduced
11359 : * capacity; kick the ILB to see if there's a better CPU to run
11360 : * on.
11361 : */
11362 : if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11363 : flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11364 : goto unlock;
11365 : }
11366 : }
11367 :
11368 : sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11369 : if (sd) {
11370 : /*
11371 : * When ASYM_PACKING; see if there's a more preferred CPU
11372 : * currently idle; in which case, kick the ILB to move tasks
11373 : * around.
11374 : */
11375 : for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11376 : if (sched_asym_prefer(i, cpu)) {
11377 : flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11378 : goto unlock;
11379 : }
11380 : }
11381 : }
11382 :
11383 : sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11384 : if (sd) {
11385 : /*
11386 : * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11387 : * to run the misfit task on.
11388 : */
11389 : if (check_misfit_status(rq, sd)) {
11390 : flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11391 : goto unlock;
11392 : }
11393 :
11394 : /*
11395 : * For asymmetric systems, we do not want to nicely balance
11396 : * cache use, instead we want to embrace asymmetry and only
11397 : * ensure tasks have enough CPU capacity.
11398 : *
11399 : * Skip the LLC logic because it's not relevant in that case.
11400 : */
11401 : goto unlock;
11402 : }
11403 :
11404 : sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11405 : if (sds) {
11406 : /*
11407 : * If there is an imbalance between LLC domains (IOW we could
11408 : * increase the overall cache use), we need some less-loaded LLC
11409 : * domain to pull some load. Likewise, we may need to spread
11410 : * load within the current LLC domain (e.g. packed SMT cores but
11411 : * other CPUs are idle). We can't really know from here how busy
11412 : * the others are - so just get a nohz balance going if it looks
11413 : * like this LLC domain has tasks we could move.
11414 : */
11415 : nr_busy = atomic_read(&sds->nr_busy_cpus);
11416 : if (nr_busy > 1) {
11417 : flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11418 : goto unlock;
11419 : }
11420 : }
11421 : unlock:
11422 : rcu_read_unlock();
11423 : out:
11424 : if (READ_ONCE(nohz.needs_update))
11425 : flags |= NOHZ_NEXT_KICK;
11426 :
11427 : if (flags)
11428 : kick_ilb(flags);
11429 : }
11430 :
11431 : static void set_cpu_sd_state_busy(int cpu)
11432 : {
11433 : struct sched_domain *sd;
11434 :
11435 : rcu_read_lock();
11436 : sd = rcu_dereference(per_cpu(sd_llc, cpu));
11437 :
11438 : if (!sd || !sd->nohz_idle)
11439 : goto unlock;
11440 : sd->nohz_idle = 0;
11441 :
11442 : atomic_inc(&sd->shared->nr_busy_cpus);
11443 : unlock:
11444 : rcu_read_unlock();
11445 : }
11446 :
11447 : void nohz_balance_exit_idle(struct rq *rq)
11448 : {
11449 : SCHED_WARN_ON(rq != this_rq());
11450 :
11451 : if (likely(!rq->nohz_tick_stopped))
11452 : return;
11453 :
11454 : rq->nohz_tick_stopped = 0;
11455 : cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11456 : atomic_dec(&nohz.nr_cpus);
11457 :
11458 : set_cpu_sd_state_busy(rq->cpu);
11459 : }
11460 :
11461 : static void set_cpu_sd_state_idle(int cpu)
11462 : {
11463 : struct sched_domain *sd;
11464 :
11465 : rcu_read_lock();
11466 : sd = rcu_dereference(per_cpu(sd_llc, cpu));
11467 :
11468 : if (!sd || sd->nohz_idle)
11469 : goto unlock;
11470 : sd->nohz_idle = 1;
11471 :
11472 : atomic_dec(&sd->shared->nr_busy_cpus);
11473 : unlock:
11474 : rcu_read_unlock();
11475 : }
11476 :
11477 : /*
11478 : * This routine will record that the CPU is going idle with tick stopped.
11479 : * This info will be used in performing idle load balancing in the future.
11480 : */
11481 : void nohz_balance_enter_idle(int cpu)
11482 : {
11483 : struct rq *rq = cpu_rq(cpu);
11484 :
11485 : SCHED_WARN_ON(cpu != smp_processor_id());
11486 :
11487 : /* If this CPU is going down, then nothing needs to be done: */
11488 : if (!cpu_active(cpu))
11489 : return;
11490 :
11491 : /* Spare idle load balancing on CPUs that don't want to be disturbed: */
11492 : if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11493 : return;
11494 :
11495 : /*
11496 : * Can be set safely without rq->lock held
11497 : * If a clear happens, it will have evaluated last additions because
11498 : * rq->lock is held during the check and the clear
11499 : */
11500 : rq->has_blocked_load = 1;
11501 :
11502 : /*
11503 : * The tick is still stopped but load could have been added in the
11504 : * meantime. We set the nohz.has_blocked flag to trig a check of the
11505 : * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11506 : * of nohz.has_blocked can only happen after checking the new load
11507 : */
11508 : if (rq->nohz_tick_stopped)
11509 : goto out;
11510 :
11511 : /* If we're a completely isolated CPU, we don't play: */
11512 : if (on_null_domain(rq))
11513 : return;
11514 :
11515 : rq->nohz_tick_stopped = 1;
11516 :
11517 : cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11518 : atomic_inc(&nohz.nr_cpus);
11519 :
11520 : /*
11521 : * Ensures that if nohz_idle_balance() fails to observe our
11522 : * @idle_cpus_mask store, it must observe the @has_blocked
11523 : * and @needs_update stores.
11524 : */
11525 : smp_mb__after_atomic();
11526 :
11527 : set_cpu_sd_state_idle(cpu);
11528 :
11529 : WRITE_ONCE(nohz.needs_update, 1);
11530 : out:
11531 : /*
11532 : * Each time a cpu enter idle, we assume that it has blocked load and
11533 : * enable the periodic update of the load of idle cpus
11534 : */
11535 : WRITE_ONCE(nohz.has_blocked, 1);
11536 : }
11537 :
11538 : static bool update_nohz_stats(struct rq *rq)
11539 : {
11540 : unsigned int cpu = rq->cpu;
11541 :
11542 : if (!rq->has_blocked_load)
11543 : return false;
11544 :
11545 : if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11546 : return false;
11547 :
11548 : if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11549 : return true;
11550 :
11551 : update_blocked_averages(cpu);
11552 :
11553 : return rq->has_blocked_load;
11554 : }
11555 :
11556 : /*
11557 : * Internal function that runs load balance for all idle cpus. The load balance
11558 : * can be a simple update of blocked load or a complete load balance with
11559 : * tasks movement depending of flags.
11560 : */
11561 : static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11562 : {
11563 : /* Earliest time when we have to do rebalance again */
11564 : unsigned long now = jiffies;
11565 : unsigned long next_balance = now + 60*HZ;
11566 : bool has_blocked_load = false;
11567 : int update_next_balance = 0;
11568 : int this_cpu = this_rq->cpu;
11569 : int balance_cpu;
11570 : struct rq *rq;
11571 :
11572 : SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11573 :
11574 : /*
11575 : * We assume there will be no idle load after this update and clear
11576 : * the has_blocked flag. If a cpu enters idle in the mean time, it will
11577 : * set the has_blocked flag and trigger another update of idle load.
11578 : * Because a cpu that becomes idle, is added to idle_cpus_mask before
11579 : * setting the flag, we are sure to not clear the state and not
11580 : * check the load of an idle cpu.
11581 : *
11582 : * Same applies to idle_cpus_mask vs needs_update.
11583 : */
11584 : if (flags & NOHZ_STATS_KICK)
11585 : WRITE_ONCE(nohz.has_blocked, 0);
11586 : if (flags & NOHZ_NEXT_KICK)
11587 : WRITE_ONCE(nohz.needs_update, 0);
11588 :
11589 : /*
11590 : * Ensures that if we miss the CPU, we must see the has_blocked
11591 : * store from nohz_balance_enter_idle().
11592 : */
11593 : smp_mb();
11594 :
11595 : /*
11596 : * Start with the next CPU after this_cpu so we will end with this_cpu and let a
11597 : * chance for other idle cpu to pull load.
11598 : */
11599 : for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
11600 : if (!idle_cpu(balance_cpu))
11601 : continue;
11602 :
11603 : /*
11604 : * If this CPU gets work to do, stop the load balancing
11605 : * work being done for other CPUs. Next load
11606 : * balancing owner will pick it up.
11607 : */
11608 : if (need_resched()) {
11609 : if (flags & NOHZ_STATS_KICK)
11610 : has_blocked_load = true;
11611 : if (flags & NOHZ_NEXT_KICK)
11612 : WRITE_ONCE(nohz.needs_update, 1);
11613 : goto abort;
11614 : }
11615 :
11616 : rq = cpu_rq(balance_cpu);
11617 :
11618 : if (flags & NOHZ_STATS_KICK)
11619 : has_blocked_load |= update_nohz_stats(rq);
11620 :
11621 : /*
11622 : * If time for next balance is due,
11623 : * do the balance.
11624 : */
11625 : if (time_after_eq(jiffies, rq->next_balance)) {
11626 : struct rq_flags rf;
11627 :
11628 : rq_lock_irqsave(rq, &rf);
11629 : update_rq_clock(rq);
11630 : rq_unlock_irqrestore(rq, &rf);
11631 :
11632 : if (flags & NOHZ_BALANCE_KICK)
11633 : rebalance_domains(rq, CPU_IDLE);
11634 : }
11635 :
11636 : if (time_after(next_balance, rq->next_balance)) {
11637 : next_balance = rq->next_balance;
11638 : update_next_balance = 1;
11639 : }
11640 : }
11641 :
11642 : /*
11643 : * next_balance will be updated only when there is a need.
11644 : * When the CPU is attached to null domain for ex, it will not be
11645 : * updated.
11646 : */
11647 : if (likely(update_next_balance))
11648 : nohz.next_balance = next_balance;
11649 :
11650 : if (flags & NOHZ_STATS_KICK)
11651 : WRITE_ONCE(nohz.next_blocked,
11652 : now + msecs_to_jiffies(LOAD_AVG_PERIOD));
11653 :
11654 : abort:
11655 : /* There is still blocked load, enable periodic update */
11656 : if (has_blocked_load)
11657 : WRITE_ONCE(nohz.has_blocked, 1);
11658 : }
11659 :
11660 : /*
11661 : * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
11662 : * rebalancing for all the cpus for whom scheduler ticks are stopped.
11663 : */
11664 : static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11665 : {
11666 : unsigned int flags = this_rq->nohz_idle_balance;
11667 :
11668 : if (!flags)
11669 : return false;
11670 :
11671 : this_rq->nohz_idle_balance = 0;
11672 :
11673 : if (idle != CPU_IDLE)
11674 : return false;
11675 :
11676 : _nohz_idle_balance(this_rq, flags);
11677 :
11678 : return true;
11679 : }
11680 :
11681 : /*
11682 : * Check if we need to run the ILB for updating blocked load before entering
11683 : * idle state.
11684 : */
11685 : void nohz_run_idle_balance(int cpu)
11686 : {
11687 : unsigned int flags;
11688 :
11689 : flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
11690 :
11691 : /*
11692 : * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
11693 : * (ie NOHZ_STATS_KICK set) and will do the same.
11694 : */
11695 : if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
11696 : _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
11697 : }
11698 :
11699 : static void nohz_newidle_balance(struct rq *this_rq)
11700 : {
11701 : int this_cpu = this_rq->cpu;
11702 :
11703 : /*
11704 : * This CPU doesn't want to be disturbed by scheduler
11705 : * housekeeping
11706 : */
11707 : if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
11708 : return;
11709 :
11710 : /* Will wake up very soon. No time for doing anything else*/
11711 : if (this_rq->avg_idle < sysctl_sched_migration_cost)
11712 : return;
11713 :
11714 : /* Don't need to update blocked load of idle CPUs*/
11715 : if (!READ_ONCE(nohz.has_blocked) ||
11716 : time_before(jiffies, READ_ONCE(nohz.next_blocked)))
11717 : return;
11718 :
11719 : /*
11720 : * Set the need to trigger ILB in order to update blocked load
11721 : * before entering idle state.
11722 : */
11723 : atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
11724 : }
11725 :
11726 : #else /* !CONFIG_NO_HZ_COMMON */
11727 : static inline void nohz_balancer_kick(struct rq *rq) { }
11728 :
11729 : static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11730 : {
11731 : return false;
11732 : }
11733 :
11734 : static inline void nohz_newidle_balance(struct rq *this_rq) { }
11735 : #endif /* CONFIG_NO_HZ_COMMON */
11736 :
11737 : /*
11738 : * newidle_balance is called by schedule() if this_cpu is about to become
11739 : * idle. Attempts to pull tasks from other CPUs.
11740 : *
11741 : * Returns:
11742 : * < 0 - we released the lock and there are !fair tasks present
11743 : * 0 - failed, no new tasks
11744 : * > 0 - success, new (fair) tasks present
11745 : */
11746 : static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
11747 : {
11748 : unsigned long next_balance = jiffies + HZ;
11749 : int this_cpu = this_rq->cpu;
11750 : u64 t0, t1, curr_cost = 0;
11751 : struct sched_domain *sd;
11752 : int pulled_task = 0;
11753 :
11754 : update_misfit_status(NULL, this_rq);
11755 :
11756 : /*
11757 : * There is a task waiting to run. No need to search for one.
11758 : * Return 0; the task will be enqueued when switching to idle.
11759 : */
11760 : if (this_rq->ttwu_pending)
11761 : return 0;
11762 :
11763 : /*
11764 : * We must set idle_stamp _before_ calling idle_balance(), such that we
11765 : * measure the duration of idle_balance() as idle time.
11766 : */
11767 : this_rq->idle_stamp = rq_clock(this_rq);
11768 :
11769 : /*
11770 : * Do not pull tasks towards !active CPUs...
11771 : */
11772 : if (!cpu_active(this_cpu))
11773 : return 0;
11774 :
11775 : /*
11776 : * This is OK, because current is on_cpu, which avoids it being picked
11777 : * for load-balance and preemption/IRQs are still disabled avoiding
11778 : * further scheduler activity on it and we're being very careful to
11779 : * re-start the picking loop.
11780 : */
11781 : rq_unpin_lock(this_rq, rf);
11782 :
11783 : rcu_read_lock();
11784 : sd = rcu_dereference_check_sched_domain(this_rq->sd);
11785 :
11786 : if (!READ_ONCE(this_rq->rd->overload) ||
11787 : (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
11788 :
11789 : if (sd)
11790 : update_next_balance(sd, &next_balance);
11791 : rcu_read_unlock();
11792 :
11793 : goto out;
11794 : }
11795 : rcu_read_unlock();
11796 :
11797 : raw_spin_rq_unlock(this_rq);
11798 :
11799 : t0 = sched_clock_cpu(this_cpu);
11800 : update_blocked_averages(this_cpu);
11801 :
11802 : rcu_read_lock();
11803 : for_each_domain(this_cpu, sd) {
11804 : int continue_balancing = 1;
11805 : u64 domain_cost;
11806 :
11807 : update_next_balance(sd, &next_balance);
11808 :
11809 : if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
11810 : break;
11811 :
11812 : if (sd->flags & SD_BALANCE_NEWIDLE) {
11813 :
11814 : pulled_task = load_balance(this_cpu, this_rq,
11815 : sd, CPU_NEWLY_IDLE,
11816 : &continue_balancing);
11817 :
11818 : t1 = sched_clock_cpu(this_cpu);
11819 : domain_cost = t1 - t0;
11820 : update_newidle_cost(sd, domain_cost);
11821 :
11822 : curr_cost += domain_cost;
11823 : t0 = t1;
11824 : }
11825 :
11826 : /*
11827 : * Stop searching for tasks to pull if there are
11828 : * now runnable tasks on this rq.
11829 : */
11830 : if (pulled_task || this_rq->nr_running > 0 ||
11831 : this_rq->ttwu_pending)
11832 : break;
11833 : }
11834 : rcu_read_unlock();
11835 :
11836 : raw_spin_rq_lock(this_rq);
11837 :
11838 : if (curr_cost > this_rq->max_idle_balance_cost)
11839 : this_rq->max_idle_balance_cost = curr_cost;
11840 :
11841 : /*
11842 : * While browsing the domains, we released the rq lock, a task could
11843 : * have been enqueued in the meantime. Since we're not going idle,
11844 : * pretend we pulled a task.
11845 : */
11846 : if (this_rq->cfs.h_nr_running && !pulled_task)
11847 : pulled_task = 1;
11848 :
11849 : /* Is there a task of a high priority class? */
11850 : if (this_rq->nr_running != this_rq->cfs.h_nr_running)
11851 : pulled_task = -1;
11852 :
11853 : out:
11854 : /* Move the next balance forward */
11855 : if (time_after(this_rq->next_balance, next_balance))
11856 : this_rq->next_balance = next_balance;
11857 :
11858 : if (pulled_task)
11859 : this_rq->idle_stamp = 0;
11860 : else
11861 : nohz_newidle_balance(this_rq);
11862 :
11863 : rq_repin_lock(this_rq, rf);
11864 :
11865 : return pulled_task;
11866 : }
11867 :
11868 : /*
11869 : * run_rebalance_domains is triggered when needed from the scheduler tick.
11870 : * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
11871 : */
11872 : static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
11873 : {
11874 : struct rq *this_rq = this_rq();
11875 : enum cpu_idle_type idle = this_rq->idle_balance ?
11876 : CPU_IDLE : CPU_NOT_IDLE;
11877 :
11878 : /*
11879 : * If this CPU has a pending nohz_balance_kick, then do the
11880 : * balancing on behalf of the other idle CPUs whose ticks are
11881 : * stopped. Do nohz_idle_balance *before* rebalance_domains to
11882 : * give the idle CPUs a chance to load balance. Else we may
11883 : * load balance only within the local sched_domain hierarchy
11884 : * and abort nohz_idle_balance altogether if we pull some load.
11885 : */
11886 : if (nohz_idle_balance(this_rq, idle))
11887 : return;
11888 :
11889 : /* normal load balance */
11890 : update_blocked_averages(this_rq->cpu);
11891 : rebalance_domains(this_rq, idle);
11892 : }
11893 :
11894 : /*
11895 : * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
11896 : */
11897 : void trigger_load_balance(struct rq *rq)
11898 : {
11899 : /*
11900 : * Don't need to rebalance while attached to NULL domain or
11901 : * runqueue CPU is not active
11902 : */
11903 : if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
11904 : return;
11905 :
11906 : if (time_after_eq(jiffies, rq->next_balance))
11907 : raise_softirq(SCHED_SOFTIRQ);
11908 :
11909 : nohz_balancer_kick(rq);
11910 : }
11911 :
11912 : static void rq_online_fair(struct rq *rq)
11913 : {
11914 : update_sysctl();
11915 :
11916 : update_runtime_enabled(rq);
11917 : }
11918 :
11919 : static void rq_offline_fair(struct rq *rq)
11920 : {
11921 : update_sysctl();
11922 :
11923 : /* Ensure any throttled groups are reachable by pick_next_task */
11924 : unthrottle_offline_cfs_rqs(rq);
11925 : }
11926 :
11927 : #endif /* CONFIG_SMP */
11928 :
11929 : #ifdef CONFIG_SCHED_CORE
11930 : static inline bool
11931 : __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
11932 : {
11933 : u64 slice = sched_slice(cfs_rq_of(se), se);
11934 : u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
11935 :
11936 : return (rtime * min_nr_tasks > slice);
11937 : }
11938 :
11939 : #define MIN_NR_TASKS_DURING_FORCEIDLE 2
11940 : static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
11941 : {
11942 : if (!sched_core_enabled(rq))
11943 : return;
11944 :
11945 : /*
11946 : * If runqueue has only one task which used up its slice and
11947 : * if the sibling is forced idle, then trigger schedule to
11948 : * give forced idle task a chance.
11949 : *
11950 : * sched_slice() considers only this active rq and it gets the
11951 : * whole slice. But during force idle, we have siblings acting
11952 : * like a single runqueue and hence we need to consider runnable
11953 : * tasks on this CPU and the forced idle CPU. Ideally, we should
11954 : * go through the forced idle rq, but that would be a perf hit.
11955 : * We can assume that the forced idle CPU has at least
11956 : * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11957 : * if we need to give up the CPU.
11958 : */
11959 : if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
11960 : __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
11961 : resched_curr(rq);
11962 : }
11963 :
11964 : /*
11965 : * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11966 : */
11967 : static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
11968 : bool forceidle)
11969 : {
11970 : for_each_sched_entity(se) {
11971 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
11972 :
11973 : if (forceidle) {
11974 : if (cfs_rq->forceidle_seq == fi_seq)
11975 : break;
11976 : cfs_rq->forceidle_seq = fi_seq;
11977 : }
11978 :
11979 : cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
11980 : }
11981 : }
11982 :
11983 : void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
11984 : {
11985 : struct sched_entity *se = &p->se;
11986 :
11987 : if (p->sched_class != &fair_sched_class)
11988 : return;
11989 :
11990 : se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
11991 : }
11992 :
11993 : bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
11994 : bool in_fi)
11995 : {
11996 : struct rq *rq = task_rq(a);
11997 : const struct sched_entity *sea = &a->se;
11998 : const struct sched_entity *seb = &b->se;
11999 : struct cfs_rq *cfs_rqa;
12000 : struct cfs_rq *cfs_rqb;
12001 : s64 delta;
12002 :
12003 : SCHED_WARN_ON(task_rq(b)->core != rq->core);
12004 :
12005 : #ifdef CONFIG_FAIR_GROUP_SCHED
12006 : /*
12007 : * Find an se in the hierarchy for tasks a and b, such that the se's
12008 : * are immediate siblings.
12009 : */
12010 : while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12011 : int sea_depth = sea->depth;
12012 : int seb_depth = seb->depth;
12013 :
12014 : if (sea_depth >= seb_depth)
12015 : sea = parent_entity(sea);
12016 : if (sea_depth <= seb_depth)
12017 : seb = parent_entity(seb);
12018 : }
12019 :
12020 : se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12021 : se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12022 :
12023 : cfs_rqa = sea->cfs_rq;
12024 : cfs_rqb = seb->cfs_rq;
12025 : #else
12026 : cfs_rqa = &task_rq(a)->cfs;
12027 : cfs_rqb = &task_rq(b)->cfs;
12028 : #endif
12029 :
12030 : /*
12031 : * Find delta after normalizing se's vruntime with its cfs_rq's
12032 : * min_vruntime_fi, which would have been updated in prior calls
12033 : * to se_fi_update().
12034 : */
12035 : delta = (s64)(sea->vruntime - seb->vruntime) +
12036 : (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12037 :
12038 : return delta > 0;
12039 : }
12040 :
12041 : static int task_is_throttled_fair(struct task_struct *p, int cpu)
12042 : {
12043 : struct cfs_rq *cfs_rq;
12044 :
12045 : #ifdef CONFIG_FAIR_GROUP_SCHED
12046 : cfs_rq = task_group(p)->cfs_rq[cpu];
12047 : #else
12048 : cfs_rq = &cpu_rq(cpu)->cfs;
12049 : #endif
12050 : return throttled_hierarchy(cfs_rq);
12051 : }
12052 : #else
12053 : static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12054 : #endif
12055 :
12056 : /*
12057 : * scheduler tick hitting a task of our scheduling class.
12058 : *
12059 : * NOTE: This function can be called remotely by the tick offload that
12060 : * goes along full dynticks. Therefore no local assumption can be made
12061 : * and everything must be accessed through the @rq and @curr passed in
12062 : * parameters.
12063 : */
12064 2930 : static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12065 : {
12066 : struct cfs_rq *cfs_rq;
12067 2930 : struct sched_entity *se = &curr->se;
12068 :
12069 5860 : for_each_sched_entity(se) {
12070 5860 : cfs_rq = cfs_rq_of(se);
12071 2930 : entity_tick(cfs_rq, se, queued);
12072 : }
12073 :
12074 2930 : if (static_branch_unlikely(&sched_numa_balancing))
12075 : task_tick_numa(rq, curr);
12076 :
12077 2930 : update_misfit_status(curr, rq);
12078 2930 : update_overutilized_status(task_rq(curr));
12079 :
12080 2930 : task_tick_core(rq, curr);
12081 2930 : }
12082 :
12083 : /*
12084 : * called on fork with the child task as argument from the parent's context
12085 : * - child not yet on the tasklist
12086 : * - preemption disabled
12087 : */
12088 382 : static void task_fork_fair(struct task_struct *p)
12089 : {
12090 : struct cfs_rq *cfs_rq;
12091 382 : struct sched_entity *se = &p->se, *curr;
12092 382 : struct rq *rq = this_rq();
12093 : struct rq_flags rf;
12094 :
12095 764 : rq_lock(rq, &rf);
12096 382 : update_rq_clock(rq);
12097 :
12098 764 : cfs_rq = task_cfs_rq(current);
12099 382 : curr = cfs_rq->curr;
12100 382 : if (curr) {
12101 380 : update_curr(cfs_rq);
12102 380 : se->vruntime = curr->vruntime;
12103 : }
12104 382 : place_entity(cfs_rq, se, 1);
12105 :
12106 382 : if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
12107 : /*
12108 : * Upon rescheduling, sched_class::put_prev_task() will place
12109 : * 'current' within the tree based on its new key value.
12110 : */
12111 0 : swap(curr->vruntime, se->vruntime);
12112 0 : resched_curr(rq);
12113 : }
12114 :
12115 382 : se->vruntime -= cfs_rq->min_vruntime;
12116 764 : rq_unlock(rq, &rf);
12117 382 : }
12118 :
12119 : /*
12120 : * Priority of the task has changed. Check to see if we preempt
12121 : * the current task.
12122 : */
12123 : static void
12124 5 : prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12125 : {
12126 5 : if (!task_on_rq_queued(p))
12127 : return;
12128 :
12129 4 : if (rq->cfs.nr_running == 1)
12130 : return;
12131 :
12132 : /*
12133 : * Reschedule if we are currently running on this runqueue and
12134 : * our priority decreased, or if we are not currently running on
12135 : * this runqueue and our priority is higher than the current's
12136 : */
12137 4 : if (task_current(rq, p)) {
12138 4 : if (p->prio > oldprio)
12139 0 : resched_curr(rq);
12140 : } else
12141 0 : check_preempt_curr(rq, p, 0);
12142 : }
12143 :
12144 : static inline bool vruntime_normalized(struct task_struct *p)
12145 : {
12146 0 : struct sched_entity *se = &p->se;
12147 :
12148 : /*
12149 : * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
12150 : * the dequeue_entity(.flags=0) will already have normalized the
12151 : * vruntime.
12152 : */
12153 0 : if (p->on_rq)
12154 : return true;
12155 :
12156 : /*
12157 : * When !on_rq, vruntime of the task has usually NOT been normalized.
12158 : * But there are some cases where it has already been normalized:
12159 : *
12160 : * - A forked child which is waiting for being woken up by
12161 : * wake_up_new_task().
12162 : * - A task which has been woken up by try_to_wake_up() and
12163 : * waiting for actually being woken up by sched_ttwu_pending().
12164 : */
12165 0 : if (!se->sum_exec_runtime ||
12166 0 : (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
12167 : return true;
12168 :
12169 : return false;
12170 : }
12171 :
12172 : #ifdef CONFIG_FAIR_GROUP_SCHED
12173 : /*
12174 : * Propagate the changes of the sched_entity across the tg tree to make it
12175 : * visible to the root
12176 : */
12177 : static void propagate_entity_cfs_rq(struct sched_entity *se)
12178 : {
12179 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12180 :
12181 : if (cfs_rq_throttled(cfs_rq))
12182 : return;
12183 :
12184 : if (!throttled_hierarchy(cfs_rq))
12185 : list_add_leaf_cfs_rq(cfs_rq);
12186 :
12187 : /* Start to propagate at parent */
12188 : se = se->parent;
12189 :
12190 : for_each_sched_entity(se) {
12191 : cfs_rq = cfs_rq_of(se);
12192 :
12193 : update_load_avg(cfs_rq, se, UPDATE_TG);
12194 :
12195 : if (cfs_rq_throttled(cfs_rq))
12196 : break;
12197 :
12198 : if (!throttled_hierarchy(cfs_rq))
12199 : list_add_leaf_cfs_rq(cfs_rq);
12200 : }
12201 : }
12202 : #else
12203 : static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12204 : #endif
12205 :
12206 : static void detach_entity_cfs_rq(struct sched_entity *se)
12207 : {
12208 0 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12209 :
12210 : #ifdef CONFIG_SMP
12211 : /*
12212 : * In case the task sched_avg hasn't been attached:
12213 : * - A forked task which hasn't been woken up by wake_up_new_task().
12214 : * - A task which has been woken up by try_to_wake_up() but is
12215 : * waiting for actually being woken up by sched_ttwu_pending().
12216 : */
12217 : if (!se->avg.last_update_time)
12218 : return;
12219 : #endif
12220 :
12221 : /* Catch up with the cfs_rq and remove our load when we leave */
12222 0 : update_load_avg(cfs_rq, se, 0);
12223 0 : detach_entity_load_avg(cfs_rq, se);
12224 : update_tg_load_avg(cfs_rq);
12225 0 : propagate_entity_cfs_rq(se);
12226 : }
12227 :
12228 : static void attach_entity_cfs_rq(struct sched_entity *se)
12229 : {
12230 0 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12231 :
12232 : /* Synchronize entity with its cfs_rq */
12233 0 : update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12234 0 : attach_entity_load_avg(cfs_rq, se);
12235 : update_tg_load_avg(cfs_rq);
12236 0 : propagate_entity_cfs_rq(se);
12237 : }
12238 :
12239 0 : static void detach_task_cfs_rq(struct task_struct *p)
12240 : {
12241 0 : struct sched_entity *se = &p->se;
12242 0 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12243 :
12244 0 : if (!vruntime_normalized(p)) {
12245 : /*
12246 : * Fix up our vruntime so that the current sleep doesn't
12247 : * cause 'unlimited' sleep bonus.
12248 : */
12249 0 : place_entity(cfs_rq, se, 0);
12250 0 : se->vruntime -= cfs_rq->min_vruntime;
12251 : }
12252 :
12253 0 : detach_entity_cfs_rq(se);
12254 0 : }
12255 :
12256 : static void attach_task_cfs_rq(struct task_struct *p)
12257 : {
12258 0 : struct sched_entity *se = &p->se;
12259 0 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12260 :
12261 0 : attach_entity_cfs_rq(se);
12262 :
12263 0 : if (!vruntime_normalized(p))
12264 0 : se->vruntime += cfs_rq->min_vruntime;
12265 : }
12266 :
12267 0 : static void switched_from_fair(struct rq *rq, struct task_struct *p)
12268 : {
12269 0 : detach_task_cfs_rq(p);
12270 0 : }
12271 :
12272 0 : static void switched_to_fair(struct rq *rq, struct task_struct *p)
12273 : {
12274 0 : attach_task_cfs_rq(p);
12275 :
12276 0 : if (task_on_rq_queued(p)) {
12277 : /*
12278 : * We were most likely switched from sched_rt, so
12279 : * kick off the schedule if running, otherwise just see
12280 : * if we can still preempt the current task.
12281 : */
12282 0 : if (task_current(rq, p))
12283 0 : resched_curr(rq);
12284 : else
12285 0 : check_preempt_curr(rq, p, 0);
12286 : }
12287 0 : }
12288 :
12289 : /* Account for a task changing its policy or group.
12290 : *
12291 : * This routine is mostly called to set cfs_rq->curr field when a task
12292 : * migrates between groups/classes.
12293 : */
12294 4 : static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12295 : {
12296 4 : struct sched_entity *se = &p->se;
12297 :
12298 : #ifdef CONFIG_SMP
12299 : if (task_on_rq_queued(p)) {
12300 : /*
12301 : * Move the next running task to the front of the list, so our
12302 : * cfs_tasks list becomes MRU one.
12303 : */
12304 : list_move(&se->group_node, &rq->cfs_tasks);
12305 : }
12306 : #endif
12307 :
12308 8 : for_each_sched_entity(se) {
12309 8 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12310 :
12311 4 : set_next_entity(cfs_rq, se);
12312 : /* ensure bandwidth has been allocated on our new cfs_rq */
12313 4 : account_cfs_rq_runtime(cfs_rq, 0);
12314 : }
12315 4 : }
12316 :
12317 1 : void init_cfs_rq(struct cfs_rq *cfs_rq)
12318 : {
12319 1 : cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12320 1 : u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12321 : #ifdef CONFIG_SMP
12322 : raw_spin_lock_init(&cfs_rq->removed.lock);
12323 : #endif
12324 1 : }
12325 :
12326 : #ifdef CONFIG_FAIR_GROUP_SCHED
12327 : static void task_change_group_fair(struct task_struct *p)
12328 : {
12329 : /*
12330 : * We couldn't detach or attach a forked task which
12331 : * hasn't been woken up by wake_up_new_task().
12332 : */
12333 : if (READ_ONCE(p->__state) == TASK_NEW)
12334 : return;
12335 :
12336 : detach_task_cfs_rq(p);
12337 :
12338 : #ifdef CONFIG_SMP
12339 : /* Tell se's cfs_rq has been changed -- migrated */
12340 : p->se.avg.last_update_time = 0;
12341 : #endif
12342 : set_task_rq(p, task_cpu(p));
12343 : attach_task_cfs_rq(p);
12344 : }
12345 :
12346 : void free_fair_sched_group(struct task_group *tg)
12347 : {
12348 : int i;
12349 :
12350 : for_each_possible_cpu(i) {
12351 : if (tg->cfs_rq)
12352 : kfree(tg->cfs_rq[i]);
12353 : if (tg->se)
12354 : kfree(tg->se[i]);
12355 : }
12356 :
12357 : kfree(tg->cfs_rq);
12358 : kfree(tg->se);
12359 : }
12360 :
12361 : int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12362 : {
12363 : struct sched_entity *se;
12364 : struct cfs_rq *cfs_rq;
12365 : int i;
12366 :
12367 : tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12368 : if (!tg->cfs_rq)
12369 : goto err;
12370 : tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12371 : if (!tg->se)
12372 : goto err;
12373 :
12374 : tg->shares = NICE_0_LOAD;
12375 :
12376 : init_cfs_bandwidth(tg_cfs_bandwidth(tg));
12377 :
12378 : for_each_possible_cpu(i) {
12379 : cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12380 : GFP_KERNEL, cpu_to_node(i));
12381 : if (!cfs_rq)
12382 : goto err;
12383 :
12384 : se = kzalloc_node(sizeof(struct sched_entity_stats),
12385 : GFP_KERNEL, cpu_to_node(i));
12386 : if (!se)
12387 : goto err_free_rq;
12388 :
12389 : init_cfs_rq(cfs_rq);
12390 : init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12391 : init_entity_runnable_average(se);
12392 : }
12393 :
12394 : return 1;
12395 :
12396 : err_free_rq:
12397 : kfree(cfs_rq);
12398 : err:
12399 : return 0;
12400 : }
12401 :
12402 : void online_fair_sched_group(struct task_group *tg)
12403 : {
12404 : struct sched_entity *se;
12405 : struct rq_flags rf;
12406 : struct rq *rq;
12407 : int i;
12408 :
12409 : for_each_possible_cpu(i) {
12410 : rq = cpu_rq(i);
12411 : se = tg->se[i];
12412 : rq_lock_irq(rq, &rf);
12413 : update_rq_clock(rq);
12414 : attach_entity_cfs_rq(se);
12415 : sync_throttle(tg, i);
12416 : rq_unlock_irq(rq, &rf);
12417 : }
12418 : }
12419 :
12420 : void unregister_fair_sched_group(struct task_group *tg)
12421 : {
12422 : unsigned long flags;
12423 : struct rq *rq;
12424 : int cpu;
12425 :
12426 : destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12427 :
12428 : for_each_possible_cpu(cpu) {
12429 : if (tg->se[cpu])
12430 : remove_entity_load_avg(tg->se[cpu]);
12431 :
12432 : /*
12433 : * Only empty task groups can be destroyed; so we can speculatively
12434 : * check on_list without danger of it being re-added.
12435 : */
12436 : if (!tg->cfs_rq[cpu]->on_list)
12437 : continue;
12438 :
12439 : rq = cpu_rq(cpu);
12440 :
12441 : raw_spin_rq_lock_irqsave(rq, flags);
12442 : list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12443 : raw_spin_rq_unlock_irqrestore(rq, flags);
12444 : }
12445 : }
12446 :
12447 : void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12448 : struct sched_entity *se, int cpu,
12449 : struct sched_entity *parent)
12450 : {
12451 : struct rq *rq = cpu_rq(cpu);
12452 :
12453 : cfs_rq->tg = tg;
12454 : cfs_rq->rq = rq;
12455 : init_cfs_rq_runtime(cfs_rq);
12456 :
12457 : tg->cfs_rq[cpu] = cfs_rq;
12458 : tg->se[cpu] = se;
12459 :
12460 : /* se could be NULL for root_task_group */
12461 : if (!se)
12462 : return;
12463 :
12464 : if (!parent) {
12465 : se->cfs_rq = &rq->cfs;
12466 : se->depth = 0;
12467 : } else {
12468 : se->cfs_rq = parent->my_q;
12469 : se->depth = parent->depth + 1;
12470 : }
12471 :
12472 : se->my_q = cfs_rq;
12473 : /* guarantee group entities always have weight */
12474 : update_load_set(&se->load, NICE_0_LOAD);
12475 : se->parent = parent;
12476 : }
12477 :
12478 : static DEFINE_MUTEX(shares_mutex);
12479 :
12480 : static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12481 : {
12482 : int i;
12483 :
12484 : lockdep_assert_held(&shares_mutex);
12485 :
12486 : /*
12487 : * We can't change the weight of the root cgroup.
12488 : */
12489 : if (!tg->se[0])
12490 : return -EINVAL;
12491 :
12492 : shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12493 :
12494 : if (tg->shares == shares)
12495 : return 0;
12496 :
12497 : tg->shares = shares;
12498 : for_each_possible_cpu(i) {
12499 : struct rq *rq = cpu_rq(i);
12500 : struct sched_entity *se = tg->se[i];
12501 : struct rq_flags rf;
12502 :
12503 : /* Propagate contribution to hierarchy */
12504 : rq_lock_irqsave(rq, &rf);
12505 : update_rq_clock(rq);
12506 : for_each_sched_entity(se) {
12507 : update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12508 : update_cfs_group(se);
12509 : }
12510 : rq_unlock_irqrestore(rq, &rf);
12511 : }
12512 :
12513 : return 0;
12514 : }
12515 :
12516 : int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12517 : {
12518 : int ret;
12519 :
12520 : mutex_lock(&shares_mutex);
12521 : if (tg_is_idle(tg))
12522 : ret = -EINVAL;
12523 : else
12524 : ret = __sched_group_set_shares(tg, shares);
12525 : mutex_unlock(&shares_mutex);
12526 :
12527 : return ret;
12528 : }
12529 :
12530 : int sched_group_set_idle(struct task_group *tg, long idle)
12531 : {
12532 : int i;
12533 :
12534 : if (tg == &root_task_group)
12535 : return -EINVAL;
12536 :
12537 : if (idle < 0 || idle > 1)
12538 : return -EINVAL;
12539 :
12540 : mutex_lock(&shares_mutex);
12541 :
12542 : if (tg->idle == idle) {
12543 : mutex_unlock(&shares_mutex);
12544 : return 0;
12545 : }
12546 :
12547 : tg->idle = idle;
12548 :
12549 : for_each_possible_cpu(i) {
12550 : struct rq *rq = cpu_rq(i);
12551 : struct sched_entity *se = tg->se[i];
12552 : struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12553 : bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12554 : long idle_task_delta;
12555 : struct rq_flags rf;
12556 :
12557 : rq_lock_irqsave(rq, &rf);
12558 :
12559 : grp_cfs_rq->idle = idle;
12560 : if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12561 : goto next_cpu;
12562 :
12563 : if (se->on_rq) {
12564 : parent_cfs_rq = cfs_rq_of(se);
12565 : if (cfs_rq_is_idle(grp_cfs_rq))
12566 : parent_cfs_rq->idle_nr_running++;
12567 : else
12568 : parent_cfs_rq->idle_nr_running--;
12569 : }
12570 :
12571 : idle_task_delta = grp_cfs_rq->h_nr_running -
12572 : grp_cfs_rq->idle_h_nr_running;
12573 : if (!cfs_rq_is_idle(grp_cfs_rq))
12574 : idle_task_delta *= -1;
12575 :
12576 : for_each_sched_entity(se) {
12577 : struct cfs_rq *cfs_rq = cfs_rq_of(se);
12578 :
12579 : if (!se->on_rq)
12580 : break;
12581 :
12582 : cfs_rq->idle_h_nr_running += idle_task_delta;
12583 :
12584 : /* Already accounted at parent level and above. */
12585 : if (cfs_rq_is_idle(cfs_rq))
12586 : break;
12587 : }
12588 :
12589 : next_cpu:
12590 : rq_unlock_irqrestore(rq, &rf);
12591 : }
12592 :
12593 : /* Idle groups have minimum weight. */
12594 : if (tg_is_idle(tg))
12595 : __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12596 : else
12597 : __sched_group_set_shares(tg, NICE_0_LOAD);
12598 :
12599 : mutex_unlock(&shares_mutex);
12600 : return 0;
12601 : }
12602 :
12603 : #else /* CONFIG_FAIR_GROUP_SCHED */
12604 :
12605 0 : void free_fair_sched_group(struct task_group *tg) { }
12606 :
12607 0 : int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12608 : {
12609 0 : return 1;
12610 : }
12611 :
12612 0 : void online_fair_sched_group(struct task_group *tg) { }
12613 :
12614 0 : void unregister_fair_sched_group(struct task_group *tg) { }
12615 :
12616 : #endif /* CONFIG_FAIR_GROUP_SCHED */
12617 :
12618 :
12619 0 : static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12620 : {
12621 0 : struct sched_entity *se = &task->se;
12622 0 : unsigned int rr_interval = 0;
12623 :
12624 : /*
12625 : * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12626 : * idle runqueue:
12627 : */
12628 0 : if (rq->cfs.load.weight)
12629 0 : rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
12630 :
12631 0 : return rr_interval;
12632 : }
12633 :
12634 : /*
12635 : * All the scheduling class methods:
12636 : */
12637 : DEFINE_SCHED_CLASS(fair) = {
12638 :
12639 : .enqueue_task = enqueue_task_fair,
12640 : .dequeue_task = dequeue_task_fair,
12641 : .yield_task = yield_task_fair,
12642 : .yield_to_task = yield_to_task_fair,
12643 :
12644 : .check_preempt_curr = check_preempt_wakeup,
12645 :
12646 : .pick_next_task = __pick_next_task_fair,
12647 : .put_prev_task = put_prev_task_fair,
12648 : .set_next_task = set_next_task_fair,
12649 :
12650 : #ifdef CONFIG_SMP
12651 : .balance = balance_fair,
12652 : .pick_task = pick_task_fair,
12653 : .select_task_rq = select_task_rq_fair,
12654 : .migrate_task_rq = migrate_task_rq_fair,
12655 :
12656 : .rq_online = rq_online_fair,
12657 : .rq_offline = rq_offline_fair,
12658 :
12659 : .task_dead = task_dead_fair,
12660 : .set_cpus_allowed = set_cpus_allowed_common,
12661 : #endif
12662 :
12663 : .task_tick = task_tick_fair,
12664 : .task_fork = task_fork_fair,
12665 :
12666 : .prio_changed = prio_changed_fair,
12667 : .switched_from = switched_from_fair,
12668 : .switched_to = switched_to_fair,
12669 :
12670 : .get_rr_interval = get_rr_interval_fair,
12671 :
12672 : .update_curr = update_curr_fair,
12673 :
12674 : #ifdef CONFIG_FAIR_GROUP_SCHED
12675 : .task_change_group = task_change_group_fair,
12676 : #endif
12677 :
12678 : #ifdef CONFIG_SCHED_CORE
12679 : .task_is_throttled = task_is_throttled_fair,
12680 : #endif
12681 :
12682 : #ifdef CONFIG_UCLAMP_TASK
12683 : .uclamp_enabled = 1,
12684 : #endif
12685 : };
12686 :
12687 : #ifdef CONFIG_SCHED_DEBUG
12688 : void print_cfs_stats(struct seq_file *m, int cpu)
12689 : {
12690 : struct cfs_rq *cfs_rq, *pos;
12691 :
12692 : rcu_read_lock();
12693 : for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12694 : print_cfs_rq(m, cpu, cfs_rq);
12695 : rcu_read_unlock();
12696 : }
12697 :
12698 : #ifdef CONFIG_NUMA_BALANCING
12699 : void show_numa_stats(struct task_struct *p, struct seq_file *m)
12700 : {
12701 : int node;
12702 : unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12703 : struct numa_group *ng;
12704 :
12705 : rcu_read_lock();
12706 : ng = rcu_dereference(p->numa_group);
12707 : for_each_online_node(node) {
12708 : if (p->numa_faults) {
12709 : tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12710 : tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12711 : }
12712 : if (ng) {
12713 : gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12714 : gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12715 : }
12716 : print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12717 : }
12718 : rcu_read_unlock();
12719 : }
12720 : #endif /* CONFIG_NUMA_BALANCING */
12721 : #endif /* CONFIG_SCHED_DEBUG */
12722 :
12723 1 : __init void init_sched_fair_class(void)
12724 : {
12725 : #ifdef CONFIG_SMP
12726 : int i;
12727 :
12728 : for_each_possible_cpu(i) {
12729 : zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
12730 : zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
12731 :
12732 : #ifdef CONFIG_CFS_BANDWIDTH
12733 : INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
12734 : INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
12735 : #endif
12736 : }
12737 :
12738 : open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
12739 :
12740 : #ifdef CONFIG_NO_HZ_COMMON
12741 : nohz.next_balance = jiffies;
12742 : nohz.next_blocked = jiffies;
12743 : zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
12744 : #endif
12745 : #endif /* SMP */
12746 :
12747 1 : }
|
