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 #include <linux/rbtree_augmented.h>
51
52 #include <asm/switch_to.h>
53
54 #include <linux/sched/cond_resched.h>
55
56 #include "sched.h"
57 #include "stats.h"
58 #include "autogroup.h"
59
60 /*
61 * The initial- and re-scaling of tunables is configurable
62 *
63 * Options are:
64 *
65 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
66 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
67 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
68 *
69 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
70 */
71 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
72
73 /*
74 * Minimal preemption granularity for CPU-bound tasks:
75 *
76 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
77 */
78 unsigned int sysctl_sched_base_slice = 750000ULL;
79 static unsigned int normalized_sysctl_sched_base_slice = 750000ULL;
80
81 /*
82 * After fork, child runs first. If set to 0 (default) then
83 * parent will (try to) run first.
84 */
85 unsigned int sysctl_sched_child_runs_first __read_mostly;
86
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
88
89 int sched_thermal_decay_shift;
setup_sched_thermal_decay_shift(char * str)90 static int __init setup_sched_thermal_decay_shift(char *str)
91 {
92 int _shift = 0;
93
94 if (kstrtoint(str, 0, &_shift))
95 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
96
97 sched_thermal_decay_shift = clamp(_shift, 0, 10);
98 return 1;
99 }
100 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
101
102 #ifdef CONFIG_SMP
103 /*
104 * For asym packing, by default the lower numbered CPU has higher priority.
105 */
arch_asym_cpu_priority(int cpu)106 int __weak arch_asym_cpu_priority(int cpu)
107 {
108 return -cpu;
109 }
110
111 /*
112 * The margin used when comparing utilization with CPU capacity.
113 *
114 * (default: ~20%)
115 */
116 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
117
118 /*
119 * The margin used when comparing CPU capacities.
120 * is 'cap1' noticeably greater than 'cap2'
121 *
122 * (default: ~5%)
123 */
124 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
125 #endif
126
127 #ifdef CONFIG_CFS_BANDWIDTH
128 /*
129 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
130 * each time a cfs_rq requests quota.
131 *
132 * Note: in the case that the slice exceeds the runtime remaining (either due
133 * to consumption or the quota being specified to be smaller than the slice)
134 * we will always only issue the remaining available time.
135 *
136 * (default: 5 msec, units: microseconds)
137 */
138 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
139 #endif
140
141 #ifdef CONFIG_NUMA_BALANCING
142 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
143 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
144 #endif
145
146 #ifdef CONFIG_SYSCTL
147 static struct ctl_table sched_fair_sysctls[] = {
148 {
149 .procname = "sched_child_runs_first",
150 .data = &sysctl_sched_child_runs_first,
151 .maxlen = sizeof(unsigned int),
152 .mode = 0644,
153 .proc_handler = proc_dointvec,
154 },
155 #ifdef CONFIG_CFS_BANDWIDTH
156 {
157 .procname = "sched_cfs_bandwidth_slice_us",
158 .data = &sysctl_sched_cfs_bandwidth_slice,
159 .maxlen = sizeof(unsigned int),
160 .mode = 0644,
161 .proc_handler = proc_dointvec_minmax,
162 .extra1 = SYSCTL_ONE,
163 },
164 #endif
165 #ifdef CONFIG_NUMA_BALANCING
166 {
167 .procname = "numa_balancing_promote_rate_limit_MBps",
168 .data = &sysctl_numa_balancing_promote_rate_limit,
169 .maxlen = sizeof(unsigned int),
170 .mode = 0644,
171 .proc_handler = proc_dointvec_minmax,
172 .extra1 = SYSCTL_ZERO,
173 },
174 #endif /* CONFIG_NUMA_BALANCING */
175 {}
176 };
177
sched_fair_sysctl_init(void)178 static int __init sched_fair_sysctl_init(void)
179 {
180 register_sysctl_init("kernel", sched_fair_sysctls);
181 return 0;
182 }
183 late_initcall(sched_fair_sysctl_init);
184 #endif
185
update_load_add(struct load_weight * lw,unsigned long inc)186 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
187 {
188 lw->weight += inc;
189 lw->inv_weight = 0;
190 }
191
update_load_sub(struct load_weight * lw,unsigned long dec)192 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
193 {
194 lw->weight -= dec;
195 lw->inv_weight = 0;
196 }
197
update_load_set(struct load_weight * lw,unsigned long w)198 static inline void update_load_set(struct load_weight *lw, unsigned long w)
199 {
200 lw->weight = w;
201 lw->inv_weight = 0;
202 }
203
204 /*
205 * Increase the granularity value when there are more CPUs,
206 * because with more CPUs the 'effective latency' as visible
207 * to users decreases. But the relationship is not linear,
208 * so pick a second-best guess by going with the log2 of the
209 * number of CPUs.
210 *
211 * This idea comes from the SD scheduler of Con Kolivas:
212 */
get_update_sysctl_factor(void)213 static unsigned int get_update_sysctl_factor(void)
214 {
215 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
216 unsigned int factor;
217
218 switch (sysctl_sched_tunable_scaling) {
219 case SCHED_TUNABLESCALING_NONE:
220 factor = 1;
221 break;
222 case SCHED_TUNABLESCALING_LINEAR:
223 factor = cpus;
224 break;
225 case SCHED_TUNABLESCALING_LOG:
226 default:
227 factor = 1 + ilog2(cpus);
228 break;
229 }
230
231 return factor;
232 }
233
update_sysctl(void)234 static void update_sysctl(void)
235 {
236 unsigned int factor = get_update_sysctl_factor();
237
238 #define SET_SYSCTL(name) \
239 (sysctl_##name = (factor) * normalized_sysctl_##name)
240 SET_SYSCTL(sched_base_slice);
241 #undef SET_SYSCTL
242 }
243
sched_init_granularity(void)244 void __init sched_init_granularity(void)
245 {
246 update_sysctl();
247 }
248
249 #define WMULT_CONST (~0U)
250 #define WMULT_SHIFT 32
251
__update_inv_weight(struct load_weight * lw)252 static void __update_inv_weight(struct load_weight *lw)
253 {
254 unsigned long w;
255
256 if (likely(lw->inv_weight))
257 return;
258
259 w = scale_load_down(lw->weight);
260
261 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
262 lw->inv_weight = 1;
263 else if (unlikely(!w))
264 lw->inv_weight = WMULT_CONST;
265 else
266 lw->inv_weight = WMULT_CONST / w;
267 }
268
269 /*
270 * delta_exec * weight / lw.weight
271 * OR
272 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
273 *
274 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
275 * we're guaranteed shift stays positive because inv_weight is guaranteed to
276 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
277 *
278 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
279 * weight/lw.weight <= 1, and therefore our shift will also be positive.
280 */
__calc_delta(u64 delta_exec,unsigned long weight,struct load_weight * lw)281 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
282 {
283 u64 fact = scale_load_down(weight);
284 u32 fact_hi = (u32)(fact >> 32);
285 int shift = WMULT_SHIFT;
286 int fs;
287
288 __update_inv_weight(lw);
289
290 if (unlikely(fact_hi)) {
291 fs = fls(fact_hi);
292 shift -= fs;
293 fact >>= fs;
294 }
295
296 fact = mul_u32_u32(fact, lw->inv_weight);
297
298 fact_hi = (u32)(fact >> 32);
299 if (fact_hi) {
300 fs = fls(fact_hi);
301 shift -= fs;
302 fact >>= fs;
303 }
304
305 return mul_u64_u32_shr(delta_exec, fact, shift);
306 }
307
308 /*
309 * delta /= w
310 */
calc_delta_fair(u64 delta,struct sched_entity * se)311 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
312 {
313 if (unlikely(se->load.weight != NICE_0_LOAD))
314 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
315
316 return delta;
317 }
318
319 const struct sched_class fair_sched_class;
320
321 /**************************************************************
322 * CFS operations on generic schedulable entities:
323 */
324
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326
327 /* Walk up scheduling entities hierarchy */
328 #define for_each_sched_entity(se) \
329 for (; se; se = se->parent)
330
list_add_leaf_cfs_rq(struct cfs_rq * cfs_rq)331 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
332 {
333 struct rq *rq = rq_of(cfs_rq);
334 int cpu = cpu_of(rq);
335
336 if (cfs_rq->on_list)
337 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
338
339 cfs_rq->on_list = 1;
340
341 /*
342 * Ensure we either appear before our parent (if already
343 * enqueued) or force our parent to appear after us when it is
344 * enqueued. The fact that we always enqueue bottom-up
345 * reduces this to two cases and a special case for the root
346 * cfs_rq. Furthermore, it also means that we will always reset
347 * tmp_alone_branch either when the branch is connected
348 * to a tree or when we reach the top of the tree
349 */
350 if (cfs_rq->tg->parent &&
351 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
352 /*
353 * If parent is already on the list, we add the child
354 * just before. Thanks to circular linked property of
355 * the list, this means to put the child at the tail
356 * of the list that starts by parent.
357 */
358 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
359 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
360 /*
361 * The branch is now connected to its tree so we can
362 * reset tmp_alone_branch to the beginning of the
363 * list.
364 */
365 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
366 return true;
367 }
368
369 if (!cfs_rq->tg->parent) {
370 /*
371 * cfs rq without parent should be put
372 * at the tail of the list.
373 */
374 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
375 &rq->leaf_cfs_rq_list);
376 /*
377 * We have reach the top of a tree so we can reset
378 * tmp_alone_branch to the beginning of the list.
379 */
380 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
381 return true;
382 }
383
384 /*
385 * The parent has not already been added so we want to
386 * make sure that it will be put after us.
387 * tmp_alone_branch points to the begin of the branch
388 * where we will add parent.
389 */
390 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
391 /*
392 * update tmp_alone_branch to points to the new begin
393 * of the branch
394 */
395 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
396 return false;
397 }
398
list_del_leaf_cfs_rq(struct cfs_rq * cfs_rq)399 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
400 {
401 if (cfs_rq->on_list) {
402 struct rq *rq = rq_of(cfs_rq);
403
404 /*
405 * With cfs_rq being unthrottled/throttled during an enqueue,
406 * it can happen the tmp_alone_branch points the a leaf that
407 * we finally want to del. In this case, tmp_alone_branch moves
408 * to the prev element but it will point to rq->leaf_cfs_rq_list
409 * at the end of the enqueue.
410 */
411 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
412 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
413
414 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
415 cfs_rq->on_list = 0;
416 }
417 }
418
assert_list_leaf_cfs_rq(struct rq * rq)419 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
420 {
421 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
422 }
423
424 /* Iterate thr' all leaf cfs_rq's on a runqueue */
425 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
426 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
427 leaf_cfs_rq_list)
428
429 /* Do the two (enqueued) entities belong to the same group ? */
430 static inline struct cfs_rq *
is_same_group(struct sched_entity * se,struct sched_entity * pse)431 is_same_group(struct sched_entity *se, struct sched_entity *pse)
432 {
433 if (se->cfs_rq == pse->cfs_rq)
434 return se->cfs_rq;
435
436 return NULL;
437 }
438
parent_entity(const struct sched_entity * se)439 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
440 {
441 return se->parent;
442 }
443
444 static void
find_matching_se(struct sched_entity ** se,struct sched_entity ** pse)445 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
446 {
447 int se_depth, pse_depth;
448
449 /*
450 * preemption test can be made between sibling entities who are in the
451 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
452 * both tasks until we find their ancestors who are siblings of common
453 * parent.
454 */
455
456 /* First walk up until both entities are at same depth */
457 se_depth = (*se)->depth;
458 pse_depth = (*pse)->depth;
459
460 while (se_depth > pse_depth) {
461 se_depth--;
462 *se = parent_entity(*se);
463 }
464
465 while (pse_depth > se_depth) {
466 pse_depth--;
467 *pse = parent_entity(*pse);
468 }
469
470 while (!is_same_group(*se, *pse)) {
471 *se = parent_entity(*se);
472 *pse = parent_entity(*pse);
473 }
474 }
475
tg_is_idle(struct task_group * tg)476 static int tg_is_idle(struct task_group *tg)
477 {
478 return tg->idle > 0;
479 }
480
cfs_rq_is_idle(struct cfs_rq * cfs_rq)481 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
482 {
483 return cfs_rq->idle > 0;
484 }
485
se_is_idle(struct sched_entity * se)486 static int se_is_idle(struct sched_entity *se)
487 {
488 if (entity_is_task(se))
489 return task_has_idle_policy(task_of(se));
490 return cfs_rq_is_idle(group_cfs_rq(se));
491 }
492
493 #else /* !CONFIG_FAIR_GROUP_SCHED */
494
495 #define for_each_sched_entity(se) \
496 for (; se; se = NULL)
497
list_add_leaf_cfs_rq(struct cfs_rq * cfs_rq)498 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
499 {
500 return true;
501 }
502
list_del_leaf_cfs_rq(struct cfs_rq * cfs_rq)503 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
504 {
505 }
506
assert_list_leaf_cfs_rq(struct rq * rq)507 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
508 {
509 }
510
511 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
512 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
513
parent_entity(struct sched_entity * se)514 static inline struct sched_entity *parent_entity(struct sched_entity *se)
515 {
516 return NULL;
517 }
518
519 static inline void
find_matching_se(struct sched_entity ** se,struct sched_entity ** pse)520 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
521 {
522 }
523
tg_is_idle(struct task_group * tg)524 static inline int tg_is_idle(struct task_group *tg)
525 {
526 return 0;
527 }
528
cfs_rq_is_idle(struct cfs_rq * cfs_rq)529 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
530 {
531 return 0;
532 }
533
se_is_idle(struct sched_entity * se)534 static int se_is_idle(struct sched_entity *se)
535 {
536 return task_has_idle_policy(task_of(se));
537 }
538
539 #endif /* CONFIG_FAIR_GROUP_SCHED */
540
541 static __always_inline
542 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
543
544 /**************************************************************
545 * Scheduling class tree data structure manipulation methods:
546 */
547
max_vruntime(u64 max_vruntime,u64 vruntime)548 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
549 {
550 s64 delta = (s64)(vruntime - max_vruntime);
551 if (delta > 0)
552 max_vruntime = vruntime;
553
554 return max_vruntime;
555 }
556
min_vruntime(u64 min_vruntime,u64 vruntime)557 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
558 {
559 s64 delta = (s64)(vruntime - min_vruntime);
560 if (delta < 0)
561 min_vruntime = vruntime;
562
563 return min_vruntime;
564 }
565
entity_before(const struct sched_entity * a,const struct sched_entity * b)566 static inline bool entity_before(const struct sched_entity *a,
567 const struct sched_entity *b)
568 {
569 return (s64)(a->vruntime - b->vruntime) < 0;
570 }
571
entity_key(struct cfs_rq * cfs_rq,struct sched_entity * se)572 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
573 {
574 return (s64)(se->vruntime - cfs_rq->min_vruntime);
575 }
576
577 #define __node_2_se(node) \
578 rb_entry((node), struct sched_entity, run_node)
579
580 /*
581 * Compute virtual time from the per-task service numbers:
582 *
583 * Fair schedulers conserve lag:
584 *
585 * \Sum lag_i = 0
586 *
587 * Where lag_i is given by:
588 *
589 * lag_i = S - s_i = w_i * (V - v_i)
590 *
591 * Where S is the ideal service time and V is it's virtual time counterpart.
592 * Therefore:
593 *
594 * \Sum lag_i = 0
595 * \Sum w_i * (V - v_i) = 0
596 * \Sum w_i * V - w_i * v_i = 0
597 *
598 * From which we can solve an expression for V in v_i (which we have in
599 * se->vruntime):
600 *
601 * \Sum v_i * w_i \Sum v_i * w_i
602 * V = -------------- = --------------
603 * \Sum w_i W
604 *
605 * Specifically, this is the weighted average of all entity virtual runtimes.
606 *
607 * [[ NOTE: this is only equal to the ideal scheduler under the condition
608 * that join/leave operations happen at lag_i = 0, otherwise the
609 * virtual time has non-continguous motion equivalent to:
610 *
611 * V +-= lag_i / W
612 *
613 * Also see the comment in place_entity() that deals with this. ]]
614 *
615 * However, since v_i is u64, and the multiplcation could easily overflow
616 * transform it into a relative form that uses smaller quantities:
617 *
618 * Substitute: v_i == (v_i - v0) + v0
619 *
620 * \Sum ((v_i - v0) + v0) * w_i \Sum (v_i - v0) * w_i
621 * V = ---------------------------- = --------------------- + v0
622 * W W
623 *
624 * Which we track using:
625 *
626 * v0 := cfs_rq->min_vruntime
627 * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
628 * \Sum w_i := cfs_rq->avg_load
629 *
630 * Since min_vruntime is a monotonic increasing variable that closely tracks
631 * the per-task service, these deltas: (v_i - v), will be in the order of the
632 * maximal (virtual) lag induced in the system due to quantisation.
633 *
634 * Also, we use scale_load_down() to reduce the size.
635 *
636 * As measured, the max (key * weight) value was ~44 bits for a kernel build.
637 */
638 static void
avg_vruntime_add(struct cfs_rq * cfs_rq,struct sched_entity * se)639 avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 {
641 unsigned long weight = scale_load_down(se->load.weight);
642 s64 key = entity_key(cfs_rq, se);
643
644 cfs_rq->avg_vruntime += key * weight;
645 cfs_rq->avg_load += weight;
646 }
647
648 static void
avg_vruntime_sub(struct cfs_rq * cfs_rq,struct sched_entity * se)649 avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
650 {
651 unsigned long weight = scale_load_down(se->load.weight);
652 s64 key = entity_key(cfs_rq, se);
653
654 cfs_rq->avg_vruntime -= key * weight;
655 cfs_rq->avg_load -= weight;
656 }
657
658 static inline
avg_vruntime_update(struct cfs_rq * cfs_rq,s64 delta)659 void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta)
660 {
661 /*
662 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load
663 */
664 cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;
665 }
666
667 /*
668 * Specifically: avg_runtime() + 0 must result in entity_eligible() := true
669 * For this to be so, the result of this function must have a left bias.
670 */
avg_vruntime(struct cfs_rq * cfs_rq)671 u64 avg_vruntime(struct cfs_rq *cfs_rq)
672 {
673 struct sched_entity *curr = cfs_rq->curr;
674 s64 avg = cfs_rq->avg_vruntime;
675 long load = cfs_rq->avg_load;
676
677 if (curr && curr->on_rq) {
678 unsigned long weight = scale_load_down(curr->load.weight);
679
680 avg += entity_key(cfs_rq, curr) * weight;
681 load += weight;
682 }
683
684 if (load) {
685 /* sign flips effective floor / ceil */
686 if (avg < 0)
687 avg -= (load - 1);
688 avg = div_s64(avg, load);
689 }
690
691 return cfs_rq->min_vruntime + avg;
692 }
693
694 /*
695 * lag_i = S - s_i = w_i * (V - v_i)
696 *
697 * However, since V is approximated by the weighted average of all entities it
698 * is possible -- by addition/removal/reweight to the tree -- to move V around
699 * and end up with a larger lag than we started with.
700 *
701 * Limit this to either double the slice length with a minimum of TICK_NSEC
702 * since that is the timing granularity.
703 *
704 * EEVDF gives the following limit for a steady state system:
705 *
706 * -r_max < lag < max(r_max, q)
707 *
708 * XXX could add max_slice to the augmented data to track this.
709 */
entity_lag(u64 avruntime,struct sched_entity * se)710 static s64 entity_lag(u64 avruntime, struct sched_entity *se)
711 {
712 s64 vlag, limit;
713
714 vlag = avruntime - se->vruntime;
715 limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
716
717 return clamp(vlag, -limit, limit);
718 }
719
update_entity_lag(struct cfs_rq * cfs_rq,struct sched_entity * se)720 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
721 {
722 SCHED_WARN_ON(!se->on_rq);
723
724 se->vlag = entity_lag(avg_vruntime(cfs_rq), se);
725 }
726
727 /*
728 * Entity is eligible once it received less service than it ought to have,
729 * eg. lag >= 0.
730 *
731 * lag_i = S - s_i = w_i*(V - v_i)
732 *
733 * lag_i >= 0 -> V >= v_i
734 *
735 * \Sum (v_i - v)*w_i
736 * V = ------------------ + v
737 * \Sum w_i
738 *
739 * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
740 *
741 * Note: using 'avg_vruntime() > se->vruntime' is inacurate due
742 * to the loss in precision caused by the division.
743 */
entity_eligible(struct cfs_rq * cfs_rq,struct sched_entity * se)744 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 {
746 struct sched_entity *curr = cfs_rq->curr;
747 s64 avg = cfs_rq->avg_vruntime;
748 long load = cfs_rq->avg_load;
749
750 if (curr && curr->on_rq) {
751 unsigned long weight = scale_load_down(curr->load.weight);
752
753 avg += entity_key(cfs_rq, curr) * weight;
754 load += weight;
755 }
756
757 return avg >= entity_key(cfs_rq, se) * load;
758 }
759
__update_min_vruntime(struct cfs_rq * cfs_rq,u64 vruntime)760 static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime)
761 {
762 u64 min_vruntime = cfs_rq->min_vruntime;
763 /*
764 * open coded max_vruntime() to allow updating avg_vruntime
765 */
766 s64 delta = (s64)(vruntime - min_vruntime);
767 if (delta > 0) {
768 avg_vruntime_update(cfs_rq, delta);
769 min_vruntime = vruntime;
770 }
771 return min_vruntime;
772 }
773
update_min_vruntime(struct cfs_rq * cfs_rq)774 static void update_min_vruntime(struct cfs_rq *cfs_rq)
775 {
776 struct sched_entity *se = __pick_first_entity(cfs_rq);
777 struct sched_entity *curr = cfs_rq->curr;
778
779 u64 vruntime = cfs_rq->min_vruntime;
780
781 if (curr) {
782 if (curr->on_rq)
783 vruntime = curr->vruntime;
784 else
785 curr = NULL;
786 }
787
788 if (se) {
789 if (!curr)
790 vruntime = se->vruntime;
791 else
792 vruntime = min_vruntime(vruntime, se->vruntime);
793 }
794
795 /* ensure we never gain time by being placed backwards. */
796 u64_u32_store(cfs_rq->min_vruntime,
797 __update_min_vruntime(cfs_rq, vruntime));
798 }
799
__entity_less(struct rb_node * a,const struct rb_node * b)800 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
801 {
802 return entity_before(__node_2_se(a), __node_2_se(b));
803 }
804
805 #define deadline_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
806
__update_min_deadline(struct sched_entity * se,struct rb_node * node)807 static inline void __update_min_deadline(struct sched_entity *se, struct rb_node *node)
808 {
809 if (node) {
810 struct sched_entity *rse = __node_2_se(node);
811 if (deadline_gt(min_deadline, se, rse))
812 se->min_deadline = rse->min_deadline;
813 }
814 }
815
816 /*
817 * se->min_deadline = min(se->deadline, left->min_deadline, right->min_deadline)
818 */
min_deadline_update(struct sched_entity * se,bool exit)819 static inline bool min_deadline_update(struct sched_entity *se, bool exit)
820 {
821 u64 old_min_deadline = se->min_deadline;
822 struct rb_node *node = &se->run_node;
823
824 se->min_deadline = se->deadline;
825 __update_min_deadline(se, node->rb_right);
826 __update_min_deadline(se, node->rb_left);
827
828 return se->min_deadline == old_min_deadline;
829 }
830
831 RB_DECLARE_CALLBACKS(static, min_deadline_cb, struct sched_entity,
832 run_node, min_deadline, min_deadline_update);
833
834 /*
835 * Enqueue an entity into the rb-tree:
836 */
__enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)837 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
838 {
839 avg_vruntime_add(cfs_rq, se);
840 se->min_deadline = se->deadline;
841 rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
842 __entity_less, &min_deadline_cb);
843 }
844
__dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)845 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
846 {
847 rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
848 &min_deadline_cb);
849 avg_vruntime_sub(cfs_rq, se);
850 }
851
__pick_first_entity(struct cfs_rq * cfs_rq)852 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
853 {
854 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
855
856 if (!left)
857 return NULL;
858
859 return __node_2_se(left);
860 }
861
862 /*
863 * Earliest Eligible Virtual Deadline First
864 *
865 * In order to provide latency guarantees for different request sizes
866 * EEVDF selects the best runnable task from two criteria:
867 *
868 * 1) the task must be eligible (must be owed service)
869 *
870 * 2) from those tasks that meet 1), we select the one
871 * with the earliest virtual deadline.
872 *
873 * We can do this in O(log n) time due to an augmented RB-tree. The
874 * tree keeps the entries sorted on service, but also functions as a
875 * heap based on the deadline by keeping:
876 *
877 * se->min_deadline = min(se->deadline, se->{left,right}->min_deadline)
878 *
879 * Which allows an EDF like search on (sub)trees.
880 */
__pick_eevdf(struct cfs_rq * cfs_rq)881 static struct sched_entity *__pick_eevdf(struct cfs_rq *cfs_rq)
882 {
883 struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
884 struct sched_entity *curr = cfs_rq->curr;
885 struct sched_entity *best = NULL;
886 struct sched_entity *best_left = NULL;
887
888 if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
889 curr = NULL;
890 best = curr;
891
892 /*
893 * Once selected, run a task until it either becomes non-eligible or
894 * until it gets a new slice. See the HACK in set_next_entity().
895 */
896 if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline)
897 return curr;
898
899 while (node) {
900 struct sched_entity *se = __node_2_se(node);
901
902 /*
903 * If this entity is not eligible, try the left subtree.
904 */
905 if (!entity_eligible(cfs_rq, se)) {
906 node = node->rb_left;
907 continue;
908 }
909
910 /*
911 * Now we heap search eligible trees for the best (min_)deadline
912 */
913 if (!best || deadline_gt(deadline, best, se))
914 best = se;
915
916 /*
917 * Every se in a left branch is eligible, keep track of the
918 * branch with the best min_deadline
919 */
920 if (node->rb_left) {
921 struct sched_entity *left = __node_2_se(node->rb_left);
922
923 if (!best_left || deadline_gt(min_deadline, best_left, left))
924 best_left = left;
925
926 /*
927 * min_deadline is in the left branch. rb_left and all
928 * descendants are eligible, so immediately switch to the second
929 * loop.
930 */
931 if (left->min_deadline == se->min_deadline)
932 break;
933 }
934
935 /* min_deadline is at this node, no need to look right */
936 if (se->deadline == se->min_deadline)
937 break;
938
939 /* else min_deadline is in the right branch. */
940 node = node->rb_right;
941 }
942
943 /*
944 * We ran into an eligible node which is itself the best.
945 * (Or nr_running == 0 and both are NULL)
946 */
947 if (!best_left || (s64)(best_left->min_deadline - best->deadline) > 0)
948 return best;
949
950 /*
951 * Now best_left and all of its children are eligible, and we are just
952 * looking for deadline == min_deadline
953 */
954 node = &best_left->run_node;
955 while (node) {
956 struct sched_entity *se = __node_2_se(node);
957
958 /* min_deadline is the current node */
959 if (se->deadline == se->min_deadline)
960 return se;
961
962 /* min_deadline is in the left branch */
963 if (node->rb_left &&
964 __node_2_se(node->rb_left)->min_deadline == se->min_deadline) {
965 node = node->rb_left;
966 continue;
967 }
968
969 /* else min_deadline is in the right branch */
970 node = node->rb_right;
971 }
972 return NULL;
973 }
974
pick_eevdf(struct cfs_rq * cfs_rq)975 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
976 {
977 struct sched_entity *se = __pick_eevdf(cfs_rq);
978
979 if (!se) {
980 struct sched_entity *left = __pick_first_entity(cfs_rq);
981 if (left) {
982 pr_err("EEVDF scheduling fail, picking leftmost\n");
983 return left;
984 }
985 }
986
987 return se;
988 }
989
990 #ifdef CONFIG_SCHED_DEBUG
__pick_last_entity(struct cfs_rq * cfs_rq)991 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
992 {
993 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
994
995 if (!last)
996 return NULL;
997
998 return __node_2_se(last);
999 }
1000
1001 /**************************************************************
1002 * Scheduling class statistics methods:
1003 */
1004 #ifdef CONFIG_SMP
sched_update_scaling(void)1005 int sched_update_scaling(void)
1006 {
1007 unsigned int factor = get_update_sysctl_factor();
1008
1009 #define WRT_SYSCTL(name) \
1010 (normalized_sysctl_##name = sysctl_##name / (factor))
1011 WRT_SYSCTL(sched_base_slice);
1012 #undef WRT_SYSCTL
1013
1014 return 0;
1015 }
1016 #endif
1017 #endif
1018
1019 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
1020
1021 /*
1022 * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
1023 * this is probably good enough.
1024 */
update_deadline(struct cfs_rq * cfs_rq,struct sched_entity * se)1025 static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
1026 {
1027 if ((s64)(se->vruntime - se->deadline) < 0)
1028 return;
1029
1030 /*
1031 * For EEVDF the virtual time slope is determined by w_i (iow.
1032 * nice) while the request time r_i is determined by
1033 * sysctl_sched_base_slice.
1034 */
1035 se->slice = sysctl_sched_base_slice;
1036
1037 /*
1038 * EEVDF: vd_i = ve_i + r_i / w_i
1039 */
1040 se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
1041
1042 /*
1043 * The task has consumed its request, reschedule.
1044 */
1045 if (cfs_rq->nr_running > 1) {
1046 resched_curr(rq_of(cfs_rq));
1047 clear_buddies(cfs_rq, se);
1048 }
1049 }
1050
1051 #include "pelt.h"
1052 #ifdef CONFIG_SMP
1053
1054 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1055 static unsigned long task_h_load(struct task_struct *p);
1056 static unsigned long capacity_of(int cpu);
1057
1058 /* Give new sched_entity start runnable values to heavy its load in infant time */
init_entity_runnable_average(struct sched_entity * se)1059 void init_entity_runnable_average(struct sched_entity *se)
1060 {
1061 struct sched_avg *sa = &se->avg;
1062
1063 memset(sa, 0, sizeof(*sa));
1064
1065 /*
1066 * Tasks are initialized with full load to be seen as heavy tasks until
1067 * they get a chance to stabilize to their real load level.
1068 * Group entities are initialized with zero load to reflect the fact that
1069 * nothing has been attached to the task group yet.
1070 */
1071 if (entity_is_task(se))
1072 sa->load_avg = scale_load_down(se->load.weight);
1073
1074 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
1075 }
1076
1077 /*
1078 * With new tasks being created, their initial util_avgs are extrapolated
1079 * based on the cfs_rq's current util_avg:
1080 *
1081 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
1082 *
1083 * However, in many cases, the above util_avg does not give a desired
1084 * value. Moreover, the sum of the util_avgs may be divergent, such
1085 * as when the series is a harmonic series.
1086 *
1087 * To solve this problem, we also cap the util_avg of successive tasks to
1088 * only 1/2 of the left utilization budget:
1089 *
1090 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1091 *
1092 * where n denotes the nth task and cpu_scale the CPU capacity.
1093 *
1094 * For example, for a CPU with 1024 of capacity, a simplest series from
1095 * the beginning would be like:
1096 *
1097 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
1098 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1099 *
1100 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1101 * if util_avg > util_avg_cap.
1102 */
post_init_entity_util_avg(struct task_struct * p)1103 void post_init_entity_util_avg(struct task_struct *p)
1104 {
1105 struct sched_entity *se = &p->se;
1106 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1107 struct sched_avg *sa = &se->avg;
1108 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1109 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1110
1111 if (p->sched_class != &fair_sched_class) {
1112 /*
1113 * For !fair tasks do:
1114 *
1115 update_cfs_rq_load_avg(now, cfs_rq);
1116 attach_entity_load_avg(cfs_rq, se);
1117 switched_from_fair(rq, p);
1118 *
1119 * such that the next switched_to_fair() has the
1120 * expected state.
1121 */
1122 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1123 return;
1124 }
1125
1126 if (cap > 0) {
1127 if (cfs_rq->avg.util_avg != 0) {
1128 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
1129 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1130
1131 if (sa->util_avg > cap)
1132 sa->util_avg = cap;
1133 } else {
1134 sa->util_avg = cap;
1135 }
1136 }
1137
1138 sa->runnable_avg = sa->util_avg;
1139 }
1140
1141 #else /* !CONFIG_SMP */
init_entity_runnable_average(struct sched_entity * se)1142 void init_entity_runnable_average(struct sched_entity *se)
1143 {
1144 }
post_init_entity_util_avg(struct task_struct * p)1145 void post_init_entity_util_avg(struct task_struct *p)
1146 {
1147 }
update_tg_load_avg(struct cfs_rq * cfs_rq)1148 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
1149 {
1150 }
1151 #endif /* CONFIG_SMP */
1152
update_curr_se(struct rq * rq,struct sched_entity * curr)1153 static s64 update_curr_se(struct rq *rq, struct sched_entity *curr)
1154 {
1155 u64 now = rq_clock_task(rq);
1156 s64 delta_exec;
1157
1158 delta_exec = now - curr->exec_start;
1159 if (unlikely(delta_exec <= 0))
1160 return delta_exec;
1161
1162 curr->exec_start = now;
1163 curr->sum_exec_runtime += delta_exec;
1164
1165 if (schedstat_enabled()) {
1166 struct sched_statistics *stats;
1167
1168 stats = __schedstats_from_se(curr);
1169 __schedstat_set(stats->exec_max,
1170 max(delta_exec, stats->exec_max));
1171 }
1172
1173 return delta_exec;
1174 }
1175
update_curr_task(struct task_struct * p,s64 delta_exec)1176 static inline void update_curr_task(struct task_struct *p, s64 delta_exec)
1177 {
1178 trace_sched_stat_runtime(p, delta_exec);
1179 account_group_exec_runtime(p, delta_exec);
1180 cgroup_account_cputime(p, delta_exec);
1181 }
1182
1183 /*
1184 * Used by other classes to account runtime.
1185 */
update_curr_common(struct rq * rq)1186 s64 update_curr_common(struct rq *rq)
1187 {
1188 struct task_struct *curr = rq->curr;
1189 s64 delta_exec;
1190
1191 delta_exec = update_curr_se(rq, &curr->se);
1192 if (likely(delta_exec > 0))
1193 update_curr_task(curr, delta_exec);
1194
1195 return delta_exec;
1196 }
1197
1198 /*
1199 * Update the current task's runtime statistics.
1200 */
update_curr(struct cfs_rq * cfs_rq)1201 static void update_curr(struct cfs_rq *cfs_rq)
1202 {
1203 struct sched_entity *curr = cfs_rq->curr;
1204 s64 delta_exec;
1205
1206 if (unlikely(!curr))
1207 return;
1208
1209 delta_exec = update_curr_se(rq_of(cfs_rq), curr);
1210 if (unlikely(delta_exec <= 0))
1211 return;
1212
1213 curr->vruntime += calc_delta_fair(delta_exec, curr);
1214 update_deadline(cfs_rq, curr);
1215 update_min_vruntime(cfs_rq);
1216
1217 if (entity_is_task(curr))
1218 update_curr_task(task_of(curr), delta_exec);
1219
1220 account_cfs_rq_runtime(cfs_rq, delta_exec);
1221 }
1222
update_curr_fair(struct rq * rq)1223 static void update_curr_fair(struct rq *rq)
1224 {
1225 update_curr(cfs_rq_of(&rq->curr->se));
1226 }
1227
1228 static inline void
update_stats_wait_start_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1229 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1230 {
1231 struct sched_statistics *stats;
1232 struct task_struct *p = NULL;
1233
1234 if (!schedstat_enabled())
1235 return;
1236
1237 stats = __schedstats_from_se(se);
1238
1239 if (entity_is_task(se))
1240 p = task_of(se);
1241
1242 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
1243 }
1244
1245 static inline void
update_stats_wait_end_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1246 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1247 {
1248 struct sched_statistics *stats;
1249 struct task_struct *p = NULL;
1250
1251 if (!schedstat_enabled())
1252 return;
1253
1254 stats = __schedstats_from_se(se);
1255
1256 /*
1257 * When the sched_schedstat changes from 0 to 1, some sched se
1258 * maybe already in the runqueue, the se->statistics.wait_start
1259 * will be 0.So it will let the delta wrong. We need to avoid this
1260 * scenario.
1261 */
1262 if (unlikely(!schedstat_val(stats->wait_start)))
1263 return;
1264
1265 if (entity_is_task(se))
1266 p = task_of(se);
1267
1268 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
1269 }
1270
1271 static inline void
update_stats_enqueue_sleeper_fair(struct cfs_rq * cfs_rq,struct sched_entity * se)1272 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1273 {
1274 struct sched_statistics *stats;
1275 struct task_struct *tsk = NULL;
1276
1277 if (!schedstat_enabled())
1278 return;
1279
1280 stats = __schedstats_from_se(se);
1281
1282 if (entity_is_task(se))
1283 tsk = task_of(se);
1284
1285 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1286 }
1287
1288 /*
1289 * Task is being enqueued - update stats:
1290 */
1291 static inline void
update_stats_enqueue_fair(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)1292 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1293 {
1294 if (!schedstat_enabled())
1295 return;
1296
1297 /*
1298 * Are we enqueueing a waiting task? (for current tasks
1299 * a dequeue/enqueue event is a NOP)
1300 */
1301 if (se != cfs_rq->curr)
1302 update_stats_wait_start_fair(cfs_rq, se);
1303
1304 if (flags & ENQUEUE_WAKEUP)
1305 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1306 }
1307
1308 static inline void
update_stats_dequeue_fair(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)1309 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1310 {
1311
1312 if (!schedstat_enabled())
1313 return;
1314
1315 /*
1316 * Mark the end of the wait period if dequeueing a
1317 * waiting task:
1318 */
1319 if (se != cfs_rq->curr)
1320 update_stats_wait_end_fair(cfs_rq, se);
1321
1322 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1323 struct task_struct *tsk = task_of(se);
1324 unsigned int state;
1325
1326 /* XXX racy against TTWU */
1327 state = READ_ONCE(tsk->__state);
1328 if (state & TASK_INTERRUPTIBLE)
1329 __schedstat_set(tsk->stats.sleep_start,
1330 rq_clock(rq_of(cfs_rq)));
1331 if (state & TASK_UNINTERRUPTIBLE)
1332 __schedstat_set(tsk->stats.block_start,
1333 rq_clock(rq_of(cfs_rq)));
1334 }
1335 }
1336
1337 /*
1338 * We are picking a new current task - update its stats:
1339 */
1340 static inline void
update_stats_curr_start(struct cfs_rq * cfs_rq,struct sched_entity * se)1341 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1342 {
1343 /*
1344 * We are starting a new run period:
1345 */
1346 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1347 }
1348
1349 /**************************************************
1350 * Scheduling class queueing methods:
1351 */
1352
is_core_idle(int cpu)1353 static inline bool is_core_idle(int cpu)
1354 {
1355 #ifdef CONFIG_SCHED_SMT
1356 int sibling;
1357
1358 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1359 if (cpu == sibling)
1360 continue;
1361
1362 if (!idle_cpu(sibling))
1363 return false;
1364 }
1365 #endif
1366
1367 return true;
1368 }
1369
1370 #ifdef CONFIG_NUMA
1371 #define NUMA_IMBALANCE_MIN 2
1372
1373 static inline long
adjust_numa_imbalance(int imbalance,int dst_running,int imb_numa_nr)1374 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1375 {
1376 /*
1377 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1378 * threshold. Above this threshold, individual tasks may be contending
1379 * for both memory bandwidth and any shared HT resources. This is an
1380 * approximation as the number of running tasks may not be related to
1381 * the number of busy CPUs due to sched_setaffinity.
1382 */
1383 if (dst_running > imb_numa_nr)
1384 return imbalance;
1385
1386 /*
1387 * Allow a small imbalance based on a simple pair of communicating
1388 * tasks that remain local when the destination is lightly loaded.
1389 */
1390 if (imbalance <= NUMA_IMBALANCE_MIN)
1391 return 0;
1392
1393 return imbalance;
1394 }
1395 #endif /* CONFIG_NUMA */
1396
1397 #ifdef CONFIG_NUMA_BALANCING
1398 /*
1399 * Approximate time to scan a full NUMA task in ms. The task scan period is
1400 * calculated based on the tasks virtual memory size and
1401 * numa_balancing_scan_size.
1402 */
1403 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1404 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1405
1406 /* Portion of address space to scan in MB */
1407 unsigned int sysctl_numa_balancing_scan_size = 256;
1408
1409 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1410 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1411
1412 /* The page with hint page fault latency < threshold in ms is considered hot */
1413 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1414
1415 struct numa_group {
1416 refcount_t refcount;
1417
1418 spinlock_t lock; /* nr_tasks, tasks */
1419 int nr_tasks;
1420 pid_t gid;
1421 int active_nodes;
1422
1423 struct rcu_head rcu;
1424 unsigned long total_faults;
1425 unsigned long max_faults_cpu;
1426 /*
1427 * faults[] array is split into two regions: faults_mem and faults_cpu.
1428 *
1429 * Faults_cpu is used to decide whether memory should move
1430 * towards the CPU. As a consequence, these stats are weighted
1431 * more by CPU use than by memory faults.
1432 */
1433 unsigned long faults[];
1434 };
1435
1436 /*
1437 * For functions that can be called in multiple contexts that permit reading
1438 * ->numa_group (see struct task_struct for locking rules).
1439 */
deref_task_numa_group(struct task_struct * p)1440 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1441 {
1442 return rcu_dereference_check(p->numa_group, p == current ||
1443 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1444 }
1445
deref_curr_numa_group(struct task_struct * p)1446 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1447 {
1448 return rcu_dereference_protected(p->numa_group, p == current);
1449 }
1450
1451 static inline unsigned long group_faults_priv(struct numa_group *ng);
1452 static inline unsigned long group_faults_shared(struct numa_group *ng);
1453
task_nr_scan_windows(struct task_struct * p)1454 static unsigned int task_nr_scan_windows(struct task_struct *p)
1455 {
1456 unsigned long rss = 0;
1457 unsigned long nr_scan_pages;
1458
1459 /*
1460 * Calculations based on RSS as non-present and empty pages are skipped
1461 * by the PTE scanner and NUMA hinting faults should be trapped based
1462 * on resident pages
1463 */
1464 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1465 rss = get_mm_rss(p->mm);
1466 if (!rss)
1467 rss = nr_scan_pages;
1468
1469 rss = round_up(rss, nr_scan_pages);
1470 return rss / nr_scan_pages;
1471 }
1472
1473 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1474 #define MAX_SCAN_WINDOW 2560
1475
task_scan_min(struct task_struct * p)1476 static unsigned int task_scan_min(struct task_struct *p)
1477 {
1478 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1479 unsigned int scan, floor;
1480 unsigned int windows = 1;
1481
1482 if (scan_size < MAX_SCAN_WINDOW)
1483 windows = MAX_SCAN_WINDOW / scan_size;
1484 floor = 1000 / windows;
1485
1486 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1487 return max_t(unsigned int, floor, scan);
1488 }
1489
task_scan_start(struct task_struct * p)1490 static unsigned int task_scan_start(struct task_struct *p)
1491 {
1492 unsigned long smin = task_scan_min(p);
1493 unsigned long period = smin;
1494 struct numa_group *ng;
1495
1496 /* Scale the maximum scan period with the amount of shared memory. */
1497 rcu_read_lock();
1498 ng = rcu_dereference(p->numa_group);
1499 if (ng) {
1500 unsigned long shared = group_faults_shared(ng);
1501 unsigned long private = group_faults_priv(ng);
1502
1503 period *= refcount_read(&ng->refcount);
1504 period *= shared + 1;
1505 period /= private + shared + 1;
1506 }
1507 rcu_read_unlock();
1508
1509 return max(smin, period);
1510 }
1511
task_scan_max(struct task_struct * p)1512 static unsigned int task_scan_max(struct task_struct *p)
1513 {
1514 unsigned long smin = task_scan_min(p);
1515 unsigned long smax;
1516 struct numa_group *ng;
1517
1518 /* Watch for min being lower than max due to floor calculations */
1519 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1520
1521 /* Scale the maximum scan period with the amount of shared memory. */
1522 ng = deref_curr_numa_group(p);
1523 if (ng) {
1524 unsigned long shared = group_faults_shared(ng);
1525 unsigned long private = group_faults_priv(ng);
1526 unsigned long period = smax;
1527
1528 period *= refcount_read(&ng->refcount);
1529 period *= shared + 1;
1530 period /= private + shared + 1;
1531
1532 smax = max(smax, period);
1533 }
1534
1535 return max(smin, smax);
1536 }
1537
account_numa_enqueue(struct rq * rq,struct task_struct * p)1538 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1539 {
1540 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1541 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1542 }
1543
account_numa_dequeue(struct rq * rq,struct task_struct * p)1544 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1545 {
1546 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1547 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1548 }
1549
1550 /* Shared or private faults. */
1551 #define NR_NUMA_HINT_FAULT_TYPES 2
1552
1553 /* Memory and CPU locality */
1554 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1555
1556 /* Averaged statistics, and temporary buffers. */
1557 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1558
task_numa_group_id(struct task_struct * p)1559 pid_t task_numa_group_id(struct task_struct *p)
1560 {
1561 struct numa_group *ng;
1562 pid_t gid = 0;
1563
1564 rcu_read_lock();
1565 ng = rcu_dereference(p->numa_group);
1566 if (ng)
1567 gid = ng->gid;
1568 rcu_read_unlock();
1569
1570 return gid;
1571 }
1572
1573 /*
1574 * The averaged statistics, shared & private, memory & CPU,
1575 * occupy the first half of the array. The second half of the
1576 * array is for current counters, which are averaged into the
1577 * first set by task_numa_placement.
1578 */
task_faults_idx(enum numa_faults_stats s,int nid,int priv)1579 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1580 {
1581 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1582 }
1583
task_faults(struct task_struct * p,int nid)1584 static inline unsigned long task_faults(struct task_struct *p, int nid)
1585 {
1586 if (!p->numa_faults)
1587 return 0;
1588
1589 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1590 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1591 }
1592
group_faults(struct task_struct * p,int nid)1593 static inline unsigned long group_faults(struct task_struct *p, int nid)
1594 {
1595 struct numa_group *ng = deref_task_numa_group(p);
1596
1597 if (!ng)
1598 return 0;
1599
1600 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1601 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1602 }
1603
group_faults_cpu(struct numa_group * group,int nid)1604 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1605 {
1606 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1607 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1608 }
1609
group_faults_priv(struct numa_group * ng)1610 static inline unsigned long group_faults_priv(struct numa_group *ng)
1611 {
1612 unsigned long faults = 0;
1613 int node;
1614
1615 for_each_online_node(node) {
1616 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1617 }
1618
1619 return faults;
1620 }
1621
group_faults_shared(struct numa_group * ng)1622 static inline unsigned long group_faults_shared(struct numa_group *ng)
1623 {
1624 unsigned long faults = 0;
1625 int node;
1626
1627 for_each_online_node(node) {
1628 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1629 }
1630
1631 return faults;
1632 }
1633
1634 /*
1635 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1636 * considered part of a numa group's pseudo-interleaving set. Migrations
1637 * between these nodes are slowed down, to allow things to settle down.
1638 */
1639 #define ACTIVE_NODE_FRACTION 3
1640
numa_is_active_node(int nid,struct numa_group * ng)1641 static bool numa_is_active_node(int nid, struct numa_group *ng)
1642 {
1643 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1644 }
1645
1646 /* Handle placement on systems where not all nodes are directly connected. */
score_nearby_nodes(struct task_struct * p,int nid,int lim_dist,bool task)1647 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1648 int lim_dist, bool task)
1649 {
1650 unsigned long score = 0;
1651 int node, max_dist;
1652
1653 /*
1654 * All nodes are directly connected, and the same distance
1655 * from each other. No need for fancy placement algorithms.
1656 */
1657 if (sched_numa_topology_type == NUMA_DIRECT)
1658 return 0;
1659
1660 /* sched_max_numa_distance may be changed in parallel. */
1661 max_dist = READ_ONCE(sched_max_numa_distance);
1662 /*
1663 * This code is called for each node, introducing N^2 complexity,
1664 * which should be ok given the number of nodes rarely exceeds 8.
1665 */
1666 for_each_online_node(node) {
1667 unsigned long faults;
1668 int dist = node_distance(nid, node);
1669
1670 /*
1671 * The furthest away nodes in the system are not interesting
1672 * for placement; nid was already counted.
1673 */
1674 if (dist >= max_dist || node == nid)
1675 continue;
1676
1677 /*
1678 * On systems with a backplane NUMA topology, compare groups
1679 * of nodes, and move tasks towards the group with the most
1680 * memory accesses. When comparing two nodes at distance
1681 * "hoplimit", only nodes closer by than "hoplimit" are part
1682 * of each group. Skip other nodes.
1683 */
1684 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1685 continue;
1686
1687 /* Add up the faults from nearby nodes. */
1688 if (task)
1689 faults = task_faults(p, node);
1690 else
1691 faults = group_faults(p, node);
1692
1693 /*
1694 * On systems with a glueless mesh NUMA topology, there are
1695 * no fixed "groups of nodes". Instead, nodes that are not
1696 * directly connected bounce traffic through intermediate
1697 * nodes; a numa_group can occupy any set of nodes.
1698 * The further away a node is, the less the faults count.
1699 * This seems to result in good task placement.
1700 */
1701 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1702 faults *= (max_dist - dist);
1703 faults /= (max_dist - LOCAL_DISTANCE);
1704 }
1705
1706 score += faults;
1707 }
1708
1709 return score;
1710 }
1711
1712 /*
1713 * These return the fraction of accesses done by a particular task, or
1714 * task group, on a particular numa node. The group weight is given a
1715 * larger multiplier, in order to group tasks together that are almost
1716 * evenly spread out between numa nodes.
1717 */
task_weight(struct task_struct * p,int nid,int dist)1718 static inline unsigned long task_weight(struct task_struct *p, int nid,
1719 int dist)
1720 {
1721 unsigned long faults, total_faults;
1722
1723 if (!p->numa_faults)
1724 return 0;
1725
1726 total_faults = p->total_numa_faults;
1727
1728 if (!total_faults)
1729 return 0;
1730
1731 faults = task_faults(p, nid);
1732 faults += score_nearby_nodes(p, nid, dist, true);
1733
1734 return 1000 * faults / total_faults;
1735 }
1736
group_weight(struct task_struct * p,int nid,int dist)1737 static inline unsigned long group_weight(struct task_struct *p, int nid,
1738 int dist)
1739 {
1740 struct numa_group *ng = deref_task_numa_group(p);
1741 unsigned long faults, total_faults;
1742
1743 if (!ng)
1744 return 0;
1745
1746 total_faults = ng->total_faults;
1747
1748 if (!total_faults)
1749 return 0;
1750
1751 faults = group_faults(p, nid);
1752 faults += score_nearby_nodes(p, nid, dist, false);
1753
1754 return 1000 * faults / total_faults;
1755 }
1756
1757 /*
1758 * If memory tiering mode is enabled, cpupid of slow memory page is
1759 * used to record scan time instead of CPU and PID. When tiering mode
1760 * is disabled at run time, the scan time (in cpupid) will be
1761 * interpreted as CPU and PID. So CPU needs to be checked to avoid to
1762 * access out of array bound.
1763 */
cpupid_valid(int cpupid)1764 static inline bool cpupid_valid(int cpupid)
1765 {
1766 return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1767 }
1768
1769 /*
1770 * For memory tiering mode, if there are enough free pages (more than
1771 * enough watermark defined here) in fast memory node, to take full
1772 * advantage of fast memory capacity, all recently accessed slow
1773 * memory pages will be migrated to fast memory node without
1774 * considering hot threshold.
1775 */
pgdat_free_space_enough(struct pglist_data * pgdat)1776 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1777 {
1778 int z;
1779 unsigned long enough_wmark;
1780
1781 enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1782 pgdat->node_present_pages >> 4);
1783 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1784 struct zone *zone = pgdat->node_zones + z;
1785
1786 if (!populated_zone(zone))
1787 continue;
1788
1789 if (zone_watermark_ok(zone, 0,
1790 wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1791 ZONE_MOVABLE, 0))
1792 return true;
1793 }
1794 return false;
1795 }
1796
1797 /*
1798 * For memory tiering mode, when page tables are scanned, the scan
1799 * time will be recorded in struct page in addition to make page
1800 * PROT_NONE for slow memory page. So when the page is accessed, in
1801 * hint page fault handler, the hint page fault latency is calculated
1802 * via,
1803 *
1804 * hint page fault latency = hint page fault time - scan time
1805 *
1806 * The smaller the hint page fault latency, the higher the possibility
1807 * for the page to be hot.
1808 */
numa_hint_fault_latency(struct page * page)1809 static int numa_hint_fault_latency(struct page *page)
1810 {
1811 int last_time, time;
1812
1813 time = jiffies_to_msecs(jiffies);
1814 last_time = xchg_page_access_time(page, time);
1815
1816 return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1817 }
1818
1819 /*
1820 * For memory tiering mode, too high promotion/demotion throughput may
1821 * hurt application latency. So we provide a mechanism to rate limit
1822 * the number of pages that are tried to be promoted.
1823 */
numa_promotion_rate_limit(struct pglist_data * pgdat,unsigned long rate_limit,int nr)1824 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1825 unsigned long rate_limit, int nr)
1826 {
1827 unsigned long nr_cand;
1828 unsigned int now, start;
1829
1830 now = jiffies_to_msecs(jiffies);
1831 mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1832 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1833 start = pgdat->nbp_rl_start;
1834 if (now - start > MSEC_PER_SEC &&
1835 cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1836 pgdat->nbp_rl_nr_cand = nr_cand;
1837 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1838 return true;
1839 return false;
1840 }
1841
1842 #define NUMA_MIGRATION_ADJUST_STEPS 16
1843
numa_promotion_adjust_threshold(struct pglist_data * pgdat,unsigned long rate_limit,unsigned int ref_th)1844 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1845 unsigned long rate_limit,
1846 unsigned int ref_th)
1847 {
1848 unsigned int now, start, th_period, unit_th, th;
1849 unsigned long nr_cand, ref_cand, diff_cand;
1850
1851 now = jiffies_to_msecs(jiffies);
1852 th_period = sysctl_numa_balancing_scan_period_max;
1853 start = pgdat->nbp_th_start;
1854 if (now - start > th_period &&
1855 cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1856 ref_cand = rate_limit *
1857 sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1858 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1859 diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1860 unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1861 th = pgdat->nbp_threshold ? : ref_th;
1862 if (diff_cand > ref_cand * 11 / 10)
1863 th = max(th - unit_th, unit_th);
1864 else if (diff_cand < ref_cand * 9 / 10)
1865 th = min(th + unit_th, ref_th * 2);
1866 pgdat->nbp_th_nr_cand = nr_cand;
1867 pgdat->nbp_threshold = th;
1868 }
1869 }
1870
should_numa_migrate_memory(struct task_struct * p,struct page * page,int src_nid,int dst_cpu)1871 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1872 int src_nid, int dst_cpu)
1873 {
1874 struct numa_group *ng = deref_curr_numa_group(p);
1875 int dst_nid = cpu_to_node(dst_cpu);
1876 int last_cpupid, this_cpupid;
1877
1878 /*
1879 * The pages in slow memory node should be migrated according
1880 * to hot/cold instead of private/shared.
1881 */
1882 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1883 !node_is_toptier(src_nid)) {
1884 struct pglist_data *pgdat;
1885 unsigned long rate_limit;
1886 unsigned int latency, th, def_th;
1887
1888 pgdat = NODE_DATA(dst_nid);
1889 if (pgdat_free_space_enough(pgdat)) {
1890 /* workload changed, reset hot threshold */
1891 pgdat->nbp_threshold = 0;
1892 return true;
1893 }
1894
1895 def_th = sysctl_numa_balancing_hot_threshold;
1896 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1897 (20 - PAGE_SHIFT);
1898 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1899
1900 th = pgdat->nbp_threshold ? : def_th;
1901 latency = numa_hint_fault_latency(page);
1902 if (latency >= th)
1903 return false;
1904
1905 return !numa_promotion_rate_limit(pgdat, rate_limit,
1906 thp_nr_pages(page));
1907 }
1908
1909 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1910 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1911
1912 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1913 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1914 return false;
1915
1916 /*
1917 * Allow first faults or private faults to migrate immediately early in
1918 * the lifetime of a task. The magic number 4 is based on waiting for
1919 * two full passes of the "multi-stage node selection" test that is
1920 * executed below.
1921 */
1922 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1923 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1924 return true;
1925
1926 /*
1927 * Multi-stage node selection is used in conjunction with a periodic
1928 * migration fault to build a temporal task<->page relation. By using
1929 * a two-stage filter we remove short/unlikely relations.
1930 *
1931 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1932 * a task's usage of a particular page (n_p) per total usage of this
1933 * page (n_t) (in a given time-span) to a probability.
1934 *
1935 * Our periodic faults will sample this probability and getting the
1936 * same result twice in a row, given these samples are fully
1937 * independent, is then given by P(n)^2, provided our sample period
1938 * is sufficiently short compared to the usage pattern.
1939 *
1940 * This quadric squishes small probabilities, making it less likely we
1941 * act on an unlikely task<->page relation.
1942 */
1943 if (!cpupid_pid_unset(last_cpupid) &&
1944 cpupid_to_nid(last_cpupid) != dst_nid)
1945 return false;
1946
1947 /* Always allow migrate on private faults */
1948 if (cpupid_match_pid(p, last_cpupid))
1949 return true;
1950
1951 /* A shared fault, but p->numa_group has not been set up yet. */
1952 if (!ng)
1953 return true;
1954
1955 /*
1956 * Destination node is much more heavily used than the source
1957 * node? Allow migration.
1958 */
1959 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1960 ACTIVE_NODE_FRACTION)
1961 return true;
1962
1963 /*
1964 * Distribute memory according to CPU & memory use on each node,
1965 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1966 *
1967 * faults_cpu(dst) 3 faults_cpu(src)
1968 * --------------- * - > ---------------
1969 * faults_mem(dst) 4 faults_mem(src)
1970 */
1971 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1972 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1973 }
1974
1975 /*
1976 * 'numa_type' describes the node at the moment of load balancing.
1977 */
1978 enum numa_type {
1979 /* The node has spare capacity that can be used to run more tasks. */
1980 node_has_spare = 0,
1981 /*
1982 * The node is fully used and the tasks don't compete for more CPU
1983 * cycles. Nevertheless, some tasks might wait before running.
1984 */
1985 node_fully_busy,
1986 /*
1987 * The node is overloaded and can't provide expected CPU cycles to all
1988 * tasks.
1989 */
1990 node_overloaded
1991 };
1992
1993 /* Cached statistics for all CPUs within a node */
1994 struct numa_stats {
1995 unsigned long load;
1996 unsigned long runnable;
1997 unsigned long util;
1998 /* Total compute capacity of CPUs on a node */
1999 unsigned long compute_capacity;
2000 unsigned int nr_running;
2001 unsigned int weight;
2002 enum numa_type node_type;
2003 int idle_cpu;
2004 };
2005
2006 struct task_numa_env {
2007 struct task_struct *p;
2008
2009 int src_cpu, src_nid;
2010 int dst_cpu, dst_nid;
2011 int imb_numa_nr;
2012
2013 struct numa_stats src_stats, dst_stats;
2014
2015 int imbalance_pct;
2016 int dist;
2017
2018 struct task_struct *best_task;
2019 long best_imp;
2020 int best_cpu;
2021 };
2022
2023 static unsigned long cpu_load(struct rq *rq);
2024 static unsigned long cpu_runnable(struct rq *rq);
2025
2026 static inline enum
numa_classify(unsigned int imbalance_pct,struct numa_stats * ns)2027 numa_type numa_classify(unsigned int imbalance_pct,
2028 struct numa_stats *ns)
2029 {
2030 if ((ns->nr_running > ns->weight) &&
2031 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
2032 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
2033 return node_overloaded;
2034
2035 if ((ns->nr_running < ns->weight) ||
2036 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
2037 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
2038 return node_has_spare;
2039
2040 return node_fully_busy;
2041 }
2042
2043 #ifdef CONFIG_SCHED_SMT
2044 /* Forward declarations of select_idle_sibling helpers */
2045 static inline bool test_idle_cores(int cpu);
numa_idle_core(int idle_core,int cpu)2046 static inline int numa_idle_core(int idle_core, int cpu)
2047 {
2048 if (!static_branch_likely(&sched_smt_present) ||
2049 idle_core >= 0 || !test_idle_cores(cpu))
2050 return idle_core;
2051
2052 /*
2053 * Prefer cores instead of packing HT siblings
2054 * and triggering future load balancing.
2055 */
2056 if (is_core_idle(cpu))
2057 idle_core = cpu;
2058
2059 return idle_core;
2060 }
2061 #else
numa_idle_core(int idle_core,int cpu)2062 static inline int numa_idle_core(int idle_core, int cpu)
2063 {
2064 return idle_core;
2065 }
2066 #endif
2067
2068 /*
2069 * Gather all necessary information to make NUMA balancing placement
2070 * decisions that are compatible with standard load balancer. This
2071 * borrows code and logic from update_sg_lb_stats but sharing a
2072 * common implementation is impractical.
2073 */
update_numa_stats(struct task_numa_env * env,struct numa_stats * ns,int nid,bool find_idle)2074 static void update_numa_stats(struct task_numa_env *env,
2075 struct numa_stats *ns, int nid,
2076 bool find_idle)
2077 {
2078 int cpu, idle_core = -1;
2079
2080 memset(ns, 0, sizeof(*ns));
2081 ns->idle_cpu = -1;
2082
2083 rcu_read_lock();
2084 for_each_cpu(cpu, cpumask_of_node(nid)) {
2085 struct rq *rq = cpu_rq(cpu);
2086
2087 ns->load += cpu_load(rq);
2088 ns->runnable += cpu_runnable(rq);
2089 ns->util += cpu_util_cfs(cpu);
2090 ns->nr_running += rq->cfs.h_nr_running;
2091 ns->compute_capacity += capacity_of(cpu);
2092
2093 if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2094 if (READ_ONCE(rq->numa_migrate_on) ||
2095 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2096 continue;
2097
2098 if (ns->idle_cpu == -1)
2099 ns->idle_cpu = cpu;
2100
2101 idle_core = numa_idle_core(idle_core, cpu);
2102 }
2103 }
2104 rcu_read_unlock();
2105
2106 ns->weight = cpumask_weight(cpumask_of_node(nid));
2107
2108 ns->node_type = numa_classify(env->imbalance_pct, ns);
2109
2110 if (idle_core >= 0)
2111 ns->idle_cpu = idle_core;
2112 }
2113
task_numa_assign(struct task_numa_env * env,struct task_struct * p,long imp)2114 static void task_numa_assign(struct task_numa_env *env,
2115 struct task_struct *p, long imp)
2116 {
2117 struct rq *rq = cpu_rq(env->dst_cpu);
2118
2119 /* Check if run-queue part of active NUMA balance. */
2120 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2121 int cpu;
2122 int start = env->dst_cpu;
2123
2124 /* Find alternative idle CPU. */
2125 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2126 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2127 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2128 continue;
2129 }
2130
2131 env->dst_cpu = cpu;
2132 rq = cpu_rq(env->dst_cpu);
2133 if (!xchg(&rq->numa_migrate_on, 1))
2134 goto assign;
2135 }
2136
2137 /* Failed to find an alternative idle CPU */
2138 return;
2139 }
2140
2141 assign:
2142 /*
2143 * Clear previous best_cpu/rq numa-migrate flag, since task now
2144 * found a better CPU to move/swap.
2145 */
2146 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2147 rq = cpu_rq(env->best_cpu);
2148 WRITE_ONCE(rq->numa_migrate_on, 0);
2149 }
2150
2151 if (env->best_task)
2152 put_task_struct(env->best_task);
2153 if (p)
2154 get_task_struct(p);
2155
2156 env->best_task = p;
2157 env->best_imp = imp;
2158 env->best_cpu = env->dst_cpu;
2159 }
2160
load_too_imbalanced(long src_load,long dst_load,struct task_numa_env * env)2161 static bool load_too_imbalanced(long src_load, long dst_load,
2162 struct task_numa_env *env)
2163 {
2164 long imb, old_imb;
2165 long orig_src_load, orig_dst_load;
2166 long src_capacity, dst_capacity;
2167
2168 /*
2169 * The load is corrected for the CPU capacity available on each node.
2170 *
2171 * src_load dst_load
2172 * ------------ vs ---------
2173 * src_capacity dst_capacity
2174 */
2175 src_capacity = env->src_stats.compute_capacity;
2176 dst_capacity = env->dst_stats.compute_capacity;
2177
2178 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2179
2180 orig_src_load = env->src_stats.load;
2181 orig_dst_load = env->dst_stats.load;
2182
2183 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2184
2185 /* Would this change make things worse? */
2186 return (imb > old_imb);
2187 }
2188
2189 /*
2190 * Maximum NUMA importance can be 1998 (2*999);
2191 * SMALLIMP @ 30 would be close to 1998/64.
2192 * Used to deter task migration.
2193 */
2194 #define SMALLIMP 30
2195
2196 /*
2197 * This checks if the overall compute and NUMA accesses of the system would
2198 * be improved if the source tasks was migrated to the target dst_cpu taking
2199 * into account that it might be best if task running on the dst_cpu should
2200 * be exchanged with the source task
2201 */
task_numa_compare(struct task_numa_env * env,long taskimp,long groupimp,bool maymove)2202 static bool task_numa_compare(struct task_numa_env *env,
2203 long taskimp, long groupimp, bool maymove)
2204 {
2205 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2206 struct rq *dst_rq = cpu_rq(env->dst_cpu);
2207 long imp = p_ng ? groupimp : taskimp;
2208 struct task_struct *cur;
2209 long src_load, dst_load;
2210 int dist = env->dist;
2211 long moveimp = imp;
2212 long load;
2213 bool stopsearch = false;
2214
2215 if (READ_ONCE(dst_rq->numa_migrate_on))
2216 return false;
2217
2218 rcu_read_lock();
2219 cur = rcu_dereference(dst_rq->curr);
2220 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2221 cur = NULL;
2222
2223 /*
2224 * Because we have preemption enabled we can get migrated around and
2225 * end try selecting ourselves (current == env->p) as a swap candidate.
2226 */
2227 if (cur == env->p) {
2228 stopsearch = true;
2229 goto unlock;
2230 }
2231
2232 if (!cur) {
2233 if (maymove && moveimp >= env->best_imp)
2234 goto assign;
2235 else
2236 goto unlock;
2237 }
2238
2239 /* Skip this swap candidate if cannot move to the source cpu. */
2240 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2241 goto unlock;
2242
2243 /*
2244 * Skip this swap candidate if it is not moving to its preferred
2245 * node and the best task is.
2246 */
2247 if (env->best_task &&
2248 env->best_task->numa_preferred_nid == env->src_nid &&
2249 cur->numa_preferred_nid != env->src_nid) {
2250 goto unlock;
2251 }
2252
2253 /*
2254 * "imp" is the fault differential for the source task between the
2255 * source and destination node. Calculate the total differential for
2256 * the source task and potential destination task. The more negative
2257 * the value is, the more remote accesses that would be expected to
2258 * be incurred if the tasks were swapped.
2259 *
2260 * If dst and source tasks are in the same NUMA group, or not
2261 * in any group then look only at task weights.
2262 */
2263 cur_ng = rcu_dereference(cur->numa_group);
2264 if (cur_ng == p_ng) {
2265 /*
2266 * Do not swap within a group or between tasks that have
2267 * no group if there is spare capacity. Swapping does
2268 * not address the load imbalance and helps one task at
2269 * the cost of punishing another.
2270 */
2271 if (env->dst_stats.node_type == node_has_spare)
2272 goto unlock;
2273
2274 imp = taskimp + task_weight(cur, env->src_nid, dist) -
2275 task_weight(cur, env->dst_nid, dist);
2276 /*
2277 * Add some hysteresis to prevent swapping the
2278 * tasks within a group over tiny differences.
2279 */
2280 if (cur_ng)
2281 imp -= imp / 16;
2282 } else {
2283 /*
2284 * Compare the group weights. If a task is all by itself
2285 * (not part of a group), use the task weight instead.
2286 */
2287 if (cur_ng && p_ng)
2288 imp += group_weight(cur, env->src_nid, dist) -
2289 group_weight(cur, env->dst_nid, dist);
2290 else
2291 imp += task_weight(cur, env->src_nid, dist) -
2292 task_weight(cur, env->dst_nid, dist);
2293 }
2294
2295 /* Discourage picking a task already on its preferred node */
2296 if (cur->numa_preferred_nid == env->dst_nid)
2297 imp -= imp / 16;
2298
2299 /*
2300 * Encourage picking a task that moves to its preferred node.
2301 * This potentially makes imp larger than it's maximum of
2302 * 1998 (see SMALLIMP and task_weight for why) but in this
2303 * case, it does not matter.
2304 */
2305 if (cur->numa_preferred_nid == env->src_nid)
2306 imp += imp / 8;
2307
2308 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2309 imp = moveimp;
2310 cur = NULL;
2311 goto assign;
2312 }
2313
2314 /*
2315 * Prefer swapping with a task moving to its preferred node over a
2316 * task that is not.
2317 */
2318 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2319 env->best_task->numa_preferred_nid != env->src_nid) {
2320 goto assign;
2321 }
2322
2323 /*
2324 * If the NUMA importance is less than SMALLIMP,
2325 * task migration might only result in ping pong
2326 * of tasks and also hurt performance due to cache
2327 * misses.
2328 */
2329 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2330 goto unlock;
2331
2332 /*
2333 * In the overloaded case, try and keep the load balanced.
2334 */
2335 load = task_h_load(env->p) - task_h_load(cur);
2336 if (!load)
2337 goto assign;
2338
2339 dst_load = env->dst_stats.load + load;
2340 src_load = env->src_stats.load - load;
2341
2342 if (load_too_imbalanced(src_load, dst_load, env))
2343 goto unlock;
2344
2345 assign:
2346 /* Evaluate an idle CPU for a task numa move. */
2347 if (!cur) {
2348 int cpu = env->dst_stats.idle_cpu;
2349
2350 /* Nothing cached so current CPU went idle since the search. */
2351 if (cpu < 0)
2352 cpu = env->dst_cpu;
2353
2354 /*
2355 * If the CPU is no longer truly idle and the previous best CPU
2356 * is, keep using it.
2357 */
2358 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2359 idle_cpu(env->best_cpu)) {
2360 cpu = env->best_cpu;
2361 }
2362
2363 env->dst_cpu = cpu;
2364 }
2365
2366 task_numa_assign(env, cur, imp);
2367
2368 /*
2369 * If a move to idle is allowed because there is capacity or load
2370 * balance improves then stop the search. While a better swap
2371 * candidate may exist, a search is not free.
2372 */
2373 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2374 stopsearch = true;
2375
2376 /*
2377 * If a swap candidate must be identified and the current best task
2378 * moves its preferred node then stop the search.
2379 */
2380 if (!maymove && env->best_task &&
2381 env->best_task->numa_preferred_nid == env->src_nid) {
2382 stopsearch = true;
2383 }
2384 unlock:
2385 rcu_read_unlock();
2386
2387 return stopsearch;
2388 }
2389
task_numa_find_cpu(struct task_numa_env * env,long taskimp,long groupimp)2390 static void task_numa_find_cpu(struct task_numa_env *env,
2391 long taskimp, long groupimp)
2392 {
2393 bool maymove = false;
2394 int cpu;
2395
2396 /*
2397 * If dst node has spare capacity, then check if there is an
2398 * imbalance that would be overruled by the load balancer.
2399 */
2400 if (env->dst_stats.node_type == node_has_spare) {
2401 unsigned int imbalance;
2402 int src_running, dst_running;
2403
2404 /*
2405 * Would movement cause an imbalance? Note that if src has
2406 * more running tasks that the imbalance is ignored as the
2407 * move improves the imbalance from the perspective of the
2408 * CPU load balancer.
2409 * */
2410 src_running = env->src_stats.nr_running - 1;
2411 dst_running = env->dst_stats.nr_running + 1;
2412 imbalance = max(0, dst_running - src_running);
2413 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2414 env->imb_numa_nr);
2415
2416 /* Use idle CPU if there is no imbalance */
2417 if (!imbalance) {
2418 maymove = true;
2419 if (env->dst_stats.idle_cpu >= 0) {
2420 env->dst_cpu = env->dst_stats.idle_cpu;
2421 task_numa_assign(env, NULL, 0);
2422 return;
2423 }
2424 }
2425 } else {
2426 long src_load, dst_load, load;
2427 /*
2428 * If the improvement from just moving env->p direction is better
2429 * than swapping tasks around, check if a move is possible.
2430 */
2431 load = task_h_load(env->p);
2432 dst_load = env->dst_stats.load + load;
2433 src_load = env->src_stats.load - load;
2434 maymove = !load_too_imbalanced(src_load, dst_load, env);
2435 }
2436
2437 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2438 /* Skip this CPU if the source task cannot migrate */
2439 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2440 continue;
2441
2442 env->dst_cpu = cpu;
2443 if (task_numa_compare(env, taskimp, groupimp, maymove))
2444 break;
2445 }
2446 }
2447
task_numa_migrate(struct task_struct * p)2448 static int task_numa_migrate(struct task_struct *p)
2449 {
2450 struct task_numa_env env = {
2451 .p = p,
2452
2453 .src_cpu = task_cpu(p),
2454 .src_nid = task_node(p),
2455
2456 .imbalance_pct = 112,
2457
2458 .best_task = NULL,
2459 .best_imp = 0,
2460 .best_cpu = -1,
2461 };
2462 unsigned long taskweight, groupweight;
2463 struct sched_domain *sd;
2464 long taskimp, groupimp;
2465 struct numa_group *ng;
2466 struct rq *best_rq;
2467 int nid, ret, dist;
2468
2469 /*
2470 * Pick the lowest SD_NUMA domain, as that would have the smallest
2471 * imbalance and would be the first to start moving tasks about.
2472 *
2473 * And we want to avoid any moving of tasks about, as that would create
2474 * random movement of tasks -- counter the numa conditions we're trying
2475 * to satisfy here.
2476 */
2477 rcu_read_lock();
2478 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2479 if (sd) {
2480 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2481 env.imb_numa_nr = sd->imb_numa_nr;
2482 }
2483 rcu_read_unlock();
2484
2485 /*
2486 * Cpusets can break the scheduler domain tree into smaller
2487 * balance domains, some of which do not cross NUMA boundaries.
2488 * Tasks that are "trapped" in such domains cannot be migrated
2489 * elsewhere, so there is no point in (re)trying.
2490 */
2491 if (unlikely(!sd)) {
2492 sched_setnuma(p, task_node(p));
2493 return -EINVAL;
2494 }
2495
2496 env.dst_nid = p->numa_preferred_nid;
2497 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2498 taskweight = task_weight(p, env.src_nid, dist);
2499 groupweight = group_weight(p, env.src_nid, dist);
2500 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2501 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2502 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2503 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2504
2505 /* Try to find a spot on the preferred nid. */
2506 task_numa_find_cpu(&env, taskimp, groupimp);
2507
2508 /*
2509 * Look at other nodes in these cases:
2510 * - there is no space available on the preferred_nid
2511 * - the task is part of a numa_group that is interleaved across
2512 * multiple NUMA nodes; in order to better consolidate the group,
2513 * we need to check other locations.
2514 */
2515 ng = deref_curr_numa_group(p);
2516 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2517 for_each_node_state(nid, N_CPU) {
2518 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2519 continue;
2520
2521 dist = node_distance(env.src_nid, env.dst_nid);
2522 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2523 dist != env.dist) {
2524 taskweight = task_weight(p, env.src_nid, dist);
2525 groupweight = group_weight(p, env.src_nid, dist);
2526 }
2527
2528 /* Only consider nodes where both task and groups benefit */
2529 taskimp = task_weight(p, nid, dist) - taskweight;
2530 groupimp = group_weight(p, nid, dist) - groupweight;
2531 if (taskimp < 0 && groupimp < 0)
2532 continue;
2533
2534 env.dist = dist;
2535 env.dst_nid = nid;
2536 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2537 task_numa_find_cpu(&env, taskimp, groupimp);
2538 }
2539 }
2540
2541 /*
2542 * If the task is part of a workload that spans multiple NUMA nodes,
2543 * and is migrating into one of the workload's active nodes, remember
2544 * this node as the task's preferred numa node, so the workload can
2545 * settle down.
2546 * A task that migrated to a second choice node will be better off
2547 * trying for a better one later. Do not set the preferred node here.
2548 */
2549 if (ng) {
2550 if (env.best_cpu == -1)
2551 nid = env.src_nid;
2552 else
2553 nid = cpu_to_node(env.best_cpu);
2554
2555 if (nid != p->numa_preferred_nid)
2556 sched_setnuma(p, nid);
2557 }
2558
2559 /* No better CPU than the current one was found. */
2560 if (env.best_cpu == -1) {
2561 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2562 return -EAGAIN;
2563 }
2564
2565 best_rq = cpu_rq(env.best_cpu);
2566 if (env.best_task == NULL) {
2567 ret = migrate_task_to(p, env.best_cpu);
2568 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2569 if (ret != 0)
2570 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2571 return ret;
2572 }
2573
2574 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2575 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2576
2577 if (ret != 0)
2578 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2579 put_task_struct(env.best_task);
2580 return ret;
2581 }
2582
2583 /* Attempt to migrate a task to a CPU on the preferred node. */
numa_migrate_preferred(struct task_struct * p)2584 static void numa_migrate_preferred(struct task_struct *p)
2585 {
2586 unsigned long interval = HZ;
2587
2588 /* This task has no NUMA fault statistics yet */
2589 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2590 return;
2591
2592 /* Periodically retry migrating the task to the preferred node */
2593 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2594 p->numa_migrate_retry = jiffies + interval;
2595
2596 /* Success if task is already running on preferred CPU */
2597 if (task_node(p) == p->numa_preferred_nid)
2598 return;
2599
2600 /* Otherwise, try migrate to a CPU on the preferred node */
2601 task_numa_migrate(p);
2602 }
2603
2604 /*
2605 * Find out how many nodes the workload is actively running on. Do this by
2606 * tracking the nodes from which NUMA hinting faults are triggered. This can
2607 * be different from the set of nodes where the workload's memory is currently
2608 * located.
2609 */
numa_group_count_active_nodes(struct numa_group * numa_group)2610 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2611 {
2612 unsigned long faults, max_faults = 0;
2613 int nid, active_nodes = 0;
2614
2615 for_each_node_state(nid, N_CPU) {
2616 faults = group_faults_cpu(numa_group, nid);
2617 if (faults > max_faults)
2618 max_faults = faults;
2619 }
2620
2621 for_each_node_state(nid, N_CPU) {
2622 faults = group_faults_cpu(numa_group, nid);
2623 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2624 active_nodes++;
2625 }
2626
2627 numa_group->max_faults_cpu = max_faults;
2628 numa_group->active_nodes = active_nodes;
2629 }
2630
2631 /*
2632 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2633 * increments. The more local the fault statistics are, the higher the scan
2634 * period will be for the next scan window. If local/(local+remote) ratio is
2635 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2636 * the scan period will decrease. Aim for 70% local accesses.
2637 */
2638 #define NUMA_PERIOD_SLOTS 10
2639 #define NUMA_PERIOD_THRESHOLD 7
2640
2641 /*
2642 * Increase the scan period (slow down scanning) if the majority of
2643 * our memory is already on our local node, or if the majority of
2644 * the page accesses are shared with other processes.
2645 * Otherwise, decrease the scan period.
2646 */
update_task_scan_period(struct task_struct * p,unsigned long shared,unsigned long private)2647 static void update_task_scan_period(struct task_struct *p,
2648 unsigned long shared, unsigned long private)
2649 {
2650 unsigned int period_slot;
2651 int lr_ratio, ps_ratio;
2652 int diff;
2653
2654 unsigned long remote = p->numa_faults_locality[0];
2655 unsigned long local = p->numa_faults_locality[1];
2656
2657 /*
2658 * If there were no record hinting faults then either the task is
2659 * completely idle or all activity is in areas that are not of interest
2660 * to automatic numa balancing. Related to that, if there were failed
2661 * migration then it implies we are migrating too quickly or the local
2662 * node is overloaded. In either case, scan slower
2663 */
2664 if (local + shared == 0 || p->numa_faults_locality[2]) {
2665 p->numa_scan_period = min(p->numa_scan_period_max,
2666 p->numa_scan_period << 1);
2667
2668 p->mm->numa_next_scan = jiffies +
2669 msecs_to_jiffies(p->numa_scan_period);
2670
2671 return;
2672 }
2673
2674 /*
2675 * Prepare to scale scan period relative to the current period.
2676 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2677 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2678 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2679 */
2680 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2681 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2682 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2683
2684 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2685 /*
2686 * Most memory accesses are local. There is no need to
2687 * do fast NUMA scanning, since memory is already local.
2688 */
2689 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2690 if (!slot)
2691 slot = 1;
2692 diff = slot * period_slot;
2693 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2694 /*
2695 * Most memory accesses are shared with other tasks.
2696 * There is no point in continuing fast NUMA scanning,
2697 * since other tasks may just move the memory elsewhere.
2698 */
2699 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2700 if (!slot)
2701 slot = 1;
2702 diff = slot * period_slot;
2703 } else {
2704 /*
2705 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2706 * yet they are not on the local NUMA node. Speed up
2707 * NUMA scanning to get the memory moved over.
2708 */
2709 int ratio = max(lr_ratio, ps_ratio);
2710 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2711 }
2712
2713 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2714 task_scan_min(p), task_scan_max(p));
2715 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2716 }
2717
2718 /*
2719 * Get the fraction of time the task has been running since the last
2720 * NUMA placement cycle. The scheduler keeps similar statistics, but
2721 * decays those on a 32ms period, which is orders of magnitude off
2722 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2723 * stats only if the task is so new there are no NUMA statistics yet.
2724 */
numa_get_avg_runtime(struct task_struct * p,u64 * period)2725 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2726 {
2727 u64 runtime, delta, now;
2728 /* Use the start of this time slice to avoid calculations. */
2729 now = p->se.exec_start;
2730 runtime = p->se.sum_exec_runtime;
2731
2732 if (p->last_task_numa_placement) {
2733 delta = runtime - p->last_sum_exec_runtime;
2734 *period = now - p->last_task_numa_placement;
2735
2736 /* Avoid time going backwards, prevent potential divide error: */
2737 if (unlikely((s64)*period < 0))
2738 *period = 0;
2739 } else {
2740 delta = p->se.avg.load_sum;
2741 *period = LOAD_AVG_MAX;
2742 }
2743
2744 p->last_sum_exec_runtime = runtime;
2745 p->last_task_numa_placement = now;
2746
2747 return delta;
2748 }
2749
2750 /*
2751 * Determine the preferred nid for a task in a numa_group. This needs to
2752 * be done in a way that produces consistent results with group_weight,
2753 * otherwise workloads might not converge.
2754 */
preferred_group_nid(struct task_struct * p,int nid)2755 static int preferred_group_nid(struct task_struct *p, int nid)
2756 {
2757 nodemask_t nodes;
2758 int dist;
2759
2760 /* Direct connections between all NUMA nodes. */
2761 if (sched_numa_topology_type == NUMA_DIRECT)
2762 return nid;
2763
2764 /*
2765 * On a system with glueless mesh NUMA topology, group_weight
2766 * scores nodes according to the number of NUMA hinting faults on
2767 * both the node itself, and on nearby nodes.
2768 */
2769 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2770 unsigned long score, max_score = 0;
2771 int node, max_node = nid;
2772
2773 dist = sched_max_numa_distance;
2774
2775 for_each_node_state(node, N_CPU) {
2776 score = group_weight(p, node, dist);
2777 if (score > max_score) {
2778 max_score = score;
2779 max_node = node;
2780 }
2781 }
2782 return max_node;
2783 }
2784
2785 /*
2786 * Finding the preferred nid in a system with NUMA backplane
2787 * interconnect topology is more involved. The goal is to locate
2788 * tasks from numa_groups near each other in the system, and
2789 * untangle workloads from different sides of the system. This requires
2790 * searching down the hierarchy of node groups, recursively searching
2791 * inside the highest scoring group of nodes. The nodemask tricks
2792 * keep the complexity of the search down.
2793 */
2794 nodes = node_states[N_CPU];
2795 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2796 unsigned long max_faults = 0;
2797 nodemask_t max_group = NODE_MASK_NONE;
2798 int a, b;
2799
2800 /* Are there nodes at this distance from each other? */
2801 if (!find_numa_distance(dist))
2802 continue;
2803
2804 for_each_node_mask(a, nodes) {
2805 unsigned long faults = 0;
2806 nodemask_t this_group;
2807 nodes_clear(this_group);
2808
2809 /* Sum group's NUMA faults; includes a==b case. */
2810 for_each_node_mask(b, nodes) {
2811 if (node_distance(a, b) < dist) {
2812 faults += group_faults(p, b);
2813 node_set(b, this_group);
2814 node_clear(b, nodes);
2815 }
2816 }
2817
2818 /* Remember the top group. */
2819 if (faults > max_faults) {
2820 max_faults = faults;
2821 max_group = this_group;
2822 /*
2823 * subtle: at the smallest distance there is
2824 * just one node left in each "group", the
2825 * winner is the preferred nid.
2826 */
2827 nid = a;
2828 }
2829 }
2830 /* Next round, evaluate the nodes within max_group. */
2831 if (!max_faults)
2832 break;
2833 nodes = max_group;
2834 }
2835 return nid;
2836 }
2837
task_numa_placement(struct task_struct * p)2838 static void task_numa_placement(struct task_struct *p)
2839 {
2840 int seq, nid, max_nid = NUMA_NO_NODE;
2841 unsigned long max_faults = 0;
2842 unsigned long fault_types[2] = { 0, 0 };
2843 unsigned long total_faults;
2844 u64 runtime, period;
2845 spinlock_t *group_lock = NULL;
2846 struct numa_group *ng;
2847
2848 /*
2849 * The p->mm->numa_scan_seq field gets updated without
2850 * exclusive access. Use READ_ONCE() here to ensure
2851 * that the field is read in a single access:
2852 */
2853 seq = READ_ONCE(p->mm->numa_scan_seq);
2854 if (p->numa_scan_seq == seq)
2855 return;
2856 p->numa_scan_seq = seq;
2857 p->numa_scan_period_max = task_scan_max(p);
2858
2859 total_faults = p->numa_faults_locality[0] +
2860 p->numa_faults_locality[1];
2861 runtime = numa_get_avg_runtime(p, &period);
2862
2863 /* If the task is part of a group prevent parallel updates to group stats */
2864 ng = deref_curr_numa_group(p);
2865 if (ng) {
2866 group_lock = &ng->lock;
2867 spin_lock_irq(group_lock);
2868 }
2869
2870 /* Find the node with the highest number of faults */
2871 for_each_online_node(nid) {
2872 /* Keep track of the offsets in numa_faults array */
2873 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2874 unsigned long faults = 0, group_faults = 0;
2875 int priv;
2876
2877 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2878 long diff, f_diff, f_weight;
2879
2880 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2881 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2882 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2883 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2884
2885 /* Decay existing window, copy faults since last scan */
2886 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2887 fault_types[priv] += p->numa_faults[membuf_idx];
2888 p->numa_faults[membuf_idx] = 0;
2889
2890 /*
2891 * Normalize the faults_from, so all tasks in a group
2892 * count according to CPU use, instead of by the raw
2893 * number of faults. Tasks with little runtime have
2894 * little over-all impact on throughput, and thus their
2895 * faults are less important.
2896 */
2897 f_weight = div64_u64(runtime << 16, period + 1);
2898 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2899 (total_faults + 1);
2900 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2901 p->numa_faults[cpubuf_idx] = 0;
2902
2903 p->numa_faults[mem_idx] += diff;
2904 p->numa_faults[cpu_idx] += f_diff;
2905 faults += p->numa_faults[mem_idx];
2906 p->total_numa_faults += diff;
2907 if (ng) {
2908 /*
2909 * safe because we can only change our own group
2910 *
2911 * mem_idx represents the offset for a given
2912 * nid and priv in a specific region because it
2913 * is at the beginning of the numa_faults array.
2914 */
2915 ng->faults[mem_idx] += diff;
2916 ng->faults[cpu_idx] += f_diff;
2917 ng->total_faults += diff;
2918 group_faults += ng->faults[mem_idx];
2919 }
2920 }
2921
2922 if (!ng) {
2923 if (faults > max_faults) {
2924 max_faults = faults;
2925 max_nid = nid;
2926 }
2927 } else if (group_faults > max_faults) {
2928 max_faults = group_faults;
2929 max_nid = nid;
2930 }
2931 }
2932
2933 /* Cannot migrate task to CPU-less node */
2934 if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2935 int near_nid = max_nid;
2936 int distance, near_distance = INT_MAX;
2937
2938 for_each_node_state(nid, N_CPU) {
2939 distance = node_distance(max_nid, nid);
2940 if (distance < near_distance) {
2941 near_nid = nid;
2942 near_distance = distance;
2943 }
2944 }
2945 max_nid = near_nid;
2946 }
2947
2948 if (ng) {
2949 numa_group_count_active_nodes(ng);
2950 spin_unlock_irq(group_lock);
2951 max_nid = preferred_group_nid(p, max_nid);
2952 }
2953
2954 if (max_faults) {
2955 /* Set the new preferred node */
2956 if (max_nid != p->numa_preferred_nid)
2957 sched_setnuma(p, max_nid);
2958 }
2959
2960 update_task_scan_period(p, fault_types[0], fault_types[1]);
2961 }
2962
get_numa_group(struct numa_group * grp)2963 static inline int get_numa_group(struct numa_group *grp)
2964 {
2965 return refcount_inc_not_zero(&grp->refcount);
2966 }
2967
put_numa_group(struct numa_group * grp)2968 static inline void put_numa_group(struct numa_group *grp)
2969 {
2970 if (refcount_dec_and_test(&grp->refcount))
2971 kfree_rcu(grp, rcu);
2972 }
2973
task_numa_group(struct task_struct * p,int cpupid,int flags,int * priv)2974 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2975 int *priv)
2976 {
2977 struct numa_group *grp, *my_grp;
2978 struct task_struct *tsk;
2979 bool join = false;
2980 int cpu = cpupid_to_cpu(cpupid);
2981 int i;
2982
2983 if (unlikely(!deref_curr_numa_group(p))) {
2984 unsigned int size = sizeof(struct numa_group) +
2985 NR_NUMA_HINT_FAULT_STATS *
2986 nr_node_ids * sizeof(unsigned long);
2987
2988 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2989 if (!grp)
2990 return;
2991
2992 refcount_set(&grp->refcount, 1);
2993 grp->active_nodes = 1;
2994 grp->max_faults_cpu = 0;
2995 spin_lock_init(&grp->lock);
2996 grp->gid = p->pid;
2997
2998 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2999 grp->faults[i] = p->numa_faults[i];
3000
3001 grp->total_faults = p->total_numa_faults;
3002
3003 grp->nr_tasks++;
3004 rcu_assign_pointer(p->numa_group, grp);
3005 }
3006
3007 rcu_read_lock();
3008 tsk = READ_ONCE(cpu_rq(cpu)->curr);
3009
3010 if (!cpupid_match_pid(tsk, cpupid))
3011 goto no_join;
3012
3013 grp = rcu_dereference(tsk->numa_group);
3014 if (!grp)
3015 goto no_join;
3016
3017 my_grp = deref_curr_numa_group(p);
3018 if (grp == my_grp)
3019 goto no_join;
3020
3021 /*
3022 * Only join the other group if its bigger; if we're the bigger group,
3023 * the other task will join us.
3024 */
3025 if (my_grp->nr_tasks > grp->nr_tasks)
3026 goto no_join;
3027
3028 /*
3029 * Tie-break on the grp address.
3030 */
3031 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
3032 goto no_join;
3033
3034 /* Always join threads in the same process. */
3035 if (tsk->mm == current->mm)
3036 join = true;
3037
3038 /* Simple filter to avoid false positives due to PID collisions */
3039 if (flags & TNF_SHARED)
3040 join = true;
3041
3042 /* Update priv based on whether false sharing was detected */
3043 *priv = !join;
3044
3045 if (join && !get_numa_group(grp))
3046 goto no_join;
3047
3048 rcu_read_unlock();
3049
3050 if (!join)
3051 return;
3052
3053 WARN_ON_ONCE(irqs_disabled());
3054 double_lock_irq(&my_grp->lock, &grp->lock);
3055
3056 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3057 my_grp->faults[i] -= p->numa_faults[i];
3058 grp->faults[i] += p->numa_faults[i];
3059 }
3060 my_grp->total_faults -= p->total_numa_faults;
3061 grp->total_faults += p->total_numa_faults;
3062
3063 my_grp->nr_tasks--;
3064 grp->nr_tasks++;
3065
3066 spin_unlock(&my_grp->lock);
3067 spin_unlock_irq(&grp->lock);
3068
3069 rcu_assign_pointer(p->numa_group, grp);
3070
3071 put_numa_group(my_grp);
3072 return;
3073
3074 no_join:
3075 rcu_read_unlock();
3076 return;
3077 }
3078
3079 /*
3080 * Get rid of NUMA statistics associated with a task (either current or dead).
3081 * If @final is set, the task is dead and has reached refcount zero, so we can
3082 * safely free all relevant data structures. Otherwise, there might be
3083 * concurrent reads from places like load balancing and procfs, and we should
3084 * reset the data back to default state without freeing ->numa_faults.
3085 */
task_numa_free(struct task_struct * p,bool final)3086 void task_numa_free(struct task_struct *p, bool final)
3087 {
3088 /* safe: p either is current or is being freed by current */
3089 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3090 unsigned long *numa_faults = p->numa_faults;
3091 unsigned long flags;
3092 int i;
3093
3094 if (!numa_faults)
3095 return;
3096
3097 if (grp) {
3098 spin_lock_irqsave(&grp->lock, flags);
3099 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3100 grp->faults[i] -= p->numa_faults[i];
3101 grp->total_faults -= p->total_numa_faults;
3102
3103 grp->nr_tasks--;
3104 spin_unlock_irqrestore(&grp->lock, flags);
3105 RCU_INIT_POINTER(p->numa_group, NULL);
3106 put_numa_group(grp);
3107 }
3108
3109 if (final) {
3110 p->numa_faults = NULL;
3111 kfree(numa_faults);
3112 } else {
3113 p->total_numa_faults = 0;
3114 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3115 numa_faults[i] = 0;
3116 }
3117 }
3118
3119 /*
3120 * Got a PROT_NONE fault for a page on @node.
3121 */
task_numa_fault(int last_cpupid,int mem_node,int pages,int flags)3122 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3123 {
3124 struct task_struct *p = current;
3125 bool migrated = flags & TNF_MIGRATED;
3126 int cpu_node = task_node(current);
3127 int local = !!(flags & TNF_FAULT_LOCAL);
3128 struct numa_group *ng;
3129 int priv;
3130
3131 if (!static_branch_likely(&sched_numa_balancing))
3132 return;
3133
3134 /* for example, ksmd faulting in a user's mm */
3135 if (!p->mm)
3136 return;
3137
3138 /*
3139 * NUMA faults statistics are unnecessary for the slow memory
3140 * node for memory tiering mode.
3141 */
3142 if (!node_is_toptier(mem_node) &&
3143 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3144 !cpupid_valid(last_cpupid)))
3145 return;
3146
3147 /* Allocate buffer to track faults on a per-node basis */
3148 if (unlikely(!p->numa_faults)) {
3149 int size = sizeof(*p->numa_faults) *
3150 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3151
3152 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3153 if (!p->numa_faults)
3154 return;
3155
3156 p->total_numa_faults = 0;
3157 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3158 }
3159
3160 /*
3161 * First accesses are treated as private, otherwise consider accesses
3162 * to be private if the accessing pid has not changed
3163 */
3164 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3165 priv = 1;
3166 } else {
3167 priv = cpupid_match_pid(p, last_cpupid);
3168 if (!priv && !(flags & TNF_NO_GROUP))
3169 task_numa_group(p, last_cpupid, flags, &priv);
3170 }
3171
3172 /*
3173 * If a workload spans multiple NUMA nodes, a shared fault that
3174 * occurs wholly within the set of nodes that the workload is
3175 * actively using should be counted as local. This allows the
3176 * scan rate to slow down when a workload has settled down.
3177 */
3178 ng = deref_curr_numa_group(p);
3179 if (!priv && !local && ng && ng->active_nodes > 1 &&
3180 numa_is_active_node(cpu_node, ng) &&
3181 numa_is_active_node(mem_node, ng))
3182 local = 1;
3183
3184 /*
3185 * Retry to migrate task to preferred node periodically, in case it
3186 * previously failed, or the scheduler moved us.
3187 */
3188 if (time_after(jiffies, p->numa_migrate_retry)) {
3189 task_numa_placement(p);
3190 numa_migrate_preferred(p);
3191 }
3192
3193 if (migrated)
3194 p->numa_pages_migrated += pages;
3195 if (flags & TNF_MIGRATE_FAIL)
3196 p->numa_faults_locality[2] += pages;
3197
3198 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3199 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3200 p->numa_faults_locality[local] += pages;
3201 }
3202
reset_ptenuma_scan(struct task_struct * p)3203 static void reset_ptenuma_scan(struct task_struct *p)
3204 {
3205 /*
3206 * We only did a read acquisition of the mmap sem, so
3207 * p->mm->numa_scan_seq is written to without exclusive access
3208 * and the update is not guaranteed to be atomic. That's not
3209 * much of an issue though, since this is just used for
3210 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3211 * expensive, to avoid any form of compiler optimizations:
3212 */
3213 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3214 p->mm->numa_scan_offset = 0;
3215 }
3216
vma_is_accessed(struct mm_struct * mm,struct vm_area_struct * vma)3217 static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma)
3218 {
3219 unsigned long pids;
3220 /*
3221 * Allow unconditional access first two times, so that all the (pages)
3222 * of VMAs get prot_none fault introduced irrespective of accesses.
3223 * This is also done to avoid any side effect of task scanning
3224 * amplifying the unfairness of disjoint set of VMAs' access.
3225 */
3226 if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)
3227 return true;
3228
3229 pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];
3230 if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))
3231 return true;
3232
3233 /*
3234 * Complete a scan that has already started regardless of PID access, or
3235 * some VMAs may never be scanned in multi-threaded applications:
3236 */
3237 if (mm->numa_scan_offset > vma->vm_start) {
3238 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID);
3239 return true;
3240 }
3241
3242 /*
3243 * This vma has not been accessed for a while, and if the number
3244 * the threads in the same process is low, which means no other
3245 * threads can help scan this vma, force a vma scan.
3246 */
3247 if (READ_ONCE(mm->numa_scan_seq) >
3248 (vma->numab_state->prev_scan_seq + get_nr_threads(current)))
3249 return true;
3250
3251 return false;
3252 }
3253
3254 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3255
3256 /*
3257 * The expensive part of numa migration is done from task_work context.
3258 * Triggered from task_tick_numa().
3259 */
task_numa_work(struct callback_head * work)3260 static void task_numa_work(struct callback_head *work)
3261 {
3262 unsigned long migrate, next_scan, now = jiffies;
3263 struct task_struct *p = current;
3264 struct mm_struct *mm = p->mm;
3265 u64 runtime = p->se.sum_exec_runtime;
3266 struct vm_area_struct *vma;
3267 unsigned long start, end;
3268 unsigned long nr_pte_updates = 0;
3269 long pages, virtpages;
3270 struct vma_iterator vmi;
3271 bool vma_pids_skipped;
3272 bool vma_pids_forced = false;
3273
3274 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
3275
3276 work->next = work;
3277 /*
3278 * Who cares about NUMA placement when they're dying.
3279 *
3280 * NOTE: make sure not to dereference p->mm before this check,
3281 * exit_task_work() happens _after_ exit_mm() so we could be called
3282 * without p->mm even though we still had it when we enqueued this
3283 * work.
3284 */
3285 if (p->flags & PF_EXITING)
3286 return;
3287
3288 if (!mm->numa_next_scan) {
3289 mm->numa_next_scan = now +
3290 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3291 }
3292
3293 /*
3294 * Enforce maximal scan/migration frequency..
3295 */
3296 migrate = mm->numa_next_scan;
3297 if (time_before(now, migrate))
3298 return;
3299
3300 if (p->numa_scan_period == 0) {
3301 p->numa_scan_period_max = task_scan_max(p);
3302 p->numa_scan_period = task_scan_start(p);
3303 }
3304
3305 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3306 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3307 return;
3308
3309 /*
3310 * Delay this task enough that another task of this mm will likely win
3311 * the next time around.
3312 */
3313 p->node_stamp += 2 * TICK_NSEC;
3314
3315 pages = sysctl_numa_balancing_scan_size;
3316 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3317 virtpages = pages * 8; /* Scan up to this much virtual space */
3318 if (!pages)
3319 return;
3320
3321
3322 if (!mmap_read_trylock(mm))
3323 return;
3324
3325 /*
3326 * VMAs are skipped if the current PID has not trapped a fault within
3327 * the VMA recently. Allow scanning to be forced if there is no
3328 * suitable VMA remaining.
3329 */
3330 vma_pids_skipped = false;
3331
3332 retry_pids:
3333 start = mm->numa_scan_offset;
3334 vma_iter_init(&vmi, mm, start);
3335 vma = vma_next(&vmi);
3336 if (!vma) {
3337 reset_ptenuma_scan(p);
3338 start = 0;
3339 vma_iter_set(&vmi, start);
3340 vma = vma_next(&vmi);
3341 }
3342
3343 for (; vma; vma = vma_next(&vmi)) {
3344 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3345 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3346 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE);
3347 continue;
3348 }
3349
3350 /*
3351 * Shared library pages mapped by multiple processes are not
3352 * migrated as it is expected they are cache replicated. Avoid
3353 * hinting faults in read-only file-backed mappings or the vdso
3354 * as migrating the pages will be of marginal benefit.
3355 */
3356 if (!vma->vm_mm ||
3357 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {
3358 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO);
3359 continue;
3360 }
3361
3362 /*
3363 * Skip inaccessible VMAs to avoid any confusion between
3364 * PROT_NONE and NUMA hinting ptes
3365 */
3366 if (!vma_is_accessible(vma)) {
3367 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE);
3368 continue;
3369 }
3370
3371 /* Initialise new per-VMA NUMAB state. */
3372 if (!vma->numab_state) {
3373 struct vma_numab_state *ptr;
3374
3375 ptr = kzalloc(sizeof(*ptr), GFP_KERNEL);
3376 if (!ptr)
3377 continue;
3378
3379 if (cmpxchg(&vma->numab_state, NULL, ptr)) {
3380 kfree(ptr);
3381 continue;
3382 }
3383
3384 vma->numab_state->start_scan_seq = mm->numa_scan_seq;
3385
3386 vma->numab_state->next_scan = now +
3387 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3388
3389 /* Reset happens after 4 times scan delay of scan start */
3390 vma->numab_state->pids_active_reset = vma->numab_state->next_scan +
3391 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3392
3393 /*
3394 * Ensure prev_scan_seq does not match numa_scan_seq,
3395 * to prevent VMAs being skipped prematurely on the
3396 * first scan:
3397 */
3398 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;
3399 }
3400
3401 /*
3402 * Scanning the VMA's of short lived tasks add more overhead. So
3403 * delay the scan for new VMAs.
3404 */
3405 if (mm->numa_scan_seq && time_before(jiffies,
3406 vma->numab_state->next_scan)) {
3407 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY);
3408 continue;
3409 }
3410
3411 /* RESET access PIDs regularly for old VMAs. */
3412 if (mm->numa_scan_seq &&
3413 time_after(jiffies, vma->numab_state->pids_active_reset)) {
3414 vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +
3415 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3416 vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);
3417 vma->numab_state->pids_active[1] = 0;
3418 }
3419
3420 /* Do not rescan VMAs twice within the same sequence. */
3421 if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {
3422 mm->numa_scan_offset = vma->vm_end;
3423 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED);
3424 continue;
3425 }
3426
3427 /*
3428 * Do not scan the VMA if task has not accessed it, unless no other
3429 * VMA candidate exists.
3430 */
3431 if (!vma_pids_forced && !vma_is_accessed(mm, vma)) {
3432 vma_pids_skipped = true;
3433 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE);
3434 continue;
3435 }
3436
3437 do {
3438 start = max(start, vma->vm_start);
3439 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3440 end = min(end, vma->vm_end);
3441 nr_pte_updates = change_prot_numa(vma, start, end);
3442
3443 /*
3444 * Try to scan sysctl_numa_balancing_size worth of
3445 * hpages that have at least one present PTE that
3446 * is not already pte-numa. If the VMA contains
3447 * areas that are unused or already full of prot_numa
3448 * PTEs, scan up to virtpages, to skip through those
3449 * areas faster.
3450 */
3451 if (nr_pte_updates)
3452 pages -= (end - start) >> PAGE_SHIFT;
3453 virtpages -= (end - start) >> PAGE_SHIFT;
3454
3455 start = end;
3456 if (pages <= 0 || virtpages <= 0)
3457 goto out;
3458
3459 cond_resched();
3460 } while (end != vma->vm_end);
3461
3462 /* VMA scan is complete, do not scan until next sequence. */
3463 vma->numab_state->prev_scan_seq = mm->numa_scan_seq;
3464
3465 /*
3466 * Only force scan within one VMA at a time, to limit the
3467 * cost of scanning a potentially uninteresting VMA.
3468 */
3469 if (vma_pids_forced)
3470 break;
3471 }
3472
3473 /*
3474 * If no VMAs are remaining and VMAs were skipped due to the PID
3475 * not accessing the VMA previously, then force a scan to ensure
3476 * forward progress:
3477 */
3478 if (!vma && !vma_pids_forced && vma_pids_skipped) {
3479 vma_pids_forced = true;
3480 goto retry_pids;
3481 }
3482
3483 out:
3484 /*
3485 * It is possible to reach the end of the VMA list but the last few
3486 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3487 * would find the !migratable VMA on the next scan but not reset the
3488 * scanner to the start so check it now.
3489 */
3490 if (vma)
3491 mm->numa_scan_offset = start;
3492 else
3493 reset_ptenuma_scan(p);
3494 mmap_read_unlock(mm);
3495
3496 /*
3497 * Make sure tasks use at least 32x as much time to run other code
3498 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3499 * Usually update_task_scan_period slows down scanning enough; on an
3500 * overloaded system we need to limit overhead on a per task basis.
3501 */
3502 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3503 u64 diff = p->se.sum_exec_runtime - runtime;
3504 p->node_stamp += 32 * diff;
3505 }
3506 }
3507
init_numa_balancing(unsigned long clone_flags,struct task_struct * p)3508 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3509 {
3510 int mm_users = 0;
3511 struct mm_struct *mm = p->mm;
3512
3513 if (mm) {
3514 mm_users = atomic_read(&mm->mm_users);
3515 if (mm_users == 1) {
3516 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3517 mm->numa_scan_seq = 0;
3518 }
3519 }
3520 p->node_stamp = 0;
3521 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3522 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3523 p->numa_migrate_retry = 0;
3524 /* Protect against double add, see task_tick_numa and task_numa_work */
3525 p->numa_work.next = &p->numa_work;
3526 p->numa_faults = NULL;
3527 p->numa_pages_migrated = 0;
3528 p->total_numa_faults = 0;
3529 RCU_INIT_POINTER(p->numa_group, NULL);
3530 p->last_task_numa_placement = 0;
3531 p->last_sum_exec_runtime = 0;
3532
3533 init_task_work(&p->numa_work, task_numa_work);
3534
3535 /* New address space, reset the preferred nid */
3536 if (!(clone_flags & CLONE_VM)) {
3537 p->numa_preferred_nid = NUMA_NO_NODE;
3538 return;
3539 }
3540
3541 /*
3542 * New thread, keep existing numa_preferred_nid which should be copied
3543 * already by arch_dup_task_struct but stagger when scans start.
3544 */
3545 if (mm) {
3546 unsigned int delay;
3547
3548 delay = min_t(unsigned int, task_scan_max(current),
3549 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3550 delay += 2 * TICK_NSEC;
3551 p->node_stamp = delay;
3552 }
3553 }
3554
3555 /*
3556 * Drive the periodic memory faults..
3557 */
task_tick_numa(struct rq * rq,struct task_struct * curr)3558 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3559 {
3560 struct callback_head *work = &curr->numa_work;
3561 u64 period, now;
3562
3563 /*
3564 * We don't care about NUMA placement if we don't have memory.
3565 */
3566 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3567 return;
3568
3569 /*
3570 * Using runtime rather than walltime has the dual advantage that
3571 * we (mostly) drive the selection from busy threads and that the
3572 * task needs to have done some actual work before we bother with
3573 * NUMA placement.
3574 */
3575 now = curr->se.sum_exec_runtime;
3576 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3577
3578 if (now > curr->node_stamp + period) {
3579 if (!curr->node_stamp)
3580 curr->numa_scan_period = task_scan_start(curr);
3581 curr->node_stamp += period;
3582
3583 if (!time_before(jiffies, curr->mm->numa_next_scan))
3584 task_work_add(curr, work, TWA_RESUME);
3585 }
3586 }
3587
update_scan_period(struct task_struct * p,int new_cpu)3588 static void update_scan_period(struct task_struct *p, int new_cpu)
3589 {
3590 int src_nid = cpu_to_node(task_cpu(p));
3591 int dst_nid = cpu_to_node(new_cpu);
3592
3593 if (!static_branch_likely(&sched_numa_balancing))
3594 return;
3595
3596 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3597 return;
3598
3599 if (src_nid == dst_nid)
3600 return;
3601
3602 /*
3603 * Allow resets if faults have been trapped before one scan
3604 * has completed. This is most likely due to a new task that
3605 * is pulled cross-node due to wakeups or load balancing.
3606 */
3607 if (p->numa_scan_seq) {
3608 /*
3609 * Avoid scan adjustments if moving to the preferred
3610 * node or if the task was not previously running on
3611 * the preferred node.
3612 */
3613 if (dst_nid == p->numa_preferred_nid ||
3614 (p->numa_preferred_nid != NUMA_NO_NODE &&
3615 src_nid != p->numa_preferred_nid))
3616 return;
3617 }
3618
3619 p->numa_scan_period = task_scan_start(p);
3620 }
3621
3622 #else
task_tick_numa(struct rq * rq,struct task_struct * curr)3623 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3624 {
3625 }
3626
account_numa_enqueue(struct rq * rq,struct task_struct * p)3627 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3628 {
3629 }
3630
account_numa_dequeue(struct rq * rq,struct task_struct * p)3631 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3632 {
3633 }
3634
update_scan_period(struct task_struct * p,int new_cpu)3635 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3636 {
3637 }
3638
3639 #endif /* CONFIG_NUMA_BALANCING */
3640
3641 static void
account_entity_enqueue(struct cfs_rq * cfs_rq,struct sched_entity * se)3642 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3643 {
3644 update_load_add(&cfs_rq->load, se->load.weight);
3645 #ifdef CONFIG_SMP
3646 if (entity_is_task(se)) {
3647 struct rq *rq = rq_of(cfs_rq);
3648
3649 account_numa_enqueue(rq, task_of(se));
3650 list_add(&se->group_node, &rq->cfs_tasks);
3651 }
3652 #endif
3653 cfs_rq->nr_running++;
3654 if (se_is_idle(se))
3655 cfs_rq->idle_nr_running++;
3656 }
3657
3658 static void
account_entity_dequeue(struct cfs_rq * cfs_rq,struct sched_entity * se)3659 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3660 {
3661 update_load_sub(&cfs_rq->load, se->load.weight);
3662 #ifdef CONFIG_SMP
3663 if (entity_is_task(se)) {
3664 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3665 list_del_init(&se->group_node);
3666 }
3667 #endif
3668 cfs_rq->nr_running--;
3669 if (se_is_idle(se))
3670 cfs_rq->idle_nr_running--;
3671 }
3672
3673 /*
3674 * Signed add and clamp on underflow.
3675 *
3676 * Explicitly do a load-store to ensure the intermediate value never hits
3677 * memory. This allows lockless observations without ever seeing the negative
3678 * values.
3679 */
3680 #define add_positive(_ptr, _val) do { \
3681 typeof(_ptr) ptr = (_ptr); \
3682 typeof(_val) val = (_val); \
3683 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3684 \
3685 res = var + val; \
3686 \
3687 if (val < 0 && res > var) \
3688 res = 0; \
3689 \
3690 WRITE_ONCE(*ptr, res); \
3691 } while (0)
3692
3693 /*
3694 * Unsigned subtract and clamp on underflow.
3695 *
3696 * Explicitly do a load-store to ensure the intermediate value never hits
3697 * memory. This allows lockless observations without ever seeing the negative
3698 * values.
3699 */
3700 #define sub_positive(_ptr, _val) do { \
3701 typeof(_ptr) ptr = (_ptr); \
3702 typeof(*ptr) val = (_val); \
3703 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3704 res = var - val; \
3705 if (res > var) \
3706 res = 0; \
3707 WRITE_ONCE(*ptr, res); \
3708 } while (0)
3709
3710 /*
3711 * Remove and clamp on negative, from a local variable.
3712 *
3713 * A variant of sub_positive(), which does not use explicit load-store
3714 * and is thus optimized for local variable updates.
3715 */
3716 #define lsub_positive(_ptr, _val) do { \
3717 typeof(_ptr) ptr = (_ptr); \
3718 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3719 } while (0)
3720
3721 #ifdef CONFIG_SMP
3722 static inline void
enqueue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3723 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3724 {
3725 cfs_rq->avg.load_avg += se->avg.load_avg;
3726 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3727 }
3728
3729 static inline void
dequeue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3730 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3731 {
3732 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3733 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3734 /* See update_cfs_rq_load_avg() */
3735 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3736 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3737 }
3738 #else
3739 static inline void
enqueue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3740 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3741 static inline void
dequeue_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)3742 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3743 #endif
3744
reweight_eevdf(struct sched_entity * se,u64 avruntime,unsigned long weight)3745 static void reweight_eevdf(struct sched_entity *se, u64 avruntime,
3746 unsigned long weight)
3747 {
3748 unsigned long old_weight = se->load.weight;
3749 s64 vlag, vslice;
3750
3751 /*
3752 * VRUNTIME
3753 * ========
3754 *
3755 * COROLLARY #1: The virtual runtime of the entity needs to be
3756 * adjusted if re-weight at !0-lag point.
3757 *
3758 * Proof: For contradiction assume this is not true, so we can
3759 * re-weight without changing vruntime at !0-lag point.
3760 *
3761 * Weight VRuntime Avg-VRuntime
3762 * before w v V
3763 * after w' v' V'
3764 *
3765 * Since lag needs to be preserved through re-weight:
3766 *
3767 * lag = (V - v)*w = (V'- v')*w', where v = v'
3768 * ==> V' = (V - v)*w/w' + v (1)
3769 *
3770 * Let W be the total weight of the entities before reweight,
3771 * since V' is the new weighted average of entities:
3772 *
3773 * V' = (WV + w'v - wv) / (W + w' - w) (2)
3774 *
3775 * by using (1) & (2) we obtain:
3776 *
3777 * (WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
3778 * ==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
3779 * ==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
3780 * ==> (V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
3781 *
3782 * Since we are doing at !0-lag point which means V != v, we
3783 * can simplify (3):
3784 *
3785 * ==> W / (W + w' - w) = w / w'
3786 * ==> Ww' = Ww + ww' - ww
3787 * ==> W * (w' - w) = w * (w' - w)
3788 * ==> W = w (re-weight indicates w' != w)
3789 *
3790 * So the cfs_rq contains only one entity, hence vruntime of
3791 * the entity @v should always equal to the cfs_rq's weighted
3792 * average vruntime @V, which means we will always re-weight
3793 * at 0-lag point, thus breach assumption. Proof completed.
3794 *
3795 *
3796 * COROLLARY #2: Re-weight does NOT affect weighted average
3797 * vruntime of all the entities.
3798 *
3799 * Proof: According to corollary #1, Eq. (1) should be:
3800 *
3801 * (V - v)*w = (V' - v')*w'
3802 * ==> v' = V' - (V - v)*w/w' (4)
3803 *
3804 * According to the weighted average formula, we have:
3805 *
3806 * V' = (WV - wv + w'v') / (W - w + w')
3807 * = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
3808 * = (WV - wv + w'V' - Vw + wv) / (W - w + w')
3809 * = (WV + w'V' - Vw) / (W - w + w')
3810 *
3811 * ==> V'*(W - w + w') = WV + w'V' - Vw
3812 * ==> V' * (W - w) = (W - w) * V (5)
3813 *
3814 * If the entity is the only one in the cfs_rq, then reweight
3815 * always occurs at 0-lag point, so V won't change. Or else
3816 * there are other entities, hence W != w, then Eq. (5) turns
3817 * into V' = V. So V won't change in either case, proof done.
3818 *
3819 *
3820 * So according to corollary #1 & #2, the effect of re-weight
3821 * on vruntime should be:
3822 *
3823 * v' = V' - (V - v) * w / w' (4)
3824 * = V - (V - v) * w / w'
3825 * = V - vl * w / w'
3826 * = V - vl'
3827 */
3828 if (avruntime != se->vruntime) {
3829 vlag = entity_lag(avruntime, se);
3830 vlag = div_s64(vlag * old_weight, weight);
3831 se->vruntime = avruntime - vlag;
3832 }
3833
3834 /*
3835 * DEADLINE
3836 * ========
3837 *
3838 * When the weight changes, the virtual time slope changes and
3839 * we should adjust the relative virtual deadline accordingly.
3840 *
3841 * d' = v' + (d - v)*w/w'
3842 * = V' - (V - v)*w/w' + (d - v)*w/w'
3843 * = V - (V - v)*w/w' + (d - v)*w/w'
3844 * = V + (d - V)*w/w'
3845 */
3846 vslice = (s64)(se->deadline - avruntime);
3847 vslice = div_s64(vslice * old_weight, weight);
3848 se->deadline = avruntime + vslice;
3849 }
3850
reweight_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,unsigned long weight)3851 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3852 unsigned long weight)
3853 {
3854 bool curr = cfs_rq->curr == se;
3855 u64 avruntime;
3856
3857 if (se->on_rq) {
3858 /* commit outstanding execution time */
3859 update_curr(cfs_rq);
3860 avruntime = avg_vruntime(cfs_rq);
3861 if (!curr)
3862 __dequeue_entity(cfs_rq, se);
3863 update_load_sub(&cfs_rq->load, se->load.weight);
3864 }
3865 dequeue_load_avg(cfs_rq, se);
3866
3867 if (se->on_rq) {
3868 reweight_eevdf(se, avruntime, weight);
3869 } else {
3870 /*
3871 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3872 * we need to scale se->vlag when w_i changes.
3873 */
3874 se->vlag = div_s64(se->vlag * se->load.weight, weight);
3875 }
3876
3877 update_load_set(&se->load, weight);
3878
3879 #ifdef CONFIG_SMP
3880 do {
3881 u32 divider = get_pelt_divider(&se->avg);
3882
3883 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3884 } while (0);
3885 #endif
3886
3887 enqueue_load_avg(cfs_rq, se);
3888 if (se->on_rq) {
3889 update_load_add(&cfs_rq->load, se->load.weight);
3890 if (!curr)
3891 __enqueue_entity(cfs_rq, se);
3892
3893 /*
3894 * The entity's vruntime has been adjusted, so let's check
3895 * whether the rq-wide min_vruntime needs updated too. Since
3896 * the calculations above require stable min_vruntime rather
3897 * than up-to-date one, we do the update at the end of the
3898 * reweight process.
3899 */
3900 update_min_vruntime(cfs_rq);
3901 }
3902 }
3903
reweight_task(struct task_struct * p,const struct load_weight * lw)3904 void reweight_task(struct task_struct *p, const struct load_weight *lw)
3905 {
3906 struct sched_entity *se = &p->se;
3907 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3908 struct load_weight *load = &se->load;
3909
3910 reweight_entity(cfs_rq, se, lw->weight);
3911 load->inv_weight = lw->inv_weight;
3912 }
3913
3914 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3915
3916 #ifdef CONFIG_FAIR_GROUP_SCHED
3917 #ifdef CONFIG_SMP
3918 /*
3919 * All this does is approximate the hierarchical proportion which includes that
3920 * global sum we all love to hate.
3921 *
3922 * That is, the weight of a group entity, is the proportional share of the
3923 * group weight based on the group runqueue weights. That is:
3924 *
3925 * tg->weight * grq->load.weight
3926 * ge->load.weight = ----------------------------- (1)
3927 * \Sum grq->load.weight
3928 *
3929 * Now, because computing that sum is prohibitively expensive to compute (been
3930 * there, done that) we approximate it with this average stuff. The average
3931 * moves slower and therefore the approximation is cheaper and more stable.
3932 *
3933 * So instead of the above, we substitute:
3934 *
3935 * grq->load.weight -> grq->avg.load_avg (2)
3936 *
3937 * which yields the following:
3938 *
3939 * tg->weight * grq->avg.load_avg
3940 * ge->load.weight = ------------------------------ (3)
3941 * tg->load_avg
3942 *
3943 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3944 *
3945 * That is shares_avg, and it is right (given the approximation (2)).
3946 *
3947 * The problem with it is that because the average is slow -- it was designed
3948 * to be exactly that of course -- this leads to transients in boundary
3949 * conditions. In specific, the case where the group was idle and we start the
3950 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3951 * yielding bad latency etc..
3952 *
3953 * Now, in that special case (1) reduces to:
3954 *
3955 * tg->weight * grq->load.weight
3956 * ge->load.weight = ----------------------------- = tg->weight (4)
3957 * grp->load.weight
3958 *
3959 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3960 *
3961 * So what we do is modify our approximation (3) to approach (4) in the (near)
3962 * UP case, like:
3963 *
3964 * ge->load.weight =
3965 *
3966 * tg->weight * grq->load.weight
3967 * --------------------------------------------------- (5)
3968 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3969 *
3970 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3971 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3972 *
3973 *
3974 * tg->weight * grq->load.weight
3975 * ge->load.weight = ----------------------------- (6)
3976 * tg_load_avg'
3977 *
3978 * Where:
3979 *
3980 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3981 * max(grq->load.weight, grq->avg.load_avg)
3982 *
3983 * And that is shares_weight and is icky. In the (near) UP case it approaches
3984 * (4) while in the normal case it approaches (3). It consistently
3985 * overestimates the ge->load.weight and therefore:
3986 *
3987 * \Sum ge->load.weight >= tg->weight
3988 *
3989 * hence icky!
3990 */
calc_group_shares(struct cfs_rq * cfs_rq)3991 static long calc_group_shares(struct cfs_rq *cfs_rq)
3992 {
3993 long tg_weight, tg_shares, load, shares;
3994 struct task_group *tg = cfs_rq->tg;
3995
3996 tg_shares = READ_ONCE(tg->shares);
3997
3998 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3999
4000 tg_weight = atomic_long_read(&tg->load_avg);
4001
4002 /* Ensure tg_weight >= load */
4003 tg_weight -= cfs_rq->tg_load_avg_contrib;
4004 tg_weight += load;
4005
4006 shares = (tg_shares * load);
4007 if (tg_weight)
4008 shares /= tg_weight;
4009
4010 /*
4011 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
4012 * of a group with small tg->shares value. It is a floor value which is
4013 * assigned as a minimum load.weight to the sched_entity representing
4014 * the group on a CPU.
4015 *
4016 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
4017 * on an 8-core system with 8 tasks each runnable on one CPU shares has
4018 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
4019 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
4020 * instead of 0.
4021 */
4022 return clamp_t(long, shares, MIN_SHARES, tg_shares);
4023 }
4024 #endif /* CONFIG_SMP */
4025
4026 /*
4027 * Recomputes the group entity based on the current state of its group
4028 * runqueue.
4029 */
update_cfs_group(struct sched_entity * se)4030 static void update_cfs_group(struct sched_entity *se)
4031 {
4032 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4033 long shares;
4034
4035 if (!gcfs_rq)
4036 return;
4037
4038 if (throttled_hierarchy(gcfs_rq))
4039 return;
4040
4041 #ifndef CONFIG_SMP
4042 shares = READ_ONCE(gcfs_rq->tg->shares);
4043 #else
4044 shares = calc_group_shares(gcfs_rq);
4045 #endif
4046 if (unlikely(se->load.weight != shares))
4047 reweight_entity(cfs_rq_of(se), se, shares);
4048 }
4049
4050 #else /* CONFIG_FAIR_GROUP_SCHED */
update_cfs_group(struct sched_entity * se)4051 static inline void update_cfs_group(struct sched_entity *se)
4052 {
4053 }
4054 #endif /* CONFIG_FAIR_GROUP_SCHED */
4055
cfs_rq_util_change(struct cfs_rq * cfs_rq,int flags)4056 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
4057 {
4058 struct rq *rq = rq_of(cfs_rq);
4059
4060 if (&rq->cfs == cfs_rq) {
4061 /*
4062 * There are a few boundary cases this might miss but it should
4063 * get called often enough that that should (hopefully) not be
4064 * a real problem.
4065 *
4066 * It will not get called when we go idle, because the idle
4067 * thread is a different class (!fair), nor will the utilization
4068 * number include things like RT tasks.
4069 *
4070 * As is, the util number is not freq-invariant (we'd have to
4071 * implement arch_scale_freq_capacity() for that).
4072 *
4073 * See cpu_util_cfs().
4074 */
4075 cpufreq_update_util(rq, flags);
4076 }
4077 }
4078
4079 #ifdef CONFIG_SMP
load_avg_is_decayed(struct sched_avg * sa)4080 static inline bool load_avg_is_decayed(struct sched_avg *sa)
4081 {
4082 if (sa->load_sum)
4083 return false;
4084
4085 if (sa->util_sum)
4086 return false;
4087
4088 if (sa->runnable_sum)
4089 return false;
4090
4091 /*
4092 * _avg must be null when _sum are null because _avg = _sum / divider
4093 * Make sure that rounding and/or propagation of PELT values never
4094 * break this.
4095 */
4096 SCHED_WARN_ON(sa->load_avg ||
4097 sa->util_avg ||
4098 sa->runnable_avg);
4099
4100 return true;
4101 }
4102
cfs_rq_last_update_time(struct cfs_rq * cfs_rq)4103 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
4104 {
4105 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
4106 cfs_rq->last_update_time_copy);
4107 }
4108 #ifdef CONFIG_FAIR_GROUP_SCHED
4109 /*
4110 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
4111 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
4112 * bottom-up, we only have to test whether the cfs_rq before us on the list
4113 * is our child.
4114 * If cfs_rq is not on the list, test whether a child needs its to be added to
4115 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
4116 */
child_cfs_rq_on_list(struct cfs_rq * cfs_rq)4117 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4118 {
4119 struct cfs_rq *prev_cfs_rq;
4120 struct list_head *prev;
4121 struct rq *rq = rq_of(cfs_rq);
4122
4123 if (cfs_rq->on_list) {
4124 prev = cfs_rq->leaf_cfs_rq_list.prev;
4125 } else {
4126 prev = rq->tmp_alone_branch;
4127 }
4128
4129 if (prev == &rq->leaf_cfs_rq_list)
4130 return false;
4131
4132 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4133
4134 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4135 }
4136
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)4137 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4138 {
4139 if (cfs_rq->load.weight)
4140 return false;
4141
4142 if (!load_avg_is_decayed(&cfs_rq->avg))
4143 return false;
4144
4145 if (child_cfs_rq_on_list(cfs_rq))
4146 return false;
4147
4148 return true;
4149 }
4150
4151 /**
4152 * update_tg_load_avg - update the tg's load avg
4153 * @cfs_rq: the cfs_rq whose avg changed
4154 *
4155 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4156 * However, because tg->load_avg is a global value there are performance
4157 * considerations.
4158 *
4159 * In order to avoid having to look at the other cfs_rq's, we use a
4160 * differential update where we store the last value we propagated. This in
4161 * turn allows skipping updates if the differential is 'small'.
4162 *
4163 * Updating tg's load_avg is necessary before update_cfs_share().
4164 */
update_tg_load_avg(struct cfs_rq * cfs_rq)4165 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4166 {
4167 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4168
4169 /*
4170 * No need to update load_avg for root_task_group as it is not used.
4171 */
4172 if (cfs_rq->tg == &root_task_group)
4173 return;
4174
4175 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4176 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4177 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4178 }
4179 }
4180
4181 /*
4182 * Called within set_task_rq() right before setting a task's CPU. The
4183 * caller only guarantees p->pi_lock is held; no other assumptions,
4184 * including the state of rq->lock, should be made.
4185 */
set_task_rq_fair(struct sched_entity * se,struct cfs_rq * prev,struct cfs_rq * next)4186 void set_task_rq_fair(struct sched_entity *se,
4187 struct cfs_rq *prev, struct cfs_rq *next)
4188 {
4189 u64 p_last_update_time;
4190 u64 n_last_update_time;
4191
4192 if (!sched_feat(ATTACH_AGE_LOAD))
4193 return;
4194
4195 /*
4196 * We are supposed to update the task to "current" time, then its up to
4197 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4198 * getting what current time is, so simply throw away the out-of-date
4199 * time. This will result in the wakee task is less decayed, but giving
4200 * the wakee more load sounds not bad.
4201 */
4202 if (!(se->avg.last_update_time && prev))
4203 return;
4204
4205 p_last_update_time = cfs_rq_last_update_time(prev);
4206 n_last_update_time = cfs_rq_last_update_time(next);
4207
4208 __update_load_avg_blocked_se(p_last_update_time, se);
4209 se->avg.last_update_time = n_last_update_time;
4210 }
4211
4212 /*
4213 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4214 * propagate its contribution. The key to this propagation is the invariant
4215 * that for each group:
4216 *
4217 * ge->avg == grq->avg (1)
4218 *
4219 * _IFF_ we look at the pure running and runnable sums. Because they
4220 * represent the very same entity, just at different points in the hierarchy.
4221 *
4222 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4223 * and simply copies the running/runnable sum over (but still wrong, because
4224 * the group entity and group rq do not have their PELT windows aligned).
4225 *
4226 * However, update_tg_cfs_load() is more complex. So we have:
4227 *
4228 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4229 *
4230 * And since, like util, the runnable part should be directly transferable,
4231 * the following would _appear_ to be the straight forward approach:
4232 *
4233 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4234 *
4235 * And per (1) we have:
4236 *
4237 * ge->avg.runnable_avg == grq->avg.runnable_avg
4238 *
4239 * Which gives:
4240 *
4241 * ge->load.weight * grq->avg.load_avg
4242 * ge->avg.load_avg = ----------------------------------- (4)
4243 * grq->load.weight
4244 *
4245 * Except that is wrong!
4246 *
4247 * Because while for entities historical weight is not important and we
4248 * really only care about our future and therefore can consider a pure
4249 * runnable sum, runqueues can NOT do this.
4250 *
4251 * We specifically want runqueues to have a load_avg that includes
4252 * historical weights. Those represent the blocked load, the load we expect
4253 * to (shortly) return to us. This only works by keeping the weights as
4254 * integral part of the sum. We therefore cannot decompose as per (3).
4255 *
4256 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4257 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4258 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4259 * runnable section of these tasks overlap (or not). If they were to perfectly
4260 * align the rq as a whole would be runnable 2/3 of the time. If however we
4261 * always have at least 1 runnable task, the rq as a whole is always runnable.
4262 *
4263 * So we'll have to approximate.. :/
4264 *
4265 * Given the constraint:
4266 *
4267 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4268 *
4269 * We can construct a rule that adds runnable to a rq by assuming minimal
4270 * overlap.
4271 *
4272 * On removal, we'll assume each task is equally runnable; which yields:
4273 *
4274 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4275 *
4276 * XXX: only do this for the part of runnable > running ?
4277 *
4278 */
4279 static inline void
update_tg_cfs_util(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4280 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4281 {
4282 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4283 u32 new_sum, divider;
4284
4285 /* Nothing to update */
4286 if (!delta_avg)
4287 return;
4288
4289 /*
4290 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4291 * See ___update_load_avg() for details.
4292 */
4293 divider = get_pelt_divider(&cfs_rq->avg);
4294
4295
4296 /* Set new sched_entity's utilization */
4297 se->avg.util_avg = gcfs_rq->avg.util_avg;
4298 new_sum = se->avg.util_avg * divider;
4299 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4300 se->avg.util_sum = new_sum;
4301
4302 /* Update parent cfs_rq utilization */
4303 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4304 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4305
4306 /* See update_cfs_rq_load_avg() */
4307 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4308 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4309 }
4310
4311 static inline void
update_tg_cfs_runnable(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4312 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4313 {
4314 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4315 u32 new_sum, divider;
4316
4317 /* Nothing to update */
4318 if (!delta_avg)
4319 return;
4320
4321 /*
4322 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4323 * See ___update_load_avg() for details.
4324 */
4325 divider = get_pelt_divider(&cfs_rq->avg);
4326
4327 /* Set new sched_entity's runnable */
4328 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4329 new_sum = se->avg.runnable_avg * divider;
4330 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4331 se->avg.runnable_sum = new_sum;
4332
4333 /* Update parent cfs_rq runnable */
4334 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4335 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4336 /* See update_cfs_rq_load_avg() */
4337 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4338 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4339 }
4340
4341 static inline void
update_tg_cfs_load(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4342 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4343 {
4344 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4345 unsigned long load_avg;
4346 u64 load_sum = 0;
4347 s64 delta_sum;
4348 u32 divider;
4349
4350 if (!runnable_sum)
4351 return;
4352
4353 gcfs_rq->prop_runnable_sum = 0;
4354
4355 /*
4356 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4357 * See ___update_load_avg() for details.
4358 */
4359 divider = get_pelt_divider(&cfs_rq->avg);
4360
4361 if (runnable_sum >= 0) {
4362 /*
4363 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4364 * the CPU is saturated running == runnable.
4365 */
4366 runnable_sum += se->avg.load_sum;
4367 runnable_sum = min_t(long, runnable_sum, divider);
4368 } else {
4369 /*
4370 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4371 * assuming all tasks are equally runnable.
4372 */
4373 if (scale_load_down(gcfs_rq->load.weight)) {
4374 load_sum = div_u64(gcfs_rq->avg.load_sum,
4375 scale_load_down(gcfs_rq->load.weight));
4376 }
4377
4378 /* But make sure to not inflate se's runnable */
4379 runnable_sum = min(se->avg.load_sum, load_sum);
4380 }
4381
4382 /*
4383 * runnable_sum can't be lower than running_sum
4384 * Rescale running sum to be in the same range as runnable sum
4385 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4386 * runnable_sum is in [0 : LOAD_AVG_MAX]
4387 */
4388 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4389 runnable_sum = max(runnable_sum, running_sum);
4390
4391 load_sum = se_weight(se) * runnable_sum;
4392 load_avg = div_u64(load_sum, divider);
4393
4394 delta_avg = load_avg - se->avg.load_avg;
4395 if (!delta_avg)
4396 return;
4397
4398 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4399
4400 se->avg.load_sum = runnable_sum;
4401 se->avg.load_avg = load_avg;
4402 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4403 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4404 /* See update_cfs_rq_load_avg() */
4405 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4406 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4407 }
4408
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4409 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4410 {
4411 cfs_rq->propagate = 1;
4412 cfs_rq->prop_runnable_sum += runnable_sum;
4413 }
4414
4415 /* Update task and its cfs_rq load average */
propagate_entity_load_avg(struct sched_entity * se)4416 static inline int propagate_entity_load_avg(struct sched_entity *se)
4417 {
4418 struct cfs_rq *cfs_rq, *gcfs_rq;
4419
4420 if (entity_is_task(se))
4421 return 0;
4422
4423 gcfs_rq = group_cfs_rq(se);
4424 if (!gcfs_rq->propagate)
4425 return 0;
4426
4427 gcfs_rq->propagate = 0;
4428
4429 cfs_rq = cfs_rq_of(se);
4430
4431 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4432
4433 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4434 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4435 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4436
4437 trace_pelt_cfs_tp(cfs_rq);
4438 trace_pelt_se_tp(se);
4439
4440 return 1;
4441 }
4442
4443 /*
4444 * Check if we need to update the load and the utilization of a blocked
4445 * group_entity:
4446 */
skip_blocked_update(struct sched_entity * se)4447 static inline bool skip_blocked_update(struct sched_entity *se)
4448 {
4449 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4450
4451 /*
4452 * If sched_entity still have not zero load or utilization, we have to
4453 * decay it:
4454 */
4455 if (se->avg.load_avg || se->avg.util_avg)
4456 return false;
4457
4458 /*
4459 * If there is a pending propagation, we have to update the load and
4460 * the utilization of the sched_entity:
4461 */
4462 if (gcfs_rq->propagate)
4463 return false;
4464
4465 /*
4466 * Otherwise, the load and the utilization of the sched_entity is
4467 * already zero and there is no pending propagation, so it will be a
4468 * waste of time to try to decay it:
4469 */
4470 return true;
4471 }
4472
4473 #else /* CONFIG_FAIR_GROUP_SCHED */
4474
update_tg_load_avg(struct cfs_rq * cfs_rq)4475 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4476
propagate_entity_load_avg(struct sched_entity * se)4477 static inline int propagate_entity_load_avg(struct sched_entity *se)
4478 {
4479 return 0;
4480 }
4481
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4482 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4483
4484 #endif /* CONFIG_FAIR_GROUP_SCHED */
4485
4486 #ifdef CONFIG_NO_HZ_COMMON
migrate_se_pelt_lag(struct sched_entity * se)4487 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4488 {
4489 u64 throttled = 0, now, lut;
4490 struct cfs_rq *cfs_rq;
4491 struct rq *rq;
4492 bool is_idle;
4493
4494 if (load_avg_is_decayed(&se->avg))
4495 return;
4496
4497 cfs_rq = cfs_rq_of(se);
4498 rq = rq_of(cfs_rq);
4499
4500 rcu_read_lock();
4501 is_idle = is_idle_task(rcu_dereference(rq->curr));
4502 rcu_read_unlock();
4503
4504 /*
4505 * The lag estimation comes with a cost we don't want to pay all the
4506 * time. Hence, limiting to the case where the source CPU is idle and
4507 * we know we are at the greatest risk to have an outdated clock.
4508 */
4509 if (!is_idle)
4510 return;
4511
4512 /*
4513 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4514 *
4515 * last_update_time (the cfs_rq's last_update_time)
4516 * = cfs_rq_clock_pelt()@cfs_rq_idle
4517 * = rq_clock_pelt()@cfs_rq_idle
4518 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4519 *
4520 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4521 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4522 *
4523 * rq_idle_lag (delta between now and rq's update)
4524 * = sched_clock_cpu() - rq_clock()@rq_idle
4525 *
4526 * We can then write:
4527 *
4528 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4529 * sched_clock_cpu() - rq_clock()@rq_idle
4530 * Where:
4531 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4532 * rq_clock()@rq_idle is rq->clock_idle
4533 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4534 * is cfs_rq->throttled_pelt_idle
4535 */
4536
4537 #ifdef CONFIG_CFS_BANDWIDTH
4538 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4539 /* The clock has been stopped for throttling */
4540 if (throttled == U64_MAX)
4541 return;
4542 #endif
4543 now = u64_u32_load(rq->clock_pelt_idle);
4544 /*
4545 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4546 * is observed the old clock_pelt_idle value and the new clock_idle,
4547 * which lead to an underestimation. The opposite would lead to an
4548 * overestimation.
4549 */
4550 smp_rmb();
4551 lut = cfs_rq_last_update_time(cfs_rq);
4552
4553 now -= throttled;
4554 if (now < lut)
4555 /*
4556 * cfs_rq->avg.last_update_time is more recent than our
4557 * estimation, let's use it.
4558 */
4559 now = lut;
4560 else
4561 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4562
4563 __update_load_avg_blocked_se(now, se);
4564 }
4565 #else
migrate_se_pelt_lag(struct sched_entity * se)4566 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4567 #endif
4568
4569 /**
4570 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4571 * @now: current time, as per cfs_rq_clock_pelt()
4572 * @cfs_rq: cfs_rq to update
4573 *
4574 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4575 * avg. The immediate corollary is that all (fair) tasks must be attached.
4576 *
4577 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4578 *
4579 * Return: true if the load decayed or we removed load.
4580 *
4581 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4582 * call update_tg_load_avg() when this function returns true.
4583 */
4584 static inline int
update_cfs_rq_load_avg(u64 now,struct cfs_rq * cfs_rq)4585 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4586 {
4587 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4588 struct sched_avg *sa = &cfs_rq->avg;
4589 int decayed = 0;
4590
4591 if (cfs_rq->removed.nr) {
4592 unsigned long r;
4593 u32 divider = get_pelt_divider(&cfs_rq->avg);
4594
4595 raw_spin_lock(&cfs_rq->removed.lock);
4596 swap(cfs_rq->removed.util_avg, removed_util);
4597 swap(cfs_rq->removed.load_avg, removed_load);
4598 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4599 cfs_rq->removed.nr = 0;
4600 raw_spin_unlock(&cfs_rq->removed.lock);
4601
4602 r = removed_load;
4603 sub_positive(&sa->load_avg, r);
4604 sub_positive(&sa->load_sum, r * divider);
4605 /* See sa->util_sum below */
4606 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4607
4608 r = removed_util;
4609 sub_positive(&sa->util_avg, r);
4610 sub_positive(&sa->util_sum, r * divider);
4611 /*
4612 * Because of rounding, se->util_sum might ends up being +1 more than
4613 * cfs->util_sum. Although this is not a problem by itself, detaching
4614 * a lot of tasks with the rounding problem between 2 updates of
4615 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4616 * cfs_util_avg is not.
4617 * Check that util_sum is still above its lower bound for the new
4618 * util_avg. Given that period_contrib might have moved since the last
4619 * sync, we are only sure that util_sum must be above or equal to
4620 * util_avg * minimum possible divider
4621 */
4622 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4623
4624 r = removed_runnable;
4625 sub_positive(&sa->runnable_avg, r);
4626 sub_positive(&sa->runnable_sum, r * divider);
4627 /* See sa->util_sum above */
4628 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4629 sa->runnable_avg * PELT_MIN_DIVIDER);
4630
4631 /*
4632 * removed_runnable is the unweighted version of removed_load so we
4633 * can use it to estimate removed_load_sum.
4634 */
4635 add_tg_cfs_propagate(cfs_rq,
4636 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4637
4638 decayed = 1;
4639 }
4640
4641 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4642 u64_u32_store_copy(sa->last_update_time,
4643 cfs_rq->last_update_time_copy,
4644 sa->last_update_time);
4645 return decayed;
4646 }
4647
4648 /**
4649 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4650 * @cfs_rq: cfs_rq to attach to
4651 * @se: sched_entity to attach
4652 *
4653 * Must call update_cfs_rq_load_avg() before this, since we rely on
4654 * cfs_rq->avg.last_update_time being current.
4655 */
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4656 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4657 {
4658 /*
4659 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4660 * See ___update_load_avg() for details.
4661 */
4662 u32 divider = get_pelt_divider(&cfs_rq->avg);
4663
4664 /*
4665 * When we attach the @se to the @cfs_rq, we must align the decay
4666 * window because without that, really weird and wonderful things can
4667 * happen.
4668 *
4669 * XXX illustrate
4670 */
4671 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4672 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4673
4674 /*
4675 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4676 * period_contrib. This isn't strictly correct, but since we're
4677 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4678 * _sum a little.
4679 */
4680 se->avg.util_sum = se->avg.util_avg * divider;
4681
4682 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4683
4684 se->avg.load_sum = se->avg.load_avg * divider;
4685 if (se_weight(se) < se->avg.load_sum)
4686 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4687 else
4688 se->avg.load_sum = 1;
4689
4690 enqueue_load_avg(cfs_rq, se);
4691 cfs_rq->avg.util_avg += se->avg.util_avg;
4692 cfs_rq->avg.util_sum += se->avg.util_sum;
4693 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4694 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4695
4696 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4697
4698 cfs_rq_util_change(cfs_rq, 0);
4699
4700 trace_pelt_cfs_tp(cfs_rq);
4701 }
4702
4703 /**
4704 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4705 * @cfs_rq: cfs_rq to detach from
4706 * @se: sched_entity to detach
4707 *
4708 * Must call update_cfs_rq_load_avg() before this, since we rely on
4709 * cfs_rq->avg.last_update_time being current.
4710 */
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4711 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4712 {
4713 dequeue_load_avg(cfs_rq, se);
4714 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4715 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4716 /* See update_cfs_rq_load_avg() */
4717 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4718 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4719
4720 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4721 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4722 /* See update_cfs_rq_load_avg() */
4723 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4724 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4725
4726 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4727
4728 cfs_rq_util_change(cfs_rq, 0);
4729
4730 trace_pelt_cfs_tp(cfs_rq);
4731 }
4732
4733 /*
4734 * Optional action to be done while updating the load average
4735 */
4736 #define UPDATE_TG 0x1
4737 #define SKIP_AGE_LOAD 0x2
4738 #define DO_ATTACH 0x4
4739 #define DO_DETACH 0x8
4740
4741 /* Update task and its cfs_rq load average */
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)4742 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4743 {
4744 u64 now = cfs_rq_clock_pelt(cfs_rq);
4745 int decayed;
4746
4747 /*
4748 * Track task load average for carrying it to new CPU after migrated, and
4749 * track group sched_entity load average for task_h_load calc in migration
4750 */
4751 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4752 __update_load_avg_se(now, cfs_rq, se);
4753
4754 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4755 decayed |= propagate_entity_load_avg(se);
4756
4757 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4758
4759 /*
4760 * DO_ATTACH means we're here from enqueue_entity().
4761 * !last_update_time means we've passed through
4762 * migrate_task_rq_fair() indicating we migrated.
4763 *
4764 * IOW we're enqueueing a task on a new CPU.
4765 */
4766 attach_entity_load_avg(cfs_rq, se);
4767 update_tg_load_avg(cfs_rq);
4768
4769 } else if (flags & DO_DETACH) {
4770 /*
4771 * DO_DETACH means we're here from dequeue_entity()
4772 * and we are migrating task out of the CPU.
4773 */
4774 detach_entity_load_avg(cfs_rq, se);
4775 update_tg_load_avg(cfs_rq);
4776 } else if (decayed) {
4777 cfs_rq_util_change(cfs_rq, 0);
4778
4779 if (flags & UPDATE_TG)
4780 update_tg_load_avg(cfs_rq);
4781 }
4782 }
4783
4784 /*
4785 * Synchronize entity load avg of dequeued entity without locking
4786 * the previous rq.
4787 */
sync_entity_load_avg(struct sched_entity * se)4788 static void sync_entity_load_avg(struct sched_entity *se)
4789 {
4790 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4791 u64 last_update_time;
4792
4793 last_update_time = cfs_rq_last_update_time(cfs_rq);
4794 __update_load_avg_blocked_se(last_update_time, se);
4795 }
4796
4797 /*
4798 * Task first catches up with cfs_rq, and then subtract
4799 * itself from the cfs_rq (task must be off the queue now).
4800 */
remove_entity_load_avg(struct sched_entity * se)4801 static void remove_entity_load_avg(struct sched_entity *se)
4802 {
4803 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4804 unsigned long flags;
4805
4806 /*
4807 * tasks cannot exit without having gone through wake_up_new_task() ->
4808 * enqueue_task_fair() which will have added things to the cfs_rq,
4809 * so we can remove unconditionally.
4810 */
4811
4812 sync_entity_load_avg(se);
4813
4814 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4815 ++cfs_rq->removed.nr;
4816 cfs_rq->removed.util_avg += se->avg.util_avg;
4817 cfs_rq->removed.load_avg += se->avg.load_avg;
4818 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4819 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4820 }
4821
cfs_rq_runnable_avg(struct cfs_rq * cfs_rq)4822 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4823 {
4824 return cfs_rq->avg.runnable_avg;
4825 }
4826
cfs_rq_load_avg(struct cfs_rq * cfs_rq)4827 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4828 {
4829 return cfs_rq->avg.load_avg;
4830 }
4831
4832 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4833
task_util(struct task_struct * p)4834 static inline unsigned long task_util(struct task_struct *p)
4835 {
4836 return READ_ONCE(p->se.avg.util_avg);
4837 }
4838
_task_util_est(struct task_struct * p)4839 static inline unsigned long _task_util_est(struct task_struct *p)
4840 {
4841 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4842
4843 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4844 }
4845
task_util_est(struct task_struct * p)4846 static inline unsigned long task_util_est(struct task_struct *p)
4847 {
4848 return max(task_util(p), _task_util_est(p));
4849 }
4850
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)4851 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4852 struct task_struct *p)
4853 {
4854 unsigned int enqueued;
4855
4856 if (!sched_feat(UTIL_EST))
4857 return;
4858
4859 /* Update root cfs_rq's estimated utilization */
4860 enqueued = cfs_rq->avg.util_est.enqueued;
4861 enqueued += _task_util_est(p);
4862 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4863
4864 trace_sched_util_est_cfs_tp(cfs_rq);
4865 }
4866
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)4867 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4868 struct task_struct *p)
4869 {
4870 unsigned int enqueued;
4871
4872 if (!sched_feat(UTIL_EST))
4873 return;
4874
4875 /* Update root cfs_rq's estimated utilization */
4876 enqueued = cfs_rq->avg.util_est.enqueued;
4877 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4878 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4879
4880 trace_sched_util_est_cfs_tp(cfs_rq);
4881 }
4882
4883 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4884
4885 /*
4886 * Check if a (signed) value is within a specified (unsigned) margin,
4887 * based on the observation that:
4888 *
4889 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4890 *
4891 * NOTE: this only works when value + margin < INT_MAX.
4892 */
within_margin(int value,int margin)4893 static inline bool within_margin(int value, int margin)
4894 {
4895 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4896 }
4897
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)4898 static inline void util_est_update(struct cfs_rq *cfs_rq,
4899 struct task_struct *p,
4900 bool task_sleep)
4901 {
4902 long last_ewma_diff, last_enqueued_diff;
4903 struct util_est ue;
4904
4905 if (!sched_feat(UTIL_EST))
4906 return;
4907
4908 /*
4909 * Skip update of task's estimated utilization when the task has not
4910 * yet completed an activation, e.g. being migrated.
4911 */
4912 if (!task_sleep)
4913 return;
4914
4915 /*
4916 * If the PELT values haven't changed since enqueue time,
4917 * skip the util_est update.
4918 */
4919 ue = p->se.avg.util_est;
4920 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4921 return;
4922
4923 last_enqueued_diff = ue.enqueued;
4924
4925 /*
4926 * Reset EWMA on utilization increases, the moving average is used only
4927 * to smooth utilization decreases.
4928 */
4929 ue.enqueued = task_util(p);
4930 if (sched_feat(UTIL_EST_FASTUP)) {
4931 if (ue.ewma < ue.enqueued) {
4932 ue.ewma = ue.enqueued;
4933 goto done;
4934 }
4935 }
4936
4937 /*
4938 * Skip update of task's estimated utilization when its members are
4939 * already ~1% close to its last activation value.
4940 */
4941 last_ewma_diff = ue.enqueued - ue.ewma;
4942 last_enqueued_diff -= ue.enqueued;
4943 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4944 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4945 goto done;
4946
4947 return;
4948 }
4949
4950 /*
4951 * To avoid overestimation of actual task utilization, skip updates if
4952 * we cannot grant there is idle time in this CPU.
4953 */
4954 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4955 return;
4956
4957 /*
4958 * Update Task's estimated utilization
4959 *
4960 * When *p completes an activation we can consolidate another sample
4961 * of the task size. This is done by storing the current PELT value
4962 * as ue.enqueued and by using this value to update the Exponential
4963 * Weighted Moving Average (EWMA):
4964 *
4965 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4966 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4967 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4968 * = w * ( last_ewma_diff ) + ewma(t-1)
4969 * = w * (last_ewma_diff + ewma(t-1) / w)
4970 *
4971 * Where 'w' is the weight of new samples, which is configured to be
4972 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4973 */
4974 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4975 ue.ewma += last_ewma_diff;
4976 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4977 done:
4978 ue.enqueued |= UTIL_AVG_UNCHANGED;
4979 WRITE_ONCE(p->se.avg.util_est, ue);
4980
4981 trace_sched_util_est_se_tp(&p->se);
4982 }
4983
util_fits_cpu(unsigned long util,unsigned long uclamp_min,unsigned long uclamp_max,int cpu)4984 static inline int util_fits_cpu(unsigned long util,
4985 unsigned long uclamp_min,
4986 unsigned long uclamp_max,
4987 int cpu)
4988 {
4989 unsigned long capacity_orig, capacity_orig_thermal;
4990 unsigned long capacity = capacity_of(cpu);
4991 bool fits, uclamp_max_fits;
4992
4993 /*
4994 * Check if the real util fits without any uclamp boost/cap applied.
4995 */
4996 fits = fits_capacity(util, capacity);
4997
4998 if (!uclamp_is_used())
4999 return fits;
5000
5001 /*
5002 * We must use capacity_orig_of() for comparing against uclamp_min and
5003 * uclamp_max. We only care about capacity pressure (by using
5004 * capacity_of()) for comparing against the real util.
5005 *
5006 * If a task is boosted to 1024 for example, we don't want a tiny
5007 * pressure to skew the check whether it fits a CPU or not.
5008 *
5009 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
5010 * should fit a little cpu even if there's some pressure.
5011 *
5012 * Only exception is for thermal pressure since it has a direct impact
5013 * on available OPP of the system.
5014 *
5015 * We honour it for uclamp_min only as a drop in performance level
5016 * could result in not getting the requested minimum performance level.
5017 *
5018 * For uclamp_max, we can tolerate a drop in performance level as the
5019 * goal is to cap the task. So it's okay if it's getting less.
5020 */
5021 capacity_orig = capacity_orig_of(cpu);
5022 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
5023
5024 /*
5025 * We want to force a task to fit a cpu as implied by uclamp_max.
5026 * But we do have some corner cases to cater for..
5027 *
5028 *
5029 * C=z
5030 * | ___
5031 * | C=y | |
5032 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5033 * | C=x | | | |
5034 * | ___ | | | |
5035 * | | | | | | | (util somewhere in this region)
5036 * | | | | | | |
5037 * | | | | | | |
5038 * +----------------------------------------
5039 * cpu0 cpu1 cpu2
5040 *
5041 * In the above example if a task is capped to a specific performance
5042 * point, y, then when:
5043 *
5044 * * util = 80% of x then it does not fit on cpu0 and should migrate
5045 * to cpu1
5046 * * util = 80% of y then it is forced to fit on cpu1 to honour
5047 * uclamp_max request.
5048 *
5049 * which is what we're enforcing here. A task always fits if
5050 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
5051 * the normal upmigration rules should withhold still.
5052 *
5053 * Only exception is when we are on max capacity, then we need to be
5054 * careful not to block overutilized state. This is so because:
5055 *
5056 * 1. There's no concept of capping at max_capacity! We can't go
5057 * beyond this performance level anyway.
5058 * 2. The system is being saturated when we're operating near
5059 * max capacity, it doesn't make sense to block overutilized.
5060 */
5061 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
5062 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
5063 fits = fits || uclamp_max_fits;
5064
5065 /*
5066 *
5067 * C=z
5068 * | ___ (region a, capped, util >= uclamp_max)
5069 * | C=y | |
5070 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5071 * | C=x | | | |
5072 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
5073 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
5074 * | | | | | | |
5075 * | | | | | | | (region c, boosted, util < uclamp_min)
5076 * +----------------------------------------
5077 * cpu0 cpu1 cpu2
5078 *
5079 * a) If util > uclamp_max, then we're capped, we don't care about
5080 * actual fitness value here. We only care if uclamp_max fits
5081 * capacity without taking margin/pressure into account.
5082 * See comment above.
5083 *
5084 * b) If uclamp_min <= util <= uclamp_max, then the normal
5085 * fits_capacity() rules apply. Except we need to ensure that we
5086 * enforce we remain within uclamp_max, see comment above.
5087 *
5088 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
5089 * need to take into account the boosted value fits the CPU without
5090 * taking margin/pressure into account.
5091 *
5092 * Cases (a) and (b) are handled in the 'fits' variable already. We
5093 * just need to consider an extra check for case (c) after ensuring we
5094 * handle the case uclamp_min > uclamp_max.
5095 */
5096 uclamp_min = min(uclamp_min, uclamp_max);
5097 if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
5098 return -1;
5099
5100 return fits;
5101 }
5102
task_fits_cpu(struct task_struct * p,int cpu)5103 static inline int task_fits_cpu(struct task_struct *p, int cpu)
5104 {
5105 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
5106 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
5107 unsigned long util = task_util_est(p);
5108 /*
5109 * Return true only if the cpu fully fits the task requirements, which
5110 * include the utilization but also the performance hints.
5111 */
5112 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
5113 }
5114
update_misfit_status(struct task_struct * p,struct rq * rq)5115 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5116 {
5117 if (!sched_asym_cpucap_active())
5118 return;
5119
5120 if (!p || p->nr_cpus_allowed == 1) {
5121 rq->misfit_task_load = 0;
5122 return;
5123 }
5124
5125 if (task_fits_cpu(p, cpu_of(rq))) {
5126 rq->misfit_task_load = 0;
5127 return;
5128 }
5129
5130 /*
5131 * Make sure that misfit_task_load will not be null even if
5132 * task_h_load() returns 0.
5133 */
5134 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5135 }
5136
5137 #else /* CONFIG_SMP */
5138
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)5139 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5140 {
5141 return !cfs_rq->nr_running;
5142 }
5143
5144 #define UPDATE_TG 0x0
5145 #define SKIP_AGE_LOAD 0x0
5146 #define DO_ATTACH 0x0
5147 #define DO_DETACH 0x0
5148
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int not_used1)5149 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5150 {
5151 cfs_rq_util_change(cfs_rq, 0);
5152 }
5153
remove_entity_load_avg(struct sched_entity * se)5154 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5155
5156 static inline void
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5157 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5158 static inline void
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5159 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5160
newidle_balance(struct rq * rq,struct rq_flags * rf)5161 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
5162 {
5163 return 0;
5164 }
5165
5166 static inline void
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)5167 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5168
5169 static inline void
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)5170 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5171
5172 static inline void
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)5173 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5174 bool task_sleep) {}
update_misfit_status(struct task_struct * p,struct rq * rq)5175 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5176
5177 #endif /* CONFIG_SMP */
5178
5179 static void
place_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5180 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5181 {
5182 u64 vslice, vruntime = avg_vruntime(cfs_rq);
5183 s64 lag = 0;
5184
5185 se->slice = sysctl_sched_base_slice;
5186 vslice = calc_delta_fair(se->slice, se);
5187
5188 /*
5189 * Due to how V is constructed as the weighted average of entities,
5190 * adding tasks with positive lag, or removing tasks with negative lag
5191 * will move 'time' backwards, this can screw around with the lag of
5192 * other tasks.
5193 *
5194 * EEVDF: placement strategy #1 / #2
5195 */
5196 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5197 struct sched_entity *curr = cfs_rq->curr;
5198 unsigned long load;
5199
5200 lag = se->vlag;
5201
5202 /*
5203 * If we want to place a task and preserve lag, we have to
5204 * consider the effect of the new entity on the weighted
5205 * average and compensate for this, otherwise lag can quickly
5206 * evaporate.
5207 *
5208 * Lag is defined as:
5209 *
5210 * lag_i = S - s_i = w_i * (V - v_i)
5211 *
5212 * To avoid the 'w_i' term all over the place, we only track
5213 * the virtual lag:
5214 *
5215 * vl_i = V - v_i <=> v_i = V - vl_i
5216 *
5217 * And we take V to be the weighted average of all v:
5218 *
5219 * V = (\Sum w_j*v_j) / W
5220 *
5221 * Where W is: \Sum w_j
5222 *
5223 * Then, the weighted average after adding an entity with lag
5224 * vl_i is given by:
5225 *
5226 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5227 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5228 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5229 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5230 * = V - w_i*vl_i / (W + w_i)
5231 *
5232 * And the actual lag after adding an entity with vl_i is:
5233 *
5234 * vl'_i = V' - v_i
5235 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5236 * = vl_i - w_i*vl_i / (W + w_i)
5237 *
5238 * Which is strictly less than vl_i. So in order to preserve lag
5239 * we should inflate the lag before placement such that the
5240 * effective lag after placement comes out right.
5241 *
5242 * As such, invert the above relation for vl'_i to get the vl_i
5243 * we need to use such that the lag after placement is the lag
5244 * we computed before dequeue.
5245 *
5246 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5247 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5248 *
5249 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5250 * = W*vl_i
5251 *
5252 * vl_i = (W + w_i)*vl'_i / W
5253 */
5254 load = cfs_rq->avg_load;
5255 if (curr && curr->on_rq)
5256 load += scale_load_down(curr->load.weight);
5257
5258 lag *= load + scale_load_down(se->load.weight);
5259 if (WARN_ON_ONCE(!load))
5260 load = 1;
5261 lag = div_s64(lag, load);
5262 }
5263
5264 se->vruntime = vruntime - lag;
5265
5266 /*
5267 * When joining the competition; the exisiting tasks will be,
5268 * on average, halfway through their slice, as such start tasks
5269 * off with half a slice to ease into the competition.
5270 */
5271 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5272 vslice /= 2;
5273
5274 /*
5275 * EEVDF: vd_i = ve_i + r_i/w_i
5276 */
5277 se->deadline = se->vruntime + vslice;
5278 }
5279
5280 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5281 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5282
5283 static inline bool cfs_bandwidth_used(void);
5284
5285 static void
enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5286 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5287 {
5288 bool curr = cfs_rq->curr == se;
5289
5290 /*
5291 * If we're the current task, we must renormalise before calling
5292 * update_curr().
5293 */
5294 if (curr)
5295 place_entity(cfs_rq, se, flags);
5296
5297 update_curr(cfs_rq);
5298
5299 /*
5300 * When enqueuing a sched_entity, we must:
5301 * - Update loads to have both entity and cfs_rq synced with now.
5302 * - For group_entity, update its runnable_weight to reflect the new
5303 * h_nr_running of its group cfs_rq.
5304 * - For group_entity, update its weight to reflect the new share of
5305 * its group cfs_rq
5306 * - Add its new weight to cfs_rq->load.weight
5307 */
5308 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5309 se_update_runnable(se);
5310 /*
5311 * XXX update_load_avg() above will have attached us to the pelt sum;
5312 * but update_cfs_group() here will re-adjust the weight and have to
5313 * undo/redo all that. Seems wasteful.
5314 */
5315 update_cfs_group(se);
5316
5317 /*
5318 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5319 * we can place the entity.
5320 */
5321 if (!curr)
5322 place_entity(cfs_rq, se, flags);
5323
5324 account_entity_enqueue(cfs_rq, se);
5325
5326 /* Entity has migrated, no longer consider this task hot */
5327 if (flags & ENQUEUE_MIGRATED)
5328 se->exec_start = 0;
5329
5330 check_schedstat_required();
5331 update_stats_enqueue_fair(cfs_rq, se, flags);
5332 if (!curr)
5333 __enqueue_entity(cfs_rq, se);
5334 se->on_rq = 1;
5335
5336 if (cfs_rq->nr_running == 1) {
5337 check_enqueue_throttle(cfs_rq);
5338 if (!throttled_hierarchy(cfs_rq)) {
5339 list_add_leaf_cfs_rq(cfs_rq);
5340 } else {
5341 #ifdef CONFIG_CFS_BANDWIDTH
5342 struct rq *rq = rq_of(cfs_rq);
5343
5344 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5345 cfs_rq->throttled_clock = rq_clock(rq);
5346 if (!cfs_rq->throttled_clock_self)
5347 cfs_rq->throttled_clock_self = rq_clock(rq);
5348 #endif
5349 }
5350 }
5351 }
5352
__clear_buddies_next(struct sched_entity * se)5353 static void __clear_buddies_next(struct sched_entity *se)
5354 {
5355 for_each_sched_entity(se) {
5356 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5357 if (cfs_rq->next != se)
5358 break;
5359
5360 cfs_rq->next = NULL;
5361 }
5362 }
5363
clear_buddies(struct cfs_rq * cfs_rq,struct sched_entity * se)5364 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5365 {
5366 if (cfs_rq->next == se)
5367 __clear_buddies_next(se);
5368 }
5369
5370 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5371
5372 static void
dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5373 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5374 {
5375 int action = UPDATE_TG;
5376
5377 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5378 action |= DO_DETACH;
5379
5380 /*
5381 * Update run-time statistics of the 'current'.
5382 */
5383 update_curr(cfs_rq);
5384
5385 /*
5386 * When dequeuing a sched_entity, we must:
5387 * - Update loads to have both entity and cfs_rq synced with now.
5388 * - For group_entity, update its runnable_weight to reflect the new
5389 * h_nr_running of its group cfs_rq.
5390 * - Subtract its previous weight from cfs_rq->load.weight.
5391 * - For group entity, update its weight to reflect the new share
5392 * of its group cfs_rq.
5393 */
5394 update_load_avg(cfs_rq, se, action);
5395 se_update_runnable(se);
5396
5397 update_stats_dequeue_fair(cfs_rq, se, flags);
5398
5399 clear_buddies(cfs_rq, se);
5400
5401 update_entity_lag(cfs_rq, se);
5402 if (se != cfs_rq->curr)
5403 __dequeue_entity(cfs_rq, se);
5404 se->on_rq = 0;
5405 account_entity_dequeue(cfs_rq, se);
5406
5407 /* return excess runtime on last dequeue */
5408 return_cfs_rq_runtime(cfs_rq);
5409
5410 update_cfs_group(se);
5411
5412 /*
5413 * Now advance min_vruntime if @se was the entity holding it back,
5414 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5415 * put back on, and if we advance min_vruntime, we'll be placed back
5416 * further than we started -- ie. we'll be penalized.
5417 */
5418 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5419 update_min_vruntime(cfs_rq);
5420
5421 if (cfs_rq->nr_running == 0)
5422 update_idle_cfs_rq_clock_pelt(cfs_rq);
5423 }
5424
5425 static void
set_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)5426 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5427 {
5428 clear_buddies(cfs_rq, se);
5429
5430 /* 'current' is not kept within the tree. */
5431 if (se->on_rq) {
5432 /*
5433 * Any task has to be enqueued before it get to execute on
5434 * a CPU. So account for the time it spent waiting on the
5435 * runqueue.
5436 */
5437 update_stats_wait_end_fair(cfs_rq, se);
5438 __dequeue_entity(cfs_rq, se);
5439 update_load_avg(cfs_rq, se, UPDATE_TG);
5440 /*
5441 * HACK, stash a copy of deadline at the point of pick in vlag,
5442 * which isn't used until dequeue.
5443 */
5444 se->vlag = se->deadline;
5445 }
5446
5447 update_stats_curr_start(cfs_rq, se);
5448 cfs_rq->curr = se;
5449
5450 /*
5451 * Track our maximum slice length, if the CPU's load is at
5452 * least twice that of our own weight (i.e. dont track it
5453 * when there are only lesser-weight tasks around):
5454 */
5455 if (schedstat_enabled() &&
5456 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5457 struct sched_statistics *stats;
5458
5459 stats = __schedstats_from_se(se);
5460 __schedstat_set(stats->slice_max,
5461 max((u64)stats->slice_max,
5462 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5463 }
5464
5465 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5466 }
5467
5468 /*
5469 * Pick the next process, keeping these things in mind, in this order:
5470 * 1) keep things fair between processes/task groups
5471 * 2) pick the "next" process, since someone really wants that to run
5472 * 3) pick the "last" process, for cache locality
5473 * 4) do not run the "skip" process, if something else is available
5474 */
5475 static struct sched_entity *
pick_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * curr)5476 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
5477 {
5478 /*
5479 * Enabling NEXT_BUDDY will affect latency but not fairness.
5480 */
5481 if (sched_feat(NEXT_BUDDY) &&
5482 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5483 return cfs_rq->next;
5484
5485 return pick_eevdf(cfs_rq);
5486 }
5487
5488 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5489
put_prev_entity(struct cfs_rq * cfs_rq,struct sched_entity * prev)5490 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5491 {
5492 /*
5493 * If still on the runqueue then deactivate_task()
5494 * was not called and update_curr() has to be done:
5495 */
5496 if (prev->on_rq)
5497 update_curr(cfs_rq);
5498
5499 /* throttle cfs_rqs exceeding runtime */
5500 check_cfs_rq_runtime(cfs_rq);
5501
5502 if (prev->on_rq) {
5503 update_stats_wait_start_fair(cfs_rq, prev);
5504 /* Put 'current' back into the tree. */
5505 __enqueue_entity(cfs_rq, prev);
5506 /* in !on_rq case, update occurred at dequeue */
5507 update_load_avg(cfs_rq, prev, 0);
5508 }
5509 cfs_rq->curr = NULL;
5510 }
5511
5512 static void
entity_tick(struct cfs_rq * cfs_rq,struct sched_entity * curr,int queued)5513 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5514 {
5515 /*
5516 * Update run-time statistics of the 'current'.
5517 */
5518 update_curr(cfs_rq);
5519
5520 /*
5521 * Ensure that runnable average is periodically updated.
5522 */
5523 update_load_avg(cfs_rq, curr, UPDATE_TG);
5524 update_cfs_group(curr);
5525
5526 #ifdef CONFIG_SCHED_HRTICK
5527 /*
5528 * queued ticks are scheduled to match the slice, so don't bother
5529 * validating it and just reschedule.
5530 */
5531 if (queued) {
5532 resched_curr(rq_of(cfs_rq));
5533 return;
5534 }
5535 /*
5536 * don't let the period tick interfere with the hrtick preemption
5537 */
5538 if (!sched_feat(DOUBLE_TICK) &&
5539 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5540 return;
5541 #endif
5542 }
5543
5544
5545 /**************************************************
5546 * CFS bandwidth control machinery
5547 */
5548
5549 #ifdef CONFIG_CFS_BANDWIDTH
5550
5551 #ifdef CONFIG_JUMP_LABEL
5552 static struct static_key __cfs_bandwidth_used;
5553
cfs_bandwidth_used(void)5554 static inline bool cfs_bandwidth_used(void)
5555 {
5556 return static_key_false(&__cfs_bandwidth_used);
5557 }
5558
cfs_bandwidth_usage_inc(void)5559 void cfs_bandwidth_usage_inc(void)
5560 {
5561 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5562 }
5563
cfs_bandwidth_usage_dec(void)5564 void cfs_bandwidth_usage_dec(void)
5565 {
5566 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5567 }
5568 #else /* CONFIG_JUMP_LABEL */
cfs_bandwidth_used(void)5569 static bool cfs_bandwidth_used(void)
5570 {
5571 return true;
5572 }
5573
cfs_bandwidth_usage_inc(void)5574 void cfs_bandwidth_usage_inc(void) {}
cfs_bandwidth_usage_dec(void)5575 void cfs_bandwidth_usage_dec(void) {}
5576 #endif /* CONFIG_JUMP_LABEL */
5577
5578 /*
5579 * default period for cfs group bandwidth.
5580 * default: 0.1s, units: nanoseconds
5581 */
default_cfs_period(void)5582 static inline u64 default_cfs_period(void)
5583 {
5584 return 100000000ULL;
5585 }
5586
sched_cfs_bandwidth_slice(void)5587 static inline u64 sched_cfs_bandwidth_slice(void)
5588 {
5589 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5590 }
5591
5592 /*
5593 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5594 * directly instead of rq->clock to avoid adding additional synchronization
5595 * around rq->lock.
5596 *
5597 * requires cfs_b->lock
5598 */
__refill_cfs_bandwidth_runtime(struct cfs_bandwidth * cfs_b)5599 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5600 {
5601 s64 runtime;
5602
5603 if (unlikely(cfs_b->quota == RUNTIME_INF))
5604 return;
5605
5606 cfs_b->runtime += cfs_b->quota;
5607 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5608 if (runtime > 0) {
5609 cfs_b->burst_time += runtime;
5610 cfs_b->nr_burst++;
5611 }
5612
5613 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5614 cfs_b->runtime_snap = cfs_b->runtime;
5615 }
5616
tg_cfs_bandwidth(struct task_group * tg)5617 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5618 {
5619 return &tg->cfs_bandwidth;
5620 }
5621
5622 /* returns 0 on failure to allocate runtime */
__assign_cfs_rq_runtime(struct cfs_bandwidth * cfs_b,struct cfs_rq * cfs_rq,u64 target_runtime)5623 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5624 struct cfs_rq *cfs_rq, u64 target_runtime)
5625 {
5626 u64 min_amount, amount = 0;
5627
5628 lockdep_assert_held(&cfs_b->lock);
5629
5630 /* note: this is a positive sum as runtime_remaining <= 0 */
5631 min_amount = target_runtime - cfs_rq->runtime_remaining;
5632
5633 if (cfs_b->quota == RUNTIME_INF)
5634 amount = min_amount;
5635 else {
5636 start_cfs_bandwidth(cfs_b);
5637
5638 if (cfs_b->runtime > 0) {
5639 amount = min(cfs_b->runtime, min_amount);
5640 cfs_b->runtime -= amount;
5641 cfs_b->idle = 0;
5642 }
5643 }
5644
5645 cfs_rq->runtime_remaining += amount;
5646
5647 return cfs_rq->runtime_remaining > 0;
5648 }
5649
5650 /* returns 0 on failure to allocate runtime */
assign_cfs_rq_runtime(struct cfs_rq * cfs_rq)5651 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5652 {
5653 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5654 int ret;
5655
5656 raw_spin_lock(&cfs_b->lock);
5657 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5658 raw_spin_unlock(&cfs_b->lock);
5659
5660 return ret;
5661 }
5662
__account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5663 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5664 {
5665 /* dock delta_exec before expiring quota (as it could span periods) */
5666 cfs_rq->runtime_remaining -= delta_exec;
5667
5668 if (likely(cfs_rq->runtime_remaining > 0))
5669 return;
5670
5671 if (cfs_rq->throttled)
5672 return;
5673 /*
5674 * if we're unable to extend our runtime we resched so that the active
5675 * hierarchy can be throttled
5676 */
5677 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5678 resched_curr(rq_of(cfs_rq));
5679 }
5680
5681 static __always_inline
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5682 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5683 {
5684 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5685 return;
5686
5687 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5688 }
5689
cfs_rq_throttled(struct cfs_rq * cfs_rq)5690 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5691 {
5692 return cfs_bandwidth_used() && cfs_rq->throttled;
5693 }
5694
5695 /* check whether cfs_rq, or any parent, is throttled */
throttled_hierarchy(struct cfs_rq * cfs_rq)5696 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5697 {
5698 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5699 }
5700
5701 /*
5702 * Ensure that neither of the group entities corresponding to src_cpu or
5703 * dest_cpu are members of a throttled hierarchy when performing group
5704 * load-balance operations.
5705 */
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)5706 static inline int throttled_lb_pair(struct task_group *tg,
5707 int src_cpu, int dest_cpu)
5708 {
5709 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5710
5711 src_cfs_rq = tg->cfs_rq[src_cpu];
5712 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5713
5714 return throttled_hierarchy(src_cfs_rq) ||
5715 throttled_hierarchy(dest_cfs_rq);
5716 }
5717
tg_unthrottle_up(struct task_group * tg,void * data)5718 static int tg_unthrottle_up(struct task_group *tg, void *data)
5719 {
5720 struct rq *rq = data;
5721 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5722
5723 cfs_rq->throttle_count--;
5724 if (!cfs_rq->throttle_count) {
5725 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5726 cfs_rq->throttled_clock_pelt;
5727
5728 /* Add cfs_rq with load or one or more already running entities to the list */
5729 if (!cfs_rq_is_decayed(cfs_rq))
5730 list_add_leaf_cfs_rq(cfs_rq);
5731
5732 if (cfs_rq->throttled_clock_self) {
5733 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5734
5735 cfs_rq->throttled_clock_self = 0;
5736
5737 if (SCHED_WARN_ON((s64)delta < 0))
5738 delta = 0;
5739
5740 cfs_rq->throttled_clock_self_time += delta;
5741 }
5742 }
5743
5744 return 0;
5745 }
5746
tg_throttle_down(struct task_group * tg,void * data)5747 static int tg_throttle_down(struct task_group *tg, void *data)
5748 {
5749 struct rq *rq = data;
5750 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5751
5752 /* group is entering throttled state, stop time */
5753 if (!cfs_rq->throttle_count) {
5754 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5755 list_del_leaf_cfs_rq(cfs_rq);
5756
5757 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5758 if (cfs_rq->nr_running)
5759 cfs_rq->throttled_clock_self = rq_clock(rq);
5760 }
5761 cfs_rq->throttle_count++;
5762
5763 return 0;
5764 }
5765
throttle_cfs_rq(struct cfs_rq * cfs_rq)5766 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5767 {
5768 struct rq *rq = rq_of(cfs_rq);
5769 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5770 struct sched_entity *se;
5771 long task_delta, idle_task_delta, dequeue = 1;
5772
5773 raw_spin_lock(&cfs_b->lock);
5774 /* This will start the period timer if necessary */
5775 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5776 /*
5777 * We have raced with bandwidth becoming available, and if we
5778 * actually throttled the timer might not unthrottle us for an
5779 * entire period. We additionally needed to make sure that any
5780 * subsequent check_cfs_rq_runtime calls agree not to throttle
5781 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5782 * for 1ns of runtime rather than just check cfs_b.
5783 */
5784 dequeue = 0;
5785 } else {
5786 list_add_tail_rcu(&cfs_rq->throttled_list,
5787 &cfs_b->throttled_cfs_rq);
5788 }
5789 raw_spin_unlock(&cfs_b->lock);
5790
5791 if (!dequeue)
5792 return false; /* Throttle no longer required. */
5793
5794 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5795
5796 /* freeze hierarchy runnable averages while throttled */
5797 rcu_read_lock();
5798 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5799 rcu_read_unlock();
5800
5801 task_delta = cfs_rq->h_nr_running;
5802 idle_task_delta = cfs_rq->idle_h_nr_running;
5803 for_each_sched_entity(se) {
5804 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5805 /* throttled entity or throttle-on-deactivate */
5806 if (!se->on_rq)
5807 goto done;
5808
5809 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5810
5811 if (cfs_rq_is_idle(group_cfs_rq(se)))
5812 idle_task_delta = cfs_rq->h_nr_running;
5813
5814 qcfs_rq->h_nr_running -= task_delta;
5815 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5816
5817 if (qcfs_rq->load.weight) {
5818 /* Avoid re-evaluating load for this entity: */
5819 se = parent_entity(se);
5820 break;
5821 }
5822 }
5823
5824 for_each_sched_entity(se) {
5825 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5826 /* throttled entity or throttle-on-deactivate */
5827 if (!se->on_rq)
5828 goto done;
5829
5830 update_load_avg(qcfs_rq, se, 0);
5831 se_update_runnable(se);
5832
5833 if (cfs_rq_is_idle(group_cfs_rq(se)))
5834 idle_task_delta = cfs_rq->h_nr_running;
5835
5836 qcfs_rq->h_nr_running -= task_delta;
5837 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5838 }
5839
5840 /* At this point se is NULL and we are at root level*/
5841 sub_nr_running(rq, task_delta);
5842
5843 done:
5844 /*
5845 * Note: distribution will already see us throttled via the
5846 * throttled-list. rq->lock protects completion.
5847 */
5848 cfs_rq->throttled = 1;
5849 SCHED_WARN_ON(cfs_rq->throttled_clock);
5850 if (cfs_rq->nr_running)
5851 cfs_rq->throttled_clock = rq_clock(rq);
5852 return true;
5853 }
5854
unthrottle_cfs_rq(struct cfs_rq * cfs_rq)5855 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5856 {
5857 struct rq *rq = rq_of(cfs_rq);
5858 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5859 struct sched_entity *se;
5860 long task_delta, idle_task_delta;
5861
5862 se = cfs_rq->tg->se[cpu_of(rq)];
5863
5864 cfs_rq->throttled = 0;
5865
5866 update_rq_clock(rq);
5867
5868 raw_spin_lock(&cfs_b->lock);
5869 if (cfs_rq->throttled_clock) {
5870 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5871 cfs_rq->throttled_clock = 0;
5872 }
5873 list_del_rcu(&cfs_rq->throttled_list);
5874 raw_spin_unlock(&cfs_b->lock);
5875
5876 /* update hierarchical throttle state */
5877 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5878
5879 if (!cfs_rq->load.weight) {
5880 if (!cfs_rq->on_list)
5881 return;
5882 /*
5883 * Nothing to run but something to decay (on_list)?
5884 * Complete the branch.
5885 */
5886 for_each_sched_entity(se) {
5887 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5888 break;
5889 }
5890 goto unthrottle_throttle;
5891 }
5892
5893 task_delta = cfs_rq->h_nr_running;
5894 idle_task_delta = cfs_rq->idle_h_nr_running;
5895 for_each_sched_entity(se) {
5896 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5897
5898 if (se->on_rq)
5899 break;
5900 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5901
5902 if (cfs_rq_is_idle(group_cfs_rq(se)))
5903 idle_task_delta = cfs_rq->h_nr_running;
5904
5905 qcfs_rq->h_nr_running += task_delta;
5906 qcfs_rq->idle_h_nr_running += idle_task_delta;
5907
5908 /* end evaluation on encountering a throttled cfs_rq */
5909 if (cfs_rq_throttled(qcfs_rq))
5910 goto unthrottle_throttle;
5911 }
5912
5913 for_each_sched_entity(se) {
5914 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5915
5916 update_load_avg(qcfs_rq, se, UPDATE_TG);
5917 se_update_runnable(se);
5918
5919 if (cfs_rq_is_idle(group_cfs_rq(se)))
5920 idle_task_delta = cfs_rq->h_nr_running;
5921
5922 qcfs_rq->h_nr_running += task_delta;
5923 qcfs_rq->idle_h_nr_running += idle_task_delta;
5924
5925 /* end evaluation on encountering a throttled cfs_rq */
5926 if (cfs_rq_throttled(qcfs_rq))
5927 goto unthrottle_throttle;
5928 }
5929
5930 /* At this point se is NULL and we are at root level*/
5931 add_nr_running(rq, task_delta);
5932
5933 unthrottle_throttle:
5934 assert_list_leaf_cfs_rq(rq);
5935
5936 /* Determine whether we need to wake up potentially idle CPU: */
5937 if (rq->curr == rq->idle && rq->cfs.nr_running)
5938 resched_curr(rq);
5939 }
5940
5941 #ifdef CONFIG_SMP
__cfsb_csd_unthrottle(void * arg)5942 static void __cfsb_csd_unthrottle(void *arg)
5943 {
5944 struct cfs_rq *cursor, *tmp;
5945 struct rq *rq = arg;
5946 struct rq_flags rf;
5947
5948 rq_lock(rq, &rf);
5949
5950 /*
5951 * Iterating over the list can trigger several call to
5952 * update_rq_clock() in unthrottle_cfs_rq().
5953 * Do it once and skip the potential next ones.
5954 */
5955 update_rq_clock(rq);
5956 rq_clock_start_loop_update(rq);
5957
5958 /*
5959 * Since we hold rq lock we're safe from concurrent manipulation of
5960 * the CSD list. However, this RCU critical section annotates the
5961 * fact that we pair with sched_free_group_rcu(), so that we cannot
5962 * race with group being freed in the window between removing it
5963 * from the list and advancing to the next entry in the list.
5964 */
5965 rcu_read_lock();
5966
5967 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5968 throttled_csd_list) {
5969 list_del_init(&cursor->throttled_csd_list);
5970
5971 if (cfs_rq_throttled(cursor))
5972 unthrottle_cfs_rq(cursor);
5973 }
5974
5975 rcu_read_unlock();
5976
5977 rq_clock_stop_loop_update(rq);
5978 rq_unlock(rq, &rf);
5979 }
5980
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5981 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5982 {
5983 struct rq *rq = rq_of(cfs_rq);
5984 bool first;
5985
5986 if (rq == this_rq()) {
5987 unthrottle_cfs_rq(cfs_rq);
5988 return;
5989 }
5990
5991 /* Already enqueued */
5992 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5993 return;
5994
5995 first = list_empty(&rq->cfsb_csd_list);
5996 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5997 if (first)
5998 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5999 }
6000 #else
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)6001 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6002 {
6003 unthrottle_cfs_rq(cfs_rq);
6004 }
6005 #endif
6006
unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)6007 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6008 {
6009 lockdep_assert_rq_held(rq_of(cfs_rq));
6010
6011 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
6012 cfs_rq->runtime_remaining <= 0))
6013 return;
6014
6015 __unthrottle_cfs_rq_async(cfs_rq);
6016 }
6017
distribute_cfs_runtime(struct cfs_bandwidth * cfs_b)6018 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
6019 {
6020 struct cfs_rq *local_unthrottle = NULL;
6021 int this_cpu = smp_processor_id();
6022 u64 runtime, remaining = 1;
6023 bool throttled = false;
6024 struct cfs_rq *cfs_rq;
6025 struct rq_flags rf;
6026 struct rq *rq;
6027
6028 rcu_read_lock();
6029 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
6030 throttled_list) {
6031 rq = rq_of(cfs_rq);
6032
6033 if (!remaining) {
6034 throttled = true;
6035 break;
6036 }
6037
6038 rq_lock_irqsave(rq, &rf);
6039 if (!cfs_rq_throttled(cfs_rq))
6040 goto next;
6041
6042 #ifdef CONFIG_SMP
6043 /* Already queued for async unthrottle */
6044 if (!list_empty(&cfs_rq->throttled_csd_list))
6045 goto next;
6046 #endif
6047
6048 /* By the above checks, this should never be true */
6049 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
6050
6051 raw_spin_lock(&cfs_b->lock);
6052 runtime = -cfs_rq->runtime_remaining + 1;
6053 if (runtime > cfs_b->runtime)
6054 runtime = cfs_b->runtime;
6055 cfs_b->runtime -= runtime;
6056 remaining = cfs_b->runtime;
6057 raw_spin_unlock(&cfs_b->lock);
6058
6059 cfs_rq->runtime_remaining += runtime;
6060
6061 /* we check whether we're throttled above */
6062 if (cfs_rq->runtime_remaining > 0) {
6063 if (cpu_of(rq) != this_cpu ||
6064 SCHED_WARN_ON(local_unthrottle))
6065 unthrottle_cfs_rq_async(cfs_rq);
6066 else
6067 local_unthrottle = cfs_rq;
6068 } else {
6069 throttled = true;
6070 }
6071
6072 next:
6073 rq_unlock_irqrestore(rq, &rf);
6074 }
6075 rcu_read_unlock();
6076
6077 if (local_unthrottle) {
6078 rq = cpu_rq(this_cpu);
6079 rq_lock_irqsave(rq, &rf);
6080 if (cfs_rq_throttled(local_unthrottle))
6081 unthrottle_cfs_rq(local_unthrottle);
6082 rq_unlock_irqrestore(rq, &rf);
6083 }
6084
6085 return throttled;
6086 }
6087
6088 /*
6089 * Responsible for refilling a task_group's bandwidth and unthrottling its
6090 * cfs_rqs as appropriate. If there has been no activity within the last
6091 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6092 * used to track this state.
6093 */
do_sched_cfs_period_timer(struct cfs_bandwidth * cfs_b,int overrun,unsigned long flags)6094 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
6095 {
6096 int throttled;
6097
6098 /* no need to continue the timer with no bandwidth constraint */
6099 if (cfs_b->quota == RUNTIME_INF)
6100 goto out_deactivate;
6101
6102 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
6103 cfs_b->nr_periods += overrun;
6104
6105 /* Refill extra burst quota even if cfs_b->idle */
6106 __refill_cfs_bandwidth_runtime(cfs_b);
6107
6108 /*
6109 * idle depends on !throttled (for the case of a large deficit), and if
6110 * we're going inactive then everything else can be deferred
6111 */
6112 if (cfs_b->idle && !throttled)
6113 goto out_deactivate;
6114
6115 if (!throttled) {
6116 /* mark as potentially idle for the upcoming period */
6117 cfs_b->idle = 1;
6118 return 0;
6119 }
6120
6121 /* account preceding periods in which throttling occurred */
6122 cfs_b->nr_throttled += overrun;
6123
6124 /*
6125 * This check is repeated as we release cfs_b->lock while we unthrottle.
6126 */
6127 while (throttled && cfs_b->runtime > 0) {
6128 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6129 /* we can't nest cfs_b->lock while distributing bandwidth */
6130 throttled = distribute_cfs_runtime(cfs_b);
6131 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6132 }
6133
6134 /*
6135 * While we are ensured activity in the period following an
6136 * unthrottle, this also covers the case in which the new bandwidth is
6137 * insufficient to cover the existing bandwidth deficit. (Forcing the
6138 * timer to remain active while there are any throttled entities.)
6139 */
6140 cfs_b->idle = 0;
6141
6142 return 0;
6143
6144 out_deactivate:
6145 return 1;
6146 }
6147
6148 /* a cfs_rq won't donate quota below this amount */
6149 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6150 /* minimum remaining period time to redistribute slack quota */
6151 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6152 /* how long we wait to gather additional slack before distributing */
6153 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6154
6155 /*
6156 * Are we near the end of the current quota period?
6157 *
6158 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6159 * hrtimer base being cleared by hrtimer_start. In the case of
6160 * migrate_hrtimers, base is never cleared, so we are fine.
6161 */
runtime_refresh_within(struct cfs_bandwidth * cfs_b,u64 min_expire)6162 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6163 {
6164 struct hrtimer *refresh_timer = &cfs_b->period_timer;
6165 s64 remaining;
6166
6167 /* if the call-back is running a quota refresh is already occurring */
6168 if (hrtimer_callback_running(refresh_timer))
6169 return 1;
6170
6171 /* is a quota refresh about to occur? */
6172 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6173 if (remaining < (s64)min_expire)
6174 return 1;
6175
6176 return 0;
6177 }
6178
start_cfs_slack_bandwidth(struct cfs_bandwidth * cfs_b)6179 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6180 {
6181 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6182
6183 /* if there's a quota refresh soon don't bother with slack */
6184 if (runtime_refresh_within(cfs_b, min_left))
6185 return;
6186
6187 /* don't push forwards an existing deferred unthrottle */
6188 if (cfs_b->slack_started)
6189 return;
6190 cfs_b->slack_started = true;
6191
6192 hrtimer_start(&cfs_b->slack_timer,
6193 ns_to_ktime(cfs_bandwidth_slack_period),
6194 HRTIMER_MODE_REL);
6195 }
6196
6197 /* we know any runtime found here is valid as update_curr() precedes return */
__return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6198 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6199 {
6200 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6201 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6202
6203 if (slack_runtime <= 0)
6204 return;
6205
6206 raw_spin_lock(&cfs_b->lock);
6207 if (cfs_b->quota != RUNTIME_INF) {
6208 cfs_b->runtime += slack_runtime;
6209
6210 /* we are under rq->lock, defer unthrottling using a timer */
6211 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6212 !list_empty(&cfs_b->throttled_cfs_rq))
6213 start_cfs_slack_bandwidth(cfs_b);
6214 }
6215 raw_spin_unlock(&cfs_b->lock);
6216
6217 /* even if it's not valid for return we don't want to try again */
6218 cfs_rq->runtime_remaining -= slack_runtime;
6219 }
6220
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6221 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6222 {
6223 if (!cfs_bandwidth_used())
6224 return;
6225
6226 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6227 return;
6228
6229 __return_cfs_rq_runtime(cfs_rq);
6230 }
6231
6232 /*
6233 * This is done with a timer (instead of inline with bandwidth return) since
6234 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6235 */
do_sched_cfs_slack_timer(struct cfs_bandwidth * cfs_b)6236 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6237 {
6238 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6239 unsigned long flags;
6240
6241 /* confirm we're still not at a refresh boundary */
6242 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6243 cfs_b->slack_started = false;
6244
6245 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6246 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6247 return;
6248 }
6249
6250 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6251 runtime = cfs_b->runtime;
6252
6253 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6254
6255 if (!runtime)
6256 return;
6257
6258 distribute_cfs_runtime(cfs_b);
6259 }
6260
6261 /*
6262 * When a group wakes up we want to make sure that its quota is not already
6263 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6264 * runtime as update_curr() throttling can not trigger until it's on-rq.
6265 */
check_enqueue_throttle(struct cfs_rq * cfs_rq)6266 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6267 {
6268 if (!cfs_bandwidth_used())
6269 return;
6270
6271 /* an active group must be handled by the update_curr()->put() path */
6272 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6273 return;
6274
6275 /* ensure the group is not already throttled */
6276 if (cfs_rq_throttled(cfs_rq))
6277 return;
6278
6279 /* update runtime allocation */
6280 account_cfs_rq_runtime(cfs_rq, 0);
6281 if (cfs_rq->runtime_remaining <= 0)
6282 throttle_cfs_rq(cfs_rq);
6283 }
6284
sync_throttle(struct task_group * tg,int cpu)6285 static void sync_throttle(struct task_group *tg, int cpu)
6286 {
6287 struct cfs_rq *pcfs_rq, *cfs_rq;
6288
6289 if (!cfs_bandwidth_used())
6290 return;
6291
6292 if (!tg->parent)
6293 return;
6294
6295 cfs_rq = tg->cfs_rq[cpu];
6296 pcfs_rq = tg->parent->cfs_rq[cpu];
6297
6298 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6299 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6300 }
6301
6302 /* conditionally throttle active cfs_rq's from put_prev_entity() */
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6303 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6304 {
6305 if (!cfs_bandwidth_used())
6306 return false;
6307
6308 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6309 return false;
6310
6311 /*
6312 * it's possible for a throttled entity to be forced into a running
6313 * state (e.g. set_curr_task), in this case we're finished.
6314 */
6315 if (cfs_rq_throttled(cfs_rq))
6316 return true;
6317
6318 return throttle_cfs_rq(cfs_rq);
6319 }
6320
sched_cfs_slack_timer(struct hrtimer * timer)6321 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6322 {
6323 struct cfs_bandwidth *cfs_b =
6324 container_of(timer, struct cfs_bandwidth, slack_timer);
6325
6326 do_sched_cfs_slack_timer(cfs_b);
6327
6328 return HRTIMER_NORESTART;
6329 }
6330
6331 extern const u64 max_cfs_quota_period;
6332
sched_cfs_period_timer(struct hrtimer * timer)6333 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6334 {
6335 struct cfs_bandwidth *cfs_b =
6336 container_of(timer, struct cfs_bandwidth, period_timer);
6337 unsigned long flags;
6338 int overrun;
6339 int idle = 0;
6340 int count = 0;
6341
6342 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6343 for (;;) {
6344 overrun = hrtimer_forward_now(timer, cfs_b->period);
6345 if (!overrun)
6346 break;
6347
6348 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6349
6350 if (++count > 3) {
6351 u64 new, old = ktime_to_ns(cfs_b->period);
6352
6353 /*
6354 * Grow period by a factor of 2 to avoid losing precision.
6355 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6356 * to fail.
6357 */
6358 new = old * 2;
6359 if (new < max_cfs_quota_period) {
6360 cfs_b->period = ns_to_ktime(new);
6361 cfs_b->quota *= 2;
6362 cfs_b->burst *= 2;
6363
6364 pr_warn_ratelimited(
6365 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6366 smp_processor_id(),
6367 div_u64(new, NSEC_PER_USEC),
6368 div_u64(cfs_b->quota, NSEC_PER_USEC));
6369 } else {
6370 pr_warn_ratelimited(
6371 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6372 smp_processor_id(),
6373 div_u64(old, NSEC_PER_USEC),
6374 div_u64(cfs_b->quota, NSEC_PER_USEC));
6375 }
6376
6377 /* reset count so we don't come right back in here */
6378 count = 0;
6379 }
6380 }
6381 if (idle)
6382 cfs_b->period_active = 0;
6383 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6384
6385 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6386 }
6387
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6388 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6389 {
6390 raw_spin_lock_init(&cfs_b->lock);
6391 cfs_b->runtime = 0;
6392 cfs_b->quota = RUNTIME_INF;
6393 cfs_b->period = ns_to_ktime(default_cfs_period());
6394 cfs_b->burst = 0;
6395 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6396
6397 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6398 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6399 cfs_b->period_timer.function = sched_cfs_period_timer;
6400
6401 /* Add a random offset so that timers interleave */
6402 hrtimer_set_expires(&cfs_b->period_timer,
6403 get_random_u32_below(cfs_b->period));
6404 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6405 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6406 cfs_b->slack_started = false;
6407 }
6408
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6409 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6410 {
6411 cfs_rq->runtime_enabled = 0;
6412 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6413 #ifdef CONFIG_SMP
6414 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6415 #endif
6416 }
6417
start_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6418 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6419 {
6420 lockdep_assert_held(&cfs_b->lock);
6421
6422 if (cfs_b->period_active)
6423 return;
6424
6425 cfs_b->period_active = 1;
6426 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6427 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6428 }
6429
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6430 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6431 {
6432 int __maybe_unused i;
6433
6434 /* init_cfs_bandwidth() was not called */
6435 if (!cfs_b->throttled_cfs_rq.next)
6436 return;
6437
6438 hrtimer_cancel(&cfs_b->period_timer);
6439 hrtimer_cancel(&cfs_b->slack_timer);
6440
6441 /*
6442 * It is possible that we still have some cfs_rq's pending on a CSD
6443 * list, though this race is very rare. In order for this to occur, we
6444 * must have raced with the last task leaving the group while there
6445 * exist throttled cfs_rq(s), and the period_timer must have queued the
6446 * CSD item but the remote cpu has not yet processed it. To handle this,
6447 * we can simply flush all pending CSD work inline here. We're
6448 * guaranteed at this point that no additional cfs_rq of this group can
6449 * join a CSD list.
6450 */
6451 #ifdef CONFIG_SMP
6452 for_each_possible_cpu(i) {
6453 struct rq *rq = cpu_rq(i);
6454 unsigned long flags;
6455
6456 if (list_empty(&rq->cfsb_csd_list))
6457 continue;
6458
6459 local_irq_save(flags);
6460 __cfsb_csd_unthrottle(rq);
6461 local_irq_restore(flags);
6462 }
6463 #endif
6464 }
6465
6466 /*
6467 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6468 *
6469 * The race is harmless, since modifying bandwidth settings of unhooked group
6470 * bits doesn't do much.
6471 */
6472
6473 /* cpu online callback */
update_runtime_enabled(struct rq * rq)6474 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6475 {
6476 struct task_group *tg;
6477
6478 lockdep_assert_rq_held(rq);
6479
6480 rcu_read_lock();
6481 list_for_each_entry_rcu(tg, &task_groups, list) {
6482 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6483 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6484
6485 raw_spin_lock(&cfs_b->lock);
6486 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6487 raw_spin_unlock(&cfs_b->lock);
6488 }
6489 rcu_read_unlock();
6490 }
6491
6492 /* cpu offline callback */
unthrottle_offline_cfs_rqs(struct rq * rq)6493 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6494 {
6495 struct task_group *tg;
6496
6497 lockdep_assert_rq_held(rq);
6498
6499 /*
6500 * The rq clock has already been updated in the
6501 * set_rq_offline(), so we should skip updating
6502 * the rq clock again in unthrottle_cfs_rq().
6503 */
6504 rq_clock_start_loop_update(rq);
6505
6506 rcu_read_lock();
6507 list_for_each_entry_rcu(tg, &task_groups, list) {
6508 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6509
6510 if (!cfs_rq->runtime_enabled)
6511 continue;
6512
6513 /*
6514 * clock_task is not advancing so we just need to make sure
6515 * there's some valid quota amount
6516 */
6517 cfs_rq->runtime_remaining = 1;
6518 /*
6519 * Offline rq is schedulable till CPU is completely disabled
6520 * in take_cpu_down(), so we prevent new cfs throttling here.
6521 */
6522 cfs_rq->runtime_enabled = 0;
6523
6524 if (cfs_rq_throttled(cfs_rq))
6525 unthrottle_cfs_rq(cfs_rq);
6526 }
6527 rcu_read_unlock();
6528
6529 rq_clock_stop_loop_update(rq);
6530 }
6531
cfs_task_bw_constrained(struct task_struct * p)6532 bool cfs_task_bw_constrained(struct task_struct *p)
6533 {
6534 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6535
6536 if (!cfs_bandwidth_used())
6537 return false;
6538
6539 if (cfs_rq->runtime_enabled ||
6540 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6541 return true;
6542
6543 return false;
6544 }
6545
6546 #ifdef CONFIG_NO_HZ_FULL
6547 /* called from pick_next_task_fair() */
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6548 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6549 {
6550 int cpu = cpu_of(rq);
6551
6552 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6553 return;
6554
6555 if (!tick_nohz_full_cpu(cpu))
6556 return;
6557
6558 if (rq->nr_running != 1)
6559 return;
6560
6561 /*
6562 * We know there is only one task runnable and we've just picked it. The
6563 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6564 * be otherwise able to stop the tick. Just need to check if we are using
6565 * bandwidth control.
6566 */
6567 if (cfs_task_bw_constrained(p))
6568 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6569 }
6570 #endif
6571
6572 #else /* CONFIG_CFS_BANDWIDTH */
6573
cfs_bandwidth_used(void)6574 static inline bool cfs_bandwidth_used(void)
6575 {
6576 return false;
6577 }
6578
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)6579 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6580 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
check_enqueue_throttle(struct cfs_rq * cfs_rq)6581 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
sync_throttle(struct task_group * tg,int cpu)6582 static inline void sync_throttle(struct task_group *tg, int cpu) {}
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6583 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6584
cfs_rq_throttled(struct cfs_rq * cfs_rq)6585 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6586 {
6587 return 0;
6588 }
6589
throttled_hierarchy(struct cfs_rq * cfs_rq)6590 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6591 {
6592 return 0;
6593 }
6594
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)6595 static inline int throttled_lb_pair(struct task_group *tg,
6596 int src_cpu, int dest_cpu)
6597 {
6598 return 0;
6599 }
6600
6601 #ifdef CONFIG_FAIR_GROUP_SCHED
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6602 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6603 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6604 #endif
6605
tg_cfs_bandwidth(struct task_group * tg)6606 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6607 {
6608 return NULL;
6609 }
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6610 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
update_runtime_enabled(struct rq * rq)6611 static inline void update_runtime_enabled(struct rq *rq) {}
unthrottle_offline_cfs_rqs(struct rq * rq)6612 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6613 #ifdef CONFIG_CGROUP_SCHED
cfs_task_bw_constrained(struct task_struct * p)6614 bool cfs_task_bw_constrained(struct task_struct *p)
6615 {
6616 return false;
6617 }
6618 #endif
6619 #endif /* CONFIG_CFS_BANDWIDTH */
6620
6621 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6622 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6623 #endif
6624
6625 /**************************************************
6626 * CFS operations on tasks:
6627 */
6628
6629 #ifdef CONFIG_SCHED_HRTICK
hrtick_start_fair(struct rq * rq,struct task_struct * p)6630 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6631 {
6632 struct sched_entity *se = &p->se;
6633
6634 SCHED_WARN_ON(task_rq(p) != rq);
6635
6636 if (rq->cfs.h_nr_running > 1) {
6637 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6638 u64 slice = se->slice;
6639 s64 delta = slice - ran;
6640
6641 if (delta < 0) {
6642 if (task_current(rq, p))
6643 resched_curr(rq);
6644 return;
6645 }
6646 hrtick_start(rq, delta);
6647 }
6648 }
6649
6650 /*
6651 * called from enqueue/dequeue and updates the hrtick when the
6652 * current task is from our class and nr_running is low enough
6653 * to matter.
6654 */
hrtick_update(struct rq * rq)6655 static void hrtick_update(struct rq *rq)
6656 {
6657 struct task_struct *curr = rq->curr;
6658
6659 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6660 return;
6661
6662 hrtick_start_fair(rq, curr);
6663 }
6664 #else /* !CONFIG_SCHED_HRTICK */
6665 static inline void
hrtick_start_fair(struct rq * rq,struct task_struct * p)6666 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6667 {
6668 }
6669
hrtick_update(struct rq * rq)6670 static inline void hrtick_update(struct rq *rq)
6671 {
6672 }
6673 #endif
6674
6675 #ifdef CONFIG_SMP
cpu_overutilized(int cpu)6676 static inline bool cpu_overutilized(int cpu)
6677 {
6678 unsigned long rq_util_min, rq_util_max;
6679
6680 if (!sched_energy_enabled())
6681 return false;
6682
6683 rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6684 rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6685
6686 /* Return true only if the utilization doesn't fit CPU's capacity */
6687 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6688 }
6689
set_rd_overutilized_status(struct root_domain * rd,unsigned int status)6690 static inline void set_rd_overutilized_status(struct root_domain *rd,
6691 unsigned int status)
6692 {
6693 if (!sched_energy_enabled())
6694 return;
6695
6696 WRITE_ONCE(rd->overutilized, status);
6697 trace_sched_overutilized_tp(rd, !!status);
6698 }
6699
check_update_overutilized_status(struct rq * rq)6700 static inline void check_update_overutilized_status(struct rq *rq)
6701 {
6702 /*
6703 * overutilized field is used for load balancing decisions only
6704 * if energy aware scheduler is being used
6705 */
6706 if (!sched_energy_enabled())
6707 return;
6708
6709 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
6710 set_rd_overutilized_status(rq->rd, SG_OVERUTILIZED);
6711 }
6712 #else
check_update_overutilized_status(struct rq * rq)6713 static inline void check_update_overutilized_status(struct rq *rq) { }
6714 #endif
6715
6716 /* Runqueue only has SCHED_IDLE tasks enqueued */
sched_idle_rq(struct rq * rq)6717 static int sched_idle_rq(struct rq *rq)
6718 {
6719 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6720 rq->nr_running);
6721 }
6722
6723 #ifdef CONFIG_SMP
sched_idle_cpu(int cpu)6724 static int sched_idle_cpu(int cpu)
6725 {
6726 return sched_idle_rq(cpu_rq(cpu));
6727 }
6728 #endif
6729
6730 /*
6731 * The enqueue_task method is called before nr_running is
6732 * increased. Here we update the fair scheduling stats and
6733 * then put the task into the rbtree:
6734 */
6735 static void
enqueue_task_fair(struct rq * rq,struct task_struct * p,int flags)6736 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6737 {
6738 struct cfs_rq *cfs_rq;
6739 struct sched_entity *se = &p->se;
6740 int idle_h_nr_running = task_has_idle_policy(p);
6741 int task_new = !(flags & ENQUEUE_WAKEUP);
6742
6743 /*
6744 * The code below (indirectly) updates schedutil which looks at
6745 * the cfs_rq utilization to select a frequency.
6746 * Let's add the task's estimated utilization to the cfs_rq's
6747 * estimated utilization, before we update schedutil.
6748 */
6749 util_est_enqueue(&rq->cfs, p);
6750
6751 /*
6752 * If in_iowait is set, the code below may not trigger any cpufreq
6753 * utilization updates, so do it here explicitly with the IOWAIT flag
6754 * passed.
6755 */
6756 if (p->in_iowait)
6757 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6758
6759 for_each_sched_entity(se) {
6760 if (se->on_rq)
6761 break;
6762 cfs_rq = cfs_rq_of(se);
6763 enqueue_entity(cfs_rq, se, flags);
6764
6765 cfs_rq->h_nr_running++;
6766 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6767
6768 if (cfs_rq_is_idle(cfs_rq))
6769 idle_h_nr_running = 1;
6770
6771 /* end evaluation on encountering a throttled cfs_rq */
6772 if (cfs_rq_throttled(cfs_rq))
6773 goto enqueue_throttle;
6774
6775 flags = ENQUEUE_WAKEUP;
6776 }
6777
6778 for_each_sched_entity(se) {
6779 cfs_rq = cfs_rq_of(se);
6780
6781 update_load_avg(cfs_rq, se, UPDATE_TG);
6782 se_update_runnable(se);
6783 update_cfs_group(se);
6784
6785 cfs_rq->h_nr_running++;
6786 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6787
6788 if (cfs_rq_is_idle(cfs_rq))
6789 idle_h_nr_running = 1;
6790
6791 /* end evaluation on encountering a throttled cfs_rq */
6792 if (cfs_rq_throttled(cfs_rq))
6793 goto enqueue_throttle;
6794 }
6795
6796 /* At this point se is NULL and we are at root level*/
6797 add_nr_running(rq, 1);
6798
6799 /*
6800 * Since new tasks are assigned an initial util_avg equal to
6801 * half of the spare capacity of their CPU, tiny tasks have the
6802 * ability to cross the overutilized threshold, which will
6803 * result in the load balancer ruining all the task placement
6804 * done by EAS. As a way to mitigate that effect, do not account
6805 * for the first enqueue operation of new tasks during the
6806 * overutilized flag detection.
6807 *
6808 * A better way of solving this problem would be to wait for
6809 * the PELT signals of tasks to converge before taking them
6810 * into account, but that is not straightforward to implement,
6811 * and the following generally works well enough in practice.
6812 */
6813 if (!task_new)
6814 check_update_overutilized_status(rq);
6815
6816 enqueue_throttle:
6817 assert_list_leaf_cfs_rq(rq);
6818
6819 hrtick_update(rq);
6820 }
6821
6822 static void set_next_buddy(struct sched_entity *se);
6823
6824 /*
6825 * The dequeue_task method is called before nr_running is
6826 * decreased. We remove the task from the rbtree and
6827 * update the fair scheduling stats:
6828 */
dequeue_task_fair(struct rq * rq,struct task_struct * p,int flags)6829 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6830 {
6831 struct cfs_rq *cfs_rq;
6832 struct sched_entity *se = &p->se;
6833 int task_sleep = flags & DEQUEUE_SLEEP;
6834 int idle_h_nr_running = task_has_idle_policy(p);
6835 bool was_sched_idle = sched_idle_rq(rq);
6836
6837 util_est_dequeue(&rq->cfs, p);
6838
6839 for_each_sched_entity(se) {
6840 cfs_rq = cfs_rq_of(se);
6841 dequeue_entity(cfs_rq, se, flags);
6842
6843 cfs_rq->h_nr_running--;
6844 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6845
6846 if (cfs_rq_is_idle(cfs_rq))
6847 idle_h_nr_running = 1;
6848
6849 /* end evaluation on encountering a throttled cfs_rq */
6850 if (cfs_rq_throttled(cfs_rq))
6851 goto dequeue_throttle;
6852
6853 /* Don't dequeue parent if it has other entities besides us */
6854 if (cfs_rq->load.weight) {
6855 /* Avoid re-evaluating load for this entity: */
6856 se = parent_entity(se);
6857 /*
6858 * Bias pick_next to pick a task from this cfs_rq, as
6859 * p is sleeping when it is within its sched_slice.
6860 */
6861 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6862 set_next_buddy(se);
6863 break;
6864 }
6865 flags |= DEQUEUE_SLEEP;
6866 }
6867
6868 for_each_sched_entity(se) {
6869 cfs_rq = cfs_rq_of(se);
6870
6871 update_load_avg(cfs_rq, se, UPDATE_TG);
6872 se_update_runnable(se);
6873 update_cfs_group(se);
6874
6875 cfs_rq->h_nr_running--;
6876 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6877
6878 if (cfs_rq_is_idle(cfs_rq))
6879 idle_h_nr_running = 1;
6880
6881 /* end evaluation on encountering a throttled cfs_rq */
6882 if (cfs_rq_throttled(cfs_rq))
6883 goto dequeue_throttle;
6884
6885 }
6886
6887 /* At this point se is NULL and we are at root level*/
6888 sub_nr_running(rq, 1);
6889
6890 /* balance early to pull high priority tasks */
6891 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6892 rq->next_balance = jiffies;
6893
6894 dequeue_throttle:
6895 util_est_update(&rq->cfs, p, task_sleep);
6896 hrtick_update(rq);
6897 }
6898
6899 #ifdef CONFIG_SMP
6900
6901 /* Working cpumask for: load_balance, load_balance_newidle. */
6902 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6903 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6904 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6905
6906 #ifdef CONFIG_NO_HZ_COMMON
6907
6908 static struct {
6909 cpumask_var_t idle_cpus_mask;
6910 atomic_t nr_cpus;
6911 int has_blocked; /* Idle CPUS has blocked load */
6912 int needs_update; /* Newly idle CPUs need their next_balance collated */
6913 unsigned long next_balance; /* in jiffy units */
6914 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6915 } nohz ____cacheline_aligned;
6916
6917 #endif /* CONFIG_NO_HZ_COMMON */
6918
cpu_load(struct rq * rq)6919 static unsigned long cpu_load(struct rq *rq)
6920 {
6921 return cfs_rq_load_avg(&rq->cfs);
6922 }
6923
6924 /*
6925 * cpu_load_without - compute CPU load without any contributions from *p
6926 * @cpu: the CPU which load is requested
6927 * @p: the task which load should be discounted
6928 *
6929 * The load of a CPU is defined by the load of tasks currently enqueued on that
6930 * CPU as well as tasks which are currently sleeping after an execution on that
6931 * CPU.
6932 *
6933 * This method returns the load of the specified CPU by discounting the load of
6934 * the specified task, whenever the task is currently contributing to the CPU
6935 * load.
6936 */
cpu_load_without(struct rq * rq,struct task_struct * p)6937 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6938 {
6939 struct cfs_rq *cfs_rq;
6940 unsigned int load;
6941
6942 /* Task has no contribution or is new */
6943 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6944 return cpu_load(rq);
6945
6946 cfs_rq = &rq->cfs;
6947 load = READ_ONCE(cfs_rq->avg.load_avg);
6948
6949 /* Discount task's util from CPU's util */
6950 lsub_positive(&load, task_h_load(p));
6951
6952 return load;
6953 }
6954
cpu_runnable(struct rq * rq)6955 static unsigned long cpu_runnable(struct rq *rq)
6956 {
6957 return cfs_rq_runnable_avg(&rq->cfs);
6958 }
6959
cpu_runnable_without(struct rq * rq,struct task_struct * p)6960 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6961 {
6962 struct cfs_rq *cfs_rq;
6963 unsigned int runnable;
6964
6965 /* Task has no contribution or is new */
6966 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6967 return cpu_runnable(rq);
6968
6969 cfs_rq = &rq->cfs;
6970 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6971
6972 /* Discount task's runnable from CPU's runnable */
6973 lsub_positive(&runnable, p->se.avg.runnable_avg);
6974
6975 return runnable;
6976 }
6977
capacity_of(int cpu)6978 static unsigned long capacity_of(int cpu)
6979 {
6980 return cpu_rq(cpu)->cpu_capacity;
6981 }
6982
record_wakee(struct task_struct * p)6983 static void record_wakee(struct task_struct *p)
6984 {
6985 /*
6986 * Only decay a single time; tasks that have less then 1 wakeup per
6987 * jiffy will not have built up many flips.
6988 */
6989 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6990 current->wakee_flips >>= 1;
6991 current->wakee_flip_decay_ts = jiffies;
6992 }
6993
6994 if (current->last_wakee != p) {
6995 current->last_wakee = p;
6996 current->wakee_flips++;
6997 }
6998 }
6999
7000 /*
7001 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
7002 *
7003 * A waker of many should wake a different task than the one last awakened
7004 * at a frequency roughly N times higher than one of its wakees.
7005 *
7006 * In order to determine whether we should let the load spread vs consolidating
7007 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
7008 * partner, and a factor of lls_size higher frequency in the other.
7009 *
7010 * With both conditions met, we can be relatively sure that the relationship is
7011 * non-monogamous, with partner count exceeding socket size.
7012 *
7013 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
7014 * whatever is irrelevant, spread criteria is apparent partner count exceeds
7015 * socket size.
7016 */
wake_wide(struct task_struct * p)7017 static int wake_wide(struct task_struct *p)
7018 {
7019 unsigned int master = current->wakee_flips;
7020 unsigned int slave = p->wakee_flips;
7021 int factor = __this_cpu_read(sd_llc_size);
7022
7023 if (master < slave)
7024 swap(master, slave);
7025 if (slave < factor || master < slave * factor)
7026 return 0;
7027 return 1;
7028 }
7029
7030 /*
7031 * The purpose of wake_affine() is to quickly determine on which CPU we can run
7032 * soonest. For the purpose of speed we only consider the waking and previous
7033 * CPU.
7034 *
7035 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7036 * cache-affine and is (or will be) idle.
7037 *
7038 * wake_affine_weight() - considers the weight to reflect the average
7039 * scheduling latency of the CPUs. This seems to work
7040 * for the overloaded case.
7041 */
7042 static int
wake_affine_idle(int this_cpu,int prev_cpu,int sync)7043 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
7044 {
7045 /*
7046 * If this_cpu is idle, it implies the wakeup is from interrupt
7047 * context. Only allow the move if cache is shared. Otherwise an
7048 * interrupt intensive workload could force all tasks onto one
7049 * node depending on the IO topology or IRQ affinity settings.
7050 *
7051 * If the prev_cpu is idle and cache affine then avoid a migration.
7052 * There is no guarantee that the cache hot data from an interrupt
7053 * is more important than cache hot data on the prev_cpu and from
7054 * a cpufreq perspective, it's better to have higher utilisation
7055 * on one CPU.
7056 */
7057 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
7058 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
7059
7060 if (sync && cpu_rq(this_cpu)->nr_running == 1)
7061 return this_cpu;
7062
7063 if (available_idle_cpu(prev_cpu))
7064 return prev_cpu;
7065
7066 return nr_cpumask_bits;
7067 }
7068
7069 static int
wake_affine_weight(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)7070 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
7071 int this_cpu, int prev_cpu, int sync)
7072 {
7073 s64 this_eff_load, prev_eff_load;
7074 unsigned long task_load;
7075
7076 this_eff_load = cpu_load(cpu_rq(this_cpu));
7077
7078 if (sync) {
7079 unsigned long current_load = task_h_load(current);
7080
7081 if (current_load > this_eff_load)
7082 return this_cpu;
7083
7084 this_eff_load -= current_load;
7085 }
7086
7087 task_load = task_h_load(p);
7088
7089 this_eff_load += task_load;
7090 if (sched_feat(WA_BIAS))
7091 this_eff_load *= 100;
7092 this_eff_load *= capacity_of(prev_cpu);
7093
7094 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
7095 prev_eff_load -= task_load;
7096 if (sched_feat(WA_BIAS))
7097 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
7098 prev_eff_load *= capacity_of(this_cpu);
7099
7100 /*
7101 * If sync, adjust the weight of prev_eff_load such that if
7102 * prev_eff == this_eff that select_idle_sibling() will consider
7103 * stacking the wakee on top of the waker if no other CPU is
7104 * idle.
7105 */
7106 if (sync)
7107 prev_eff_load += 1;
7108
7109 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
7110 }
7111
wake_affine(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)7112 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
7113 int this_cpu, int prev_cpu, int sync)
7114 {
7115 int target = nr_cpumask_bits;
7116
7117 if (sched_feat(WA_IDLE))
7118 target = wake_affine_idle(this_cpu, prev_cpu, sync);
7119
7120 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
7121 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
7122
7123 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
7124 if (target != this_cpu)
7125 return prev_cpu;
7126
7127 schedstat_inc(sd->ttwu_move_affine);
7128 schedstat_inc(p->stats.nr_wakeups_affine);
7129 return target;
7130 }
7131
7132 static struct sched_group *
7133 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
7134
7135 /*
7136 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
7137 */
7138 static int
find_idlest_group_cpu(struct sched_group * group,struct task_struct * p,int this_cpu)7139 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7140 {
7141 unsigned long load, min_load = ULONG_MAX;
7142 unsigned int min_exit_latency = UINT_MAX;
7143 u64 latest_idle_timestamp = 0;
7144 int least_loaded_cpu = this_cpu;
7145 int shallowest_idle_cpu = -1;
7146 int i;
7147
7148 /* Check if we have any choice: */
7149 if (group->group_weight == 1)
7150 return cpumask_first(sched_group_span(group));
7151
7152 /* Traverse only the allowed CPUs */
7153 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7154 struct rq *rq = cpu_rq(i);
7155
7156 if (!sched_core_cookie_match(rq, p))
7157 continue;
7158
7159 if (sched_idle_cpu(i))
7160 return i;
7161
7162 if (available_idle_cpu(i)) {
7163 struct cpuidle_state *idle = idle_get_state(rq);
7164 if (idle && idle->exit_latency < min_exit_latency) {
7165 /*
7166 * We give priority to a CPU whose idle state
7167 * has the smallest exit latency irrespective
7168 * of any idle timestamp.
7169 */
7170 min_exit_latency = idle->exit_latency;
7171 latest_idle_timestamp = rq->idle_stamp;
7172 shallowest_idle_cpu = i;
7173 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
7174 rq->idle_stamp > latest_idle_timestamp) {
7175 /*
7176 * If equal or no active idle state, then
7177 * the most recently idled CPU might have
7178 * a warmer cache.
7179 */
7180 latest_idle_timestamp = rq->idle_stamp;
7181 shallowest_idle_cpu = i;
7182 }
7183 } else if (shallowest_idle_cpu == -1) {
7184 load = cpu_load(cpu_rq(i));
7185 if (load < min_load) {
7186 min_load = load;
7187 least_loaded_cpu = i;
7188 }
7189 }
7190 }
7191
7192 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7193 }
7194
find_idlest_cpu(struct sched_domain * sd,struct task_struct * p,int cpu,int prev_cpu,int sd_flag)7195 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
7196 int cpu, int prev_cpu, int sd_flag)
7197 {
7198 int new_cpu = cpu;
7199
7200 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7201 return prev_cpu;
7202
7203 /*
7204 * We need task's util for cpu_util_without, sync it up to
7205 * prev_cpu's last_update_time.
7206 */
7207 if (!(sd_flag & SD_BALANCE_FORK))
7208 sync_entity_load_avg(&p->se);
7209
7210 while (sd) {
7211 struct sched_group *group;
7212 struct sched_domain *tmp;
7213 int weight;
7214
7215 if (!(sd->flags & sd_flag)) {
7216 sd = sd->child;
7217 continue;
7218 }
7219
7220 group = find_idlest_group(sd, p, cpu);
7221 if (!group) {
7222 sd = sd->child;
7223 continue;
7224 }
7225
7226 new_cpu = find_idlest_group_cpu(group, p, cpu);
7227 if (new_cpu == cpu) {
7228 /* Now try balancing at a lower domain level of 'cpu': */
7229 sd = sd->child;
7230 continue;
7231 }
7232
7233 /* Now try balancing at a lower domain level of 'new_cpu': */
7234 cpu = new_cpu;
7235 weight = sd->span_weight;
7236 sd = NULL;
7237 for_each_domain(cpu, tmp) {
7238 if (weight <= tmp->span_weight)
7239 break;
7240 if (tmp->flags & sd_flag)
7241 sd = tmp;
7242 }
7243 }
7244
7245 return new_cpu;
7246 }
7247
__select_idle_cpu(int cpu,struct task_struct * p)7248 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7249 {
7250 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7251 sched_cpu_cookie_match(cpu_rq(cpu), p))
7252 return cpu;
7253
7254 return -1;
7255 }
7256
7257 #ifdef CONFIG_SCHED_SMT
7258 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7259 EXPORT_SYMBOL_GPL(sched_smt_present);
7260
set_idle_cores(int cpu,int val)7261 static inline void set_idle_cores(int cpu, int val)
7262 {
7263 struct sched_domain_shared *sds;
7264
7265 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7266 if (sds)
7267 WRITE_ONCE(sds->has_idle_cores, val);
7268 }
7269
test_idle_cores(int cpu)7270 static inline bool test_idle_cores(int cpu)
7271 {
7272 struct sched_domain_shared *sds;
7273
7274 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7275 if (sds)
7276 return READ_ONCE(sds->has_idle_cores);
7277
7278 return false;
7279 }
7280
7281 /*
7282 * Scans the local SMT mask to see if the entire core is idle, and records this
7283 * information in sd_llc_shared->has_idle_cores.
7284 *
7285 * Since SMT siblings share all cache levels, inspecting this limited remote
7286 * state should be fairly cheap.
7287 */
__update_idle_core(struct rq * rq)7288 void __update_idle_core(struct rq *rq)
7289 {
7290 int core = cpu_of(rq);
7291 int cpu;
7292
7293 rcu_read_lock();
7294 if (test_idle_cores(core))
7295 goto unlock;
7296
7297 for_each_cpu(cpu, cpu_smt_mask(core)) {
7298 if (cpu == core)
7299 continue;
7300
7301 if (!available_idle_cpu(cpu))
7302 goto unlock;
7303 }
7304
7305 set_idle_cores(core, 1);
7306 unlock:
7307 rcu_read_unlock();
7308 }
7309
7310 /*
7311 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7312 * there are no idle cores left in the system; tracked through
7313 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7314 */
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7315 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7316 {
7317 bool idle = true;
7318 int cpu;
7319
7320 for_each_cpu(cpu, cpu_smt_mask(core)) {
7321 if (!available_idle_cpu(cpu)) {
7322 idle = false;
7323 if (*idle_cpu == -1) {
7324 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) {
7325 *idle_cpu = cpu;
7326 break;
7327 }
7328 continue;
7329 }
7330 break;
7331 }
7332 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
7333 *idle_cpu = cpu;
7334 }
7335
7336 if (idle)
7337 return core;
7338
7339 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7340 return -1;
7341 }
7342
7343 /*
7344 * Scan the local SMT mask for idle CPUs.
7345 */
select_idle_smt(struct task_struct * p,struct sched_domain * sd,int target)7346 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7347 {
7348 int cpu;
7349
7350 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7351 if (cpu == target)
7352 continue;
7353 /*
7354 * Check if the CPU is in the LLC scheduling domain of @target.
7355 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
7356 */
7357 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7358 continue;
7359 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7360 return cpu;
7361 }
7362
7363 return -1;
7364 }
7365
7366 #else /* CONFIG_SCHED_SMT */
7367
set_idle_cores(int cpu,int val)7368 static inline void set_idle_cores(int cpu, int val)
7369 {
7370 }
7371
test_idle_cores(int cpu)7372 static inline bool test_idle_cores(int cpu)
7373 {
7374 return false;
7375 }
7376
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7377 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7378 {
7379 return __select_idle_cpu(core, p);
7380 }
7381
select_idle_smt(struct task_struct * p,struct sched_domain * sd,int target)7382 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7383 {
7384 return -1;
7385 }
7386
7387 #endif /* CONFIG_SCHED_SMT */
7388
7389 /*
7390 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7391 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7392 * average idle time for this rq (as found in rq->avg_idle).
7393 */
select_idle_cpu(struct task_struct * p,struct sched_domain * sd,bool has_idle_core,int target)7394 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7395 {
7396 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7397 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7398 struct sched_domain_shared *sd_share;
7399 struct rq *this_rq = this_rq();
7400 int this = smp_processor_id();
7401 struct sched_domain *this_sd = NULL;
7402 u64 time = 0;
7403
7404 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7405
7406 if (sched_feat(SIS_PROP) && !has_idle_core) {
7407 u64 avg_cost, avg_idle, span_avg;
7408 unsigned long now = jiffies;
7409
7410 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
7411 if (!this_sd)
7412 return -1;
7413
7414 /*
7415 * If we're busy, the assumption that the last idle period
7416 * predicts the future is flawed; age away the remaining
7417 * predicted idle time.
7418 */
7419 if (unlikely(this_rq->wake_stamp < now)) {
7420 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
7421 this_rq->wake_stamp++;
7422 this_rq->wake_avg_idle >>= 1;
7423 }
7424 }
7425
7426 avg_idle = this_rq->wake_avg_idle;
7427 avg_cost = this_sd->avg_scan_cost + 1;
7428
7429 span_avg = sd->span_weight * avg_idle;
7430 if (span_avg > 4*avg_cost)
7431 nr = div_u64(span_avg, avg_cost);
7432 else
7433 nr = 4;
7434
7435 time = cpu_clock(this);
7436 }
7437
7438 if (sched_feat(SIS_UTIL)) {
7439 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7440 if (sd_share) {
7441 /* because !--nr is the condition to stop scan */
7442 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7443 /* overloaded LLC is unlikely to have idle cpu/core */
7444 if (nr == 1)
7445 return -1;
7446 }
7447 }
7448
7449 for_each_cpu_wrap(cpu, cpus, target + 1) {
7450 if (has_idle_core) {
7451 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7452 if ((unsigned int)i < nr_cpumask_bits)
7453 return i;
7454
7455 } else {
7456 if (!--nr)
7457 return -1;
7458 idle_cpu = __select_idle_cpu(cpu, p);
7459 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7460 break;
7461 }
7462 }
7463
7464 if (has_idle_core)
7465 set_idle_cores(target, false);
7466
7467 if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
7468 time = cpu_clock(this) - time;
7469
7470 /*
7471 * Account for the scan cost of wakeups against the average
7472 * idle time.
7473 */
7474 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
7475
7476 update_avg(&this_sd->avg_scan_cost, time);
7477 }
7478
7479 return idle_cpu;
7480 }
7481
7482 /*
7483 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7484 * the task fits. If no CPU is big enough, but there are idle ones, try to
7485 * maximize capacity.
7486 */
7487 static int
select_idle_capacity(struct task_struct * p,struct sched_domain * sd,int target)7488 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7489 {
7490 unsigned long task_util, util_min, util_max, best_cap = 0;
7491 int fits, best_fits = 0;
7492 int cpu, best_cpu = -1;
7493 struct cpumask *cpus;
7494
7495 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7496 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7497
7498 task_util = task_util_est(p);
7499 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7500 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7501
7502 for_each_cpu_wrap(cpu, cpus, target) {
7503 unsigned long cpu_cap = capacity_of(cpu);
7504
7505 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7506 continue;
7507
7508 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7509
7510 /* This CPU fits with all requirements */
7511 if (fits > 0)
7512 return cpu;
7513 /*
7514 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7515 * Look for the CPU with best capacity.
7516 */
7517 else if (fits < 0)
7518 cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
7519
7520 /*
7521 * First, select CPU which fits better (-1 being better than 0).
7522 * Then, select the one with best capacity at same level.
7523 */
7524 if ((fits < best_fits) ||
7525 ((fits == best_fits) && (cpu_cap > best_cap))) {
7526 best_cap = cpu_cap;
7527 best_cpu = cpu;
7528 best_fits = fits;
7529 }
7530 }
7531
7532 return best_cpu;
7533 }
7534
asym_fits_cpu(unsigned long util,unsigned long util_min,unsigned long util_max,int cpu)7535 static inline bool asym_fits_cpu(unsigned long util,
7536 unsigned long util_min,
7537 unsigned long util_max,
7538 int cpu)
7539 {
7540 if (sched_asym_cpucap_active())
7541 /*
7542 * Return true only if the cpu fully fits the task requirements
7543 * which include the utilization and the performance hints.
7544 */
7545 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7546
7547 return true;
7548 }
7549
7550 /*
7551 * Try and locate an idle core/thread in the LLC cache domain.
7552 */
select_idle_sibling(struct task_struct * p,int prev,int target)7553 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7554 {
7555 bool has_idle_core = false;
7556 struct sched_domain *sd;
7557 unsigned long task_util, util_min, util_max;
7558 int i, recent_used_cpu;
7559
7560 /*
7561 * On asymmetric system, update task utilization because we will check
7562 * that the task fits with cpu's capacity.
7563 */
7564 if (sched_asym_cpucap_active()) {
7565 sync_entity_load_avg(&p->se);
7566 task_util = task_util_est(p);
7567 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7568 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7569 }
7570
7571 /*
7572 * per-cpu select_rq_mask usage
7573 */
7574 lockdep_assert_irqs_disabled();
7575
7576 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7577 asym_fits_cpu(task_util, util_min, util_max, target))
7578 return target;
7579
7580 /*
7581 * If the previous CPU is cache affine and idle, don't be stupid:
7582 */
7583 if (prev != target && cpus_share_cache(prev, target) &&
7584 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7585 asym_fits_cpu(task_util, util_min, util_max, prev))
7586 return prev;
7587
7588 /*
7589 * Allow a per-cpu kthread to stack with the wakee if the
7590 * kworker thread and the tasks previous CPUs are the same.
7591 * The assumption is that the wakee queued work for the
7592 * per-cpu kthread that is now complete and the wakeup is
7593 * essentially a sync wakeup. An obvious example of this
7594 * pattern is IO completions.
7595 */
7596 if (is_per_cpu_kthread(current) &&
7597 in_task() &&
7598 prev == smp_processor_id() &&
7599 this_rq()->nr_running <= 1 &&
7600 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7601 return prev;
7602 }
7603
7604 /* Check a recently used CPU as a potential idle candidate: */
7605 recent_used_cpu = p->recent_used_cpu;
7606 p->recent_used_cpu = prev;
7607 if (recent_used_cpu != prev &&
7608 recent_used_cpu != target &&
7609 cpus_share_cache(recent_used_cpu, target) &&
7610 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7611 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7612 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7613 return recent_used_cpu;
7614 }
7615
7616 /*
7617 * For asymmetric CPU capacity systems, our domain of interest is
7618 * sd_asym_cpucapacity rather than sd_llc.
7619 */
7620 if (sched_asym_cpucap_active()) {
7621 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7622 /*
7623 * On an asymmetric CPU capacity system where an exclusive
7624 * cpuset defines a symmetric island (i.e. one unique
7625 * capacity_orig value through the cpuset), the key will be set
7626 * but the CPUs within that cpuset will not have a domain with
7627 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7628 * capacity path.
7629 */
7630 if (sd) {
7631 i = select_idle_capacity(p, sd, target);
7632 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7633 }
7634 }
7635
7636 sd = rcu_dereference(per_cpu(sd_llc, target));
7637 if (!sd)
7638 return target;
7639
7640 if (sched_smt_active()) {
7641 has_idle_core = test_idle_cores(target);
7642
7643 if (!has_idle_core && cpus_share_cache(prev, target)) {
7644 i = select_idle_smt(p, sd, prev);
7645 if ((unsigned int)i < nr_cpumask_bits)
7646 return i;
7647 }
7648 }
7649
7650 i = select_idle_cpu(p, sd, has_idle_core, target);
7651 if ((unsigned)i < nr_cpumask_bits)
7652 return i;
7653
7654 return target;
7655 }
7656
7657 /**
7658 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7659 * @cpu: the CPU to get the utilization for
7660 * @p: task for which the CPU utilization should be predicted or NULL
7661 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7662 * @boost: 1 to enable boosting, otherwise 0
7663 *
7664 * The unit of the return value must be the same as the one of CPU capacity
7665 * so that CPU utilization can be compared with CPU capacity.
7666 *
7667 * CPU utilization is the sum of running time of runnable tasks plus the
7668 * recent utilization of currently non-runnable tasks on that CPU.
7669 * It represents the amount of CPU capacity currently used by CFS tasks in
7670 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7671 * capacity at f_max.
7672 *
7673 * The estimated CPU utilization is defined as the maximum between CPU
7674 * utilization and sum of the estimated utilization of the currently
7675 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7676 * previously-executed tasks, which helps better deduce how busy a CPU will
7677 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7678 * of such a task would be significantly decayed at this point of time.
7679 *
7680 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7681 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7682 * utilization. Boosting is implemented in cpu_util() so that internal
7683 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7684 * latter via cpu_util_cfs_boost().
7685 *
7686 * CPU utilization can be higher than the current CPU capacity
7687 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7688 * of rounding errors as well as task migrations or wakeups of new tasks.
7689 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7690 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7691 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7692 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7693 * though since this is useful for predicting the CPU capacity required
7694 * after task migrations (scheduler-driven DVFS).
7695 *
7696 * Return: (Boosted) (estimated) utilization for the specified CPU.
7697 */
7698 static unsigned long
cpu_util(int cpu,struct task_struct * p,int dst_cpu,int boost)7699 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7700 {
7701 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7702 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7703 unsigned long runnable;
7704
7705 if (boost) {
7706 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7707 util = max(util, runnable);
7708 }
7709
7710 /*
7711 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7712 * contribution. If @p migrates from another CPU to @cpu add its
7713 * contribution. In all the other cases @cpu is not impacted by the
7714 * migration so its util_avg is already correct.
7715 */
7716 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7717 lsub_positive(&util, task_util(p));
7718 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7719 util += task_util(p);
7720
7721 if (sched_feat(UTIL_EST)) {
7722 unsigned long util_est;
7723
7724 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7725
7726 /*
7727 * During wake-up @p isn't enqueued yet and doesn't contribute
7728 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7729 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7730 * has been enqueued.
7731 *
7732 * During exec (@dst_cpu = -1) @p is enqueued and does
7733 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7734 * Remove it to "simulate" cpu_util without @p's contribution.
7735 *
7736 * Despite the task_on_rq_queued(@p) check there is still a
7737 * small window for a possible race when an exec
7738 * select_task_rq_fair() races with LB's detach_task().
7739 *
7740 * detach_task()
7741 * deactivate_task()
7742 * p->on_rq = TASK_ON_RQ_MIGRATING;
7743 * -------------------------------- A
7744 * dequeue_task() \
7745 * dequeue_task_fair() + Race Time
7746 * util_est_dequeue() /
7747 * -------------------------------- B
7748 *
7749 * The additional check "current == p" is required to further
7750 * reduce the race window.
7751 */
7752 if (dst_cpu == cpu)
7753 util_est += _task_util_est(p);
7754 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7755 lsub_positive(&util_est, _task_util_est(p));
7756
7757 util = max(util, util_est);
7758 }
7759
7760 return min(util, capacity_orig_of(cpu));
7761 }
7762
cpu_util_cfs(int cpu)7763 unsigned long cpu_util_cfs(int cpu)
7764 {
7765 return cpu_util(cpu, NULL, -1, 0);
7766 }
7767
cpu_util_cfs_boost(int cpu)7768 unsigned long cpu_util_cfs_boost(int cpu)
7769 {
7770 return cpu_util(cpu, NULL, -1, 1);
7771 }
7772
7773 /*
7774 * cpu_util_without: compute cpu utilization without any contributions from *p
7775 * @cpu: the CPU which utilization is requested
7776 * @p: the task which utilization should be discounted
7777 *
7778 * The utilization of a CPU is defined by the utilization of tasks currently
7779 * enqueued on that CPU as well as tasks which are currently sleeping after an
7780 * execution on that CPU.
7781 *
7782 * This method returns the utilization of the specified CPU by discounting the
7783 * utilization of the specified task, whenever the task is currently
7784 * contributing to the CPU utilization.
7785 */
cpu_util_without(int cpu,struct task_struct * p)7786 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7787 {
7788 /* Task has no contribution or is new */
7789 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7790 p = NULL;
7791
7792 return cpu_util(cpu, p, -1, 0);
7793 }
7794
7795 /*
7796 * energy_env - Utilization landscape for energy estimation.
7797 * @task_busy_time: Utilization contribution by the task for which we test the
7798 * placement. Given by eenv_task_busy_time().
7799 * @pd_busy_time: Utilization of the whole perf domain without the task
7800 * contribution. Given by eenv_pd_busy_time().
7801 * @cpu_cap: Maximum CPU capacity for the perf domain.
7802 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7803 */
7804 struct energy_env {
7805 unsigned long task_busy_time;
7806 unsigned long pd_busy_time;
7807 unsigned long cpu_cap;
7808 unsigned long pd_cap;
7809 };
7810
7811 /*
7812 * Compute the task busy time for compute_energy(). This time cannot be
7813 * injected directly into effective_cpu_util() because of the IRQ scaling.
7814 * The latter only makes sense with the most recent CPUs where the task has
7815 * run.
7816 */
eenv_task_busy_time(struct energy_env * eenv,struct task_struct * p,int prev_cpu)7817 static inline void eenv_task_busy_time(struct energy_env *eenv,
7818 struct task_struct *p, int prev_cpu)
7819 {
7820 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7821 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7822
7823 if (unlikely(irq >= max_cap))
7824 busy_time = max_cap;
7825 else
7826 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7827
7828 eenv->task_busy_time = busy_time;
7829 }
7830
7831 /*
7832 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7833 * utilization for each @pd_cpus, it however doesn't take into account
7834 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7835 * scale the EM reported power consumption at the (eventually clamped)
7836 * cpu_capacity.
7837 *
7838 * The contribution of the task @p for which we want to estimate the
7839 * energy cost is removed (by cpu_util()) and must be calculated
7840 * separately (see eenv_task_busy_time). This ensures:
7841 *
7842 * - A stable PD utilization, no matter which CPU of that PD we want to place
7843 * the task on.
7844 *
7845 * - A fair comparison between CPUs as the task contribution (task_util())
7846 * will always be the same no matter which CPU utilization we rely on
7847 * (util_avg or util_est).
7848 *
7849 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7850 * exceed @eenv->pd_cap.
7851 */
eenv_pd_busy_time(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p)7852 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7853 struct cpumask *pd_cpus,
7854 struct task_struct *p)
7855 {
7856 unsigned long busy_time = 0;
7857 int cpu;
7858
7859 for_each_cpu(cpu, pd_cpus) {
7860 unsigned long util = cpu_util(cpu, p, -1, 0);
7861
7862 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7863 }
7864
7865 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7866 }
7867
7868 /*
7869 * Compute the maximum utilization for compute_energy() when the task @p
7870 * is placed on the cpu @dst_cpu.
7871 *
7872 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7873 * exceed @eenv->cpu_cap.
7874 */
7875 static inline unsigned long
eenv_pd_max_util(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p,int dst_cpu)7876 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7877 struct task_struct *p, int dst_cpu)
7878 {
7879 unsigned long max_util = 0;
7880 int cpu;
7881
7882 for_each_cpu(cpu, pd_cpus) {
7883 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7884 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7885 unsigned long eff_util;
7886
7887 /*
7888 * Performance domain frequency: utilization clamping
7889 * must be considered since it affects the selection
7890 * of the performance domain frequency.
7891 * NOTE: in case RT tasks are running, by default the
7892 * FREQUENCY_UTIL's utilization can be max OPP.
7893 */
7894 eff_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7895 max_util = max(max_util, eff_util);
7896 }
7897
7898 return min(max_util, eenv->cpu_cap);
7899 }
7900
7901 /*
7902 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7903 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7904 * contribution is ignored.
7905 */
7906 static inline unsigned long
compute_energy(struct energy_env * eenv,struct perf_domain * pd,struct cpumask * pd_cpus,struct task_struct * p,int dst_cpu)7907 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7908 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7909 {
7910 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7911 unsigned long busy_time = eenv->pd_busy_time;
7912
7913 if (dst_cpu >= 0)
7914 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7915
7916 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7917 }
7918
7919 /*
7920 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7921 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7922 * spare capacity in each performance domain and uses it as a potential
7923 * candidate to execute the task. Then, it uses the Energy Model to figure
7924 * out which of the CPU candidates is the most energy-efficient.
7925 *
7926 * The rationale for this heuristic is as follows. In a performance domain,
7927 * all the most energy efficient CPU candidates (according to the Energy
7928 * Model) are those for which we'll request a low frequency. When there are
7929 * several CPUs for which the frequency request will be the same, we don't
7930 * have enough data to break the tie between them, because the Energy Model
7931 * only includes active power costs. With this model, if we assume that
7932 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7933 * the maximum spare capacity in a performance domain is guaranteed to be among
7934 * the best candidates of the performance domain.
7935 *
7936 * In practice, it could be preferable from an energy standpoint to pack
7937 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7938 * but that could also hurt our chances to go cluster idle, and we have no
7939 * ways to tell with the current Energy Model if this is actually a good
7940 * idea or not. So, find_energy_efficient_cpu() basically favors
7941 * cluster-packing, and spreading inside a cluster. That should at least be
7942 * a good thing for latency, and this is consistent with the idea that most
7943 * of the energy savings of EAS come from the asymmetry of the system, and
7944 * not so much from breaking the tie between identical CPUs. That's also the
7945 * reason why EAS is enabled in the topology code only for systems where
7946 * SD_ASYM_CPUCAPACITY is set.
7947 *
7948 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7949 * they don't have any useful utilization data yet and it's not possible to
7950 * forecast their impact on energy consumption. Consequently, they will be
7951 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7952 * to be energy-inefficient in some use-cases. The alternative would be to
7953 * bias new tasks towards specific types of CPUs first, or to try to infer
7954 * their util_avg from the parent task, but those heuristics could hurt
7955 * other use-cases too. So, until someone finds a better way to solve this,
7956 * let's keep things simple by re-using the existing slow path.
7957 */
find_energy_efficient_cpu(struct task_struct * p,int prev_cpu)7958 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7959 {
7960 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7961 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7962 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7963 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7964 struct root_domain *rd = this_rq()->rd;
7965 int cpu, best_energy_cpu, target = -1;
7966 int prev_fits = -1, best_fits = -1;
7967 unsigned long best_thermal_cap = 0;
7968 unsigned long prev_thermal_cap = 0;
7969 struct sched_domain *sd;
7970 struct perf_domain *pd;
7971 struct energy_env eenv;
7972
7973 rcu_read_lock();
7974 pd = rcu_dereference(rd->pd);
7975 if (!pd || READ_ONCE(rd->overutilized))
7976 goto unlock;
7977
7978 /*
7979 * Energy-aware wake-up happens on the lowest sched_domain starting
7980 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7981 */
7982 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7983 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7984 sd = sd->parent;
7985 if (!sd)
7986 goto unlock;
7987
7988 target = prev_cpu;
7989
7990 sync_entity_load_avg(&p->se);
7991 if (!task_util_est(p) && p_util_min == 0)
7992 goto unlock;
7993
7994 eenv_task_busy_time(&eenv, p, prev_cpu);
7995
7996 for (; pd; pd = pd->next) {
7997 unsigned long util_min = p_util_min, util_max = p_util_max;
7998 unsigned long cpu_cap, cpu_thermal_cap, util;
7999 long prev_spare_cap = -1, max_spare_cap = -1;
8000 unsigned long rq_util_min, rq_util_max;
8001 unsigned long cur_delta, base_energy;
8002 int max_spare_cap_cpu = -1;
8003 int fits, max_fits = -1;
8004
8005 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
8006
8007 if (cpumask_empty(cpus))
8008 continue;
8009
8010 /* Account thermal pressure for the energy estimation */
8011 cpu = cpumask_first(cpus);
8012 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
8013 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
8014
8015 eenv.cpu_cap = cpu_thermal_cap;
8016 eenv.pd_cap = 0;
8017
8018 for_each_cpu(cpu, cpus) {
8019 struct rq *rq = cpu_rq(cpu);
8020
8021 eenv.pd_cap += cpu_thermal_cap;
8022
8023 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
8024 continue;
8025
8026 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
8027 continue;
8028
8029 util = cpu_util(cpu, p, cpu, 0);
8030 cpu_cap = capacity_of(cpu);
8031
8032 /*
8033 * Skip CPUs that cannot satisfy the capacity request.
8034 * IOW, placing the task there would make the CPU
8035 * overutilized. Take uclamp into account to see how
8036 * much capacity we can get out of the CPU; this is
8037 * aligned with sched_cpu_util().
8038 */
8039 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
8040 /*
8041 * Open code uclamp_rq_util_with() except for
8042 * the clamp() part. Ie: apply max aggregation
8043 * only. util_fits_cpu() logic requires to
8044 * operate on non clamped util but must use the
8045 * max-aggregated uclamp_{min, max}.
8046 */
8047 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
8048 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
8049
8050 util_min = max(rq_util_min, p_util_min);
8051 util_max = max(rq_util_max, p_util_max);
8052 }
8053
8054 fits = util_fits_cpu(util, util_min, util_max, cpu);
8055 if (!fits)
8056 continue;
8057
8058 lsub_positive(&cpu_cap, util);
8059
8060 if (cpu == prev_cpu) {
8061 /* Always use prev_cpu as a candidate. */
8062 prev_spare_cap = cpu_cap;
8063 prev_fits = fits;
8064 } else if ((fits > max_fits) ||
8065 ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
8066 /*
8067 * Find the CPU with the maximum spare capacity
8068 * among the remaining CPUs in the performance
8069 * domain.
8070 */
8071 max_spare_cap = cpu_cap;
8072 max_spare_cap_cpu = cpu;
8073 max_fits = fits;
8074 }
8075 }
8076
8077 if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
8078 continue;
8079
8080 eenv_pd_busy_time(&eenv, cpus, p);
8081 /* Compute the 'base' energy of the pd, without @p */
8082 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
8083
8084 /* Evaluate the energy impact of using prev_cpu. */
8085 if (prev_spare_cap > -1) {
8086 prev_delta = compute_energy(&eenv, pd, cpus, p,
8087 prev_cpu);
8088 /* CPU utilization has changed */
8089 if (prev_delta < base_energy)
8090 goto unlock;
8091 prev_delta -= base_energy;
8092 prev_thermal_cap = cpu_thermal_cap;
8093 best_delta = min(best_delta, prev_delta);
8094 }
8095
8096 /* Evaluate the energy impact of using max_spare_cap_cpu. */
8097 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
8098 /* Current best energy cpu fits better */
8099 if (max_fits < best_fits)
8100 continue;
8101
8102 /*
8103 * Both don't fit performance hint (i.e. uclamp_min)
8104 * but best energy cpu has better capacity.
8105 */
8106 if ((max_fits < 0) &&
8107 (cpu_thermal_cap <= best_thermal_cap))
8108 continue;
8109
8110 cur_delta = compute_energy(&eenv, pd, cpus, p,
8111 max_spare_cap_cpu);
8112 /* CPU utilization has changed */
8113 if (cur_delta < base_energy)
8114 goto unlock;
8115 cur_delta -= base_energy;
8116
8117 /*
8118 * Both fit for the task but best energy cpu has lower
8119 * energy impact.
8120 */
8121 if ((max_fits > 0) && (best_fits > 0) &&
8122 (cur_delta >= best_delta))
8123 continue;
8124
8125 best_delta = cur_delta;
8126 best_energy_cpu = max_spare_cap_cpu;
8127 best_fits = max_fits;
8128 best_thermal_cap = cpu_thermal_cap;
8129 }
8130 }
8131 rcu_read_unlock();
8132
8133 if ((best_fits > prev_fits) ||
8134 ((best_fits > 0) && (best_delta < prev_delta)) ||
8135 ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
8136 target = best_energy_cpu;
8137
8138 return target;
8139
8140 unlock:
8141 rcu_read_unlock();
8142
8143 return target;
8144 }
8145
8146 /*
8147 * select_task_rq_fair: Select target runqueue for the waking task in domains
8148 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8149 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8150 *
8151 * Balances load by selecting the idlest CPU in the idlest group, or under
8152 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8153 *
8154 * Returns the target CPU number.
8155 */
8156 static int
select_task_rq_fair(struct task_struct * p,int prev_cpu,int wake_flags)8157 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8158 {
8159 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8160 struct sched_domain *tmp, *sd = NULL;
8161 int cpu = smp_processor_id();
8162 int new_cpu = prev_cpu;
8163 int want_affine = 0;
8164 /* SD_flags and WF_flags share the first nibble */
8165 int sd_flag = wake_flags & 0xF;
8166
8167 /*
8168 * required for stable ->cpus_allowed
8169 */
8170 lockdep_assert_held(&p->pi_lock);
8171 if (wake_flags & WF_TTWU) {
8172 record_wakee(p);
8173
8174 if ((wake_flags & WF_CURRENT_CPU) &&
8175 cpumask_test_cpu(cpu, p->cpus_ptr))
8176 return cpu;
8177
8178 if (sched_energy_enabled()) {
8179 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8180 if (new_cpu >= 0)
8181 return new_cpu;
8182 new_cpu = prev_cpu;
8183 }
8184
8185 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8186 }
8187
8188 rcu_read_lock();
8189 for_each_domain(cpu, tmp) {
8190 /*
8191 * If both 'cpu' and 'prev_cpu' are part of this domain,
8192 * cpu is a valid SD_WAKE_AFFINE target.
8193 */
8194 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8195 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8196 if (cpu != prev_cpu)
8197 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8198
8199 sd = NULL; /* Prefer wake_affine over balance flags */
8200 break;
8201 }
8202
8203 /*
8204 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8205 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8206 * will usually go to the fast path.
8207 */
8208 if (tmp->flags & sd_flag)
8209 sd = tmp;
8210 else if (!want_affine)
8211 break;
8212 }
8213
8214 if (unlikely(sd)) {
8215 /* Slow path */
8216 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
8217 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
8218 /* Fast path */
8219 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8220 }
8221 rcu_read_unlock();
8222
8223 return new_cpu;
8224 }
8225
8226 /*
8227 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8228 * cfs_rq_of(p) references at time of call are still valid and identify the
8229 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8230 */
migrate_task_rq_fair(struct task_struct * p,int new_cpu)8231 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8232 {
8233 struct sched_entity *se = &p->se;
8234
8235 if (!task_on_rq_migrating(p)) {
8236 remove_entity_load_avg(se);
8237
8238 /*
8239 * Here, the task's PELT values have been updated according to
8240 * the current rq's clock. But if that clock hasn't been
8241 * updated in a while, a substantial idle time will be missed,
8242 * leading to an inflation after wake-up on the new rq.
8243 *
8244 * Estimate the missing time from the cfs_rq last_update_time
8245 * and update sched_avg to improve the PELT continuity after
8246 * migration.
8247 */
8248 migrate_se_pelt_lag(se);
8249 }
8250
8251 /* Tell new CPU we are migrated */
8252 se->avg.last_update_time = 0;
8253
8254 update_scan_period(p, new_cpu);
8255 }
8256
task_dead_fair(struct task_struct * p)8257 static void task_dead_fair(struct task_struct *p)
8258 {
8259 remove_entity_load_avg(&p->se);
8260 }
8261
8262 static int
balance_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8263 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8264 {
8265 if (rq->nr_running)
8266 return 1;
8267
8268 return newidle_balance(rq, rf) != 0;
8269 }
8270 #endif /* CONFIG_SMP */
8271
set_next_buddy(struct sched_entity * se)8272 static void set_next_buddy(struct sched_entity *se)
8273 {
8274 for_each_sched_entity(se) {
8275 if (SCHED_WARN_ON(!se->on_rq))
8276 return;
8277 if (se_is_idle(se))
8278 return;
8279 cfs_rq_of(se)->next = se;
8280 }
8281 }
8282
8283 /*
8284 * Preempt the current task with a newly woken task if needed:
8285 */
check_preempt_wakeup_fair(struct rq * rq,struct task_struct * p,int wake_flags)8286 static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int wake_flags)
8287 {
8288 struct task_struct *curr = rq->curr;
8289 struct sched_entity *se = &curr->se, *pse = &p->se;
8290 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8291 int next_buddy_marked = 0;
8292 int cse_is_idle, pse_is_idle;
8293
8294 if (unlikely(se == pse))
8295 return;
8296
8297 /*
8298 * This is possible from callers such as attach_tasks(), in which we
8299 * unconditionally wakeup_preempt() after an enqueue (which may have
8300 * lead to a throttle). This both saves work and prevents false
8301 * next-buddy nomination below.
8302 */
8303 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8304 return;
8305
8306 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8307 set_next_buddy(pse);
8308 next_buddy_marked = 1;
8309 }
8310
8311 /*
8312 * We can come here with TIF_NEED_RESCHED already set from new task
8313 * wake up path.
8314 *
8315 * Note: this also catches the edge-case of curr being in a throttled
8316 * group (e.g. via set_curr_task), since update_curr() (in the
8317 * enqueue of curr) will have resulted in resched being set. This
8318 * prevents us from potentially nominating it as a false LAST_BUDDY
8319 * below.
8320 */
8321 if (test_tsk_need_resched(curr))
8322 return;
8323
8324 if (!sched_feat(WAKEUP_PREEMPTION))
8325 return;
8326
8327 find_matching_se(&se, &pse);
8328 WARN_ON_ONCE(!pse);
8329
8330 cse_is_idle = se_is_idle(se);
8331 pse_is_idle = se_is_idle(pse);
8332
8333 /*
8334 * Preempt an idle entity in favor of a non-idle entity (and don't preempt
8335 * in the inverse case).
8336 */
8337 if (cse_is_idle && !pse_is_idle)
8338 goto preempt;
8339 if (cse_is_idle != pse_is_idle)
8340 return;
8341
8342 /*
8343 * BATCH and IDLE tasks do not preempt others.
8344 */
8345 if (unlikely(p->policy != SCHED_NORMAL))
8346 return;
8347
8348 cfs_rq = cfs_rq_of(se);
8349 update_curr(cfs_rq);
8350 /*
8351 * XXX pick_eevdf(cfs_rq) != se ?
8352 */
8353 if (pick_eevdf(cfs_rq) == pse)
8354 goto preempt;
8355
8356 return;
8357
8358 preempt:
8359 resched_curr(rq);
8360 }
8361
8362 #ifdef CONFIG_SMP
pick_task_fair(struct rq * rq)8363 static struct task_struct *pick_task_fair(struct rq *rq)
8364 {
8365 struct sched_entity *se;
8366 struct cfs_rq *cfs_rq;
8367
8368 again:
8369 cfs_rq = &rq->cfs;
8370 if (!cfs_rq->nr_running)
8371 return NULL;
8372
8373 do {
8374 struct sched_entity *curr = cfs_rq->curr;
8375
8376 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8377 if (curr) {
8378 if (curr->on_rq)
8379 update_curr(cfs_rq);
8380 else
8381 curr = NULL;
8382
8383 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8384 goto again;
8385 }
8386
8387 se = pick_next_entity(cfs_rq, curr);
8388 cfs_rq = group_cfs_rq(se);
8389 } while (cfs_rq);
8390
8391 return task_of(se);
8392 }
8393 #endif
8394
8395 struct task_struct *
pick_next_task_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8396 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8397 {
8398 struct cfs_rq *cfs_rq = &rq->cfs;
8399 struct sched_entity *se;
8400 struct task_struct *p;
8401 int new_tasks;
8402
8403 again:
8404 if (!sched_fair_runnable(rq))
8405 goto idle;
8406
8407 #ifdef CONFIG_FAIR_GROUP_SCHED
8408 if (!prev || prev->sched_class != &fair_sched_class)
8409 goto simple;
8410
8411 /*
8412 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8413 * likely that a next task is from the same cgroup as the current.
8414 *
8415 * Therefore attempt to avoid putting and setting the entire cgroup
8416 * hierarchy, only change the part that actually changes.
8417 */
8418
8419 do {
8420 struct sched_entity *curr = cfs_rq->curr;
8421
8422 /*
8423 * Since we got here without doing put_prev_entity() we also
8424 * have to consider cfs_rq->curr. If it is still a runnable
8425 * entity, update_curr() will update its vruntime, otherwise
8426 * forget we've ever seen it.
8427 */
8428 if (curr) {
8429 if (curr->on_rq)
8430 update_curr(cfs_rq);
8431 else
8432 curr = NULL;
8433
8434 /*
8435 * This call to check_cfs_rq_runtime() will do the
8436 * throttle and dequeue its entity in the parent(s).
8437 * Therefore the nr_running test will indeed
8438 * be correct.
8439 */
8440 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8441 cfs_rq = &rq->cfs;
8442
8443 if (!cfs_rq->nr_running)
8444 goto idle;
8445
8446 goto simple;
8447 }
8448 }
8449
8450 se = pick_next_entity(cfs_rq, curr);
8451 cfs_rq = group_cfs_rq(se);
8452 } while (cfs_rq);
8453
8454 p = task_of(se);
8455
8456 /*
8457 * Since we haven't yet done put_prev_entity and if the selected task
8458 * is a different task than we started out with, try and touch the
8459 * least amount of cfs_rqs.
8460 */
8461 if (prev != p) {
8462 struct sched_entity *pse = &prev->se;
8463
8464 while (!(cfs_rq = is_same_group(se, pse))) {
8465 int se_depth = se->depth;
8466 int pse_depth = pse->depth;
8467
8468 if (se_depth <= pse_depth) {
8469 put_prev_entity(cfs_rq_of(pse), pse);
8470 pse = parent_entity(pse);
8471 }
8472 if (se_depth >= pse_depth) {
8473 set_next_entity(cfs_rq_of(se), se);
8474 se = parent_entity(se);
8475 }
8476 }
8477
8478 put_prev_entity(cfs_rq, pse);
8479 set_next_entity(cfs_rq, se);
8480 }
8481
8482 goto done;
8483 simple:
8484 #endif
8485 if (prev)
8486 put_prev_task(rq, prev);
8487
8488 do {
8489 se = pick_next_entity(cfs_rq, NULL);
8490 set_next_entity(cfs_rq, se);
8491 cfs_rq = group_cfs_rq(se);
8492 } while (cfs_rq);
8493
8494 p = task_of(se);
8495
8496 done: __maybe_unused;
8497 #ifdef CONFIG_SMP
8498 /*
8499 * Move the next running task to the front of
8500 * the list, so our cfs_tasks list becomes MRU
8501 * one.
8502 */
8503 list_move(&p->se.group_node, &rq->cfs_tasks);
8504 #endif
8505
8506 if (hrtick_enabled_fair(rq))
8507 hrtick_start_fair(rq, p);
8508
8509 update_misfit_status(p, rq);
8510 sched_fair_update_stop_tick(rq, p);
8511
8512 return p;
8513
8514 idle:
8515 if (!rf)
8516 return NULL;
8517
8518 new_tasks = newidle_balance(rq, rf);
8519
8520 /*
8521 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8522 * possible for any higher priority task to appear. In that case we
8523 * must re-start the pick_next_entity() loop.
8524 */
8525 if (new_tasks < 0)
8526 return RETRY_TASK;
8527
8528 if (new_tasks > 0)
8529 goto again;
8530
8531 /*
8532 * rq is about to be idle, check if we need to update the
8533 * lost_idle_time of clock_pelt
8534 */
8535 update_idle_rq_clock_pelt(rq);
8536
8537 return NULL;
8538 }
8539
__pick_next_task_fair(struct rq * rq)8540 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8541 {
8542 return pick_next_task_fair(rq, NULL, NULL);
8543 }
8544
8545 /*
8546 * Account for a descheduled task:
8547 */
put_prev_task_fair(struct rq * rq,struct task_struct * prev)8548 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8549 {
8550 struct sched_entity *se = &prev->se;
8551 struct cfs_rq *cfs_rq;
8552
8553 for_each_sched_entity(se) {
8554 cfs_rq = cfs_rq_of(se);
8555 put_prev_entity(cfs_rq, se);
8556 }
8557 }
8558
8559 /*
8560 * sched_yield() is very simple
8561 */
yield_task_fair(struct rq * rq)8562 static void yield_task_fair(struct rq *rq)
8563 {
8564 struct task_struct *curr = rq->curr;
8565 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8566 struct sched_entity *se = &curr->se;
8567
8568 /*
8569 * Are we the only task in the tree?
8570 */
8571 if (unlikely(rq->nr_running == 1))
8572 return;
8573
8574 clear_buddies(cfs_rq, se);
8575
8576 update_rq_clock(rq);
8577 /*
8578 * Update run-time statistics of the 'current'.
8579 */
8580 update_curr(cfs_rq);
8581 /*
8582 * Tell update_rq_clock() that we've just updated,
8583 * so we don't do microscopic update in schedule()
8584 * and double the fastpath cost.
8585 */
8586 rq_clock_skip_update(rq);
8587
8588 se->deadline += calc_delta_fair(se->slice, se);
8589 }
8590
yield_to_task_fair(struct rq * rq,struct task_struct * p)8591 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8592 {
8593 struct sched_entity *se = &p->se;
8594
8595 /* throttled hierarchies are not runnable */
8596 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8597 return false;
8598
8599 /* Tell the scheduler that we'd really like pse to run next. */
8600 set_next_buddy(se);
8601
8602 yield_task_fair(rq);
8603
8604 return true;
8605 }
8606
8607 #ifdef CONFIG_SMP
8608 /**************************************************
8609 * Fair scheduling class load-balancing methods.
8610 *
8611 * BASICS
8612 *
8613 * The purpose of load-balancing is to achieve the same basic fairness the
8614 * per-CPU scheduler provides, namely provide a proportional amount of compute
8615 * time to each task. This is expressed in the following equation:
8616 *
8617 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8618 *
8619 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8620 * W_i,0 is defined as:
8621 *
8622 * W_i,0 = \Sum_j w_i,j (2)
8623 *
8624 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8625 * is derived from the nice value as per sched_prio_to_weight[].
8626 *
8627 * The weight average is an exponential decay average of the instantaneous
8628 * weight:
8629 *
8630 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8631 *
8632 * C_i is the compute capacity of CPU i, typically it is the
8633 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8634 * can also include other factors [XXX].
8635 *
8636 * To achieve this balance we define a measure of imbalance which follows
8637 * directly from (1):
8638 *
8639 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8640 *
8641 * We them move tasks around to minimize the imbalance. In the continuous
8642 * function space it is obvious this converges, in the discrete case we get
8643 * a few fun cases generally called infeasible weight scenarios.
8644 *
8645 * [XXX expand on:
8646 * - infeasible weights;
8647 * - local vs global optima in the discrete case. ]
8648 *
8649 *
8650 * SCHED DOMAINS
8651 *
8652 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8653 * for all i,j solution, we create a tree of CPUs that follows the hardware
8654 * topology where each level pairs two lower groups (or better). This results
8655 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8656 * tree to only the first of the previous level and we decrease the frequency
8657 * of load-balance at each level inv. proportional to the number of CPUs in
8658 * the groups.
8659 *
8660 * This yields:
8661 *
8662 * log_2 n 1 n
8663 * \Sum { --- * --- * 2^i } = O(n) (5)
8664 * i = 0 2^i 2^i
8665 * `- size of each group
8666 * | | `- number of CPUs doing load-balance
8667 * | `- freq
8668 * `- sum over all levels
8669 *
8670 * Coupled with a limit on how many tasks we can migrate every balance pass,
8671 * this makes (5) the runtime complexity of the balancer.
8672 *
8673 * An important property here is that each CPU is still (indirectly) connected
8674 * to every other CPU in at most O(log n) steps:
8675 *
8676 * The adjacency matrix of the resulting graph is given by:
8677 *
8678 * log_2 n
8679 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8680 * k = 0
8681 *
8682 * And you'll find that:
8683 *
8684 * A^(log_2 n)_i,j != 0 for all i,j (7)
8685 *
8686 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8687 * The task movement gives a factor of O(m), giving a convergence complexity
8688 * of:
8689 *
8690 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8691 *
8692 *
8693 * WORK CONSERVING
8694 *
8695 * In order to avoid CPUs going idle while there's still work to do, new idle
8696 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8697 * tree itself instead of relying on other CPUs to bring it work.
8698 *
8699 * This adds some complexity to both (5) and (8) but it reduces the total idle
8700 * time.
8701 *
8702 * [XXX more?]
8703 *
8704 *
8705 * CGROUPS
8706 *
8707 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8708 *
8709 * s_k,i
8710 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8711 * S_k
8712 *
8713 * Where
8714 *
8715 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8716 *
8717 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8718 *
8719 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8720 * property.
8721 *
8722 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8723 * rewrite all of this once again.]
8724 */
8725
8726 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8727
8728 enum fbq_type { regular, remote, all };
8729
8730 /*
8731 * 'group_type' describes the group of CPUs at the moment of load balancing.
8732 *
8733 * The enum is ordered by pulling priority, with the group with lowest priority
8734 * first so the group_type can simply be compared when selecting the busiest
8735 * group. See update_sd_pick_busiest().
8736 */
8737 enum group_type {
8738 /* The group has spare capacity that can be used to run more tasks. */
8739 group_has_spare = 0,
8740 /*
8741 * The group is fully used and the tasks don't compete for more CPU
8742 * cycles. Nevertheless, some tasks might wait before running.
8743 */
8744 group_fully_busy,
8745 /*
8746 * One task doesn't fit with CPU's capacity and must be migrated to a
8747 * more powerful CPU.
8748 */
8749 group_misfit_task,
8750 /*
8751 * Balance SMT group that's fully busy. Can benefit from migration
8752 * a task on SMT with busy sibling to another CPU on idle core.
8753 */
8754 group_smt_balance,
8755 /*
8756 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8757 * and the task should be migrated to it instead of running on the
8758 * current CPU.
8759 */
8760 group_asym_packing,
8761 /*
8762 * The tasks' affinity constraints previously prevented the scheduler
8763 * from balancing the load across the system.
8764 */
8765 group_imbalanced,
8766 /*
8767 * The CPU is overloaded and can't provide expected CPU cycles to all
8768 * tasks.
8769 */
8770 group_overloaded
8771 };
8772
8773 enum migration_type {
8774 migrate_load = 0,
8775 migrate_util,
8776 migrate_task,
8777 migrate_misfit
8778 };
8779
8780 #define LBF_ALL_PINNED 0x01
8781 #define LBF_NEED_BREAK 0x02
8782 #define LBF_DST_PINNED 0x04
8783 #define LBF_SOME_PINNED 0x08
8784 #define LBF_ACTIVE_LB 0x10
8785
8786 struct lb_env {
8787 struct sched_domain *sd;
8788
8789 struct rq *src_rq;
8790 int src_cpu;
8791
8792 int dst_cpu;
8793 struct rq *dst_rq;
8794
8795 struct cpumask *dst_grpmask;
8796 int new_dst_cpu;
8797 enum cpu_idle_type idle;
8798 long imbalance;
8799 /* The set of CPUs under consideration for load-balancing */
8800 struct cpumask *cpus;
8801
8802 unsigned int flags;
8803
8804 unsigned int loop;
8805 unsigned int loop_break;
8806 unsigned int loop_max;
8807
8808 enum fbq_type fbq_type;
8809 enum migration_type migration_type;
8810 struct list_head tasks;
8811 };
8812
8813 /*
8814 * Is this task likely cache-hot:
8815 */
task_hot(struct task_struct * p,struct lb_env * env)8816 static int task_hot(struct task_struct *p, struct lb_env *env)
8817 {
8818 s64 delta;
8819
8820 lockdep_assert_rq_held(env->src_rq);
8821
8822 if (p->sched_class != &fair_sched_class)
8823 return 0;
8824
8825 if (unlikely(task_has_idle_policy(p)))
8826 return 0;
8827
8828 /* SMT siblings share cache */
8829 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8830 return 0;
8831
8832 /*
8833 * Buddy candidates are cache hot:
8834 */
8835 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8836 (&p->se == cfs_rq_of(&p->se)->next))
8837 return 1;
8838
8839 if (sysctl_sched_migration_cost == -1)
8840 return 1;
8841
8842 /*
8843 * Don't migrate task if the task's cookie does not match
8844 * with the destination CPU's core cookie.
8845 */
8846 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8847 return 1;
8848
8849 if (sysctl_sched_migration_cost == 0)
8850 return 0;
8851
8852 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8853
8854 return delta < (s64)sysctl_sched_migration_cost;
8855 }
8856
8857 #ifdef CONFIG_NUMA_BALANCING
8858 /*
8859 * Returns 1, if task migration degrades locality
8860 * Returns 0, if task migration improves locality i.e migration preferred.
8861 * Returns -1, if task migration is not affected by locality.
8862 */
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)8863 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8864 {
8865 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8866 unsigned long src_weight, dst_weight;
8867 int src_nid, dst_nid, dist;
8868
8869 if (!static_branch_likely(&sched_numa_balancing))
8870 return -1;
8871
8872 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8873 return -1;
8874
8875 src_nid = cpu_to_node(env->src_cpu);
8876 dst_nid = cpu_to_node(env->dst_cpu);
8877
8878 if (src_nid == dst_nid)
8879 return -1;
8880
8881 /* Migrating away from the preferred node is always bad. */
8882 if (src_nid == p->numa_preferred_nid) {
8883 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8884 return 1;
8885 else
8886 return -1;
8887 }
8888
8889 /* Encourage migration to the preferred node. */
8890 if (dst_nid == p->numa_preferred_nid)
8891 return 0;
8892
8893 /* Leaving a core idle is often worse than degrading locality. */
8894 if (env->idle == CPU_IDLE)
8895 return -1;
8896
8897 dist = node_distance(src_nid, dst_nid);
8898 if (numa_group) {
8899 src_weight = group_weight(p, src_nid, dist);
8900 dst_weight = group_weight(p, dst_nid, dist);
8901 } else {
8902 src_weight = task_weight(p, src_nid, dist);
8903 dst_weight = task_weight(p, dst_nid, dist);
8904 }
8905
8906 return dst_weight < src_weight;
8907 }
8908
8909 #else
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)8910 static inline int migrate_degrades_locality(struct task_struct *p,
8911 struct lb_env *env)
8912 {
8913 return -1;
8914 }
8915 #endif
8916
8917 /*
8918 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8919 */
8920 static
can_migrate_task(struct task_struct * p,struct lb_env * env)8921 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8922 {
8923 int tsk_cache_hot;
8924
8925 lockdep_assert_rq_held(env->src_rq);
8926 if (p->sched_task_hot)
8927 p->sched_task_hot = 0;
8928
8929 /*
8930 * We do not migrate tasks that are:
8931 * 1) throttled_lb_pair, or
8932 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8933 * 3) running (obviously), or
8934 * 4) are cache-hot on their current CPU.
8935 */
8936 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8937 return 0;
8938
8939 /* Disregard pcpu kthreads; they are where they need to be. */
8940 if (kthread_is_per_cpu(p))
8941 return 0;
8942
8943 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8944 int cpu;
8945
8946 schedstat_inc(p->stats.nr_failed_migrations_affine);
8947
8948 env->flags |= LBF_SOME_PINNED;
8949
8950 /*
8951 * Remember if this task can be migrated to any other CPU in
8952 * our sched_group. We may want to revisit it if we couldn't
8953 * meet load balance goals by pulling other tasks on src_cpu.
8954 *
8955 * Avoid computing new_dst_cpu
8956 * - for NEWLY_IDLE
8957 * - if we have already computed one in current iteration
8958 * - if it's an active balance
8959 */
8960 if (env->idle == CPU_NEWLY_IDLE ||
8961 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8962 return 0;
8963
8964 /* Prevent to re-select dst_cpu via env's CPUs: */
8965 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8966 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8967 env->flags |= LBF_DST_PINNED;
8968 env->new_dst_cpu = cpu;
8969 break;
8970 }
8971 }
8972
8973 return 0;
8974 }
8975
8976 /* Record that we found at least one task that could run on dst_cpu */
8977 env->flags &= ~LBF_ALL_PINNED;
8978
8979 if (task_on_cpu(env->src_rq, p)) {
8980 schedstat_inc(p->stats.nr_failed_migrations_running);
8981 return 0;
8982 }
8983
8984 /*
8985 * Aggressive migration if:
8986 * 1) active balance
8987 * 2) destination numa is preferred
8988 * 3) task is cache cold, or
8989 * 4) too many balance attempts have failed.
8990 */
8991 if (env->flags & LBF_ACTIVE_LB)
8992 return 1;
8993
8994 tsk_cache_hot = migrate_degrades_locality(p, env);
8995 if (tsk_cache_hot == -1)
8996 tsk_cache_hot = task_hot(p, env);
8997
8998 if (tsk_cache_hot <= 0 ||
8999 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
9000 if (tsk_cache_hot == 1)
9001 p->sched_task_hot = 1;
9002 return 1;
9003 }
9004
9005 schedstat_inc(p->stats.nr_failed_migrations_hot);
9006 return 0;
9007 }
9008
9009 /*
9010 * detach_task() -- detach the task for the migration specified in env
9011 */
detach_task(struct task_struct * p,struct lb_env * env)9012 static void detach_task(struct task_struct *p, struct lb_env *env)
9013 {
9014 lockdep_assert_rq_held(env->src_rq);
9015
9016 if (p->sched_task_hot) {
9017 p->sched_task_hot = 0;
9018 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
9019 schedstat_inc(p->stats.nr_forced_migrations);
9020 }
9021
9022 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
9023 set_task_cpu(p, env->dst_cpu);
9024 }
9025
9026 /*
9027 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9028 * part of active balancing operations within "domain".
9029 *
9030 * Returns a task if successful and NULL otherwise.
9031 */
detach_one_task(struct lb_env * env)9032 static struct task_struct *detach_one_task(struct lb_env *env)
9033 {
9034 struct task_struct *p;
9035
9036 lockdep_assert_rq_held(env->src_rq);
9037
9038 list_for_each_entry_reverse(p,
9039 &env->src_rq->cfs_tasks, se.group_node) {
9040 if (!can_migrate_task(p, env))
9041 continue;
9042
9043 detach_task(p, env);
9044
9045 /*
9046 * Right now, this is only the second place where
9047 * lb_gained[env->idle] is updated (other is detach_tasks)
9048 * so we can safely collect stats here rather than
9049 * inside detach_tasks().
9050 */
9051 schedstat_inc(env->sd->lb_gained[env->idle]);
9052 return p;
9053 }
9054 return NULL;
9055 }
9056
9057 /*
9058 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9059 * busiest_rq, as part of a balancing operation within domain "sd".
9060 *
9061 * Returns number of detached tasks if successful and 0 otherwise.
9062 */
detach_tasks(struct lb_env * env)9063 static int detach_tasks(struct lb_env *env)
9064 {
9065 struct list_head *tasks = &env->src_rq->cfs_tasks;
9066 unsigned long util, load;
9067 struct task_struct *p;
9068 int detached = 0;
9069
9070 lockdep_assert_rq_held(env->src_rq);
9071
9072 /*
9073 * Source run queue has been emptied by another CPU, clear
9074 * LBF_ALL_PINNED flag as we will not test any task.
9075 */
9076 if (env->src_rq->nr_running <= 1) {
9077 env->flags &= ~LBF_ALL_PINNED;
9078 return 0;
9079 }
9080
9081 if (env->imbalance <= 0)
9082 return 0;
9083
9084 while (!list_empty(tasks)) {
9085 /*
9086 * We don't want to steal all, otherwise we may be treated likewise,
9087 * which could at worst lead to a livelock crash.
9088 */
9089 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
9090 break;
9091
9092 env->loop++;
9093 /* We've more or less seen every task there is, call it quits */
9094 if (env->loop > env->loop_max)
9095 break;
9096
9097 /* take a breather every nr_migrate tasks */
9098 if (env->loop > env->loop_break) {
9099 env->loop_break += SCHED_NR_MIGRATE_BREAK;
9100 env->flags |= LBF_NEED_BREAK;
9101 break;
9102 }
9103
9104 p = list_last_entry(tasks, struct task_struct, se.group_node);
9105
9106 if (!can_migrate_task(p, env))
9107 goto next;
9108
9109 switch (env->migration_type) {
9110 case migrate_load:
9111 /*
9112 * Depending of the number of CPUs and tasks and the
9113 * cgroup hierarchy, task_h_load() can return a null
9114 * value. Make sure that env->imbalance decreases
9115 * otherwise detach_tasks() will stop only after
9116 * detaching up to loop_max tasks.
9117 */
9118 load = max_t(unsigned long, task_h_load(p), 1);
9119
9120 if (sched_feat(LB_MIN) &&
9121 load < 16 && !env->sd->nr_balance_failed)
9122 goto next;
9123
9124 /*
9125 * Make sure that we don't migrate too much load.
9126 * Nevertheless, let relax the constraint if
9127 * scheduler fails to find a good waiting task to
9128 * migrate.
9129 */
9130 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
9131 goto next;
9132
9133 env->imbalance -= load;
9134 break;
9135
9136 case migrate_util:
9137 util = task_util_est(p);
9138
9139 if (util > env->imbalance)
9140 goto next;
9141
9142 env->imbalance -= util;
9143 break;
9144
9145 case migrate_task:
9146 env->imbalance--;
9147 break;
9148
9149 case migrate_misfit:
9150 /* This is not a misfit task */
9151 if (task_fits_cpu(p, env->src_cpu))
9152 goto next;
9153
9154 env->imbalance = 0;
9155 break;
9156 }
9157
9158 detach_task(p, env);
9159 list_add(&p->se.group_node, &env->tasks);
9160
9161 detached++;
9162
9163 #ifdef CONFIG_PREEMPTION
9164 /*
9165 * NEWIDLE balancing is a source of latency, so preemptible
9166 * kernels will stop after the first task is detached to minimize
9167 * the critical section.
9168 */
9169 if (env->idle == CPU_NEWLY_IDLE)
9170 break;
9171 #endif
9172
9173 /*
9174 * We only want to steal up to the prescribed amount of
9175 * load/util/tasks.
9176 */
9177 if (env->imbalance <= 0)
9178 break;
9179
9180 continue;
9181 next:
9182 if (p->sched_task_hot)
9183 schedstat_inc(p->stats.nr_failed_migrations_hot);
9184
9185 list_move(&p->se.group_node, tasks);
9186 }
9187
9188 /*
9189 * Right now, this is one of only two places we collect this stat
9190 * so we can safely collect detach_one_task() stats here rather
9191 * than inside detach_one_task().
9192 */
9193 schedstat_add(env->sd->lb_gained[env->idle], detached);
9194
9195 return detached;
9196 }
9197
9198 /*
9199 * attach_task() -- attach the task detached by detach_task() to its new rq.
9200 */
attach_task(struct rq * rq,struct task_struct * p)9201 static void attach_task(struct rq *rq, struct task_struct *p)
9202 {
9203 lockdep_assert_rq_held(rq);
9204
9205 WARN_ON_ONCE(task_rq(p) != rq);
9206 activate_task(rq, p, ENQUEUE_NOCLOCK);
9207 wakeup_preempt(rq, p, 0);
9208 }
9209
9210 /*
9211 * attach_one_task() -- attaches the task returned from detach_one_task() to
9212 * its new rq.
9213 */
attach_one_task(struct rq * rq,struct task_struct * p)9214 static void attach_one_task(struct rq *rq, struct task_struct *p)
9215 {
9216 struct rq_flags rf;
9217
9218 rq_lock(rq, &rf);
9219 update_rq_clock(rq);
9220 attach_task(rq, p);
9221 rq_unlock(rq, &rf);
9222 }
9223
9224 /*
9225 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9226 * new rq.
9227 */
attach_tasks(struct lb_env * env)9228 static void attach_tasks(struct lb_env *env)
9229 {
9230 struct list_head *tasks = &env->tasks;
9231 struct task_struct *p;
9232 struct rq_flags rf;
9233
9234 rq_lock(env->dst_rq, &rf);
9235 update_rq_clock(env->dst_rq);
9236
9237 while (!list_empty(tasks)) {
9238 p = list_first_entry(tasks, struct task_struct, se.group_node);
9239 list_del_init(&p->se.group_node);
9240
9241 attach_task(env->dst_rq, p);
9242 }
9243
9244 rq_unlock(env->dst_rq, &rf);
9245 }
9246
9247 #ifdef CONFIG_NO_HZ_COMMON
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9248 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9249 {
9250 if (cfs_rq->avg.load_avg)
9251 return true;
9252
9253 if (cfs_rq->avg.util_avg)
9254 return true;
9255
9256 return false;
9257 }
9258
others_have_blocked(struct rq * rq)9259 static inline bool others_have_blocked(struct rq *rq)
9260 {
9261 if (READ_ONCE(rq->avg_rt.util_avg))
9262 return true;
9263
9264 if (READ_ONCE(rq->avg_dl.util_avg))
9265 return true;
9266
9267 if (thermal_load_avg(rq))
9268 return true;
9269
9270 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
9271 if (READ_ONCE(rq->avg_irq.util_avg))
9272 return true;
9273 #endif
9274
9275 return false;
9276 }
9277
update_blocked_load_tick(struct rq * rq)9278 static inline void update_blocked_load_tick(struct rq *rq)
9279 {
9280 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9281 }
9282
update_blocked_load_status(struct rq * rq,bool has_blocked)9283 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9284 {
9285 if (!has_blocked)
9286 rq->has_blocked_load = 0;
9287 }
9288 #else
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9289 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
others_have_blocked(struct rq * rq)9290 static inline bool others_have_blocked(struct rq *rq) { return false; }
update_blocked_load_tick(struct rq * rq)9291 static inline void update_blocked_load_tick(struct rq *rq) {}
update_blocked_load_status(struct rq * rq,bool has_blocked)9292 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9293 #endif
9294
__update_blocked_others(struct rq * rq,bool * done)9295 static bool __update_blocked_others(struct rq *rq, bool *done)
9296 {
9297 const struct sched_class *curr_class;
9298 u64 now = rq_clock_pelt(rq);
9299 unsigned long thermal_pressure;
9300 bool decayed;
9301
9302 /*
9303 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9304 * DL and IRQ signals have been updated before updating CFS.
9305 */
9306 curr_class = rq->curr->sched_class;
9307
9308 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9309
9310 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9311 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9312 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9313 update_irq_load_avg(rq, 0);
9314
9315 if (others_have_blocked(rq))
9316 *done = false;
9317
9318 return decayed;
9319 }
9320
9321 #ifdef CONFIG_FAIR_GROUP_SCHED
9322
__update_blocked_fair(struct rq * rq,bool * done)9323 static bool __update_blocked_fair(struct rq *rq, bool *done)
9324 {
9325 struct cfs_rq *cfs_rq, *pos;
9326 bool decayed = false;
9327 int cpu = cpu_of(rq);
9328
9329 /*
9330 * Iterates the task_group tree in a bottom up fashion, see
9331 * list_add_leaf_cfs_rq() for details.
9332 */
9333 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9334 struct sched_entity *se;
9335
9336 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9337 update_tg_load_avg(cfs_rq);
9338
9339 if (cfs_rq->nr_running == 0)
9340 update_idle_cfs_rq_clock_pelt(cfs_rq);
9341
9342 if (cfs_rq == &rq->cfs)
9343 decayed = true;
9344 }
9345
9346 /* Propagate pending load changes to the parent, if any: */
9347 se = cfs_rq->tg->se[cpu];
9348 if (se && !skip_blocked_update(se))
9349 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9350
9351 /*
9352 * There can be a lot of idle CPU cgroups. Don't let fully
9353 * decayed cfs_rqs linger on the list.
9354 */
9355 if (cfs_rq_is_decayed(cfs_rq))
9356 list_del_leaf_cfs_rq(cfs_rq);
9357
9358 /* Don't need periodic decay once load/util_avg are null */
9359 if (cfs_rq_has_blocked(cfs_rq))
9360 *done = false;
9361 }
9362
9363 return decayed;
9364 }
9365
9366 /*
9367 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9368 * This needs to be done in a top-down fashion because the load of a child
9369 * group is a fraction of its parents load.
9370 */
update_cfs_rq_h_load(struct cfs_rq * cfs_rq)9371 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9372 {
9373 struct rq *rq = rq_of(cfs_rq);
9374 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9375 unsigned long now = jiffies;
9376 unsigned long load;
9377
9378 if (cfs_rq->last_h_load_update == now)
9379 return;
9380
9381 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9382 for_each_sched_entity(se) {
9383 cfs_rq = cfs_rq_of(se);
9384 WRITE_ONCE(cfs_rq->h_load_next, se);
9385 if (cfs_rq->last_h_load_update == now)
9386 break;
9387 }
9388
9389 if (!se) {
9390 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9391 cfs_rq->last_h_load_update = now;
9392 }
9393
9394 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9395 load = cfs_rq->h_load;
9396 load = div64_ul(load * se->avg.load_avg,
9397 cfs_rq_load_avg(cfs_rq) + 1);
9398 cfs_rq = group_cfs_rq(se);
9399 cfs_rq->h_load = load;
9400 cfs_rq->last_h_load_update = now;
9401 }
9402 }
9403
task_h_load(struct task_struct * p)9404 static unsigned long task_h_load(struct task_struct *p)
9405 {
9406 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9407
9408 update_cfs_rq_h_load(cfs_rq);
9409 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9410 cfs_rq_load_avg(cfs_rq) + 1);
9411 }
9412 #else
__update_blocked_fair(struct rq * rq,bool * done)9413 static bool __update_blocked_fair(struct rq *rq, bool *done)
9414 {
9415 struct cfs_rq *cfs_rq = &rq->cfs;
9416 bool decayed;
9417
9418 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9419 if (cfs_rq_has_blocked(cfs_rq))
9420 *done = false;
9421
9422 return decayed;
9423 }
9424
task_h_load(struct task_struct * p)9425 static unsigned long task_h_load(struct task_struct *p)
9426 {
9427 return p->se.avg.load_avg;
9428 }
9429 #endif
9430
update_blocked_averages(int cpu)9431 static void update_blocked_averages(int cpu)
9432 {
9433 bool decayed = false, done = true;
9434 struct rq *rq = cpu_rq(cpu);
9435 struct rq_flags rf;
9436
9437 rq_lock_irqsave(rq, &rf);
9438 update_blocked_load_tick(rq);
9439 update_rq_clock(rq);
9440
9441 decayed |= __update_blocked_others(rq, &done);
9442 decayed |= __update_blocked_fair(rq, &done);
9443
9444 update_blocked_load_status(rq, !done);
9445 if (decayed)
9446 cpufreq_update_util(rq, 0);
9447 rq_unlock_irqrestore(rq, &rf);
9448 }
9449
9450 /********** Helpers for find_busiest_group ************************/
9451
9452 /*
9453 * sg_lb_stats - stats of a sched_group required for load_balancing
9454 */
9455 struct sg_lb_stats {
9456 unsigned long avg_load; /*Avg load across the CPUs of the group */
9457 unsigned long group_load; /* Total load over the CPUs of the group */
9458 unsigned long group_capacity;
9459 unsigned long group_util; /* Total utilization over the CPUs of the group */
9460 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9461 unsigned int sum_nr_running; /* Nr of tasks running in the group */
9462 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9463 unsigned int idle_cpus;
9464 unsigned int group_weight;
9465 enum group_type group_type;
9466 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9467 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9468 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9469 #ifdef CONFIG_NUMA_BALANCING
9470 unsigned int nr_numa_running;
9471 unsigned int nr_preferred_running;
9472 #endif
9473 };
9474
9475 /*
9476 * sd_lb_stats - Structure to store the statistics of a sched_domain
9477 * during load balancing.
9478 */
9479 struct sd_lb_stats {
9480 struct sched_group *busiest; /* Busiest group in this sd */
9481 struct sched_group *local; /* Local group in this sd */
9482 unsigned long total_load; /* Total load of all groups in sd */
9483 unsigned long total_capacity; /* Total capacity of all groups in sd */
9484 unsigned long avg_load; /* Average load across all groups in sd */
9485 unsigned int prefer_sibling; /* tasks should go to sibling first */
9486
9487 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9488 struct sg_lb_stats local_stat; /* Statistics of the local group */
9489 };
9490
init_sd_lb_stats(struct sd_lb_stats * sds)9491 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9492 {
9493 /*
9494 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9495 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9496 * We must however set busiest_stat::group_type and
9497 * busiest_stat::idle_cpus to the worst busiest group because
9498 * update_sd_pick_busiest() reads these before assignment.
9499 */
9500 *sds = (struct sd_lb_stats){
9501 .busiest = NULL,
9502 .local = NULL,
9503 .total_load = 0UL,
9504 .total_capacity = 0UL,
9505 .busiest_stat = {
9506 .idle_cpus = UINT_MAX,
9507 .group_type = group_has_spare,
9508 },
9509 };
9510 }
9511
scale_rt_capacity(int cpu)9512 static unsigned long scale_rt_capacity(int cpu)
9513 {
9514 struct rq *rq = cpu_rq(cpu);
9515 unsigned long max = arch_scale_cpu_capacity(cpu);
9516 unsigned long used, free;
9517 unsigned long irq;
9518
9519 irq = cpu_util_irq(rq);
9520
9521 if (unlikely(irq >= max))
9522 return 1;
9523
9524 /*
9525 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9526 * (running and not running) with weights 0 and 1024 respectively.
9527 * avg_thermal.load_avg tracks thermal pressure and the weighted
9528 * average uses the actual delta max capacity(load).
9529 */
9530 used = READ_ONCE(rq->avg_rt.util_avg);
9531 used += READ_ONCE(rq->avg_dl.util_avg);
9532 used += thermal_load_avg(rq);
9533
9534 if (unlikely(used >= max))
9535 return 1;
9536
9537 free = max - used;
9538
9539 return scale_irq_capacity(free, irq, max);
9540 }
9541
update_cpu_capacity(struct sched_domain * sd,int cpu)9542 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9543 {
9544 unsigned long capacity = scale_rt_capacity(cpu);
9545 struct sched_group *sdg = sd->groups;
9546
9547 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9548
9549 if (!capacity)
9550 capacity = 1;
9551
9552 cpu_rq(cpu)->cpu_capacity = capacity;
9553 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9554
9555 sdg->sgc->capacity = capacity;
9556 sdg->sgc->min_capacity = capacity;
9557 sdg->sgc->max_capacity = capacity;
9558 }
9559
update_group_capacity(struct sched_domain * sd,int cpu)9560 void update_group_capacity(struct sched_domain *sd, int cpu)
9561 {
9562 struct sched_domain *child = sd->child;
9563 struct sched_group *group, *sdg = sd->groups;
9564 unsigned long capacity, min_capacity, max_capacity;
9565 unsigned long interval;
9566
9567 interval = msecs_to_jiffies(sd->balance_interval);
9568 interval = clamp(interval, 1UL, max_load_balance_interval);
9569 sdg->sgc->next_update = jiffies + interval;
9570
9571 if (!child) {
9572 update_cpu_capacity(sd, cpu);
9573 return;
9574 }
9575
9576 capacity = 0;
9577 min_capacity = ULONG_MAX;
9578 max_capacity = 0;
9579
9580 if (child->flags & SD_OVERLAP) {
9581 /*
9582 * SD_OVERLAP domains cannot assume that child groups
9583 * span the current group.
9584 */
9585
9586 for_each_cpu(cpu, sched_group_span(sdg)) {
9587 unsigned long cpu_cap = capacity_of(cpu);
9588
9589 capacity += cpu_cap;
9590 min_capacity = min(cpu_cap, min_capacity);
9591 max_capacity = max(cpu_cap, max_capacity);
9592 }
9593 } else {
9594 /*
9595 * !SD_OVERLAP domains can assume that child groups
9596 * span the current group.
9597 */
9598
9599 group = child->groups;
9600 do {
9601 struct sched_group_capacity *sgc = group->sgc;
9602
9603 capacity += sgc->capacity;
9604 min_capacity = min(sgc->min_capacity, min_capacity);
9605 max_capacity = max(sgc->max_capacity, max_capacity);
9606 group = group->next;
9607 } while (group != child->groups);
9608 }
9609
9610 sdg->sgc->capacity = capacity;
9611 sdg->sgc->min_capacity = min_capacity;
9612 sdg->sgc->max_capacity = max_capacity;
9613 }
9614
9615 /*
9616 * Check whether the capacity of the rq has been noticeably reduced by side
9617 * activity. The imbalance_pct is used for the threshold.
9618 * Return true is the capacity is reduced
9619 */
9620 static inline int
check_cpu_capacity(struct rq * rq,struct sched_domain * sd)9621 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9622 {
9623 return ((rq->cpu_capacity * sd->imbalance_pct) <
9624 (rq->cpu_capacity_orig * 100));
9625 }
9626
9627 /*
9628 * Check whether a rq has a misfit task and if it looks like we can actually
9629 * help that task: we can migrate the task to a CPU of higher capacity, or
9630 * the task's current CPU is heavily pressured.
9631 */
check_misfit_status(struct rq * rq,struct sched_domain * sd)9632 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9633 {
9634 return rq->misfit_task_load &&
9635 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9636 check_cpu_capacity(rq, sd));
9637 }
9638
9639 /*
9640 * Group imbalance indicates (and tries to solve) the problem where balancing
9641 * groups is inadequate due to ->cpus_ptr constraints.
9642 *
9643 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9644 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9645 * Something like:
9646 *
9647 * { 0 1 2 3 } { 4 5 6 7 }
9648 * * * * *
9649 *
9650 * If we were to balance group-wise we'd place two tasks in the first group and
9651 * two tasks in the second group. Clearly this is undesired as it will overload
9652 * cpu 3 and leave one of the CPUs in the second group unused.
9653 *
9654 * The current solution to this issue is detecting the skew in the first group
9655 * by noticing the lower domain failed to reach balance and had difficulty
9656 * moving tasks due to affinity constraints.
9657 *
9658 * When this is so detected; this group becomes a candidate for busiest; see
9659 * update_sd_pick_busiest(). And calculate_imbalance() and
9660 * find_busiest_group() avoid some of the usual balance conditions to allow it
9661 * to create an effective group imbalance.
9662 *
9663 * This is a somewhat tricky proposition since the next run might not find the
9664 * group imbalance and decide the groups need to be balanced again. A most
9665 * subtle and fragile situation.
9666 */
9667
sg_imbalanced(struct sched_group * group)9668 static inline int sg_imbalanced(struct sched_group *group)
9669 {
9670 return group->sgc->imbalance;
9671 }
9672
9673 /*
9674 * group_has_capacity returns true if the group has spare capacity that could
9675 * be used by some tasks.
9676 * We consider that a group has spare capacity if the number of task is
9677 * smaller than the number of CPUs or if the utilization is lower than the
9678 * available capacity for CFS tasks.
9679 * For the latter, we use a threshold to stabilize the state, to take into
9680 * account the variance of the tasks' load and to return true if the available
9681 * capacity in meaningful for the load balancer.
9682 * As an example, an available capacity of 1% can appear but it doesn't make
9683 * any benefit for the load balance.
9684 */
9685 static inline bool
group_has_capacity(unsigned int imbalance_pct,struct sg_lb_stats * sgs)9686 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9687 {
9688 if (sgs->sum_nr_running < sgs->group_weight)
9689 return true;
9690
9691 if ((sgs->group_capacity * imbalance_pct) <
9692 (sgs->group_runnable * 100))
9693 return false;
9694
9695 if ((sgs->group_capacity * 100) >
9696 (sgs->group_util * imbalance_pct))
9697 return true;
9698
9699 return false;
9700 }
9701
9702 /*
9703 * group_is_overloaded returns true if the group has more tasks than it can
9704 * handle.
9705 * group_is_overloaded is not equals to !group_has_capacity because a group
9706 * with the exact right number of tasks, has no more spare capacity but is not
9707 * overloaded so both group_has_capacity and group_is_overloaded return
9708 * false.
9709 */
9710 static inline bool
group_is_overloaded(unsigned int imbalance_pct,struct sg_lb_stats * sgs)9711 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9712 {
9713 if (sgs->sum_nr_running <= sgs->group_weight)
9714 return false;
9715
9716 if ((sgs->group_capacity * 100) <
9717 (sgs->group_util * imbalance_pct))
9718 return true;
9719
9720 if ((sgs->group_capacity * imbalance_pct) <
9721 (sgs->group_runnable * 100))
9722 return true;
9723
9724 return false;
9725 }
9726
9727 static inline enum
group_classify(unsigned int imbalance_pct,struct sched_group * group,struct sg_lb_stats * sgs)9728 group_type group_classify(unsigned int imbalance_pct,
9729 struct sched_group *group,
9730 struct sg_lb_stats *sgs)
9731 {
9732 if (group_is_overloaded(imbalance_pct, sgs))
9733 return group_overloaded;
9734
9735 if (sg_imbalanced(group))
9736 return group_imbalanced;
9737
9738 if (sgs->group_asym_packing)
9739 return group_asym_packing;
9740
9741 if (sgs->group_smt_balance)
9742 return group_smt_balance;
9743
9744 if (sgs->group_misfit_task_load)
9745 return group_misfit_task;
9746
9747 if (!group_has_capacity(imbalance_pct, sgs))
9748 return group_fully_busy;
9749
9750 return group_has_spare;
9751 }
9752
9753 /**
9754 * sched_use_asym_prio - Check whether asym_packing priority must be used
9755 * @sd: The scheduling domain of the load balancing
9756 * @cpu: A CPU
9757 *
9758 * Always use CPU priority when balancing load between SMT siblings. When
9759 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9760 * use CPU priority if the whole core is idle.
9761 *
9762 * Returns: True if the priority of @cpu must be followed. False otherwise.
9763 */
sched_use_asym_prio(struct sched_domain * sd,int cpu)9764 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9765 {
9766 if (!sched_smt_active())
9767 return true;
9768
9769 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9770 }
9771
9772 /**
9773 * sched_asym - Check if the destination CPU can do asym_packing load balance
9774 * @env: The load balancing environment
9775 * @sds: Load-balancing data with statistics of the local group
9776 * @sgs: Load-balancing statistics of the candidate busiest group
9777 * @group: The candidate busiest group
9778 *
9779 * @env::dst_cpu can do asym_packing if it has higher priority than the
9780 * preferred CPU of @group.
9781 *
9782 * SMT is a special case. If we are balancing load between cores, @env::dst_cpu
9783 * can do asym_packing balance only if all its SMT siblings are idle. Also, it
9784 * can only do it if @group is an SMT group and has exactly on busy CPU. Larger
9785 * imbalances in the number of CPUS are dealt with in find_busiest_group().
9786 *
9787 * If we are balancing load within an SMT core, or at PKG domain level, always
9788 * proceed.
9789 *
9790 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9791 * otherwise.
9792 */
9793 static inline bool
sched_asym(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * sgs,struct sched_group * group)9794 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
9795 struct sched_group *group)
9796 {
9797 /* Ensure that the whole local core is idle, if applicable. */
9798 if (!sched_use_asym_prio(env->sd, env->dst_cpu))
9799 return false;
9800
9801 /*
9802 * CPU priorities does not make sense for SMT cores with more than one
9803 * busy sibling.
9804 */
9805 if (group->flags & SD_SHARE_CPUCAPACITY) {
9806 if (sgs->group_weight - sgs->idle_cpus != 1)
9807 return false;
9808 }
9809
9810 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9811 }
9812
9813 /* One group has more than one SMT CPU while the other group does not */
smt_vs_nonsmt_groups(struct sched_group * sg1,struct sched_group * sg2)9814 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9815 struct sched_group *sg2)
9816 {
9817 if (!sg1 || !sg2)
9818 return false;
9819
9820 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9821 (sg2->flags & SD_SHARE_CPUCAPACITY);
9822 }
9823
smt_balance(struct lb_env * env,struct sg_lb_stats * sgs,struct sched_group * group)9824 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9825 struct sched_group *group)
9826 {
9827 if (env->idle == CPU_NOT_IDLE)
9828 return false;
9829
9830 /*
9831 * For SMT source group, it is better to move a task
9832 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9833 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9834 * will not be on.
9835 */
9836 if (group->flags & SD_SHARE_CPUCAPACITY &&
9837 sgs->sum_h_nr_running > 1)
9838 return true;
9839
9840 return false;
9841 }
9842
sibling_imbalance(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * busiest,struct sg_lb_stats * local)9843 static inline long sibling_imbalance(struct lb_env *env,
9844 struct sd_lb_stats *sds,
9845 struct sg_lb_stats *busiest,
9846 struct sg_lb_stats *local)
9847 {
9848 int ncores_busiest, ncores_local;
9849 long imbalance;
9850
9851 if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9852 return 0;
9853
9854 ncores_busiest = sds->busiest->cores;
9855 ncores_local = sds->local->cores;
9856
9857 if (ncores_busiest == ncores_local) {
9858 imbalance = busiest->sum_nr_running;
9859 lsub_positive(&imbalance, local->sum_nr_running);
9860 return imbalance;
9861 }
9862
9863 /* Balance such that nr_running/ncores ratio are same on both groups */
9864 imbalance = ncores_local * busiest->sum_nr_running;
9865 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9866 /* Normalize imbalance and do rounding on normalization */
9867 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9868 imbalance /= ncores_local + ncores_busiest;
9869
9870 /* Take advantage of resource in an empty sched group */
9871 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9872 busiest->sum_nr_running > 1)
9873 imbalance = 2;
9874
9875 return imbalance;
9876 }
9877
9878 static inline bool
sched_reduced_capacity(struct rq * rq,struct sched_domain * sd)9879 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9880 {
9881 /*
9882 * When there is more than 1 task, the group_overloaded case already
9883 * takes care of cpu with reduced capacity
9884 */
9885 if (rq->cfs.h_nr_running != 1)
9886 return false;
9887
9888 return check_cpu_capacity(rq, sd);
9889 }
9890
9891 /**
9892 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9893 * @env: The load balancing environment.
9894 * @sds: Load-balancing data with statistics of the local group.
9895 * @group: sched_group whose statistics are to be updated.
9896 * @sgs: variable to hold the statistics for this group.
9897 * @sg_status: Holds flag indicating the status of the sched_group
9898 */
update_sg_lb_stats(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * group,struct sg_lb_stats * sgs,int * sg_status)9899 static inline void update_sg_lb_stats(struct lb_env *env,
9900 struct sd_lb_stats *sds,
9901 struct sched_group *group,
9902 struct sg_lb_stats *sgs,
9903 int *sg_status)
9904 {
9905 int i, nr_running, local_group;
9906
9907 memset(sgs, 0, sizeof(*sgs));
9908
9909 local_group = group == sds->local;
9910
9911 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9912 struct rq *rq = cpu_rq(i);
9913 unsigned long load = cpu_load(rq);
9914
9915 sgs->group_load += load;
9916 sgs->group_util += cpu_util_cfs(i);
9917 sgs->group_runnable += cpu_runnable(rq);
9918 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9919
9920 nr_running = rq->nr_running;
9921 sgs->sum_nr_running += nr_running;
9922
9923 if (nr_running > 1)
9924 *sg_status |= SG_OVERLOAD;
9925
9926 if (cpu_overutilized(i))
9927 *sg_status |= SG_OVERUTILIZED;
9928
9929 #ifdef CONFIG_NUMA_BALANCING
9930 sgs->nr_numa_running += rq->nr_numa_running;
9931 sgs->nr_preferred_running += rq->nr_preferred_running;
9932 #endif
9933 /*
9934 * No need to call idle_cpu() if nr_running is not 0
9935 */
9936 if (!nr_running && idle_cpu(i)) {
9937 sgs->idle_cpus++;
9938 /* Idle cpu can't have misfit task */
9939 continue;
9940 }
9941
9942 if (local_group)
9943 continue;
9944
9945 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9946 /* Check for a misfit task on the cpu */
9947 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9948 sgs->group_misfit_task_load = rq->misfit_task_load;
9949 *sg_status |= SG_OVERLOAD;
9950 }
9951 } else if ((env->idle != CPU_NOT_IDLE) &&
9952 sched_reduced_capacity(rq, env->sd)) {
9953 /* Check for a task running on a CPU with reduced capacity */
9954 if (sgs->group_misfit_task_load < load)
9955 sgs->group_misfit_task_load = load;
9956 }
9957 }
9958
9959 sgs->group_capacity = group->sgc->capacity;
9960
9961 sgs->group_weight = group->group_weight;
9962
9963 /* Check if dst CPU is idle and preferred to this group */
9964 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9965 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9966 sched_asym(env, sds, sgs, group)) {
9967 sgs->group_asym_packing = 1;
9968 }
9969
9970 /* Check for loaded SMT group to be balanced to dst CPU */
9971 if (!local_group && smt_balance(env, sgs, group))
9972 sgs->group_smt_balance = 1;
9973
9974 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9975
9976 /* Computing avg_load makes sense only when group is overloaded */
9977 if (sgs->group_type == group_overloaded)
9978 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9979 sgs->group_capacity;
9980 }
9981
9982 /**
9983 * update_sd_pick_busiest - return 1 on busiest group
9984 * @env: The load balancing environment.
9985 * @sds: sched_domain statistics
9986 * @sg: sched_group candidate to be checked for being the busiest
9987 * @sgs: sched_group statistics
9988 *
9989 * Determine if @sg is a busier group than the previously selected
9990 * busiest group.
9991 *
9992 * Return: %true if @sg is a busier group than the previously selected
9993 * busiest group. %false otherwise.
9994 */
update_sd_pick_busiest(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * sg,struct sg_lb_stats * sgs)9995 static bool update_sd_pick_busiest(struct lb_env *env,
9996 struct sd_lb_stats *sds,
9997 struct sched_group *sg,
9998 struct sg_lb_stats *sgs)
9999 {
10000 struct sg_lb_stats *busiest = &sds->busiest_stat;
10001
10002 /* Make sure that there is at least one task to pull */
10003 if (!sgs->sum_h_nr_running)
10004 return false;
10005
10006 /*
10007 * Don't try to pull misfit tasks we can't help.
10008 * We can use max_capacity here as reduction in capacity on some
10009 * CPUs in the group should either be possible to resolve
10010 * internally or be covered by avg_load imbalance (eventually).
10011 */
10012 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10013 (sgs->group_type == group_misfit_task) &&
10014 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
10015 sds->local_stat.group_type != group_has_spare))
10016 return false;
10017
10018 if (sgs->group_type > busiest->group_type)
10019 return true;
10020
10021 if (sgs->group_type < busiest->group_type)
10022 return false;
10023
10024 /*
10025 * The candidate and the current busiest group are the same type of
10026 * group. Let check which one is the busiest according to the type.
10027 */
10028
10029 switch (sgs->group_type) {
10030 case group_overloaded:
10031 /* Select the overloaded group with highest avg_load. */
10032 if (sgs->avg_load <= busiest->avg_load)
10033 return false;
10034 break;
10035
10036 case group_imbalanced:
10037 /*
10038 * Select the 1st imbalanced group as we don't have any way to
10039 * choose one more than another.
10040 */
10041 return false;
10042
10043 case group_asym_packing:
10044 /* Prefer to move from lowest priority CPU's work */
10045 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
10046 return false;
10047 break;
10048
10049 case group_misfit_task:
10050 /*
10051 * If we have more than one misfit sg go with the biggest
10052 * misfit.
10053 */
10054 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
10055 return false;
10056 break;
10057
10058 case group_smt_balance:
10059 /*
10060 * Check if we have spare CPUs on either SMT group to
10061 * choose has spare or fully busy handling.
10062 */
10063 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
10064 goto has_spare;
10065
10066 fallthrough;
10067
10068 case group_fully_busy:
10069 /*
10070 * Select the fully busy group with highest avg_load. In
10071 * theory, there is no need to pull task from such kind of
10072 * group because tasks have all compute capacity that they need
10073 * but we can still improve the overall throughput by reducing
10074 * contention when accessing shared HW resources.
10075 *
10076 * XXX for now avg_load is not computed and always 0 so we
10077 * select the 1st one, except if @sg is composed of SMT
10078 * siblings.
10079 */
10080
10081 if (sgs->avg_load < busiest->avg_load)
10082 return false;
10083
10084 if (sgs->avg_load == busiest->avg_load) {
10085 /*
10086 * SMT sched groups need more help than non-SMT groups.
10087 * If @sg happens to also be SMT, either choice is good.
10088 */
10089 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
10090 return false;
10091 }
10092
10093 break;
10094
10095 case group_has_spare:
10096 /*
10097 * Do not pick sg with SMT CPUs over sg with pure CPUs,
10098 * as we do not want to pull task off SMT core with one task
10099 * and make the core idle.
10100 */
10101 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
10102 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
10103 return false;
10104 else
10105 return true;
10106 }
10107 has_spare:
10108
10109 /*
10110 * Select not overloaded group with lowest number of idle cpus
10111 * and highest number of running tasks. We could also compare
10112 * the spare capacity which is more stable but it can end up
10113 * that the group has less spare capacity but finally more idle
10114 * CPUs which means less opportunity to pull tasks.
10115 */
10116 if (sgs->idle_cpus > busiest->idle_cpus)
10117 return false;
10118 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
10119 (sgs->sum_nr_running <= busiest->sum_nr_running))
10120 return false;
10121
10122 break;
10123 }
10124
10125 /*
10126 * Candidate sg has no more than one task per CPU and has higher
10127 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
10128 * throughput. Maximize throughput, power/energy consequences are not
10129 * considered.
10130 */
10131 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10132 (sgs->group_type <= group_fully_busy) &&
10133 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
10134 return false;
10135
10136 return true;
10137 }
10138
10139 #ifdef CONFIG_NUMA_BALANCING
fbq_classify_group(struct sg_lb_stats * sgs)10140 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10141 {
10142 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10143 return regular;
10144 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10145 return remote;
10146 return all;
10147 }
10148
fbq_classify_rq(struct rq * rq)10149 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10150 {
10151 if (rq->nr_running > rq->nr_numa_running)
10152 return regular;
10153 if (rq->nr_running > rq->nr_preferred_running)
10154 return remote;
10155 return all;
10156 }
10157 #else
fbq_classify_group(struct sg_lb_stats * sgs)10158 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10159 {
10160 return all;
10161 }
10162
fbq_classify_rq(struct rq * rq)10163 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10164 {
10165 return regular;
10166 }
10167 #endif /* CONFIG_NUMA_BALANCING */
10168
10169
10170 struct sg_lb_stats;
10171
10172 /*
10173 * task_running_on_cpu - return 1 if @p is running on @cpu.
10174 */
10175
task_running_on_cpu(int cpu,struct task_struct * p)10176 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10177 {
10178 /* Task has no contribution or is new */
10179 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10180 return 0;
10181
10182 if (task_on_rq_queued(p))
10183 return 1;
10184
10185 return 0;
10186 }
10187
10188 /**
10189 * idle_cpu_without - would a given CPU be idle without p ?
10190 * @cpu: the processor on which idleness is tested.
10191 * @p: task which should be ignored.
10192 *
10193 * Return: 1 if the CPU would be idle. 0 otherwise.
10194 */
idle_cpu_without(int cpu,struct task_struct * p)10195 static int idle_cpu_without(int cpu, struct task_struct *p)
10196 {
10197 struct rq *rq = cpu_rq(cpu);
10198
10199 if (rq->curr != rq->idle && rq->curr != p)
10200 return 0;
10201
10202 /*
10203 * rq->nr_running can't be used but an updated version without the
10204 * impact of p on cpu must be used instead. The updated nr_running
10205 * be computed and tested before calling idle_cpu_without().
10206 */
10207
10208 #ifdef CONFIG_SMP
10209 if (rq->ttwu_pending)
10210 return 0;
10211 #endif
10212
10213 return 1;
10214 }
10215
10216 /*
10217 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10218 * @sd: The sched_domain level to look for idlest group.
10219 * @group: sched_group whose statistics are to be updated.
10220 * @sgs: variable to hold the statistics for this group.
10221 * @p: The task for which we look for the idlest group/CPU.
10222 */
update_sg_wakeup_stats(struct sched_domain * sd,struct sched_group * group,struct sg_lb_stats * sgs,struct task_struct * p)10223 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10224 struct sched_group *group,
10225 struct sg_lb_stats *sgs,
10226 struct task_struct *p)
10227 {
10228 int i, nr_running;
10229
10230 memset(sgs, 0, sizeof(*sgs));
10231
10232 /* Assume that task can't fit any CPU of the group */
10233 if (sd->flags & SD_ASYM_CPUCAPACITY)
10234 sgs->group_misfit_task_load = 1;
10235
10236 for_each_cpu(i, sched_group_span(group)) {
10237 struct rq *rq = cpu_rq(i);
10238 unsigned int local;
10239
10240 sgs->group_load += cpu_load_without(rq, p);
10241 sgs->group_util += cpu_util_without(i, p);
10242 sgs->group_runnable += cpu_runnable_without(rq, p);
10243 local = task_running_on_cpu(i, p);
10244 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10245
10246 nr_running = rq->nr_running - local;
10247 sgs->sum_nr_running += nr_running;
10248
10249 /*
10250 * No need to call idle_cpu_without() if nr_running is not 0
10251 */
10252 if (!nr_running && idle_cpu_without(i, p))
10253 sgs->idle_cpus++;
10254
10255 /* Check if task fits in the CPU */
10256 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10257 sgs->group_misfit_task_load &&
10258 task_fits_cpu(p, i))
10259 sgs->group_misfit_task_load = 0;
10260
10261 }
10262
10263 sgs->group_capacity = group->sgc->capacity;
10264
10265 sgs->group_weight = group->group_weight;
10266
10267 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10268
10269 /*
10270 * Computing avg_load makes sense only when group is fully busy or
10271 * overloaded
10272 */
10273 if (sgs->group_type == group_fully_busy ||
10274 sgs->group_type == group_overloaded)
10275 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10276 sgs->group_capacity;
10277 }
10278
update_pick_idlest(struct sched_group * idlest,struct sg_lb_stats * idlest_sgs,struct sched_group * group,struct sg_lb_stats * sgs)10279 static bool update_pick_idlest(struct sched_group *idlest,
10280 struct sg_lb_stats *idlest_sgs,
10281 struct sched_group *group,
10282 struct sg_lb_stats *sgs)
10283 {
10284 if (sgs->group_type < idlest_sgs->group_type)
10285 return true;
10286
10287 if (sgs->group_type > idlest_sgs->group_type)
10288 return false;
10289
10290 /*
10291 * The candidate and the current idlest group are the same type of
10292 * group. Let check which one is the idlest according to the type.
10293 */
10294
10295 switch (sgs->group_type) {
10296 case group_overloaded:
10297 case group_fully_busy:
10298 /* Select the group with lowest avg_load. */
10299 if (idlest_sgs->avg_load <= sgs->avg_load)
10300 return false;
10301 break;
10302
10303 case group_imbalanced:
10304 case group_asym_packing:
10305 case group_smt_balance:
10306 /* Those types are not used in the slow wakeup path */
10307 return false;
10308
10309 case group_misfit_task:
10310 /* Select group with the highest max capacity */
10311 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10312 return false;
10313 break;
10314
10315 case group_has_spare:
10316 /* Select group with most idle CPUs */
10317 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10318 return false;
10319
10320 /* Select group with lowest group_util */
10321 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10322 idlest_sgs->group_util <= sgs->group_util)
10323 return false;
10324
10325 break;
10326 }
10327
10328 return true;
10329 }
10330
10331 /*
10332 * find_idlest_group() finds and returns the least busy CPU group within the
10333 * domain.
10334 *
10335 * Assumes p is allowed on at least one CPU in sd.
10336 */
10337 static struct sched_group *
find_idlest_group(struct sched_domain * sd,struct task_struct * p,int this_cpu)10338 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10339 {
10340 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10341 struct sg_lb_stats local_sgs, tmp_sgs;
10342 struct sg_lb_stats *sgs;
10343 unsigned long imbalance;
10344 struct sg_lb_stats idlest_sgs = {
10345 .avg_load = UINT_MAX,
10346 .group_type = group_overloaded,
10347 };
10348
10349 do {
10350 int local_group;
10351
10352 /* Skip over this group if it has no CPUs allowed */
10353 if (!cpumask_intersects(sched_group_span(group),
10354 p->cpus_ptr))
10355 continue;
10356
10357 /* Skip over this group if no cookie matched */
10358 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10359 continue;
10360
10361 local_group = cpumask_test_cpu(this_cpu,
10362 sched_group_span(group));
10363
10364 if (local_group) {
10365 sgs = &local_sgs;
10366 local = group;
10367 } else {
10368 sgs = &tmp_sgs;
10369 }
10370
10371 update_sg_wakeup_stats(sd, group, sgs, p);
10372
10373 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10374 idlest = group;
10375 idlest_sgs = *sgs;
10376 }
10377
10378 } while (group = group->next, group != sd->groups);
10379
10380
10381 /* There is no idlest group to push tasks to */
10382 if (!idlest)
10383 return NULL;
10384
10385 /* The local group has been skipped because of CPU affinity */
10386 if (!local)
10387 return idlest;
10388
10389 /*
10390 * If the local group is idler than the selected idlest group
10391 * don't try and push the task.
10392 */
10393 if (local_sgs.group_type < idlest_sgs.group_type)
10394 return NULL;
10395
10396 /*
10397 * If the local group is busier than the selected idlest group
10398 * try and push the task.
10399 */
10400 if (local_sgs.group_type > idlest_sgs.group_type)
10401 return idlest;
10402
10403 switch (local_sgs.group_type) {
10404 case group_overloaded:
10405 case group_fully_busy:
10406
10407 /* Calculate allowed imbalance based on load */
10408 imbalance = scale_load_down(NICE_0_LOAD) *
10409 (sd->imbalance_pct-100) / 100;
10410
10411 /*
10412 * When comparing groups across NUMA domains, it's possible for
10413 * the local domain to be very lightly loaded relative to the
10414 * remote domains but "imbalance" skews the comparison making
10415 * remote CPUs look much more favourable. When considering
10416 * cross-domain, add imbalance to the load on the remote node
10417 * and consider staying local.
10418 */
10419
10420 if ((sd->flags & SD_NUMA) &&
10421 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10422 return NULL;
10423
10424 /*
10425 * If the local group is less loaded than the selected
10426 * idlest group don't try and push any tasks.
10427 */
10428 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10429 return NULL;
10430
10431 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10432 return NULL;
10433 break;
10434
10435 case group_imbalanced:
10436 case group_asym_packing:
10437 case group_smt_balance:
10438 /* Those type are not used in the slow wakeup path */
10439 return NULL;
10440
10441 case group_misfit_task:
10442 /* Select group with the highest max capacity */
10443 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10444 return NULL;
10445 break;
10446
10447 case group_has_spare:
10448 #ifdef CONFIG_NUMA
10449 if (sd->flags & SD_NUMA) {
10450 int imb_numa_nr = sd->imb_numa_nr;
10451 #ifdef CONFIG_NUMA_BALANCING
10452 int idlest_cpu;
10453 /*
10454 * If there is spare capacity at NUMA, try to select
10455 * the preferred node
10456 */
10457 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10458 return NULL;
10459
10460 idlest_cpu = cpumask_first(sched_group_span(idlest));
10461 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10462 return idlest;
10463 #endif /* CONFIG_NUMA_BALANCING */
10464 /*
10465 * Otherwise, keep the task close to the wakeup source
10466 * and improve locality if the number of running tasks
10467 * would remain below threshold where an imbalance is
10468 * allowed while accounting for the possibility the
10469 * task is pinned to a subset of CPUs. If there is a
10470 * real need of migration, periodic load balance will
10471 * take care of it.
10472 */
10473 if (p->nr_cpus_allowed != NR_CPUS) {
10474 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10475
10476 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10477 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10478 }
10479
10480 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10481 if (!adjust_numa_imbalance(imbalance,
10482 local_sgs.sum_nr_running + 1,
10483 imb_numa_nr)) {
10484 return NULL;
10485 }
10486 }
10487 #endif /* CONFIG_NUMA */
10488
10489 /*
10490 * Select group with highest number of idle CPUs. We could also
10491 * compare the utilization which is more stable but it can end
10492 * up that the group has less spare capacity but finally more
10493 * idle CPUs which means more opportunity to run task.
10494 */
10495 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10496 return NULL;
10497 break;
10498 }
10499
10500 return idlest;
10501 }
10502
update_idle_cpu_scan(struct lb_env * env,unsigned long sum_util)10503 static void update_idle_cpu_scan(struct lb_env *env,
10504 unsigned long sum_util)
10505 {
10506 struct sched_domain_shared *sd_share;
10507 int llc_weight, pct;
10508 u64 x, y, tmp;
10509 /*
10510 * Update the number of CPUs to scan in LLC domain, which could
10511 * be used as a hint in select_idle_cpu(). The update of sd_share
10512 * could be expensive because it is within a shared cache line.
10513 * So the write of this hint only occurs during periodic load
10514 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10515 * can fire way more frequently than the former.
10516 */
10517 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10518 return;
10519
10520 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10521 if (env->sd->span_weight != llc_weight)
10522 return;
10523
10524 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10525 if (!sd_share)
10526 return;
10527
10528 /*
10529 * The number of CPUs to search drops as sum_util increases, when
10530 * sum_util hits 85% or above, the scan stops.
10531 * The reason to choose 85% as the threshold is because this is the
10532 * imbalance_pct(117) when a LLC sched group is overloaded.
10533 *
10534 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10535 * and y'= y / SCHED_CAPACITY_SCALE
10536 *
10537 * x is the ratio of sum_util compared to the CPU capacity:
10538 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10539 * y' is the ratio of CPUs to be scanned in the LLC domain,
10540 * and the number of CPUs to scan is calculated by:
10541 *
10542 * nr_scan = llc_weight * y' [2]
10543 *
10544 * When x hits the threshold of overloaded, AKA, when
10545 * x = 100 / pct, y drops to 0. According to [1],
10546 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10547 *
10548 * Scale x by SCHED_CAPACITY_SCALE:
10549 * x' = sum_util / llc_weight; [3]
10550 *
10551 * and finally [1] becomes:
10552 * y = SCHED_CAPACITY_SCALE -
10553 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10554 *
10555 */
10556 /* equation [3] */
10557 x = sum_util;
10558 do_div(x, llc_weight);
10559
10560 /* equation [4] */
10561 pct = env->sd->imbalance_pct;
10562 tmp = x * x * pct * pct;
10563 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10564 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10565 y = SCHED_CAPACITY_SCALE - tmp;
10566
10567 /* equation [2] */
10568 y *= llc_weight;
10569 do_div(y, SCHED_CAPACITY_SCALE);
10570 if ((int)y != sd_share->nr_idle_scan)
10571 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10572 }
10573
10574 /**
10575 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10576 * @env: The load balancing environment.
10577 * @sds: variable to hold the statistics for this sched_domain.
10578 */
10579
update_sd_lb_stats(struct lb_env * env,struct sd_lb_stats * sds)10580 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10581 {
10582 struct sched_group *sg = env->sd->groups;
10583 struct sg_lb_stats *local = &sds->local_stat;
10584 struct sg_lb_stats tmp_sgs;
10585 unsigned long sum_util = 0;
10586 int sg_status = 0;
10587
10588 do {
10589 struct sg_lb_stats *sgs = &tmp_sgs;
10590 int local_group;
10591
10592 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10593 if (local_group) {
10594 sds->local = sg;
10595 sgs = local;
10596
10597 if (env->idle != CPU_NEWLY_IDLE ||
10598 time_after_eq(jiffies, sg->sgc->next_update))
10599 update_group_capacity(env->sd, env->dst_cpu);
10600 }
10601
10602 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10603
10604 if (local_group)
10605 goto next_group;
10606
10607
10608 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10609 sds->busiest = sg;
10610 sds->busiest_stat = *sgs;
10611 }
10612
10613 next_group:
10614 /* Now, start updating sd_lb_stats */
10615 sds->total_load += sgs->group_load;
10616 sds->total_capacity += sgs->group_capacity;
10617
10618 sum_util += sgs->group_util;
10619 sg = sg->next;
10620 } while (sg != env->sd->groups);
10621
10622 /*
10623 * Indicate that the child domain of the busiest group prefers tasks
10624 * go to a child's sibling domains first. NB the flags of a sched group
10625 * are those of the child domain.
10626 */
10627 if (sds->busiest)
10628 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10629
10630
10631 if (env->sd->flags & SD_NUMA)
10632 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10633
10634 if (!env->sd->parent) {
10635 /* update overload indicator if we are at root domain */
10636 WRITE_ONCE(env->dst_rq->rd->overload, sg_status & SG_OVERLOAD);
10637
10638 /* Update over-utilization (tipping point, U >= 0) indicator */
10639 set_rd_overutilized_status(env->dst_rq->rd,
10640 sg_status & SG_OVERUTILIZED);
10641 } else if (sg_status & SG_OVERUTILIZED) {
10642 set_rd_overutilized_status(env->dst_rq->rd, SG_OVERUTILIZED);
10643 }
10644
10645 update_idle_cpu_scan(env, sum_util);
10646 }
10647
10648 /**
10649 * calculate_imbalance - Calculate the amount of imbalance present within the
10650 * groups of a given sched_domain during load balance.
10651 * @env: load balance environment
10652 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10653 */
calculate_imbalance(struct lb_env * env,struct sd_lb_stats * sds)10654 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10655 {
10656 struct sg_lb_stats *local, *busiest;
10657
10658 local = &sds->local_stat;
10659 busiest = &sds->busiest_stat;
10660
10661 if (busiest->group_type == group_misfit_task) {
10662 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10663 /* Set imbalance to allow misfit tasks to be balanced. */
10664 env->migration_type = migrate_misfit;
10665 env->imbalance = 1;
10666 } else {
10667 /*
10668 * Set load imbalance to allow moving task from cpu
10669 * with reduced capacity.
10670 */
10671 env->migration_type = migrate_load;
10672 env->imbalance = busiest->group_misfit_task_load;
10673 }
10674 return;
10675 }
10676
10677 if (busiest->group_type == group_asym_packing) {
10678 /*
10679 * In case of asym capacity, we will try to migrate all load to
10680 * the preferred CPU.
10681 */
10682 env->migration_type = migrate_task;
10683 env->imbalance = busiest->sum_h_nr_running;
10684 return;
10685 }
10686
10687 if (busiest->group_type == group_smt_balance) {
10688 /* Reduce number of tasks sharing CPU capacity */
10689 env->migration_type = migrate_task;
10690 env->imbalance = 1;
10691 return;
10692 }
10693
10694 if (busiest->group_type == group_imbalanced) {
10695 /*
10696 * In the group_imb case we cannot rely on group-wide averages
10697 * to ensure CPU-load equilibrium, try to move any task to fix
10698 * the imbalance. The next load balance will take care of
10699 * balancing back the system.
10700 */
10701 env->migration_type = migrate_task;
10702 env->imbalance = 1;
10703 return;
10704 }
10705
10706 /*
10707 * Try to use spare capacity of local group without overloading it or
10708 * emptying busiest.
10709 */
10710 if (local->group_type == group_has_spare) {
10711 if ((busiest->group_type > group_fully_busy) &&
10712 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10713 /*
10714 * If busiest is overloaded, try to fill spare
10715 * capacity. This might end up creating spare capacity
10716 * in busiest or busiest still being overloaded but
10717 * there is no simple way to directly compute the
10718 * amount of load to migrate in order to balance the
10719 * system.
10720 */
10721 env->migration_type = migrate_util;
10722 env->imbalance = max(local->group_capacity, local->group_util) -
10723 local->group_util;
10724
10725 /*
10726 * In some cases, the group's utilization is max or even
10727 * higher than capacity because of migrations but the
10728 * local CPU is (newly) idle. There is at least one
10729 * waiting task in this overloaded busiest group. Let's
10730 * try to pull it.
10731 */
10732 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10733 env->migration_type = migrate_task;
10734 env->imbalance = 1;
10735 }
10736
10737 return;
10738 }
10739
10740 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10741 /*
10742 * When prefer sibling, evenly spread running tasks on
10743 * groups.
10744 */
10745 env->migration_type = migrate_task;
10746 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10747 } else {
10748
10749 /*
10750 * If there is no overload, we just want to even the number of
10751 * idle cpus.
10752 */
10753 env->migration_type = migrate_task;
10754 env->imbalance = max_t(long, 0,
10755 (local->idle_cpus - busiest->idle_cpus));
10756 }
10757
10758 #ifdef CONFIG_NUMA
10759 /* Consider allowing a small imbalance between NUMA groups */
10760 if (env->sd->flags & SD_NUMA) {
10761 env->imbalance = adjust_numa_imbalance(env->imbalance,
10762 local->sum_nr_running + 1,
10763 env->sd->imb_numa_nr);
10764 }
10765 #endif
10766
10767 /* Number of tasks to move to restore balance */
10768 env->imbalance >>= 1;
10769
10770 return;
10771 }
10772
10773 /*
10774 * Local is fully busy but has to take more load to relieve the
10775 * busiest group
10776 */
10777 if (local->group_type < group_overloaded) {
10778 /*
10779 * Local will become overloaded so the avg_load metrics are
10780 * finally needed.
10781 */
10782
10783 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10784 local->group_capacity;
10785
10786 /*
10787 * If the local group is more loaded than the selected
10788 * busiest group don't try to pull any tasks.
10789 */
10790 if (local->avg_load >= busiest->avg_load) {
10791 env->imbalance = 0;
10792 return;
10793 }
10794
10795 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10796 sds->total_capacity;
10797
10798 /*
10799 * If the local group is more loaded than the average system
10800 * load, don't try to pull any tasks.
10801 */
10802 if (local->avg_load >= sds->avg_load) {
10803 env->imbalance = 0;
10804 return;
10805 }
10806
10807 }
10808
10809 /*
10810 * Both group are or will become overloaded and we're trying to get all
10811 * the CPUs to the average_load, so we don't want to push ourselves
10812 * above the average load, nor do we wish to reduce the max loaded CPU
10813 * below the average load. At the same time, we also don't want to
10814 * reduce the group load below the group capacity. Thus we look for
10815 * the minimum possible imbalance.
10816 */
10817 env->migration_type = migrate_load;
10818 env->imbalance = min(
10819 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10820 (sds->avg_load - local->avg_load) * local->group_capacity
10821 ) / SCHED_CAPACITY_SCALE;
10822 }
10823
10824 /******* find_busiest_group() helpers end here *********************/
10825
10826 /*
10827 * Decision matrix according to the local and busiest group type:
10828 *
10829 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10830 * has_spare nr_idle balanced N/A N/A balanced balanced
10831 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10832 * misfit_task force N/A N/A N/A N/A N/A
10833 * asym_packing force force N/A N/A force force
10834 * imbalanced force force N/A N/A force force
10835 * overloaded force force N/A N/A force avg_load
10836 *
10837 * N/A : Not Applicable because already filtered while updating
10838 * statistics.
10839 * balanced : The system is balanced for these 2 groups.
10840 * force : Calculate the imbalance as load migration is probably needed.
10841 * avg_load : Only if imbalance is significant enough.
10842 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10843 * different in groups.
10844 */
10845
10846 /**
10847 * find_busiest_group - Returns the busiest group within the sched_domain
10848 * if there is an imbalance.
10849 * @env: The load balancing environment.
10850 *
10851 * Also calculates the amount of runnable load which should be moved
10852 * to restore balance.
10853 *
10854 * Return: - The busiest group if imbalance exists.
10855 */
find_busiest_group(struct lb_env * env)10856 static struct sched_group *find_busiest_group(struct lb_env *env)
10857 {
10858 struct sg_lb_stats *local, *busiest;
10859 struct sd_lb_stats sds;
10860
10861 init_sd_lb_stats(&sds);
10862
10863 /*
10864 * Compute the various statistics relevant for load balancing at
10865 * this level.
10866 */
10867 update_sd_lb_stats(env, &sds);
10868
10869 /* There is no busy sibling group to pull tasks from */
10870 if (!sds.busiest)
10871 goto out_balanced;
10872
10873 busiest = &sds.busiest_stat;
10874
10875 /* Misfit tasks should be dealt with regardless of the avg load */
10876 if (busiest->group_type == group_misfit_task)
10877 goto force_balance;
10878
10879 if (sched_energy_enabled()) {
10880 struct root_domain *rd = env->dst_rq->rd;
10881
10882 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10883 goto out_balanced;
10884 }
10885
10886 /* ASYM feature bypasses nice load balance check */
10887 if (busiest->group_type == group_asym_packing)
10888 goto force_balance;
10889
10890 /*
10891 * If the busiest group is imbalanced the below checks don't
10892 * work because they assume all things are equal, which typically
10893 * isn't true due to cpus_ptr constraints and the like.
10894 */
10895 if (busiest->group_type == group_imbalanced)
10896 goto force_balance;
10897
10898 local = &sds.local_stat;
10899 /*
10900 * If the local group is busier than the selected busiest group
10901 * don't try and pull any tasks.
10902 */
10903 if (local->group_type > busiest->group_type)
10904 goto out_balanced;
10905
10906 /*
10907 * When groups are overloaded, use the avg_load to ensure fairness
10908 * between tasks.
10909 */
10910 if (local->group_type == group_overloaded) {
10911 /*
10912 * If the local group is more loaded than the selected
10913 * busiest group don't try to pull any tasks.
10914 */
10915 if (local->avg_load >= busiest->avg_load)
10916 goto out_balanced;
10917
10918 /* XXX broken for overlapping NUMA groups */
10919 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10920 sds.total_capacity;
10921
10922 /*
10923 * Don't pull any tasks if this group is already above the
10924 * domain average load.
10925 */
10926 if (local->avg_load >= sds.avg_load)
10927 goto out_balanced;
10928
10929 /*
10930 * If the busiest group is more loaded, use imbalance_pct to be
10931 * conservative.
10932 */
10933 if (100 * busiest->avg_load <=
10934 env->sd->imbalance_pct * local->avg_load)
10935 goto out_balanced;
10936 }
10937
10938 /*
10939 * Try to move all excess tasks to a sibling domain of the busiest
10940 * group's child domain.
10941 */
10942 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10943 sibling_imbalance(env, &sds, busiest, local) > 1)
10944 goto force_balance;
10945
10946 if (busiest->group_type != group_overloaded) {
10947 if (env->idle == CPU_NOT_IDLE) {
10948 /*
10949 * If the busiest group is not overloaded (and as a
10950 * result the local one too) but this CPU is already
10951 * busy, let another idle CPU try to pull task.
10952 */
10953 goto out_balanced;
10954 }
10955
10956 if (busiest->group_type == group_smt_balance &&
10957 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10958 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10959 goto force_balance;
10960 }
10961
10962 if (busiest->group_weight > 1 &&
10963 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10964 /*
10965 * If the busiest group is not overloaded
10966 * and there is no imbalance between this and busiest
10967 * group wrt idle CPUs, it is balanced. The imbalance
10968 * becomes significant if the diff is greater than 1
10969 * otherwise we might end up to just move the imbalance
10970 * on another group. Of course this applies only if
10971 * there is more than 1 CPU per group.
10972 */
10973 goto out_balanced;
10974 }
10975
10976 if (busiest->sum_h_nr_running == 1) {
10977 /*
10978 * busiest doesn't have any tasks waiting to run
10979 */
10980 goto out_balanced;
10981 }
10982 }
10983
10984 force_balance:
10985 /* Looks like there is an imbalance. Compute it */
10986 calculate_imbalance(env, &sds);
10987 return env->imbalance ? sds.busiest : NULL;
10988
10989 out_balanced:
10990 env->imbalance = 0;
10991 return NULL;
10992 }
10993
10994 /*
10995 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10996 */
find_busiest_queue(struct lb_env * env,struct sched_group * group)10997 static struct rq *find_busiest_queue(struct lb_env *env,
10998 struct sched_group *group)
10999 {
11000 struct rq *busiest = NULL, *rq;
11001 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
11002 unsigned int busiest_nr = 0;
11003 int i;
11004
11005 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
11006 unsigned long capacity, load, util;
11007 unsigned int nr_running;
11008 enum fbq_type rt;
11009
11010 rq = cpu_rq(i);
11011 rt = fbq_classify_rq(rq);
11012
11013 /*
11014 * We classify groups/runqueues into three groups:
11015 * - regular: there are !numa tasks
11016 * - remote: there are numa tasks that run on the 'wrong' node
11017 * - all: there is no distinction
11018 *
11019 * In order to avoid migrating ideally placed numa tasks,
11020 * ignore those when there's better options.
11021 *
11022 * If we ignore the actual busiest queue to migrate another
11023 * task, the next balance pass can still reduce the busiest
11024 * queue by moving tasks around inside the node.
11025 *
11026 * If we cannot move enough load due to this classification
11027 * the next pass will adjust the group classification and
11028 * allow migration of more tasks.
11029 *
11030 * Both cases only affect the total convergence complexity.
11031 */
11032 if (rt > env->fbq_type)
11033 continue;
11034
11035 nr_running = rq->cfs.h_nr_running;
11036 if (!nr_running)
11037 continue;
11038
11039 capacity = capacity_of(i);
11040
11041 /*
11042 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
11043 * eventually lead to active_balancing high->low capacity.
11044 * Higher per-CPU capacity is considered better than balancing
11045 * average load.
11046 */
11047 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
11048 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
11049 nr_running == 1)
11050 continue;
11051
11052 /*
11053 * Make sure we only pull tasks from a CPU of lower priority
11054 * when balancing between SMT siblings.
11055 *
11056 * If balancing between cores, let lower priority CPUs help
11057 * SMT cores with more than one busy sibling.
11058 */
11059 if ((env->sd->flags & SD_ASYM_PACKING) &&
11060 sched_use_asym_prio(env->sd, i) &&
11061 sched_asym_prefer(i, env->dst_cpu) &&
11062 nr_running == 1)
11063 continue;
11064
11065 switch (env->migration_type) {
11066 case migrate_load:
11067 /*
11068 * When comparing with load imbalance, use cpu_load()
11069 * which is not scaled with the CPU capacity.
11070 */
11071 load = cpu_load(rq);
11072
11073 if (nr_running == 1 && load > env->imbalance &&
11074 !check_cpu_capacity(rq, env->sd))
11075 break;
11076
11077 /*
11078 * For the load comparisons with the other CPUs,
11079 * consider the cpu_load() scaled with the CPU
11080 * capacity, so that the load can be moved away
11081 * from the CPU that is potentially running at a
11082 * lower capacity.
11083 *
11084 * Thus we're looking for max(load_i / capacity_i),
11085 * crosswise multiplication to rid ourselves of the
11086 * division works out to:
11087 * load_i * capacity_j > load_j * capacity_i;
11088 * where j is our previous maximum.
11089 */
11090 if (load * busiest_capacity > busiest_load * capacity) {
11091 busiest_load = load;
11092 busiest_capacity = capacity;
11093 busiest = rq;
11094 }
11095 break;
11096
11097 case migrate_util:
11098 util = cpu_util_cfs_boost(i);
11099
11100 /*
11101 * Don't try to pull utilization from a CPU with one
11102 * running task. Whatever its utilization, we will fail
11103 * detach the task.
11104 */
11105 if (nr_running <= 1)
11106 continue;
11107
11108 if (busiest_util < util) {
11109 busiest_util = util;
11110 busiest = rq;
11111 }
11112 break;
11113
11114 case migrate_task:
11115 if (busiest_nr < nr_running) {
11116 busiest_nr = nr_running;
11117 busiest = rq;
11118 }
11119 break;
11120
11121 case migrate_misfit:
11122 /*
11123 * For ASYM_CPUCAPACITY domains with misfit tasks we
11124 * simply seek the "biggest" misfit task.
11125 */
11126 if (rq->misfit_task_load > busiest_load) {
11127 busiest_load = rq->misfit_task_load;
11128 busiest = rq;
11129 }
11130
11131 break;
11132
11133 }
11134 }
11135
11136 return busiest;
11137 }
11138
11139 /*
11140 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11141 * so long as it is large enough.
11142 */
11143 #define MAX_PINNED_INTERVAL 512
11144
11145 static inline bool
asym_active_balance(struct lb_env * env)11146 asym_active_balance(struct lb_env *env)
11147 {
11148 /*
11149 * ASYM_PACKING needs to force migrate tasks from busy but lower
11150 * priority CPUs in order to pack all tasks in the highest priority
11151 * CPUs. When done between cores, do it only if the whole core if the
11152 * whole core is idle.
11153 *
11154 * If @env::src_cpu is an SMT core with busy siblings, let
11155 * the lower priority @env::dst_cpu help it. Do not follow
11156 * CPU priority.
11157 */
11158 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
11159 sched_use_asym_prio(env->sd, env->dst_cpu) &&
11160 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11161 !sched_use_asym_prio(env->sd, env->src_cpu));
11162 }
11163
11164 static inline bool
imbalanced_active_balance(struct lb_env * env)11165 imbalanced_active_balance(struct lb_env *env)
11166 {
11167 struct sched_domain *sd = env->sd;
11168
11169 /*
11170 * The imbalanced case includes the case of pinned tasks preventing a fair
11171 * distribution of the load on the system but also the even distribution of the
11172 * threads on a system with spare capacity
11173 */
11174 if ((env->migration_type == migrate_task) &&
11175 (sd->nr_balance_failed > sd->cache_nice_tries+2))
11176 return 1;
11177
11178 return 0;
11179 }
11180
need_active_balance(struct lb_env * env)11181 static int need_active_balance(struct lb_env *env)
11182 {
11183 struct sched_domain *sd = env->sd;
11184
11185 if (asym_active_balance(env))
11186 return 1;
11187
11188 if (imbalanced_active_balance(env))
11189 return 1;
11190
11191 /*
11192 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11193 * It's worth migrating the task if the src_cpu's capacity is reduced
11194 * because of other sched_class or IRQs if more capacity stays
11195 * available on dst_cpu.
11196 */
11197 if ((env->idle != CPU_NOT_IDLE) &&
11198 (env->src_rq->cfs.h_nr_running == 1)) {
11199 if ((check_cpu_capacity(env->src_rq, sd)) &&
11200 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11201 return 1;
11202 }
11203
11204 if (env->migration_type == migrate_misfit)
11205 return 1;
11206
11207 return 0;
11208 }
11209
11210 static int active_load_balance_cpu_stop(void *data);
11211
should_we_balance(struct lb_env * env)11212 static int should_we_balance(struct lb_env *env)
11213 {
11214 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11215 struct sched_group *sg = env->sd->groups;
11216 int cpu, idle_smt = -1;
11217
11218 /*
11219 * Ensure the balancing environment is consistent; can happen
11220 * when the softirq triggers 'during' hotplug.
11221 */
11222 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11223 return 0;
11224
11225 /*
11226 * In the newly idle case, we will allow all the CPUs
11227 * to do the newly idle load balance.
11228 *
11229 * However, we bail out if we already have tasks or a wakeup pending,
11230 * to optimize wakeup latency.
11231 */
11232 if (env->idle == CPU_NEWLY_IDLE) {
11233 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11234 return 0;
11235 return 1;
11236 }
11237
11238 cpumask_copy(swb_cpus, group_balance_mask(sg));
11239 /* Try to find first idle CPU */
11240 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11241 if (!idle_cpu(cpu))
11242 continue;
11243
11244 /*
11245 * Don't balance to idle SMT in busy core right away when
11246 * balancing cores, but remember the first idle SMT CPU for
11247 * later consideration. Find CPU on an idle core first.
11248 */
11249 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11250 if (idle_smt == -1)
11251 idle_smt = cpu;
11252 /*
11253 * If the core is not idle, and first SMT sibling which is
11254 * idle has been found, then its not needed to check other
11255 * SMT siblings for idleness:
11256 */
11257 #ifdef CONFIG_SCHED_SMT
11258 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11259 #endif
11260 continue;
11261 }
11262
11263 /*
11264 * Are we the first idle core in a non-SMT domain or higher,
11265 * or the first idle CPU in a SMT domain?
11266 */
11267 return cpu == env->dst_cpu;
11268 }
11269
11270 /* Are we the first idle CPU with busy siblings? */
11271 if (idle_smt != -1)
11272 return idle_smt == env->dst_cpu;
11273
11274 /* Are we the first CPU of this group ? */
11275 return group_balance_cpu(sg) == env->dst_cpu;
11276 }
11277
11278 /*
11279 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11280 * tasks if there is an imbalance.
11281 */
load_balance(int this_cpu,struct rq * this_rq,struct sched_domain * sd,enum cpu_idle_type idle,int * continue_balancing)11282 static int load_balance(int this_cpu, struct rq *this_rq,
11283 struct sched_domain *sd, enum cpu_idle_type idle,
11284 int *continue_balancing)
11285 {
11286 int ld_moved, cur_ld_moved, active_balance = 0;
11287 struct sched_domain *sd_parent = sd->parent;
11288 struct sched_group *group;
11289 struct rq *busiest;
11290 struct rq_flags rf;
11291 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11292 struct lb_env env = {
11293 .sd = sd,
11294 .dst_cpu = this_cpu,
11295 .dst_rq = this_rq,
11296 .dst_grpmask = group_balance_mask(sd->groups),
11297 .idle = idle,
11298 .loop_break = SCHED_NR_MIGRATE_BREAK,
11299 .cpus = cpus,
11300 .fbq_type = all,
11301 .tasks = LIST_HEAD_INIT(env.tasks),
11302 };
11303
11304 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11305
11306 schedstat_inc(sd->lb_count[idle]);
11307
11308 redo:
11309 if (!should_we_balance(&env)) {
11310 *continue_balancing = 0;
11311 goto out_balanced;
11312 }
11313
11314 group = find_busiest_group(&env);
11315 if (!group) {
11316 schedstat_inc(sd->lb_nobusyg[idle]);
11317 goto out_balanced;
11318 }
11319
11320 busiest = find_busiest_queue(&env, group);
11321 if (!busiest) {
11322 schedstat_inc(sd->lb_nobusyq[idle]);
11323 goto out_balanced;
11324 }
11325
11326 WARN_ON_ONCE(busiest == env.dst_rq);
11327
11328 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11329
11330 env.src_cpu = busiest->cpu;
11331 env.src_rq = busiest;
11332
11333 ld_moved = 0;
11334 /* Clear this flag as soon as we find a pullable task */
11335 env.flags |= LBF_ALL_PINNED;
11336 if (busiest->nr_running > 1) {
11337 /*
11338 * Attempt to move tasks. If find_busiest_group has found
11339 * an imbalance but busiest->nr_running <= 1, the group is
11340 * still unbalanced. ld_moved simply stays zero, so it is
11341 * correctly treated as an imbalance.
11342 */
11343 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11344
11345 more_balance:
11346 rq_lock_irqsave(busiest, &rf);
11347 update_rq_clock(busiest);
11348
11349 /*
11350 * cur_ld_moved - load moved in current iteration
11351 * ld_moved - cumulative load moved across iterations
11352 */
11353 cur_ld_moved = detach_tasks(&env);
11354
11355 /*
11356 * We've detached some tasks from busiest_rq. Every
11357 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11358 * unlock busiest->lock, and we are able to be sure
11359 * that nobody can manipulate the tasks in parallel.
11360 * See task_rq_lock() family for the details.
11361 */
11362
11363 rq_unlock(busiest, &rf);
11364
11365 if (cur_ld_moved) {
11366 attach_tasks(&env);
11367 ld_moved += cur_ld_moved;
11368 }
11369
11370 local_irq_restore(rf.flags);
11371
11372 if (env.flags & LBF_NEED_BREAK) {
11373 env.flags &= ~LBF_NEED_BREAK;
11374 goto more_balance;
11375 }
11376
11377 /*
11378 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11379 * us and move them to an alternate dst_cpu in our sched_group
11380 * where they can run. The upper limit on how many times we
11381 * iterate on same src_cpu is dependent on number of CPUs in our
11382 * sched_group.
11383 *
11384 * This changes load balance semantics a bit on who can move
11385 * load to a given_cpu. In addition to the given_cpu itself
11386 * (or a ilb_cpu acting on its behalf where given_cpu is
11387 * nohz-idle), we now have balance_cpu in a position to move
11388 * load to given_cpu. In rare situations, this may cause
11389 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11390 * _independently_ and at _same_ time to move some load to
11391 * given_cpu) causing excess load to be moved to given_cpu.
11392 * This however should not happen so much in practice and
11393 * moreover subsequent load balance cycles should correct the
11394 * excess load moved.
11395 */
11396 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11397
11398 /* Prevent to re-select dst_cpu via env's CPUs */
11399 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11400
11401 env.dst_rq = cpu_rq(env.new_dst_cpu);
11402 env.dst_cpu = env.new_dst_cpu;
11403 env.flags &= ~LBF_DST_PINNED;
11404 env.loop = 0;
11405 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11406
11407 /*
11408 * Go back to "more_balance" rather than "redo" since we
11409 * need to continue with same src_cpu.
11410 */
11411 goto more_balance;
11412 }
11413
11414 /*
11415 * We failed to reach balance because of affinity.
11416 */
11417 if (sd_parent) {
11418 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11419
11420 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11421 *group_imbalance = 1;
11422 }
11423
11424 /* All tasks on this runqueue were pinned by CPU affinity */
11425 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11426 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11427 /*
11428 * Attempting to continue load balancing at the current
11429 * sched_domain level only makes sense if there are
11430 * active CPUs remaining as possible busiest CPUs to
11431 * pull load from which are not contained within the
11432 * destination group that is receiving any migrated
11433 * load.
11434 */
11435 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11436 env.loop = 0;
11437 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11438 goto redo;
11439 }
11440 goto out_all_pinned;
11441 }
11442 }
11443
11444 if (!ld_moved) {
11445 schedstat_inc(sd->lb_failed[idle]);
11446 /*
11447 * Increment the failure counter only on periodic balance.
11448 * We do not want newidle balance, which can be very
11449 * frequent, pollute the failure counter causing
11450 * excessive cache_hot migrations and active balances.
11451 */
11452 if (idle != CPU_NEWLY_IDLE)
11453 sd->nr_balance_failed++;
11454
11455 if (need_active_balance(&env)) {
11456 unsigned long flags;
11457
11458 raw_spin_rq_lock_irqsave(busiest, flags);
11459
11460 /*
11461 * Don't kick the active_load_balance_cpu_stop,
11462 * if the curr task on busiest CPU can't be
11463 * moved to this_cpu:
11464 */
11465 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11466 raw_spin_rq_unlock_irqrestore(busiest, flags);
11467 goto out_one_pinned;
11468 }
11469
11470 /* Record that we found at least one task that could run on this_cpu */
11471 env.flags &= ~LBF_ALL_PINNED;
11472
11473 /*
11474 * ->active_balance synchronizes accesses to
11475 * ->active_balance_work. Once set, it's cleared
11476 * only after active load balance is finished.
11477 */
11478 if (!busiest->active_balance) {
11479 busiest->active_balance = 1;
11480 busiest->push_cpu = this_cpu;
11481 active_balance = 1;
11482 }
11483
11484 preempt_disable();
11485 raw_spin_rq_unlock_irqrestore(busiest, flags);
11486 if (active_balance) {
11487 stop_one_cpu_nowait(cpu_of(busiest),
11488 active_load_balance_cpu_stop, busiest,
11489 &busiest->active_balance_work);
11490 }
11491 preempt_enable();
11492 }
11493 } else {
11494 sd->nr_balance_failed = 0;
11495 }
11496
11497 if (likely(!active_balance) || need_active_balance(&env)) {
11498 /* We were unbalanced, so reset the balancing interval */
11499 sd->balance_interval = sd->min_interval;
11500 }
11501
11502 goto out;
11503
11504 out_balanced:
11505 /*
11506 * We reach balance although we may have faced some affinity
11507 * constraints. Clear the imbalance flag only if other tasks got
11508 * a chance to move and fix the imbalance.
11509 */
11510 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11511 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11512
11513 if (*group_imbalance)
11514 *group_imbalance = 0;
11515 }
11516
11517 out_all_pinned:
11518 /*
11519 * We reach balance because all tasks are pinned at this level so
11520 * we can't migrate them. Let the imbalance flag set so parent level
11521 * can try to migrate them.
11522 */
11523 schedstat_inc(sd->lb_balanced[idle]);
11524
11525 sd->nr_balance_failed = 0;
11526
11527 out_one_pinned:
11528 ld_moved = 0;
11529
11530 /*
11531 * newidle_balance() disregards balance intervals, so we could
11532 * repeatedly reach this code, which would lead to balance_interval
11533 * skyrocketing in a short amount of time. Skip the balance_interval
11534 * increase logic to avoid that.
11535 */
11536 if (env.idle == CPU_NEWLY_IDLE)
11537 goto out;
11538
11539 /* tune up the balancing interval */
11540 if ((env.flags & LBF_ALL_PINNED &&
11541 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11542 sd->balance_interval < sd->max_interval)
11543 sd->balance_interval *= 2;
11544 out:
11545 return ld_moved;
11546 }
11547
11548 static inline unsigned long
get_sd_balance_interval(struct sched_domain * sd,int cpu_busy)11549 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11550 {
11551 unsigned long interval = sd->balance_interval;
11552
11553 if (cpu_busy)
11554 interval *= sd->busy_factor;
11555
11556 /* scale ms to jiffies */
11557 interval = msecs_to_jiffies(interval);
11558
11559 /*
11560 * Reduce likelihood of busy balancing at higher domains racing with
11561 * balancing at lower domains by preventing their balancing periods
11562 * from being multiples of each other.
11563 */
11564 if (cpu_busy)
11565 interval -= 1;
11566
11567 interval = clamp(interval, 1UL, max_load_balance_interval);
11568
11569 return interval;
11570 }
11571
11572 static inline void
update_next_balance(struct sched_domain * sd,unsigned long * next_balance)11573 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11574 {
11575 unsigned long interval, next;
11576
11577 /* used by idle balance, so cpu_busy = 0 */
11578 interval = get_sd_balance_interval(sd, 0);
11579 next = sd->last_balance + interval;
11580
11581 if (time_after(*next_balance, next))
11582 *next_balance = next;
11583 }
11584
11585 /*
11586 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11587 * running tasks off the busiest CPU onto idle CPUs. It requires at
11588 * least 1 task to be running on each physical CPU where possible, and
11589 * avoids physical / logical imbalances.
11590 */
active_load_balance_cpu_stop(void * data)11591 static int active_load_balance_cpu_stop(void *data)
11592 {
11593 struct rq *busiest_rq = data;
11594 int busiest_cpu = cpu_of(busiest_rq);
11595 int target_cpu = busiest_rq->push_cpu;
11596 struct rq *target_rq = cpu_rq(target_cpu);
11597 struct sched_domain *sd;
11598 struct task_struct *p = NULL;
11599 struct rq_flags rf;
11600
11601 rq_lock_irq(busiest_rq, &rf);
11602 /*
11603 * Between queueing the stop-work and running it is a hole in which
11604 * CPUs can become inactive. We should not move tasks from or to
11605 * inactive CPUs.
11606 */
11607 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11608 goto out_unlock;
11609
11610 /* Make sure the requested CPU hasn't gone down in the meantime: */
11611 if (unlikely(busiest_cpu != smp_processor_id() ||
11612 !busiest_rq->active_balance))
11613 goto out_unlock;
11614
11615 /* Is there any task to move? */
11616 if (busiest_rq->nr_running <= 1)
11617 goto out_unlock;
11618
11619 /*
11620 * This condition is "impossible", if it occurs
11621 * we need to fix it. Originally reported by
11622 * Bjorn Helgaas on a 128-CPU setup.
11623 */
11624 WARN_ON_ONCE(busiest_rq == target_rq);
11625
11626 /* Search for an sd spanning us and the target CPU. */
11627 rcu_read_lock();
11628 for_each_domain(target_cpu, sd) {
11629 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11630 break;
11631 }
11632
11633 if (likely(sd)) {
11634 struct lb_env env = {
11635 .sd = sd,
11636 .dst_cpu = target_cpu,
11637 .dst_rq = target_rq,
11638 .src_cpu = busiest_rq->cpu,
11639 .src_rq = busiest_rq,
11640 .idle = CPU_IDLE,
11641 .flags = LBF_ACTIVE_LB,
11642 };
11643
11644 schedstat_inc(sd->alb_count);
11645 update_rq_clock(busiest_rq);
11646
11647 p = detach_one_task(&env);
11648 if (p) {
11649 schedstat_inc(sd->alb_pushed);
11650 /* Active balancing done, reset the failure counter. */
11651 sd->nr_balance_failed = 0;
11652 } else {
11653 schedstat_inc(sd->alb_failed);
11654 }
11655 }
11656 rcu_read_unlock();
11657 out_unlock:
11658 busiest_rq->active_balance = 0;
11659 rq_unlock(busiest_rq, &rf);
11660
11661 if (p)
11662 attach_one_task(target_rq, p);
11663
11664 local_irq_enable();
11665
11666 return 0;
11667 }
11668
11669 static DEFINE_SPINLOCK(balancing);
11670
11671 /*
11672 * Scale the max load_balance interval with the number of CPUs in the system.
11673 * This trades load-balance latency on larger machines for less cross talk.
11674 */
update_max_interval(void)11675 void update_max_interval(void)
11676 {
11677 max_load_balance_interval = HZ*num_online_cpus()/10;
11678 }
11679
update_newidle_cost(struct sched_domain * sd,u64 cost)11680 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11681 {
11682 if (cost > sd->max_newidle_lb_cost) {
11683 /*
11684 * Track max cost of a domain to make sure to not delay the
11685 * next wakeup on the CPU.
11686 */
11687 sd->max_newidle_lb_cost = cost;
11688 sd->last_decay_max_lb_cost = jiffies;
11689 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11690 /*
11691 * Decay the newidle max times by ~1% per second to ensure that
11692 * it is not outdated and the current max cost is actually
11693 * shorter.
11694 */
11695 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11696 sd->last_decay_max_lb_cost = jiffies;
11697
11698 return true;
11699 }
11700
11701 return false;
11702 }
11703
11704 /*
11705 * It checks each scheduling domain to see if it is due to be balanced,
11706 * and initiates a balancing operation if so.
11707 *
11708 * Balancing parameters are set up in init_sched_domains.
11709 */
rebalance_domains(struct rq * rq,enum cpu_idle_type idle)11710 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11711 {
11712 int continue_balancing = 1;
11713 int cpu = rq->cpu;
11714 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11715 unsigned long interval;
11716 struct sched_domain *sd;
11717 /* Earliest time when we have to do rebalance again */
11718 unsigned long next_balance = jiffies + 60*HZ;
11719 int update_next_balance = 0;
11720 int need_serialize, need_decay = 0;
11721 u64 max_cost = 0;
11722
11723 rcu_read_lock();
11724 for_each_domain(cpu, sd) {
11725 /*
11726 * Decay the newidle max times here because this is a regular
11727 * visit to all the domains.
11728 */
11729 need_decay = update_newidle_cost(sd, 0);
11730 max_cost += sd->max_newidle_lb_cost;
11731
11732 /*
11733 * Stop the load balance at this level. There is another
11734 * CPU in our sched group which is doing load balancing more
11735 * actively.
11736 */
11737 if (!continue_balancing) {
11738 if (need_decay)
11739 continue;
11740 break;
11741 }
11742
11743 interval = get_sd_balance_interval(sd, busy);
11744
11745 need_serialize = sd->flags & SD_SERIALIZE;
11746 if (need_serialize) {
11747 if (!spin_trylock(&balancing))
11748 goto out;
11749 }
11750
11751 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11752 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11753 /*
11754 * The LBF_DST_PINNED logic could have changed
11755 * env->dst_cpu, so we can't know our idle
11756 * state even if we migrated tasks. Update it.
11757 */
11758 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11759 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11760 }
11761 sd->last_balance = jiffies;
11762 interval = get_sd_balance_interval(sd, busy);
11763 }
11764 if (need_serialize)
11765 spin_unlock(&balancing);
11766 out:
11767 if (time_after(next_balance, sd->last_balance + interval)) {
11768 next_balance = sd->last_balance + interval;
11769 update_next_balance = 1;
11770 }
11771 }
11772 if (need_decay) {
11773 /*
11774 * Ensure the rq-wide value also decays but keep it at a
11775 * reasonable floor to avoid funnies with rq->avg_idle.
11776 */
11777 rq->max_idle_balance_cost =
11778 max((u64)sysctl_sched_migration_cost, max_cost);
11779 }
11780 rcu_read_unlock();
11781
11782 /*
11783 * next_balance will be updated only when there is a need.
11784 * When the cpu is attached to null domain for ex, it will not be
11785 * updated.
11786 */
11787 if (likely(update_next_balance))
11788 rq->next_balance = next_balance;
11789
11790 }
11791
on_null_domain(struct rq * rq)11792 static inline int on_null_domain(struct rq *rq)
11793 {
11794 return unlikely(!rcu_dereference_sched(rq->sd));
11795 }
11796
11797 #ifdef CONFIG_NO_HZ_COMMON
11798 /*
11799 * idle load balancing details
11800 * - When one of the busy CPUs notice that there may be an idle rebalancing
11801 * needed, they will kick the idle load balancer, which then does idle
11802 * load balancing for all the idle CPUs.
11803 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11804 * anywhere yet.
11805 */
11806
find_new_ilb(void)11807 static inline int find_new_ilb(void)
11808 {
11809 int ilb;
11810 const struct cpumask *hk_mask;
11811
11812 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11813
11814 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11815
11816 if (ilb == smp_processor_id())
11817 continue;
11818
11819 if (idle_cpu(ilb))
11820 return ilb;
11821 }
11822
11823 return nr_cpu_ids;
11824 }
11825
11826 /*
11827 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11828 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11829 */
kick_ilb(unsigned int flags)11830 static void kick_ilb(unsigned int flags)
11831 {
11832 int ilb_cpu;
11833
11834 /*
11835 * Increase nohz.next_balance only when if full ilb is triggered but
11836 * not if we only update stats.
11837 */
11838 if (flags & NOHZ_BALANCE_KICK)
11839 nohz.next_balance = jiffies+1;
11840
11841 ilb_cpu = find_new_ilb();
11842
11843 if (ilb_cpu >= nr_cpu_ids)
11844 return;
11845
11846 /*
11847 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11848 * the first flag owns it; cleared by nohz_csd_func().
11849 */
11850 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11851 if (flags & NOHZ_KICK_MASK)
11852 return;
11853
11854 /*
11855 * This way we generate an IPI on the target CPU which
11856 * is idle. And the softirq performing nohz idle load balance
11857 * will be run before returning from the IPI.
11858 */
11859 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11860 }
11861
11862 /*
11863 * Current decision point for kicking the idle load balancer in the presence
11864 * of idle CPUs in the system.
11865 */
nohz_balancer_kick(struct rq * rq)11866 static void nohz_balancer_kick(struct rq *rq)
11867 {
11868 unsigned long now = jiffies;
11869 struct sched_domain_shared *sds;
11870 struct sched_domain *sd;
11871 int nr_busy, i, cpu = rq->cpu;
11872 unsigned int flags = 0;
11873
11874 if (unlikely(rq->idle_balance))
11875 return;
11876
11877 /*
11878 * We may be recently in ticked or tickless idle mode. At the first
11879 * busy tick after returning from idle, we will update the busy stats.
11880 */
11881 nohz_balance_exit_idle(rq);
11882
11883 /*
11884 * None are in tickless mode and hence no need for NOHZ idle load
11885 * balancing.
11886 */
11887 if (likely(!atomic_read(&nohz.nr_cpus)))
11888 return;
11889
11890 if (READ_ONCE(nohz.has_blocked) &&
11891 time_after(now, READ_ONCE(nohz.next_blocked)))
11892 flags = NOHZ_STATS_KICK;
11893
11894 if (time_before(now, nohz.next_balance))
11895 goto out;
11896
11897 if (rq->nr_running >= 2) {
11898 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11899 goto out;
11900 }
11901
11902 rcu_read_lock();
11903
11904 sd = rcu_dereference(rq->sd);
11905 if (sd) {
11906 /*
11907 * If there's a CFS task and the current CPU has reduced
11908 * capacity; kick the ILB to see if there's a better CPU to run
11909 * on.
11910 */
11911 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11912 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11913 goto unlock;
11914 }
11915 }
11916
11917 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11918 if (sd) {
11919 /*
11920 * When ASYM_PACKING; see if there's a more preferred CPU
11921 * currently idle; in which case, kick the ILB to move tasks
11922 * around.
11923 *
11924 * When balancing betwen cores, all the SMT siblings of the
11925 * preferred CPU must be idle.
11926 */
11927 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11928 if (sched_use_asym_prio(sd, i) &&
11929 sched_asym_prefer(i, cpu)) {
11930 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11931 goto unlock;
11932 }
11933 }
11934 }
11935
11936 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11937 if (sd) {
11938 /*
11939 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11940 * to run the misfit task on.
11941 */
11942 if (check_misfit_status(rq, sd)) {
11943 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11944 goto unlock;
11945 }
11946
11947 /*
11948 * For asymmetric systems, we do not want to nicely balance
11949 * cache use, instead we want to embrace asymmetry and only
11950 * ensure tasks have enough CPU capacity.
11951 *
11952 * Skip the LLC logic because it's not relevant in that case.
11953 */
11954 goto unlock;
11955 }
11956
11957 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11958 if (sds) {
11959 /*
11960 * If there is an imbalance between LLC domains (IOW we could
11961 * increase the overall cache use), we need some less-loaded LLC
11962 * domain to pull some load. Likewise, we may need to spread
11963 * load within the current LLC domain (e.g. packed SMT cores but
11964 * other CPUs are idle). We can't really know from here how busy
11965 * the others are - so just get a nohz balance going if it looks
11966 * like this LLC domain has tasks we could move.
11967 */
11968 nr_busy = atomic_read(&sds->nr_busy_cpus);
11969 if (nr_busy > 1) {
11970 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11971 goto unlock;
11972 }
11973 }
11974 unlock:
11975 rcu_read_unlock();
11976 out:
11977 if (READ_ONCE(nohz.needs_update))
11978 flags |= NOHZ_NEXT_KICK;
11979
11980 if (flags)
11981 kick_ilb(flags);
11982 }
11983
set_cpu_sd_state_busy(int cpu)11984 static void set_cpu_sd_state_busy(int cpu)
11985 {
11986 struct sched_domain *sd;
11987
11988 rcu_read_lock();
11989 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11990
11991 if (!sd || !sd->nohz_idle)
11992 goto unlock;
11993 sd->nohz_idle = 0;
11994
11995 atomic_inc(&sd->shared->nr_busy_cpus);
11996 unlock:
11997 rcu_read_unlock();
11998 }
11999
nohz_balance_exit_idle(struct rq * rq)12000 void nohz_balance_exit_idle(struct rq *rq)
12001 {
12002 SCHED_WARN_ON(rq != this_rq());
12003
12004 if (likely(!rq->nohz_tick_stopped))
12005 return;
12006
12007 rq->nohz_tick_stopped = 0;
12008 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
12009 atomic_dec(&nohz.nr_cpus);
12010
12011 set_cpu_sd_state_busy(rq->cpu);
12012 }
12013
set_cpu_sd_state_idle(int cpu)12014 static void set_cpu_sd_state_idle(int cpu)
12015 {
12016 struct sched_domain *sd;
12017
12018 rcu_read_lock();
12019 sd = rcu_dereference(per_cpu(sd_llc, cpu));
12020
12021 if (!sd || sd->nohz_idle)
12022 goto unlock;
12023 sd->nohz_idle = 1;
12024
12025 atomic_dec(&sd->shared->nr_busy_cpus);
12026 unlock:
12027 rcu_read_unlock();
12028 }
12029
12030 /*
12031 * This routine will record that the CPU is going idle with tick stopped.
12032 * This info will be used in performing idle load balancing in the future.
12033 */
nohz_balance_enter_idle(int cpu)12034 void nohz_balance_enter_idle(int cpu)
12035 {
12036 struct rq *rq = cpu_rq(cpu);
12037
12038 SCHED_WARN_ON(cpu != smp_processor_id());
12039
12040 /* If this CPU is going down, then nothing needs to be done: */
12041 if (!cpu_active(cpu))
12042 return;
12043
12044 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
12045 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
12046 return;
12047
12048 /*
12049 * Can be set safely without rq->lock held
12050 * If a clear happens, it will have evaluated last additions because
12051 * rq->lock is held during the check and the clear
12052 */
12053 rq->has_blocked_load = 1;
12054
12055 /*
12056 * The tick is still stopped but load could have been added in the
12057 * meantime. We set the nohz.has_blocked flag to trig a check of the
12058 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
12059 * of nohz.has_blocked can only happen after checking the new load
12060 */
12061 if (rq->nohz_tick_stopped)
12062 goto out;
12063
12064 /* If we're a completely isolated CPU, we don't play: */
12065 if (on_null_domain(rq))
12066 return;
12067
12068 rq->nohz_tick_stopped = 1;
12069
12070 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
12071 atomic_inc(&nohz.nr_cpus);
12072
12073 /*
12074 * Ensures that if nohz_idle_balance() fails to observe our
12075 * @idle_cpus_mask store, it must observe the @has_blocked
12076 * and @needs_update stores.
12077 */
12078 smp_mb__after_atomic();
12079
12080 set_cpu_sd_state_idle(cpu);
12081
12082 WRITE_ONCE(nohz.needs_update, 1);
12083 out:
12084 /*
12085 * Each time a cpu enter idle, we assume that it has blocked load and
12086 * enable the periodic update of the load of idle cpus
12087 */
12088 WRITE_ONCE(nohz.has_blocked, 1);
12089 }
12090
update_nohz_stats(struct rq * rq)12091 static bool update_nohz_stats(struct rq *rq)
12092 {
12093 unsigned int cpu = rq->cpu;
12094
12095 if (!rq->has_blocked_load)
12096 return false;
12097
12098 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
12099 return false;
12100
12101 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
12102 return true;
12103
12104 update_blocked_averages(cpu);
12105
12106 return rq->has_blocked_load;
12107 }
12108
12109 /*
12110 * Internal function that runs load balance for all idle cpus. The load balance
12111 * can be a simple update of blocked load or a complete load balance with
12112 * tasks movement depending of flags.
12113 */
_nohz_idle_balance(struct rq * this_rq,unsigned int flags)12114 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
12115 {
12116 /* Earliest time when we have to do rebalance again */
12117 unsigned long now = jiffies;
12118 unsigned long next_balance = now + 60*HZ;
12119 bool has_blocked_load = false;
12120 int update_next_balance = 0;
12121 int this_cpu = this_rq->cpu;
12122 int balance_cpu;
12123 struct rq *rq;
12124
12125 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
12126
12127 /*
12128 * We assume there will be no idle load after this update and clear
12129 * the has_blocked flag. If a cpu enters idle in the mean time, it will
12130 * set the has_blocked flag and trigger another update of idle load.
12131 * Because a cpu that becomes idle, is added to idle_cpus_mask before
12132 * setting the flag, we are sure to not clear the state and not
12133 * check the load of an idle cpu.
12134 *
12135 * Same applies to idle_cpus_mask vs needs_update.
12136 */
12137 if (flags & NOHZ_STATS_KICK)
12138 WRITE_ONCE(nohz.has_blocked, 0);
12139 if (flags & NOHZ_NEXT_KICK)
12140 WRITE_ONCE(nohz.needs_update, 0);
12141
12142 /*
12143 * Ensures that if we miss the CPU, we must see the has_blocked
12144 * store from nohz_balance_enter_idle().
12145 */
12146 smp_mb();
12147
12148 /*
12149 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12150 * chance for other idle cpu to pull load.
12151 */
12152 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
12153 if (!idle_cpu(balance_cpu))
12154 continue;
12155
12156 /*
12157 * If this CPU gets work to do, stop the load balancing
12158 * work being done for other CPUs. Next load
12159 * balancing owner will pick it up.
12160 */
12161 if (!idle_cpu(this_cpu) && need_resched()) {
12162 if (flags & NOHZ_STATS_KICK)
12163 has_blocked_load = true;
12164 if (flags & NOHZ_NEXT_KICK)
12165 WRITE_ONCE(nohz.needs_update, 1);
12166 goto abort;
12167 }
12168
12169 rq = cpu_rq(balance_cpu);
12170
12171 if (flags & NOHZ_STATS_KICK)
12172 has_blocked_load |= update_nohz_stats(rq);
12173
12174 /*
12175 * If time for next balance is due,
12176 * do the balance.
12177 */
12178 if (time_after_eq(jiffies, rq->next_balance)) {
12179 struct rq_flags rf;
12180
12181 rq_lock_irqsave(rq, &rf);
12182 update_rq_clock(rq);
12183 rq_unlock_irqrestore(rq, &rf);
12184
12185 if (flags & NOHZ_BALANCE_KICK)
12186 rebalance_domains(rq, CPU_IDLE);
12187 }
12188
12189 if (time_after(next_balance, rq->next_balance)) {
12190 next_balance = rq->next_balance;
12191 update_next_balance = 1;
12192 }
12193 }
12194
12195 /*
12196 * next_balance will be updated only when there is a need.
12197 * When the CPU is attached to null domain for ex, it will not be
12198 * updated.
12199 */
12200 if (likely(update_next_balance))
12201 nohz.next_balance = next_balance;
12202
12203 if (flags & NOHZ_STATS_KICK)
12204 WRITE_ONCE(nohz.next_blocked,
12205 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12206
12207 abort:
12208 /* There is still blocked load, enable periodic update */
12209 if (has_blocked_load)
12210 WRITE_ONCE(nohz.has_blocked, 1);
12211 }
12212
12213 /*
12214 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12215 * rebalancing for all the cpus for whom scheduler ticks are stopped.
12216 */
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12217 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12218 {
12219 unsigned int flags = this_rq->nohz_idle_balance;
12220
12221 if (!flags)
12222 return false;
12223
12224 this_rq->nohz_idle_balance = 0;
12225
12226 if (idle != CPU_IDLE)
12227 return false;
12228
12229 _nohz_idle_balance(this_rq, flags);
12230
12231 return true;
12232 }
12233
12234 /*
12235 * Check if we need to run the ILB for updating blocked load before entering
12236 * idle state.
12237 */
nohz_run_idle_balance(int cpu)12238 void nohz_run_idle_balance(int cpu)
12239 {
12240 unsigned int flags;
12241
12242 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12243
12244 /*
12245 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12246 * (ie NOHZ_STATS_KICK set) and will do the same.
12247 */
12248 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12249 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12250 }
12251
nohz_newidle_balance(struct rq * this_rq)12252 static void nohz_newidle_balance(struct rq *this_rq)
12253 {
12254 int this_cpu = this_rq->cpu;
12255
12256 /*
12257 * This CPU doesn't want to be disturbed by scheduler
12258 * housekeeping
12259 */
12260 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12261 return;
12262
12263 /* Will wake up very soon. No time for doing anything else*/
12264 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12265 return;
12266
12267 /* Don't need to update blocked load of idle CPUs*/
12268 if (!READ_ONCE(nohz.has_blocked) ||
12269 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12270 return;
12271
12272 /*
12273 * Set the need to trigger ILB in order to update blocked load
12274 * before entering idle state.
12275 */
12276 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12277 }
12278
12279 #else /* !CONFIG_NO_HZ_COMMON */
nohz_balancer_kick(struct rq * rq)12280 static inline void nohz_balancer_kick(struct rq *rq) { }
12281
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12282 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12283 {
12284 return false;
12285 }
12286
nohz_newidle_balance(struct rq * this_rq)12287 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12288 #endif /* CONFIG_NO_HZ_COMMON */
12289
12290 /*
12291 * newidle_balance is called by schedule() if this_cpu is about to become
12292 * idle. Attempts to pull tasks from other CPUs.
12293 *
12294 * Returns:
12295 * < 0 - we released the lock and there are !fair tasks present
12296 * 0 - failed, no new tasks
12297 * > 0 - success, new (fair) tasks present
12298 */
newidle_balance(struct rq * this_rq,struct rq_flags * rf)12299 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12300 {
12301 unsigned long next_balance = jiffies + HZ;
12302 int this_cpu = this_rq->cpu;
12303 u64 t0, t1, curr_cost = 0;
12304 struct sched_domain *sd;
12305 int pulled_task = 0;
12306
12307 update_misfit_status(NULL, this_rq);
12308
12309 /*
12310 * There is a task waiting to run. No need to search for one.
12311 * Return 0; the task will be enqueued when switching to idle.
12312 */
12313 if (this_rq->ttwu_pending)
12314 return 0;
12315
12316 /*
12317 * We must set idle_stamp _before_ calling idle_balance(), such that we
12318 * measure the duration of idle_balance() as idle time.
12319 */
12320 this_rq->idle_stamp = rq_clock(this_rq);
12321
12322 /*
12323 * Do not pull tasks towards !active CPUs...
12324 */
12325 if (!cpu_active(this_cpu))
12326 return 0;
12327
12328 /*
12329 * This is OK, because current is on_cpu, which avoids it being picked
12330 * for load-balance and preemption/IRQs are still disabled avoiding
12331 * further scheduler activity on it and we're being very careful to
12332 * re-start the picking loop.
12333 */
12334 rq_unpin_lock(this_rq, rf);
12335
12336 rcu_read_lock();
12337 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12338
12339 if (!READ_ONCE(this_rq->rd->overload) ||
12340 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12341
12342 if (sd)
12343 update_next_balance(sd, &next_balance);
12344 rcu_read_unlock();
12345
12346 goto out;
12347 }
12348 rcu_read_unlock();
12349
12350 raw_spin_rq_unlock(this_rq);
12351
12352 t0 = sched_clock_cpu(this_cpu);
12353 update_blocked_averages(this_cpu);
12354
12355 rcu_read_lock();
12356 for_each_domain(this_cpu, sd) {
12357 int continue_balancing = 1;
12358 u64 domain_cost;
12359
12360 update_next_balance(sd, &next_balance);
12361
12362 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12363 break;
12364
12365 if (sd->flags & SD_BALANCE_NEWIDLE) {
12366
12367 pulled_task = load_balance(this_cpu, this_rq,
12368 sd, CPU_NEWLY_IDLE,
12369 &continue_balancing);
12370
12371 t1 = sched_clock_cpu(this_cpu);
12372 domain_cost = t1 - t0;
12373 update_newidle_cost(sd, domain_cost);
12374
12375 curr_cost += domain_cost;
12376 t0 = t1;
12377 }
12378
12379 /*
12380 * Stop searching for tasks to pull if there are
12381 * now runnable tasks on this rq.
12382 */
12383 if (pulled_task || this_rq->nr_running > 0 ||
12384 this_rq->ttwu_pending)
12385 break;
12386 }
12387 rcu_read_unlock();
12388
12389 raw_spin_rq_lock(this_rq);
12390
12391 if (curr_cost > this_rq->max_idle_balance_cost)
12392 this_rq->max_idle_balance_cost = curr_cost;
12393
12394 /*
12395 * While browsing the domains, we released the rq lock, a task could
12396 * have been enqueued in the meantime. Since we're not going idle,
12397 * pretend we pulled a task.
12398 */
12399 if (this_rq->cfs.h_nr_running && !pulled_task)
12400 pulled_task = 1;
12401
12402 /* Is there a task of a high priority class? */
12403 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12404 pulled_task = -1;
12405
12406 out:
12407 /* Move the next balance forward */
12408 if (time_after(this_rq->next_balance, next_balance))
12409 this_rq->next_balance = next_balance;
12410
12411 if (pulled_task)
12412 this_rq->idle_stamp = 0;
12413 else
12414 nohz_newidle_balance(this_rq);
12415
12416 rq_repin_lock(this_rq, rf);
12417
12418 return pulled_task;
12419 }
12420
12421 /*
12422 * run_rebalance_domains is triggered when needed from the scheduler tick.
12423 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12424 */
run_rebalance_domains(struct softirq_action * h)12425 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12426 {
12427 struct rq *this_rq = this_rq();
12428 enum cpu_idle_type idle = this_rq->idle_balance ?
12429 CPU_IDLE : CPU_NOT_IDLE;
12430
12431 /*
12432 * If this CPU has a pending nohz_balance_kick, then do the
12433 * balancing on behalf of the other idle CPUs whose ticks are
12434 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12435 * give the idle CPUs a chance to load balance. Else we may
12436 * load balance only within the local sched_domain hierarchy
12437 * and abort nohz_idle_balance altogether if we pull some load.
12438 */
12439 if (nohz_idle_balance(this_rq, idle))
12440 return;
12441
12442 /* normal load balance */
12443 update_blocked_averages(this_rq->cpu);
12444 rebalance_domains(this_rq, idle);
12445 }
12446
12447 /*
12448 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12449 */
trigger_load_balance(struct rq * rq)12450 void trigger_load_balance(struct rq *rq)
12451 {
12452 /*
12453 * Don't need to rebalance while attached to NULL domain or
12454 * runqueue CPU is not active
12455 */
12456 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12457 return;
12458
12459 if (time_after_eq(jiffies, rq->next_balance))
12460 raise_softirq(SCHED_SOFTIRQ);
12461
12462 nohz_balancer_kick(rq);
12463 }
12464
rq_online_fair(struct rq * rq)12465 static void rq_online_fair(struct rq *rq)
12466 {
12467 update_sysctl();
12468
12469 update_runtime_enabled(rq);
12470 }
12471
rq_offline_fair(struct rq * rq)12472 static void rq_offline_fair(struct rq *rq)
12473 {
12474 update_sysctl();
12475
12476 /* Ensure any throttled groups are reachable by pick_next_task */
12477 unthrottle_offline_cfs_rqs(rq);
12478 }
12479
12480 #endif /* CONFIG_SMP */
12481
12482 #ifdef CONFIG_SCHED_CORE
12483 static inline bool
__entity_slice_used(struct sched_entity * se,int min_nr_tasks)12484 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12485 {
12486 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12487 u64 slice = se->slice;
12488
12489 return (rtime * min_nr_tasks > slice);
12490 }
12491
12492 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
task_tick_core(struct rq * rq,struct task_struct * curr)12493 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12494 {
12495 if (!sched_core_enabled(rq))
12496 return;
12497
12498 /*
12499 * If runqueue has only one task which used up its slice and
12500 * if the sibling is forced idle, then trigger schedule to
12501 * give forced idle task a chance.
12502 *
12503 * sched_slice() considers only this active rq and it gets the
12504 * whole slice. But during force idle, we have siblings acting
12505 * like a single runqueue and hence we need to consider runnable
12506 * tasks on this CPU and the forced idle CPU. Ideally, we should
12507 * go through the forced idle rq, but that would be a perf hit.
12508 * We can assume that the forced idle CPU has at least
12509 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12510 * if we need to give up the CPU.
12511 */
12512 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12513 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12514 resched_curr(rq);
12515 }
12516
12517 /*
12518 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12519 */
se_fi_update(const struct sched_entity * se,unsigned int fi_seq,bool forceidle)12520 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12521 bool forceidle)
12522 {
12523 for_each_sched_entity(se) {
12524 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12525
12526 if (forceidle) {
12527 if (cfs_rq->forceidle_seq == fi_seq)
12528 break;
12529 cfs_rq->forceidle_seq = fi_seq;
12530 }
12531
12532 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12533 }
12534 }
12535
task_vruntime_update(struct rq * rq,struct task_struct * p,bool in_fi)12536 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12537 {
12538 struct sched_entity *se = &p->se;
12539
12540 if (p->sched_class != &fair_sched_class)
12541 return;
12542
12543 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12544 }
12545
cfs_prio_less(const struct task_struct * a,const struct task_struct * b,bool in_fi)12546 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12547 bool in_fi)
12548 {
12549 struct rq *rq = task_rq(a);
12550 const struct sched_entity *sea = &a->se;
12551 const struct sched_entity *seb = &b->se;
12552 struct cfs_rq *cfs_rqa;
12553 struct cfs_rq *cfs_rqb;
12554 s64 delta;
12555
12556 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12557
12558 #ifdef CONFIG_FAIR_GROUP_SCHED
12559 /*
12560 * Find an se in the hierarchy for tasks a and b, such that the se's
12561 * are immediate siblings.
12562 */
12563 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12564 int sea_depth = sea->depth;
12565 int seb_depth = seb->depth;
12566
12567 if (sea_depth >= seb_depth)
12568 sea = parent_entity(sea);
12569 if (sea_depth <= seb_depth)
12570 seb = parent_entity(seb);
12571 }
12572
12573 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12574 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12575
12576 cfs_rqa = sea->cfs_rq;
12577 cfs_rqb = seb->cfs_rq;
12578 #else
12579 cfs_rqa = &task_rq(a)->cfs;
12580 cfs_rqb = &task_rq(b)->cfs;
12581 #endif
12582
12583 /*
12584 * Find delta after normalizing se's vruntime with its cfs_rq's
12585 * min_vruntime_fi, which would have been updated in prior calls
12586 * to se_fi_update().
12587 */
12588 delta = (s64)(sea->vruntime - seb->vruntime) +
12589 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12590
12591 return delta > 0;
12592 }
12593
task_is_throttled_fair(struct task_struct * p,int cpu)12594 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12595 {
12596 struct cfs_rq *cfs_rq;
12597
12598 #ifdef CONFIG_FAIR_GROUP_SCHED
12599 cfs_rq = task_group(p)->cfs_rq[cpu];
12600 #else
12601 cfs_rq = &cpu_rq(cpu)->cfs;
12602 #endif
12603 return throttled_hierarchy(cfs_rq);
12604 }
12605 #else
task_tick_core(struct rq * rq,struct task_struct * curr)12606 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12607 #endif
12608
12609 /*
12610 * scheduler tick hitting a task of our scheduling class.
12611 *
12612 * NOTE: This function can be called remotely by the tick offload that
12613 * goes along full dynticks. Therefore no local assumption can be made
12614 * and everything must be accessed through the @rq and @curr passed in
12615 * parameters.
12616 */
task_tick_fair(struct rq * rq,struct task_struct * curr,int queued)12617 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12618 {
12619 struct cfs_rq *cfs_rq;
12620 struct sched_entity *se = &curr->se;
12621
12622 for_each_sched_entity(se) {
12623 cfs_rq = cfs_rq_of(se);
12624 entity_tick(cfs_rq, se, queued);
12625 }
12626
12627 if (static_branch_unlikely(&sched_numa_balancing))
12628 task_tick_numa(rq, curr);
12629
12630 update_misfit_status(curr, rq);
12631 check_update_overutilized_status(task_rq(curr));
12632
12633 task_tick_core(rq, curr);
12634 }
12635
12636 /*
12637 * called on fork with the child task as argument from the parent's context
12638 * - child not yet on the tasklist
12639 * - preemption disabled
12640 */
task_fork_fair(struct task_struct * p)12641 static void task_fork_fair(struct task_struct *p)
12642 {
12643 struct sched_entity *se = &p->se, *curr;
12644 struct cfs_rq *cfs_rq;
12645 struct rq *rq = this_rq();
12646 struct rq_flags rf;
12647
12648 rq_lock(rq, &rf);
12649 update_rq_clock(rq);
12650
12651 cfs_rq = task_cfs_rq(current);
12652 curr = cfs_rq->curr;
12653 if (curr)
12654 update_curr(cfs_rq);
12655 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12656 rq_unlock(rq, &rf);
12657 }
12658
12659 /*
12660 * Priority of the task has changed. Check to see if we preempt
12661 * the current task.
12662 */
12663 static void
prio_changed_fair(struct rq * rq,struct task_struct * p,int oldprio)12664 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12665 {
12666 if (!task_on_rq_queued(p))
12667 return;
12668
12669 if (rq->cfs.nr_running == 1)
12670 return;
12671
12672 /*
12673 * Reschedule if we are currently running on this runqueue and
12674 * our priority decreased, or if we are not currently running on
12675 * this runqueue and our priority is higher than the current's
12676 */
12677 if (task_current(rq, p)) {
12678 if (p->prio > oldprio)
12679 resched_curr(rq);
12680 } else
12681 wakeup_preempt(rq, p, 0);
12682 }
12683
12684 #ifdef CONFIG_FAIR_GROUP_SCHED
12685 /*
12686 * Propagate the changes of the sched_entity across the tg tree to make it
12687 * visible to the root
12688 */
propagate_entity_cfs_rq(struct sched_entity * se)12689 static void propagate_entity_cfs_rq(struct sched_entity *se)
12690 {
12691 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12692
12693 if (cfs_rq_throttled(cfs_rq))
12694 return;
12695
12696 if (!throttled_hierarchy(cfs_rq))
12697 list_add_leaf_cfs_rq(cfs_rq);
12698
12699 /* Start to propagate at parent */
12700 se = se->parent;
12701
12702 for_each_sched_entity(se) {
12703 cfs_rq = cfs_rq_of(se);
12704
12705 update_load_avg(cfs_rq, se, UPDATE_TG);
12706
12707 if (cfs_rq_throttled(cfs_rq))
12708 break;
12709
12710 if (!throttled_hierarchy(cfs_rq))
12711 list_add_leaf_cfs_rq(cfs_rq);
12712 }
12713 }
12714 #else
propagate_entity_cfs_rq(struct sched_entity * se)12715 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12716 #endif
12717
detach_entity_cfs_rq(struct sched_entity * se)12718 static void detach_entity_cfs_rq(struct sched_entity *se)
12719 {
12720 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12721
12722 #ifdef CONFIG_SMP
12723 /*
12724 * In case the task sched_avg hasn't been attached:
12725 * - A forked task which hasn't been woken up by wake_up_new_task().
12726 * - A task which has been woken up by try_to_wake_up() but is
12727 * waiting for actually being woken up by sched_ttwu_pending().
12728 */
12729 if (!se->avg.last_update_time)
12730 return;
12731 #endif
12732
12733 /* Catch up with the cfs_rq and remove our load when we leave */
12734 update_load_avg(cfs_rq, se, 0);
12735 detach_entity_load_avg(cfs_rq, se);
12736 update_tg_load_avg(cfs_rq);
12737 propagate_entity_cfs_rq(se);
12738 }
12739
attach_entity_cfs_rq(struct sched_entity * se)12740 static void attach_entity_cfs_rq(struct sched_entity *se)
12741 {
12742 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12743
12744 /* Synchronize entity with its cfs_rq */
12745 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12746 attach_entity_load_avg(cfs_rq, se);
12747 update_tg_load_avg(cfs_rq);
12748 propagate_entity_cfs_rq(se);
12749 }
12750
detach_task_cfs_rq(struct task_struct * p)12751 static void detach_task_cfs_rq(struct task_struct *p)
12752 {
12753 struct sched_entity *se = &p->se;
12754
12755 detach_entity_cfs_rq(se);
12756 }
12757
attach_task_cfs_rq(struct task_struct * p)12758 static void attach_task_cfs_rq(struct task_struct *p)
12759 {
12760 struct sched_entity *se = &p->se;
12761
12762 attach_entity_cfs_rq(se);
12763 }
12764
switched_from_fair(struct rq * rq,struct task_struct * p)12765 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12766 {
12767 detach_task_cfs_rq(p);
12768 }
12769
switched_to_fair(struct rq * rq,struct task_struct * p)12770 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12771 {
12772 attach_task_cfs_rq(p);
12773
12774 if (task_on_rq_queued(p)) {
12775 /*
12776 * We were most likely switched from sched_rt, so
12777 * kick off the schedule if running, otherwise just see
12778 * if we can still preempt the current task.
12779 */
12780 if (task_current(rq, p))
12781 resched_curr(rq);
12782 else
12783 wakeup_preempt(rq, p, 0);
12784 }
12785 }
12786
12787 /* Account for a task changing its policy or group.
12788 *
12789 * This routine is mostly called to set cfs_rq->curr field when a task
12790 * migrates between groups/classes.
12791 */
set_next_task_fair(struct rq * rq,struct task_struct * p,bool first)12792 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12793 {
12794 struct sched_entity *se = &p->se;
12795
12796 #ifdef CONFIG_SMP
12797 if (task_on_rq_queued(p)) {
12798 /*
12799 * Move the next running task to the front of the list, so our
12800 * cfs_tasks list becomes MRU one.
12801 */
12802 list_move(&se->group_node, &rq->cfs_tasks);
12803 }
12804 #endif
12805
12806 for_each_sched_entity(se) {
12807 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12808
12809 set_next_entity(cfs_rq, se);
12810 /* ensure bandwidth has been allocated on our new cfs_rq */
12811 account_cfs_rq_runtime(cfs_rq, 0);
12812 }
12813 }
12814
init_cfs_rq(struct cfs_rq * cfs_rq)12815 void init_cfs_rq(struct cfs_rq *cfs_rq)
12816 {
12817 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12818 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12819 #ifdef CONFIG_SMP
12820 raw_spin_lock_init(&cfs_rq->removed.lock);
12821 #endif
12822 }
12823
12824 #ifdef CONFIG_FAIR_GROUP_SCHED
task_change_group_fair(struct task_struct * p)12825 static void task_change_group_fair(struct task_struct *p)
12826 {
12827 /*
12828 * We couldn't detach or attach a forked task which
12829 * hasn't been woken up by wake_up_new_task().
12830 */
12831 if (READ_ONCE(p->__state) == TASK_NEW)
12832 return;
12833
12834 detach_task_cfs_rq(p);
12835
12836 #ifdef CONFIG_SMP
12837 /* Tell se's cfs_rq has been changed -- migrated */
12838 p->se.avg.last_update_time = 0;
12839 #endif
12840 set_task_rq(p, task_cpu(p));
12841 attach_task_cfs_rq(p);
12842 }
12843
free_fair_sched_group(struct task_group * tg)12844 void free_fair_sched_group(struct task_group *tg)
12845 {
12846 int i;
12847
12848 for_each_possible_cpu(i) {
12849 if (tg->cfs_rq)
12850 kfree(tg->cfs_rq[i]);
12851 if (tg->se)
12852 kfree(tg->se[i]);
12853 }
12854
12855 kfree(tg->cfs_rq);
12856 kfree(tg->se);
12857 }
12858
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)12859 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12860 {
12861 struct sched_entity *se;
12862 struct cfs_rq *cfs_rq;
12863 int i;
12864
12865 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12866 if (!tg->cfs_rq)
12867 goto err;
12868 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12869 if (!tg->se)
12870 goto err;
12871
12872 tg->shares = NICE_0_LOAD;
12873
12874 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12875
12876 for_each_possible_cpu(i) {
12877 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12878 GFP_KERNEL, cpu_to_node(i));
12879 if (!cfs_rq)
12880 goto err;
12881
12882 se = kzalloc_node(sizeof(struct sched_entity_stats),
12883 GFP_KERNEL, cpu_to_node(i));
12884 if (!se)
12885 goto err_free_rq;
12886
12887 init_cfs_rq(cfs_rq);
12888 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12889 init_entity_runnable_average(se);
12890 }
12891
12892 return 1;
12893
12894 err_free_rq:
12895 kfree(cfs_rq);
12896 err:
12897 return 0;
12898 }
12899
online_fair_sched_group(struct task_group * tg)12900 void online_fair_sched_group(struct task_group *tg)
12901 {
12902 struct sched_entity *se;
12903 struct rq_flags rf;
12904 struct rq *rq;
12905 int i;
12906
12907 for_each_possible_cpu(i) {
12908 rq = cpu_rq(i);
12909 se = tg->se[i];
12910 rq_lock_irq(rq, &rf);
12911 update_rq_clock(rq);
12912 attach_entity_cfs_rq(se);
12913 sync_throttle(tg, i);
12914 rq_unlock_irq(rq, &rf);
12915 }
12916 }
12917
unregister_fair_sched_group(struct task_group * tg)12918 void unregister_fair_sched_group(struct task_group *tg)
12919 {
12920 unsigned long flags;
12921 struct rq *rq;
12922 int cpu;
12923
12924 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12925
12926 for_each_possible_cpu(cpu) {
12927 if (tg->se[cpu])
12928 remove_entity_load_avg(tg->se[cpu]);
12929
12930 /*
12931 * Only empty task groups can be destroyed; so we can speculatively
12932 * check on_list without danger of it being re-added.
12933 */
12934 if (!tg->cfs_rq[cpu]->on_list)
12935 continue;
12936
12937 rq = cpu_rq(cpu);
12938
12939 raw_spin_rq_lock_irqsave(rq, flags);
12940 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12941 raw_spin_rq_unlock_irqrestore(rq, flags);
12942 }
12943 }
12944
init_tg_cfs_entry(struct task_group * tg,struct cfs_rq * cfs_rq,struct sched_entity * se,int cpu,struct sched_entity * parent)12945 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12946 struct sched_entity *se, int cpu,
12947 struct sched_entity *parent)
12948 {
12949 struct rq *rq = cpu_rq(cpu);
12950
12951 cfs_rq->tg = tg;
12952 cfs_rq->rq = rq;
12953 init_cfs_rq_runtime(cfs_rq);
12954
12955 tg->cfs_rq[cpu] = cfs_rq;
12956 tg->se[cpu] = se;
12957
12958 /* se could be NULL for root_task_group */
12959 if (!se)
12960 return;
12961
12962 if (!parent) {
12963 se->cfs_rq = &rq->cfs;
12964 se->depth = 0;
12965 } else {
12966 se->cfs_rq = parent->my_q;
12967 se->depth = parent->depth + 1;
12968 }
12969
12970 se->my_q = cfs_rq;
12971 /* guarantee group entities always have weight */
12972 update_load_set(&se->load, NICE_0_LOAD);
12973 se->parent = parent;
12974 }
12975
12976 static DEFINE_MUTEX(shares_mutex);
12977
__sched_group_set_shares(struct task_group * tg,unsigned long shares)12978 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12979 {
12980 int i;
12981
12982 lockdep_assert_held(&shares_mutex);
12983
12984 /*
12985 * We can't change the weight of the root cgroup.
12986 */
12987 if (!tg->se[0])
12988 return -EINVAL;
12989
12990 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12991
12992 if (tg->shares == shares)
12993 return 0;
12994
12995 tg->shares = shares;
12996 for_each_possible_cpu(i) {
12997 struct rq *rq = cpu_rq(i);
12998 struct sched_entity *se = tg->se[i];
12999 struct rq_flags rf;
13000
13001 /* Propagate contribution to hierarchy */
13002 rq_lock_irqsave(rq, &rf);
13003 update_rq_clock(rq);
13004 for_each_sched_entity(se) {
13005 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
13006 update_cfs_group(se);
13007 }
13008 rq_unlock_irqrestore(rq, &rf);
13009 }
13010
13011 return 0;
13012 }
13013
sched_group_set_shares(struct task_group * tg,unsigned long shares)13014 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
13015 {
13016 int ret;
13017
13018 mutex_lock(&shares_mutex);
13019 if (tg_is_idle(tg))
13020 ret = -EINVAL;
13021 else
13022 ret = __sched_group_set_shares(tg, shares);
13023 mutex_unlock(&shares_mutex);
13024
13025 return ret;
13026 }
13027
sched_group_set_idle(struct task_group * tg,long idle)13028 int sched_group_set_idle(struct task_group *tg, long idle)
13029 {
13030 int i;
13031
13032 if (tg == &root_task_group)
13033 return -EINVAL;
13034
13035 if (idle < 0 || idle > 1)
13036 return -EINVAL;
13037
13038 mutex_lock(&shares_mutex);
13039
13040 if (tg->idle == idle) {
13041 mutex_unlock(&shares_mutex);
13042 return 0;
13043 }
13044
13045 tg->idle = idle;
13046
13047 for_each_possible_cpu(i) {
13048 struct rq *rq = cpu_rq(i);
13049 struct sched_entity *se = tg->se[i];
13050 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
13051 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
13052 long idle_task_delta;
13053 struct rq_flags rf;
13054
13055 rq_lock_irqsave(rq, &rf);
13056
13057 grp_cfs_rq->idle = idle;
13058 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
13059 goto next_cpu;
13060
13061 if (se->on_rq) {
13062 parent_cfs_rq = cfs_rq_of(se);
13063 if (cfs_rq_is_idle(grp_cfs_rq))
13064 parent_cfs_rq->idle_nr_running++;
13065 else
13066 parent_cfs_rq->idle_nr_running--;
13067 }
13068
13069 idle_task_delta = grp_cfs_rq->h_nr_running -
13070 grp_cfs_rq->idle_h_nr_running;
13071 if (!cfs_rq_is_idle(grp_cfs_rq))
13072 idle_task_delta *= -1;
13073
13074 for_each_sched_entity(se) {
13075 struct cfs_rq *cfs_rq = cfs_rq_of(se);
13076
13077 if (!se->on_rq)
13078 break;
13079
13080 cfs_rq->idle_h_nr_running += idle_task_delta;
13081
13082 /* Already accounted at parent level and above. */
13083 if (cfs_rq_is_idle(cfs_rq))
13084 break;
13085 }
13086
13087 next_cpu:
13088 rq_unlock_irqrestore(rq, &rf);
13089 }
13090
13091 /* Idle groups have minimum weight. */
13092 if (tg_is_idle(tg))
13093 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
13094 else
13095 __sched_group_set_shares(tg, NICE_0_LOAD);
13096
13097 mutex_unlock(&shares_mutex);
13098 return 0;
13099 }
13100
13101 #else /* CONFIG_FAIR_GROUP_SCHED */
13102
free_fair_sched_group(struct task_group * tg)13103 void free_fair_sched_group(struct task_group *tg) { }
13104
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)13105 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
13106 {
13107 return 1;
13108 }
13109
online_fair_sched_group(struct task_group * tg)13110 void online_fair_sched_group(struct task_group *tg) { }
13111
unregister_fair_sched_group(struct task_group * tg)13112 void unregister_fair_sched_group(struct task_group *tg) { }
13113
13114 #endif /* CONFIG_FAIR_GROUP_SCHED */
13115
13116
get_rr_interval_fair(struct rq * rq,struct task_struct * task)13117 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
13118 {
13119 struct sched_entity *se = &task->se;
13120 unsigned int rr_interval = 0;
13121
13122 /*
13123 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
13124 * idle runqueue:
13125 */
13126 if (rq->cfs.load.weight)
13127 rr_interval = NS_TO_JIFFIES(se->slice);
13128
13129 return rr_interval;
13130 }
13131
13132 /*
13133 * All the scheduling class methods:
13134 */
13135 DEFINE_SCHED_CLASS(fair) = {
13136
13137 .enqueue_task = enqueue_task_fair,
13138 .dequeue_task = dequeue_task_fair,
13139 .yield_task = yield_task_fair,
13140 .yield_to_task = yield_to_task_fair,
13141
13142 .wakeup_preempt = check_preempt_wakeup_fair,
13143
13144 .pick_next_task = __pick_next_task_fair,
13145 .put_prev_task = put_prev_task_fair,
13146 .set_next_task = set_next_task_fair,
13147
13148 #ifdef CONFIG_SMP
13149 .balance = balance_fair,
13150 .pick_task = pick_task_fair,
13151 .select_task_rq = select_task_rq_fair,
13152 .migrate_task_rq = migrate_task_rq_fair,
13153
13154 .rq_online = rq_online_fair,
13155 .rq_offline = rq_offline_fair,
13156
13157 .task_dead = task_dead_fair,
13158 .set_cpus_allowed = set_cpus_allowed_common,
13159 #endif
13160
13161 .task_tick = task_tick_fair,
13162 .task_fork = task_fork_fair,
13163
13164 .prio_changed = prio_changed_fair,
13165 .switched_from = switched_from_fair,
13166 .switched_to = switched_to_fair,
13167
13168 .get_rr_interval = get_rr_interval_fair,
13169
13170 .update_curr = update_curr_fair,
13171
13172 #ifdef CONFIG_FAIR_GROUP_SCHED
13173 .task_change_group = task_change_group_fair,
13174 #endif
13175
13176 #ifdef CONFIG_SCHED_CORE
13177 .task_is_throttled = task_is_throttled_fair,
13178 #endif
13179
13180 #ifdef CONFIG_UCLAMP_TASK
13181 .uclamp_enabled = 1,
13182 #endif
13183 };
13184
13185 #ifdef CONFIG_SCHED_DEBUG
print_cfs_stats(struct seq_file * m,int cpu)13186 void print_cfs_stats(struct seq_file *m, int cpu)
13187 {
13188 struct cfs_rq *cfs_rq, *pos;
13189
13190 rcu_read_lock();
13191 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13192 print_cfs_rq(m, cpu, cfs_rq);
13193 rcu_read_unlock();
13194 }
13195
13196 #ifdef CONFIG_NUMA_BALANCING
show_numa_stats(struct task_struct * p,struct seq_file * m)13197 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13198 {
13199 int node;
13200 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13201 struct numa_group *ng;
13202
13203 rcu_read_lock();
13204 ng = rcu_dereference(p->numa_group);
13205 for_each_online_node(node) {
13206 if (p->numa_faults) {
13207 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13208 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13209 }
13210 if (ng) {
13211 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13212 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13213 }
13214 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13215 }
13216 rcu_read_unlock();
13217 }
13218 #endif /* CONFIG_NUMA_BALANCING */
13219 #endif /* CONFIG_SCHED_DEBUG */
13220
init_sched_fair_class(void)13221 __init void init_sched_fair_class(void)
13222 {
13223 #ifdef CONFIG_SMP
13224 int i;
13225
13226 for_each_possible_cpu(i) {
13227 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13228 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
13229 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13230 GFP_KERNEL, cpu_to_node(i));
13231
13232 #ifdef CONFIG_CFS_BANDWIDTH
13233 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13234 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13235 #endif
13236 }
13237
13238 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13239
13240 #ifdef CONFIG_NO_HZ_COMMON
13241 nohz.next_balance = jiffies;
13242 nohz.next_blocked = jiffies;
13243 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
13244 #endif
13245 #endif /* SMP */
13246
13247 }
13248