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